1: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2: %% Continuous And Discontinuous Phase Transitions %%
3: %% In Hypergraph Processes %%
4: %% %%
5: %% Darling, Levin, Norris %%
6: %% %%
7: %% Version March 2 2004 %%
8: %% %%
9: %% rwrd@afterlife.ncsc.mil, levin@math.utah.edu, %%
10: %% j.r.norris@statslab.cam.ac.uk %%
11: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
12: \documentclass[12pt]{amsart}
13: \usepackage{amssymb,xspace,graphicx,enumerate}
14: \usepackage[author-year]{amsrefs}
15: %
16: % theorem environments
17: \newtheorem{mainthm}{Theorem}
18: \newtheorem{thm}{Theorem}[section]
19: \newtheorem{prop}[thm]{Proposition}
20: \newtheorem{lem}[thm]{Lemma}
21: \newtheorem{cor}[thm]{Corollary}
22: \newtheorem{conj}{Conjecture}
23: \theoremstyle{definition}
24: \newtheorem{defn}[thm]{Definition}
25: \theoremstyle{remark}
26: \newtheorem{rmk}{Remark}[section]
27: \newtheorem{rmks}[rmk]{Remarks}
28: \newtheorem*{rmk*}{Remark}
29: \newtheorem*{rmks*}{Remarks}
30: % enumerate environments
31: \newenvironment{enumeratei}{\begin{enumerate}[\upshape (i)]}
32: {\end{enumerate}}
33: \newenvironment{enumeratea}{\begin{enumerate}[\upshape (a)]}
34: {\end{enumerate}}
35: \newenvironment{enumeraten}{\begin{enumerate}[{\tt 1.}]}
36: {\end{enumerate}}
37: % formatting
38: % margins
39: \addtolength{\textwidth}{1in}
40: \addtolength{\evensidemargin}{-0.5in}
41: \addtolength{\oddsidemargin}{-0.5in}
42: \renewcommand{\baselinestretch}{1.2}
43: \renewcommand{\qedsymbol}{\ensuremath{\blacksquare}}
44: %%%%% macros %%%%%%
45: % general math and probability
46: \newcommand{\deq}{\overset{{\rm def}}{=}} % defined equal
47: \newcommand{\diseq}{\, {\stackrel {{\cal L}} {=}}} % equal in distn
48: \newcommand{\dconv}{\, {\stackrel {{\cal L}} {\to}}} % conv in distn
49: \newcommand{\sif}{\sigma \mbox{\rm -field}} % sigma-field
50: \newcommand{\E}{E} % expectation
51: \newcommand{\Eb}[1]{\E\!\left( #1 \right)} % with brackets
52: \newcommand{\Emc}[2]{\E_{#1}\left(#2\right)}
53: \renewcommand{\P}{P} % probability
54: \newcommand{\Pb}[1]{\P\!\left( #1 \right)} % with brackets
55: \newcommand{\Pmc}[2]{\P_{#1}\left(#2\right)}
56: \newcommand{\Pm}{{\boldsymbol P}}
57: \newcommand{\N}{{\mathbb N}} % natural numbers
58: \newcommand{\C}{{\mathcal C}}
59: \newcommand{\Z}{{\ensuremath{\mathbb Z}}} % integers
60: \newcommand{\Q}{{\mathbb Q}} % rationals
61: \newcommand{\R}{{\mathbb R}} % real numbers
62: \newcommand{\F}{{\mathcal F}} % sigma-field
63: \newcommand{\G}{{\mathcal G}} % another sigma-field
64: \newcommand{\B}{{\mathcal{B}}} % Borel sigma-field
65: \newcommand{\X}{{\mathcal X}}
66: \newcommand{\given}{\mid}
67: \newcommand{\givent}{\; \vline \;}
68: \newcommand{\Given}{\; \vline \;}
69: \newcommand{\pr}{{\mathsf Pr}}
70: \newcommand{\Lp}[2]{\|#2\|_{#1}}
71: \newcommand{\one}{{\bf 1}} % indicator function
72: \newcommand{\wh}{\widehat}
73: \renewcommand{\and}{\; \mbox{ and } \;}
74: \newcommand{\iid}{i.i.d.\ }
75: \DeclareMathOperator{\var}{Var}
76: \DeclareMathOperator{\cov}{Cov}
77: \newcommand{\shift}{\Theta}
78: \newcommand{\bin}{\{0,1\}}
79: \newcommand{\inprob}{\stackrel{\rm Pr} {\longrightarrow}}
80: \renewcommand{\l}{\ell}
81: \newcommand{\sequ}[2]{\{ #1_{#2} \}_{#2=0}^\infty}
82: \newcommand{\fc}{\frac{1}{2\pi}}
83: \newcommand{\vect}[1]{{\boldsymbol #1}}
84: \newcommand{\law}{{\mathcal L}}
85: \newcommand{\ito}{It\^{o}\xspace}
86: \renewcommand{\emptyset}{\varnothing}
87: % graph theory
88: \newcommand{\graph}{{\mathcal G}}
89: \newcommand{\ve}{V} % vertices
90: \newcommand{\ed}{{\mathcal E}} % edges
91: % greek
92: \newcommand{\de}{\delta}
93: \newcommand{\ep}{\epsilon}
94: \newcommand{\al}{\alpha}
95: \newcommand{\Ga}{\Gamma}
96: \newcommand{\la}{\lambda}
97: \renewcommand{\th}{\theta}
98: \newcommand{\om}{\omega}
99: \renewcommand{\phi}{\varphi}
100: % this paper
101: \newcommand{\convfdd}{\stackrel{{\rm f.d.d.}}{\longrightarrow}}
102: \newcommand{\weakconv}{\stackrel{{\rm d}}{\longrightarrow}}
103: \newcommand{\hc}{{\sf hypergraph collapse}\xspace}
104: %%%
105:
106: \newtheorem{rmksss}{Remarks}[subsection]
107: \newcommand{\Gr}{{\mathcal G}}
108: \begin{document}
109: \title[Hypergraph Processes]{Continuous And Discontinuous Phase
110: Transitions In Hypergraph Processes}
111:
112: %
113: \author[R.W.R.\@ Darling]{R.W.R.\@ Darling}
114: \address{National\@ Security\@ Agency\\
115: P.O.\ Box 535\\ Annapolis Junction, MD 20701}
116: \email{rwrd@afterlife.ncsc.mil}
117: %\urladdr{}
118: %
119: \author[D.A.\@ Levin]{David A.\@ Levin}
120: \address{Department of Mathematics\\The University\@ of Utah\\
121: 155 S.\@ 1400 E.\\Salt Lake City, UT 84112--0090}
122: \email{levin@math.utah.edu}
123: %
124: \author[J.R.\@ Norris]{James R.\@ Norris}
125: \address{Statistical Laboratory\\ Centre For Mathematical Sciences\\
126: Wilberforce Road, Cambridge, CB3 0WB}
127: \email{j.r.norris@statslab.cam.ac.uk}
128:
129:
130:
131: \maketitle
132:
133: \vspace{-0.25in}
134:
135: \centerline{\emph{National Security Agency, University of Utah,
136: University of Cambridge}}
137:
138: \begin{abstract}
139: Let $V$ denote a set of $N$ vertices. To construct a \emph{hypergraph
140: process}, create a new hyperedge at each event time of a Poisson
141: process; the cardinality $K$ of this hyperedge is random, with
142: generating function $\rho(x) \deq \sum \rho_k x^k$,
143: where $\Pb{K = k} = \rho_k$; given $K=k$, the $k$ vertices appearing
144: in the new hyperedge are selected uniformly at random from
145: $V$. Assume $\rho_1 + \rho_2 > 0$. Hyperedges of cardinality $1$
146: are called \emph{patches}, and serve as a way of selecting root
147: vertices. Identifiable vertices are those which are reachable
148: from these root vertices, in a strong sense which generalizes the
149: notion of graph component. Hyperedges are called identifiable if all
150: of their vertices are identifiable. We use ``fluid limit'' scaling:
151: hyperedges arrive at rate $N$, and we study structures of size
152: $O(1)$ and $O(N)$. After division by $N$, numbers of identifiable
153: vertices and hyperedges exhibit phase transitions, which
154: may be continuous or discontinuous depending on the shape of the
155: structure function $-\log(1 - x)/\rho'(x), \ x\in(0,1)$.
156: Both the case $\rho_1 > 0$, and the case $\rho_1 = 0 < \rho_2$
157: are considered; for the latter, a single extraneous patch is added
158: to mark the root vertex.
159: \end{abstract}
160:
161: \newpage
162: %\tableofcontents
163:
164: \section{Introduction}
165:
166: The \emph{$k$-core} of a graph is the largest subgraph with
167: minimum degree at least $k$.
168: \ocite{PSW:core} study the following algorithm for
169: finding the $2$-core of a graph:
170: \begin{enumeraten}
171: \item If vertices of degree one exist, select one
172: and remove the edge incident to it. This may
173: cause the degree of other vertices to drop.
174: \item If there are no degree one vertices remaining,
175: stop.
176: \item Repeat.
177: \end{enumeraten}
178: The graph obtained at the conclusion of this
179: algorithm is the $2$-core.
180:
181: This algorithm is a special case of another,
182: run on hypergraphs, called \emph{hypergraph collapse} and first
183: studied in \ocite{DN:HG}.
184: By a \emph{hypergraph} we shall mean a map $\Lambda:2^V\to\{0,1,2,\dots\}$,
185: where $V$ is a finite set of \emph{vertices} and $2^V$ is the set of subsets
186: of $V$.
187: It will sometimes be helpful to think in terms of an {\em edge-labelling} of
188: $\Lambda$, which is a choice of a set $I$ and a map
189: $e:I\to 2^V$ such that $\Lambda(A)=|\{i\in I: e(i)=A\}|$ for all $A$.
190: Thus $e$ describes a set of labelled subsets of $V$, which we call
191: \emph{hyperedges} and then $\Lambda$ gives
192: the number of hyperedges at each subset of $V$.
193: Hyperedges of unit cardinality are called \emph{patches}. Hypergraph
194: collapse is the following algorithm:
195: \begin{enumeraten}
196: \item If a patch exists, select one and remove it together with the unique
197: vertex $v$ it contains. This will cause
198: any other hyperedge $e(i)$ containing $v$ to be replaced by
199: $e(i) \setminus \{v\}$.
200: \item If there are no patches remaining, stop.
201: \item Repeat.
202: \end{enumeraten}
203: Although we have described the algorithm in terms of an edge-labelled
204: hypergraph,
205: the possible moves for $\Lambda$ do not depend on the edge-labelling chosen.
206: The vertices which are removed by hypergraph collapse are called
207: \emph{identifiable}, and hyperedges which contain only identifiable
208: vertices are also called \emph{identifiable}. These definitions do not depend
209: on the order in which patches are chosen during hypergraph collapse;
210: see \ocite{DN:HG}.
211:
212: The core-finding algorithm of \ocite{PSW:core} is hypergraph collapse
213: applied to the \emph{dual} hypergraph. To obtain the dual, note that we can
214: think of $e$ as a subset of $V\times I$. The roles of $V$ and $I$ are now
215: symmetric, so $e$ also corresponds to an edge-labelling of a hypergraph
216: $\Lambda'$
217: in which the status of vertices and hyperedges is reversed.
218: Vertices (resp.\ hyperedges) of
219: $\Lambda$ not in the core correspond to identifiable hyperedges (resp.\
220: identifiable vertices) of $\Lambda'$. More information about graph
221: cores can be found in \ocite{F:core}, and hypergraph cores are
222: considered by \ocite{C:core}.
223:
224: The identifiable vertices obtained by hypergraph collapse
225: also serve to generalize to hypergraphs the definition of
226: graph component. A graph is a hypergraph having edges only of
227: cardinality two, and consequently has no patches. However, if the
228: single hyperedge $\{v\}$ is added to the graph [making it a
229: hypergraph], then the identifiable vertices obtained by running
230: hypergraph collapse on the augmented graph are exactly the vertices in
231: the graph component containing $v$. The identifiable edges are all the
232: edges of this graph component.
233:
234: This motivates the following definition for patch-free hypergraphs: A
235: vertex is \emph{in the domain}
236: of $v$ if it is in the set of identifiable vertices when
237: the hypergraph is augmented by the addition of the
238: hyperedge $\{v\}$.
239:
240: The purpose of this paper is to study the time-evolution of the
241: set of identifiable vertices and the set of identifiable edges
242: in a \emph{Poisson hypergraph process}, which is a hypergraph-valued,
243: continuous-time stochastic process. The vertex set is $V =
244: \{1,2,\ldots,N\}$, and the process depends on parameters
245: $\{\rho_j\}_{j=1}^N$. Attached to each subset $A$ of $V$ is a Poisson
246: clock run at rate $N \rho_{|A|}/ \binom{N}{|A|}$, and these
247: clocks are independent of one another. [Here $|A|$ denotes
248: the cardinality of $A$.] When the clock associated to $A$ ``rings'',
249: a new hyperedge equal to $A$ is added to the hypergraph. The overall rate
250: at which hyperedges of cardinality $k$ are added is then $N \rho_k$.
251: We will call this process the Poisson($\rho$) hypergraph process.
252: This is a generalization of the ordinary random graph process, in
253: which edges form between each pair of vertices independently at a fixed rate.
254:
255: While for $N$ fixed, this process depends only on the finite sequence
256: $\{\rho_k\}_{k=1}^N$, we will be interested in the asymptotic behavior
257: as $N \rightarrow \infty$, so we will assume that always the infinite
258: sequence $\{\rho_k\}_{k=1}^\infty$ is given. Moreover, this sequence
259: is required to be a probability distribution on $\{1,2,\ldots\}$ with
260: finite expectation and satisfying $\rho_1 + \rho_2 > 0$. The
261: generating function $x \mapsto \sum_{k=1}^\infty \rho_k x^k$ will
262: be denoted by $\rho$.
263:
264: In \ocite{DN:HG}, the \emph{Poisson($\beta$) random hypergraph} is
265: defined, where $\{\beta_k\}$ is a sequence of positive
266: real numbers. This is a random hypergraph with vertex set
267: V = $\{1,2,\ldots,N\}$, so that for each $A \subset V$,
268: the number of occurrences of the hyperedge $A$ is a
269: Poisson random variable with expectation $N\beta_{|A|}/\binom{N}{|A|}$,
270: and these random variables are independent for different subsets of $V$.
271: If $\{\Lambda_t\}_{t \geq 0}$ is a Poisson($\rho$) hypergraph
272: process, then for fixed $t \geq 0$, $\Lambda_t$ is a Poisson($t\rho$)
273: random hypergraph.
274:
275: We separate out two distinct cases in the study of Poisson hypergraph
276: processes, depending on whether $\rho_1 > 0$ or $\rho_1 = 0$. When
277: $\rho_1 = 0$, the hypergraph never acquires patches, and provided the
278: initial hypergraph is patch-free, the set of identifiable vertices
279: is forever void. As the previous discussion of ordinary graphs
280: suggests, it is natural to consider in such cases the set of vertices
281: in the domain of a distinguished vertex.
282:
283: We discuss now the case $\rho_1 > 0$.
284: Our first result describes the evolution of the rescaled number of
285: identifiable vertices and hyperedges in the Poisson($\rho$)
286: hypergraph process $\{\Lambda_t\}_{t \geq 0}$. Let
287: \begin{equation} \label{eq:T_tilde_N_and_Z_tilde_N_defn}
288: \begin{split}
289: \tilde{T}^N_t & = \frac{|\text{identifiable vertices in } \Lambda_t|}{N}\\
290: \tilde{Z}^N_t & = \frac{|\text{identifiable hyperedges in } \Lambda_t|}{N} \,.
291: \end{split}
292: \end{equation}
293: The \emph{structure function} $t$, defined as
294: \begin{equation} \label{eq:structure_function}
295: t(x) \deq \frac{ - \log(1-x) } { \rho'(x) } \,, \quad
296: x \in (0,1) \,,
297: \end{equation}
298: plays a central role for hypergraph processes.
299: [Recall that $\rho(x) = \sum_{k=1}^\infty \rho_k x^k$.]
300: Typically $t$ is not invertible, but there
301: is a right-continuous monotonic function called the
302: \emph{lower envelope}:
303: \begin{equation} \label{eq:gsdefinition}
304: g(s) \deq \inf\{ x \in (0,1) \;:\; t(x) > s\}\,,
305: \quad s \geq 0 \,.
306: \end{equation}
307: Also important for hypergraph processes is
308: the \emph{upper envelope}:
309: \begin{equation} \label{eq:g_star_definition}
310: g^\star(s) \deq
311: \sup\{x \in (0, 1) \;:\; t(x) < s\} \vee 0 \,,
312: \quad s \geq 0 \,.
313: \end{equation}
314:
315: We classify structure functions into three types:
316: graph-like, bicritical, and exceptional. This
317: taxonomy is given in Table \ref{tab:class}. Figure
318: \ref{fig:bicritical} shows a bicritical structure function and the
319: corresponding lower envelope.
320:
321: \begin{table}[h]
322: \begin{center}
323: \begin{tabular}{c|c|c}
324: \textsc{Type} & \textsc{Description} & \textsc{Example of $\rho(x)$} \\
325: \hline
326: graph-like & \parbox{2.00in}{\vspace{0.1in}
327: $t$ is strictly increasing, and \\
328: $g$ and $g^\star$ are continuous.\vspace{0.1in}}
329: & cubic with $3 \rho_3 \leq \rho_2$ \\
330: \hline
331: bicritical & \parbox{2.00in}{\vspace{0.1in}
332: $g$ and $g^\star$ each have \\
333: exactly one discontinuity.
334: \vspace{0.1in} }
335: & cubic with $3 \rho_3 > \rho_2$ \\
336: \hline
337: exceptional & \parbox{2.00in}{\vspace{0.1in}
338: $g$ or $g^\star$ has two or more \\
339: discontinuities.
340: \vspace{0.1in} }
341: & $\frac{x + 5x^3 + 994x^{200}}{1000}$
342: \end{tabular}
343:
344: \vspace{0.1in}
345:
346: \caption{Classification of structure functions \label{tab:class}}
347: \end{center}
348: \end{table}
349:
350: \begin{figure}
351: \begin{center}
352: \includegraphics{bicritical}
353: \caption{\emph{Left}:
354: Bicritical structure function, with $t(x)$ on the horizontal axis,
355: corresponding to a quartic polynomial $\rho(x)$ with
356: $0 < \rho_1 < \rho_2 < \rho_3 < \rho_4$. \emph{Right}:
357: Lower envelope, showing the single discontinuity.
358: \label{fig:bicritical}}
359: \end{center}
360: \end{figure}
361:
362: Let $\Xi \subset \R_+$ denote the discontinuity set of
363: $g$:
364: \begin{equation} \label{eq:discontinuity_set}
365: \Xi \deq \{ s > 0 \,:\, g(s-) \neq g(s) \} \,,
366: \end{equation}
367: where $g(s-) \deq \lim_{t \uparrow s} g(t)$.
368:
369: For $s \in \Xi$, both $g(s-)$ and $g(s)$ are zeros of the function
370: $x \mapsto \rho'(x) + \log(1 - x)$. For the sake of simplicity of
371: exposition, we shall assume below that there are never any zeros
372: of this function strictly between $g(s-)$ and $g(s)$:
373: \begin{equation} \label{eq:onlytwozeros}
374: \{x \,:\, s \rho'(x) + \log(1 - x) = 0 \}
375: \bigcap (g(s-), g(s)) = \emptyset \,,
376: \quad \text{for all } s \in \Xi \,.
377: \end{equation}
378: Also assume that $\Xi$ has no accumulation points. This is true, for
379: example, if $\sum_k k^2 \rho_k < \infty$.
380:
381: Let $\{B_s, \ s \in \Xi\}$ denote a collection of
382: independent Bernoulli($1/2$) random variables, indexed by the
383: discontinuity set \eqref{eq:discontinuity_set}.
384: Define
385: \begin{equation} \label{eq:ifl}
386: \begin{split}
387: \tilde{T}_t & \deq g(t-) + B_t(g(t) - g(t-)) \,, \quad t \in \Xi \\
388: \tilde{T}_t & \deq g(t), \ t \not\in \Xi \,.
389: \end{split}
390: \end{equation}
391: In other words, at each point of discontinuity we choose the left
392: limit or the right limit of $g$ according to the flip of a fair coin.
393: Finally, let
394: \begin{equation} \label{eq:z_tilde_definition}
395: \tilde{Z}_t \deq t\rho(\tilde{T}_t)
396: - (1 - \tilde{T}_t) \log(1 - \tilde{T}_t) \,.
397: \end{equation}
398:
399: For a sequence of stochastic processes $\{X^N\}_{N=1}^\infty$,
400: where $X^N = \{X^N_t\}_{t \geq 0}$, and a
401: stochastic process $X = \{X_t\}_{t \geq 0}$, we write $X^N
402: \convfdd X$ if the finite-dimensional distributions of $X^N$
403: converge to those of $X$. For a sequence of random variables
404: (or vectors) $\{X^N\}$, we write $X^N \weakconv X$ to indicate that $X^N$
405: converges in distribution to $X$.
406: %
407: %
408: \begin{mainthm} \label{thm:intro_lambda_geq_one}
409: Consider a Poisson hypergraph process such that $\rho_1 > 0$, and
410: suppose \eqref{eq:onlytwozeros} holds. As $N \rightarrow \infty$,
411: \begin{equation} \label{eq:fluid_limit_with_patches}
412: \{(\tilde{T}^N_t, \tilde{Z}^N_t)\}_{t \geq 0}
413: \convfdd
414: \{(\tilde{T}_t, \tilde{Z}_t)\}_{t \geq 0} \,.
415: \end{equation}
416: Furthermore for any compact interval $I \subset [0,\infty) \setminus
417: \Xi$,
418: \begin{equation} \label{eq:convinprob}
419: \sup_{t \in I}\left|(\tilde{T}^N_t, \tilde{Z}^N_t) -
420: \left( g(t), t\rho(g(t)) - [1 - g(t)]\log(1 - g(t)) \right) \right|
421: \rightarrow 0
422: \end{equation}
423: in probability as $N \rightarrow \infty$.
424: \end{mainthm}
425: %
426: %
427:
428: We now turn to the case of patch-free hypergraph processes,
429: i.e.\ the regime where $\rho_1 = 0 < \rho_2$.
430: %By virtue of \eqref{eq:rho1_plus_rho2},
431: In this case $g(s) = 0$ for all
432: $s \in [0, (2\rho_2)^{-1})$. There are three
433: possibilities for the behavior of $g$ at
434: $(2\rho_2)^{-1}$, enumerated in Table \ref{tab:rho_1_equals_zero}.
435:
436: \begin{table}[h]
437: \begin{center}
438: \begin{tabular}{cc}
439: \textsc{Sub-case of $\rho_1 = 0 < \rho_2$} & \textsc{
440: Behavior of $g$} \\
441: \hline
442: $3\rho_3 < \rho_2$ & \parbox{2.0in}{\vspace{0.1in} $g$ is
443: continuous at $(2 \rho_2)^{-1}$, \\ and right derivative is
444: finite \vspace{0.1in}} \\
445: \hline
446: $3\rho_3 = \rho_2$ & \parbox{2.0in}{\vspace{0.1in}
447: $\rho_4,\rho_5,\ldots$ determine whether \\
448: $g$ is continuous at $(2 \rho_2 )^{-1}$ \vspace{0.1in}} \\
449: \hline
450: $3 \rho_3 > \rho_2$ & \parbox{2.0in}{\vspace{0.1in}
451: $g$ is discontinuous at $(2 \rho_2)^{-1}$}
452: \end{tabular}
453:
454: \vspace{0.1in}
455:
456: \caption{The $0 = \rho_1 < \rho_2$ regime. \label{tab:rho_1_equals_zero}}
457: \end{center}
458: \end{table}
459:
460: For simplicity, we focus on the case where $g$ has a single
461: discontinuity, located at $(2\rho_2)^{-1}$; i.e.\
462: $\Xi = \{(2\rho_2)^{-1}\}$. The general case follows the same pattern as
463: Theorem \ref{thm:intro_lambda_geq_one}, because
464: after the number of identifiable vertices has reached $O(N)$, the
465: subsequent evolution is much the same as the $\rho_1 > 0$ case.
466:
467: In the $\rho_1 = 0$ and $\rho_2 > 0$ regime, another structure function
468: besides \eqref{eq:structure_function} comes into
469: play, namely the structure function $t_2$ of the graph which
470: results from discarding all hyperedges of cardinality more than two:
471: \begin{equation} \label{eq:t2}
472: t_2(x) \deq \frac{-\log(1 - x)}{2 \rho_2x} \,,
473: \quad x \in (0, 1) \,.
474: \end{equation}
475: Since $t_2$ is monotonic, the corresponding lower envelope
476: $g_2$ defined as
477: \begin{equation} \label{eq:g2s}
478: g_2(s) \deq \inf\{ x \in (0, 1) \,:\, t_2(x) > s \}\,,
479: \quad s \geq 0\,,
480: \end{equation}
481: is continuous. As before, $g_2(s) = 0$ for $0 \leq s \leq
482: (2\rho_2)^{-1}$, and $g_2(s) \rightarrow 1$ as $s\rightarrow \infty$;
483: it describes the asymptotic proportion of vertices in the giant
484: component of a random graph where the ratio of edges to vertices is
485: $s\rho_2$.
486:
487: We will construct in Section \ref{sec:patch-free_process} an increasing
488: process $\{M_t\}$ so that the distribution of $M_t$ is
489: \begin{equation} \label{eq:bp_population_0}
490: \Pb{M_t = n} =
491: \begin{cases}
492: e^{-2 \beta_2 n} \left( 2 t \rho_2 n \right)^{n - 1} / n!
493: & \text{if } n \in \N \,, \\
494: \phi_t & \text{if } n = \infty \,,
495: \end{cases}
496: \end{equation}
497: where $\phi_t$ is the largest solution $x$ in $[0,1]$ of $2 t \rho_2 x +
498: \log(1 - x) = 0$. [Notice that $\phi_t = 0$ for $2 t \rho_2 \leq 1$, and
499: $0 < \phi_t < 1$ otherwise.]
500:
501: Write $T^N_t$ for the number of vertices in the domain of $v_0$ in
502: $\Lambda_t$, and write $Z^N_t$ for the number of hyperedges
503: identifiable from $v_0$ in $\Lambda_t$. Set
504: $\bar{T}^N_t \deq N^{-1}T^N_t$ and
505: $\bar{Z}^N_t \deq N^{-1}Z^N_t$.
506: Also, define
507: \begin{align*}
508: \bar{T}_t & \deq g(t) \one_{\{M_t = \infty\}} \,; \\
509: \bar{Z}_t & \deq \left\{ t\rho(g(t)) - [1 - g(t)]\log(1 - g(t))
510: \right\} \one_{\{M_t = \infty\}} \,.
511: \end{align*}
512: %
513: \begin{mainthm} \label{thm:intro_no_patch_process}
514: Consider a Poisson hypergraph process such that
515: $\rho_1 = 0 < \rho_2$, and suppose $g$ has a
516: single discontinuity located at $(2\rho_2)^{-1}$. Fix a distinguished
517: vertex $v_0$. The number of vertices in the domain of $v_0$, and
518: number of hyperedges identifiable from $v_0$, obey the following limits
519: in distribution as $N \rightarrow \infty$:
520: \begin{equation} \label{eq:micro_no_patch_fluid_limit}
521: \{(T^N_t, Z^N_t)\}_{t \geq 0}
522: \text{ converges weakly in $D\left([0, \infty), (\N \cup
523: \{\infty\})^2 \right)$ to }
524: \{(M_t, M_t)\}_{t \geq 0} \,,
525: \end{equation}
526: where we adjoin $\infty$ to $\N$ as a
527: compactifying point. Also
528: \begin{equation} \label{eq:macro_no_patch_fluid_limit}
529: \{(\bar{T}^N_t, \bar{Z}^N_t)\}_{t \geq 0}
530: \convfdd \{ (\bar{T}_t,\bar{Z}_t) \}_{t \geq 0} \,.
531: \end{equation}
532: \end{mainthm}
533:
534: \begin{rmk}
535: Observe the difference between the limit law
536: $\{(\bar{T}_t,\bar{Z}_t)\}_{t \geq 0}$ in
537: \eqref{eq:macro_no_patch_fluid_limit} and the limit law $\{
538: (\tilde{T}_t, \tilde{Z}_t)\}_{t \geq 0}$ in
539: \eqref{eq:fluid_limit_with_patches}: $\tilde{T}_t$ conforms to the
540: deterministic lower envelope $g(t)$, except at points in the finite
541: discontinuity set, whereas $\bar{T}_t$ waits until the random time
542: $\chi \deq \inf\{t \geq 0 \,:\, M_t = \infty\}$, with distribution
543: function $g_2(t)$, before jumping from $0$ up to $g(t)$.
544: \end{rmk}
545: \begin{rmk}
546: See Remark \ref{rmk:no_conv_in_D} as to whether the convergence
547: \eqref{eq:macro_no_patch_fluid_limit} extends to weak convergence in
548: the Skorohod space $D([0, \infty),\R_+^2)$.
549: \end{rmk}
550:
551: The rest of this paper is organized as follows: Some definitions
552: concerning hypergraphs are given in Section \ref{sec:definitions}. We
553: establish that certain key processes are Markov in Section
554: \ref{sec:markov}. The case of hypergraphs and hypergraph processes
555: with patches are treated in Section \ref{sec:review} and Section
556: \ref{sec:patches} respectively. Theorem \ref{thm:intro_lambda_geq_one}
557: is proved in Section \ref{sec:patches}. Patch-free random hypergraphs
558: and hypergraph processes are treated in Section
559: \ref{sec:patch-free_static} and Section \ref{sec:patch-free_process},
560: respectively. Theorem 2 is proved in Section
561: \ref{sec:patch-free_process}. Finally, we mention future directions in
562: Section \ref{sec:future}.
563:
564: \section{Hypergraph definitions} \label{sec:definitions}
565:
566: Recall from the Introduction that the identifiable vertices are those vertices
567: removed by the hypergraph collapse algorithm described there, and the
568: identifiable hyperedges are those hyperedges consisting only of
569: identifiable vertices.
570:
571: Given a hypergraph $\Lambda$ and a subset $S \subset V$,
572: $\Lambda^S$ denotes the hypergraph after all vertices in $S$ are
573: deleted. More precisely,
574: \begin{equation} \label{eq:LambdaS}
575: \Lambda^S(A) \deq \sum_{B \supset A, \ B\setminus S = A}
576: \Lambda(B), \quad A \subset V\setminus S \,.
577: \end{equation}
578: We now more exactly specify the hypergraph collapse algorithm: select
579: if possible a vertex $v$ with
580: $\Lambda(\{v\}) \geq 1$; replace $V$ by $V \setminus \{v\}$ and
581: $\Lambda$ by $\Lambda^{\{v\}}$; then repeat. When
582: the algorithm terminates, we obtain a set
583: $V^\star$ consisting of the identifiable
584: vertices, and a patch-free hypergraph $\Lambda^{V^\star}$ on
585: $V \setminus V^\star$.
586:
587: Suppose $\Lambda$ is a patch-free hypergraph, and thus
588: having no identifiable vertices. Given such
589: a hypergraph $\Lambda$ and a distinguished vertex $v_0$, we
590: say that $v$ is in the \emph{domain} of
591: $v_0$ in $\Lambda$ if $v$ is identifiable in the hypergraph
592: $\Lambda + \one_{\{v_0\}}$ obtained by augmenting $\Lambda$
593: by the hyperedge $\{v_0\}$. A hyperedge is said to be
594: \emph{identifiable} from $v_0$ if it is identifiable in $\Lambda +
595: \one_{\{v_0\}}$.
596:
597: \emph{Warning}: For a general patch-free hypergraph, it is possible for
598: vertex $u$ to be in the domain of $v$, while $v$ is not in the domain of $u$,
599: although this cannot happen in graphs; see Figure
600: \ref{fig:identifiable}.
601:
602: \begin{figure}
603: \begin{center}
604: \includegraphics{identifiable}
605: \caption{Adding a patch on $v$ makes $u$ identifiable, but not
606: vice versa. \label{fig:identifiable}}
607: \end{center}
608: \end{figure}
609:
610: %
611: %
612: \section{Poisson Hypergraph Processes: Markov Properties}
613: \label{sec:markov}
614: %
615: %
616:
617: For $\{\beta_k\}_{k=1}^\infty$ a sequence of non-negative
618: numbers,
619: a \emph{Poisson($\beta$) random hypergraph} is a random hypergraph
620: $\Lambda$ with vertex set $V = \{1,\ldots,N\}$ so
621: that for $A \subset V$,
622: \begin{enumeratei}
623: \item the random variable $\Lambda(A)$ has
624: a Poisson distribution with mean $N \beta_{|A|}/\binom{N}{|A|}$, and
625: \item $\{ \Lambda(A) \,:\, A \subset V\}$ is
626: a collection of independent random variables.
627: \end{enumeratei}
628:
629: In what follows, $\{\rho_k\}_{k=1}^\infty$ will be a probability
630: distribution on the positive integers which has finite mean and
631: \begin{equation} \label{eq:rho1_plus_rho2}
632: \rho_1 + \rho_2 > 0 \,.
633: \end{equation}
634: We now give an explicit construction of the hypergraph-valued
635: stochastic process described in the introduction.
636: Let $K_1, K_2, \ldots$ be a sequence of independent random variables
637: in $\{1,2,3,\ldots\}$ with common distribution
638: $\Pb{K_n = k} = \rho_k$, for all $n,k \in \N$. Denote by
639: $A_1, A_2, \ldots$ a sequence of independent random subsets of $V$,
640: such that $A_n$ is chosen uniformly at random from the subsets of $V$ of
641: size $K_n$ whenever $K_n \leq N$; the set $A_n$ is not defined when
642: $K_n > N$. Let $\{E_t\}_{t \geq 0}$ be a Poisson process, run at rate
643: $N$, having arrival times $\tau_1, \tau_2, \ldots$. Define a stochastic
644: process $\{\Lambda_t\}_{t \geq 0}$ with values in the set of
645: hypergraphs with vertex set $V$ by
646: \begin{equation*}
647: \Lambda_t(A) \deq \sum_{n \,:\, \tau_n \leq t}\one_{\{A = A_n\}} \,.
648: \end{equation*}
649: Interpret $\Lambda_t(A)$ as the number of
650: occurrences of hyperedge $A$ by time $t$.
651: In summary, for each $A \subset V$,
652: \begin{equation} \label{eq:pprate}
653: \{\Lambda_t(A)\} \text{ is a Poisson process of rate }
654: N \frac{\rho_{|A|} }{ \binom{N}{|A|} } \,,
655: \end{equation}
656: and all these Poisson processes are independent. We call
657: $\{\Lambda_t\}_{t \geq 0}$ a \emph{Poisson($\rho$) hypergraph process},
658: where $\rho$ denotes the generating function
659: \begin{equation}
660: \rho(x) \deq \sum_{k \geq 1} \rho_k x^k \,.
661: \end{equation}
662: The finite mean assumption is equivalent to $\rho'(1) < \infty$.
663: For fixed $t \geq 0$, $\Lambda_t$ is a Poisson($t \rho$) random
664: hypergraph.
665:
666: Whereas the hypergraph literature has tended to concentrate on
667: the ``$k$-uniform'' case (i.e.\ $\rho_k = 1$ for some $k$), we
668: find the superposition of $k$-uniform random hypergraphs for
669: various different values of $k$ can be handled without special effort,
670: and leads to asymptotic properties absent from the $k$-uniform
671: case. Moreover the Poisson structure simplifies our arguments, for
672: example by allowing some summary statistics of $\{\Lambda_t\}_{t \geq
673: 0}$ to be Markov processes in their own right: see Proposition
674: \ref{prop:markov_property}. Poissonization is, of course,
675: a well-established procedure -- see \ocite{A:PCH}.
676:
677: Previous literature has also concentrated on the case $\Lambda\le1$.
678: We now sketch a way to deduce from our results for a Poisson($\beta$) random
679: hypergraph $\Lambda$ some corresponding results for $\Lambda \wedge 1$.
680: We note moreover that if $\rho_k = 1$ for some $k$ then $\Lambda \wedge 1$
681: is exactly a $k$-uniform hypergraph.
682: The set of identifiable vertices is the same for $\Lambda$ and
683: $\Lambda \wedge 1$ but $\Lambda$ may have additional identifiable hyperedges.
684: First consider patches. Throwing a Poisson($N\beta$) number of balls
685: (i.e.\ patches) uniformly at random into $N$ urns yields a
686: Binomial($N$,$1 - e^{-\beta_1}$) number of occupied urns (i.e.\
687: vertices covered by at least one patch). Hence the number of patches
688: in $\Lambda$, less the number in $\Lambda \wedge 1$, divided by $N$,
689: has limit in probability $\beta_1 + e^{-\beta_1} - 1$. On the other hand,
690: the expected number of subsets of size at least 2 receiving at least 2
691: hyperedges is bounded uniformly in $N$.
692: Hence, after rescaling by $N^{-1}$, only the extra patches in $\Lambda$
693: can contribute in the limit and of course all of these do so.
694:
695: \begin{prop} \label{prop:markov_property}
696: Let $T_t$ and $Z_t$ denote the numbers of identifiable vertices and
697: identifiable hyperedges for $\Lambda_t$. Both $\{T_t\}_{t \geq 0}$ and
698: $\{(T_t,Z_t)\}_{t \geq 0}$ are Markov processes. The
699: number of non-identifiable hyperedges in $\Lambda_t$, given
700: that $T_t = m$, is conditionally Poisson,
701: with mean
702: \begin{equation} \label{eq:non_identifiable_exactly}
703: N t \left[ 1 - \sum_{k \geq 1} \rho_k
704: \frac{\binom{m}{k} + (N-m) \binom{m}{k-1}}
705: {\binom{N}{k}} \right] \,.
706: \end{equation}
707: When $m - N\gamma = o(N)$, for $\gamma \in [0,1]$, this reduces as $N
708: \rightarrow \infty$ to
709: \begin{equation} \label{eq:nonidentifiable}
710: N t \left[1 - \rho(\gamma) - (1-\gamma)\rho'(\gamma) \right]
711: + o(N) \,.
712: \end{equation}
713: \end{prop}
714: %
715: \begin{rmk}
716: Because the total number of hyperedges in $\Lambda_t$ is
717: Poisson($Nt$), %\eqref{eq:nonidentifiable}
718: Proposition \ref{prop:markov_property} reduces the study of limits
719: of identifiable hyperedges to study of limits of identifiable
720: vertices. In particular, if $N^{-1} T_t$ converges in distribution
721: as $N \rightarrow \infty$ to a random variable $\tilde{T_t}$, then
722: necessarily
723: \begin{equation} \label{eq:zt}
724: N^{-1} Z_t \weakconv t\left[ \rho( \tilde{T}_t )
725: + (1 - \tilde{T}_t) \rho'(\tilde{T}_t) \right] \,.
726: \end{equation}
727: \end{rmk}
728: \begin{rmk}
729: It is easy to identify the generator of $\{T_t\}_{t \geq 0}$,
730: rescale by division by $N$, and take a limit on any compact
731: interval $I \subset \R_+ \setminus \Xi$ (see
732: \eqref{eq:discontinuity_set}); however this approach did not lead to
733: a proof of Theorem \ref{thm:intro_lambda_geq_one}, because of the
734: difficulty of passing through discontinuous phase transitions.
735: \end{rmk}
736: To prepare for the proof, some measure-theoretic
737: apparatus is needed. Let $(\Omega,\F,\P)$ be the
738: probability space on which the process $\{\Lambda_t\}_{t \geq 0}$
739: is defined.
740: For any set $S \subset V$, and any $t \geq 0$,
741: define the $\sigma$-field $\F_t^S \subset \F$ as
742: \begin{equation*}
743: \F_t^S \deq \bigvee_{0 \leq s \leq t}
744: \sigma\{ \Lambda_s(A) \; : \; |A \setminus S| \leq 1 \} \,.
745: \end{equation*}
746: Let $V_t^\star$ denote the set of vertices identifiable at time $t$. By
747: construction, the event $\{ V_t^\star = S \}$ occurs if and only if,
748: %among supersets of the union of all vertices covered by patches,
749: among all sets containing all vertices covered by patches,
750: $S$ is
751: the minimal subset of $V$ for which $\Lambda_t(A) = 0$
752: whenever $|A \setminus S| = 1$. Thus $\{V^\star_t = S\} \in \F_t^S$.
753:
754: When we consider $V^\star_t$ as a ``stopping set'' for a
755: set-indexed process, it becomes natural to define another
756: $\sigma$-field:
757: \begin{equation*}
758: \F_{V^\star_t} \deq \left\{ B \in \F \,:\,
759: B \cap \{V^\star_t = S\} \in \F_t^S
760: \text{ for all } S \subset V \right\} \,.
761: \end{equation*}
762: $T_s$ and $Z_s$ are $\F_{V^\star_t}$-measurable, for all
763: $0 \leq s \leq t$. We may describe $\F_{V^\star_t}$ informally
764: as the knowledge we have about $\{\Lambda_s\}_{0 \leq s \leq t}$
765: after performing hypergraph collapse at each time $s \in [0,t]$.
766:
767: \begin{lem} \label{lem:stopset}
768: \hspace{0.1in}
769:
770: \begin{enumeratei}
771: \item \label{it:stopset1}
772: Fix any $t > 0$. Pick any collection of non-negative integers
773: $\{k_A \,:\, A \subset V \}$,
774: and set
775: \begin{equation*}
776: p(S) \deq \P\!\left( \bigcap_{ A \,:\, |A \setminus S| > 1}
777: \left\{ \Lambda_t(A) = k_A \right\} \right) \,.
778: \end{equation*}
779: Then
780: \begin{equation*}
781: \P\!\left( \bigcap_{A \,:\, |A \setminus V^\star_t| > 1}
782: \left\{ \Lambda_t(A) = k_a \right\} \Given
783: \F_{V^\star_t} \right) = p(V^\star_t) \,.
784: \end{equation*}
785: \item \label{it:stopset2}
786: Fix any $t > 0$. The conditional distribution of the random hypergraph
787: $\Lambda^S_t$ (in the notation of \eqref{eq:LambdaS}),
788: given $\F_{V^\star_t}$, on the event $\{V^\star_t = S\}$, where
789: $|S| = m$, is that of a Poisson($\beta$) random hypergraph on $N - m$
790: vertices with parameters
791: \begin{align}
792: \beta_1 & \deq 0 \nonumber \\
793: \beta_j & \deq \frac{t}{1 - m/N}
794: \binom{N-m}{j} \sum_{i \geq 0} \rho_{i+j}
795: \frac{\binom{m}{i}}{\binom{N}{j+i}} \,, \quad j \geq 2
796: \,. \label{eq:betaj}
797: \end{align}
798: \end{enumeratei}
799: \end{lem}
800: For a random variable $X$, we write $X \sim \text{Poisson}(\mu)$
801: to indicate that the distribution of $X$ is Poisson with
802: expectation $\mu$. Also we will write $X \sim \text{Binomial}(n,p)$
803: to indicate that $X$ is a Binomial random variable with
804: parameters $n$ and $p$.
805: \begin{proof}[Proof of \eqref{it:stopset1}]
806: Certainly $p(V^\star_t)$ is $\F_{V^\star_t}$-measurable. It remains to
807: show that, for any $B \in \F_{V^\star_t}$,
808: \begin{equation*}
809: \int_B p(V^\star_t) d\P = \P\!\left( B \cap \bigcap_{A \,:\,
810: |A \setminus V^\star_t| > 1} \{ \Lambda_t(A) = k_a \} \right) \,.
811: \end{equation*}
812: Split the event on the right into disjoint events by intersecting with
813: $\{V^\star_t = S\}$ for each $S \subset V$. For each $S$,
814: $B \cap \{V^\star_t = S \}$ lies in $\F^S_t$, and therefore
815: is independent of $\{\Lambda_t(A) = k_a\}$ for every $A$ such that
816: $|A \setminus S| > 1$, by construction of a Poisson hypergraph
817: process. The right side becomes
818: \begin{equation*}
819: \sum_{S \subset V} p(S) \P\!\left(B \cap \{V^\star_t = S \}\right)
820: \end{equation*}
821: which is equal to the left side; \eqref{it:stopset1} follows.
822: \end{proof}
823:
824: \begin{proof}[Proof of \eqref{it:stopset2}]
825: Suppose $S \subset V$ and $A \subset V \setminus S$ with
826: $|A| = j \geq 2$. For any $C \subset S$ with $|C|=i$,
827: %Condition on $\{V^\star_t = S\}$, and fix $A \subset V \setminus S$
828: %with $|A| = j \geq 2$. For any $C \subset S$ with $|C|=i$,
829: \eqref{eq:pprate} implies that
830: \begin{equation*}
831: \Lambda_t(A \cup C) \sim
832: {\rm Poisson}\left( t \rho_{j+i} N / \binom{N}{j+i} \right) \,.
833: \end{equation*}
834: The result of part \eqref{it:stopset1} implies that
835: the random variables $\Lambda_t(A \cup C)$
836: are conditionally independent for different choices of
837: $C$, given $\{V^\star_t = S \} \cap \F_{V^\star_t}$.
838:
839: If $|S| = m$, there are $\binom{m}{i}$ choices of $C$, and
840: following the notation of \eqref{eq:LambdaS},
841: \[
842: \Lambda^S_t(A) = \sum_{C \subset S} \Lambda_t(A \cup C)
843: \sim {\rm Poisson}\left( t N \sum_{i \geq 0} \rho_{j+i}
844: \binom{m}{i}/\binom{N}{j+i} \right) \,.
845: \]
846: In a Poisson($\beta$) random hypergraph on $(N-m)$ vertices, the
847: number of occurrences of $A$, where $|A|=j$, is Poisson with parameter
848: \[
849: (N-m) \beta_j/\binom{N-m}{k} \,.
850: \]
851: On comparison with the previous line, this verifies the formula
852: \eqref{eq:betaj} for $\beta_j$, when $j \geq 2$. Clearly there
853: are no $1$-hyperedges in $\Lambda^S_t$ when
854: $\{V^\star_t = S \}$, by definition of identifiability. Hence
855: \eqref{it:stopset2} is established.
856: \end{proof}
857:
858: \begin{proof}[Proof of Proposition \ref{prop:markov_property}]
859: Fix any $t > 0$. Suppose that $T_t = m$. The first jump in
860: the process $\{(T_s,Z_s)\}_{s \geq t}$ can occur only when a new hyperedge
861: arrives, and the arrival time is independent of the past. The law of
862: the jump depends only on two things: the set $A$ of vertices in the new
863: hyperedge (which is independent of the past), and on the hypergraph
864: $\Lambda^S_t$, where $S \deq V^\star_t$.
865: Lemma \ref{lem:stopset}\eqref{it:stopset2}
866: establishes that the law of $\Lambda^S_t$,
867: conditional on $\F_{V^\star_t}$ is fully determined
868: by $m$, $t$, and the parameters $\{\rho_i\}_{i \geq 1}$;
869: in particular it is conditionally independent of
870: $\{(T_s,Z_s)\}_{0 \leq s \leq t}$
871: given that $\{T_t = m\}$. Hence the Markovian
872: property of $\{T_t\}_{t \geq 0}$ and
873: $\{(T_t,Z_t)\}_{t \geq 0}$ is established.
874:
875: It follows from Lemma \ref{lem:stopset} that the total number of non-identifiable
876: hyperedges in $\Lambda_t$, given that $\{T_t = m\}$, is conditionally
877: Poisson, with mean $(N - m) \sum \beta_j$,
878: for $\beta_j$ as in \eqref{eq:betaj}.
879: Write $k \deq i+j$, and switch the order
880: of summation, to obtain
881: \[
882: \left( 1 - \frac{m}{N} \right) \sum \beta_j
883: = t \sum_{k \geq 2} \rho_k \sum_{j=2}^k
884: \binom{N-m}{j}\binom{m}{k-j} / \binom{N}{k} \,.
885: \]
886: On considering the Hypergeometric($(N,N-m,k)$) distribution,
887: we see that the inner sum is
888: \[
889: 1 - \left[ \binom{m}{k} + (N-m) \binom{m}{k-1} \right]
890: / \binom{N}{k} \,.
891: \]
892: The last expression is zero when $k=1$, so $(N-m)\sum \beta_j$ takes
893: the form \eqref{eq:non_identifiable_exactly}. When
894: $m - N \gamma = O(N)$, the last expression converges,
895: as $N \rightarrow \infty$, to $1 - \gamma^k -
896: k\gamma^{k-1}(1-\gamma)$, and is bounded between $0$ and $1$.
897: The Bounded Convergence Theorem yields \eqref{eq:nonidentifiable}.
898: \end{proof}
899:
900: %
901: %
902: \section{Identifiability In Random Hypergraphs With Patches}
903: \label{sec:review}
904: %
905: %
906: In this section we review some material from \ocite{DN:HG}.
907:
908: Fix $t > 0$, and set $\Lambda \deq \Lambda_t$, $\beta_k
909: \deq t\rho_k$. In this case, $\Lambda$ is a Poisson($\beta$) random
910: hypergraph. Suppose we perform hypergraph collapse, described
911: above, in the following special way: at each step the next vertex $v$ to
912: be deleted is selected with a probability proportional to the number
913: of patches on $v$. This is called \emph{randomized collapse}.
914: The \emph{debris} of a hypergraph is the number of hyperedges
915: equal to the empty set.
916: Set
917: $\Lambda_0 \deq \Lambda$, and let
918: $\{\Lambda_n\}_{n \in \N}$ denote the sequence of hypergraphs
919: obtained. Set $Y_n$ and $Z_n$ to be the amount of patches and debris,
920: respectively, in $\Lambda_n$; formally
921: \[
922: Y_n \deq \sum_{v \in V} \Lambda_n(\{v\})\,, \quad \text{and} \quad
923: Z_n \deq \Lambda_n(\emptyset) \,.
924: \]
925: The key observation in \ocite{DN:HG} is that
926: $\{(Y_n,Z_n)\}_{n \in \N}$ is a Markov chain (but not the same one as in
927: Proposition \ref{prop:markov_property}, for here $t$
928: is fixed!), which stops at
929: \begin{equation} \label{eq:T_definition}
930: T \deq \inf\{n \,:\, Y_n=0\} \,.
931: \end{equation}
932: Moreover, conditional on $\{Y_n = m, Z_n = k\}$,
933: \begin{equation} \begin{split} \label{eq:y_and_z}
934: Z_{n + 1} & = k + 1 + W_{n + 1} \,, \\
935: Y_{n + 1} & = m - 1 - W_{n + 1} + U_{n + 1} \,.
936: \end{split} \end{equation}
937: Here $W_{n+1}$ and $U_{n+1}$ are independent, with
938: \begin{equation} \begin{split} \label{eq:u_and_w}
939: W_{n + 1} & \sim {\rm Binomial}\left( m - 1, \frac{1}{N-n} \right) \\
940: U_{n + 1} & \sim {\rm Poisson}\left( (N-n-1)t\lambda_2(N,n) \right)
941: \end{split} \end{equation}
942: where
943: \begin{equation} \label{eq:lambda2}
944: \lambda_2(N,n) \deq N \sum_{i=0}^n \rho_{2 + i}
945: \binom{n}{i} / \binom{N}{i+2} \,.
946: \end{equation}
947:
948: By construction, $T = |V^\star|$, the number of
949: identifiable vertices, and $Z \deq Z_T \deq \Lambda_T(\emptyset)$
950: is the number of identifiable hyperedges. For comparison, note that, by
951: %\eqref{eq:nonreducible},
952: Proposition \ref{prop:markov_property}
953: the number of
954: non-identifiable hyperedges in $\Lambda$, given that
955: $T = N\gamma$, is conditionally Poisson, with mean
956: \begin{equation} \label{eq:poisson_mean_asymp}
957: N(t - \beta(\gamma) -
958: (1 - \gamma)\beta'(\gamma)) + o(N) \,.
959: \end{equation}
960:
961: %By passing to the limit as $N \rightarrow \infty$ for the Markov
962: %chain $\{Y_n\}_{n \in \N}$ after rescaling,
963: We obtained a limit theorem for
964: $\tilde{T}^N \deq N^{-1}T$ and $\tilde{Z}^N \deq N^{-1}Z$,
965: where $Z$ is the number of identifiable hyperedges. We state the result
966: in a simple case. Set
967: \[
968: \beta(x) \deq \sum_k \beta_k x^k \,, \quad
969: x \in [0, 1] \,.
970: \]
971: Assume that $\beta_1 > 0$ and that the derivative $\beta'(1)
972: <\infty$.
973: Then
974: \begin{equation} \label{eq:18}
975: \{x \in [0, 1) \,:\, \beta'(x) + \log(1 - x) < 0 \}
976: \end{equation}
977: is non-empty, and its infimum is $g(t)$, as defined in
978: \eqref{eq:gsdefinition}. By our assumption
979: \eqref{eq:onlytwozeros}, there is at most
980: one $x \in [0,g(t))$ such that $\beta'(x) + \log(1 - x) = 0$,
981: namely $g(t-)$; this is different to $g(t)$ only if
982: $t \in \Xi$, the set of discontinuity points of the lower envelope
983: $g$.
984:
985: Let $\tilde{T}$ be a random variable taking values $g(t)$ and
986: $g(t-)$, each with probability $1/2$. As a special case of
987: of \ocite{DN:HG}*{Theorem 2.2} we know:
988:
989: \begin{thm} \label{thm:fixedtime}
990: The following limit in distribution holds as $N \rightarrow \infty$:
991: \begin{equation} \label{eq:tandzlimits}
992: \left( \tilde{T}^N,\tilde{Z}^N \right) \weakconv
993: \left( \tilde{T}, \beta(\tilde{T}) - (1 - \tilde{T})
994: \log(1 - \tilde{T}) \right) \,.
995: \end{equation}
996: \end{thm}
997:
998: \begin{rmk}
999: \ocite{GN:EE} have shown that the limit for the rescaled number of
1000: identifiable hyperedges can be decomposed as follows: $(1 -
1001: \tilde{T})\log(1 - \tilde{T})$ counts the \emph{essential
1002: hyperedges}, i.e.\ those whose absence would have reduced the set of
1003: identifiable vertices, and $\beta(\tilde{T})$ counts the remainder.
1004: \end{rmk}
1005:
1006: \begin{rmk}
1007: Suppose in particular that $\Lambda \deq \Lambda_t$
1008: and $\beta(x) \deq t\rho(x)$ for some $t \in \Xi$, the
1009: discontinuity set of $g$. Then \eqref{eq:tandzlimits} implies that
1010: the proportion of identifiable vertices has a limit in distribution
1011: which is random, taking the values $g(t)$ and $g(t-)$ each with
1012: probability $1/2$.
1013: \end{rmk}
1014:
1015: \begin{rmk}
1016: It suffices to derive the limit for $\tilde{T}^N$,
1017: since the limit for $\tilde{Z}^N$ follows from Proposition
1018: \ref{prop:markov_property}. To check this, recall that, by
1019: \eqref{eq:zt}, if $\tilde{T}^N$ converges to $g(t)$, then the number
1020: of identifiable hyperedges, divided by $N$, converges to
1021: \begin{equation} \label{eq:formula1}
1022: t \left\{ \rho( g(t) ) + [1 - g(t)]\rho'( g(t) ) \right\} \,.
1023: \end{equation}
1024: However by definition of $g(t)$,
1025: $t\rho'(g(t)) = -\log(1 - g(t))$, so we have recovered the
1026: formula $\beta(\tilde{T}) - (1 - \tilde{T})\log(1 - \tilde{T})$.
1027: \end{rmk}
1028:
1029: %
1030: %
1031: \section{Identifiability In Hypergraph Processes With Patches}
1032: \label{sec:patches}
1033: %
1034: %
1035:
1036: In this section we move from the static random hypergraph model of Theorem
1037: \ref{thm:fixedtime} to the Poisson($\rho$) hypergraph process
1038: $\{\Lambda_t\}_{t \geq 0}$, providing here a proof of
1039: Theorem \ref{thm:intro_lambda_geq_one}.
1040:
1041: Extending the notation of the previous section, let $\tilde{T}^N_t$
1042: and $\tilde{Z}^N_t$ denote the rescaled numbers of identifiable
1043: vertices and hyperedges for $\Lambda_t$, respectively, as
1044: defined in \eqref{eq:T_tilde_N_and_Z_tilde_N_defn}. Note that $t
1045: \mapsto \tilde{T}^N_t$ and $t \mapsto \tilde{Z}^N_t$ are increasing,
1046: right-continuous, stochastic processes. It follows from Proposition
1047: \ref{prop:markov_property} that $\{(\tilde{T}^N_t, \tilde{Z}^N_t)\}_{t
1048: \geq 0}$ is a Markov process.
1049:
1050: \begin{proof}[Proof of Theorem \ref{thm:intro_lambda_geq_one}]
1051: Fix $0 \leq t_1 < \ldots < t_r$. We have to show the
1052: convergence in distribution
1053: \begin{equation} \label{eq:tz_limit_law}
1054: \{ (\tilde{T}^N_{t_i}, \tilde{Z}^N_{t_i}) \}_{i = 1, \ldots, r}
1055: \weakconv \{ (\tilde{T}_{t_i}, \tilde{Z}_{t_i}) \}_{i = 1,
1056: \ldots, r} \,.
1057: \end{equation}
1058: It suffices to do so when at least one of $\{t_i, t_{i + 1}\}$ is not a
1059: discontinuity point, for every $i \in \{1,\ldots,r-1\}$.
1060: Proposition \ref{prop:markov_property} showed that
1061: $\{(\tilde{T}^N_t,\tilde{Z}^N_t)\}_{t \geq 0}$ is Markov,
1062: and for any Markov process $\{Y_t\}_{t \geq 0}$ the conditional
1063: law of $Y_{t_r}$ given $(Y_{t_1},\ldots,Y_{t_{r-1}})$ is the same
1064: as the conditional law given $Y_{t_{r-1}}$. Hence it suffices
1065: to consider the case $r=2$ such that $t_1 \not\in \Xi$ or
1066: $t_2 \not\in \Xi$, and these possibilities are both subsumed in the
1067: case $r=3$ with $t_1,t_3 \not\in \Xi$.
1068: Then only the marginal limit at time
1069: $t_2$, as given in Theorem \ref{thm:fixedtime} is random,
1070: so Theorem \ref{thm:fixedtime} implies the
1071: full convergence in distribution.
1072:
1073: The second assertion follows from the first since all processes are
1074: increasing, and the limit is deterministic and continuous on
1075: $I$.
1076: \end{proof}
1077:
1078: \begin{rmk} \label{rmk:no_conv_in_D}
1079: The rescaled number of essential hyperedges, as studied by
1080: \ocite{GN:EE}, has a limit
1081: $\{-(1 - \tilde{T}_t)\log(1 - \tilde{T}_t) \}_{t \geq 0}$ in the
1082: same sense as \eqref{eq:fluid_limit_with_patches} and
1083: \eqref{eq:convinprob}.
1084: \end{rmk}
1085:
1086: \begin{rmk}
1087: One may ask whether the convergence
1088: \eqref{eq:fluid_limit_with_patches} extends to weak convergence in
1089: the Skorohod space $D([0,\infty),\R_+^2)$. Since $t \mapsto
1090: \tilde{T}^N_t$ and $t \mapsto \tilde{Z}^N_t$ are non-decreasing, the
1091: necessary and sufficient condition of \ocite{JS:LT1}*{p.\ 306}
1092: may be applied, which would require that the sum of squared jumps of
1093: $\{\tilde{T}^N_t\}$ converges in law to the sum of squared jumps of
1094: $\{\tilde{T}_t\}$, and similarly for
1095: $\{\tilde{Z}^N_t\}$. Unfortunately the techniques presented in this
1096: paper do not seem to be able to confirm this; indeed, it seems
1097: plausible that, for arbitrarily large $N$, and for $t \in \Xi$,
1098: there is a probability bounded away from zero that $\tilde{T}^N_s$
1099: makes more than one jump in going from $\approx g(t-)$ to $\approx
1100: g(t)$ at time $s \approx t$, and this would contradict the condition
1101: stated.
1102: \end{rmk}
1103:
1104: \begin{rmk}
1105: If \eqref{eq:onlytwozeros} is false, one can reformulate the process
1106: \eqref{eq:ifl}, by consulting \ocite{DN:HG}*{Theorem 2.2} and prove
1107: a corresponding version of Theorem \ref{thm:intro_lambda_geq_one}.
1108: \end{rmk}
1109:
1110: \section{Domain Of A Vertex In A Hypergraph Without Patches}
1111: \label{sec:patch-free_static}
1112:
1113: We revert to the fixed-time setting of Section
1114: \ref{sec:review}. Suppose $\Lambda$ is a Poisson($\beta$) random
1115: hypergraph, such that
1116: \[
1117: \beta_0 = \beta_1 = 0 < \beta_2\,, \quad
1118: \beta(x) \deq \sum_{k \geq 2}\beta_k x^k\,, \
1119: x \in [0, 1]\,.
1120: \]
1121:
1122: Fix a vertex $v_0$. Write $T^N$ for the number of vertices in the domain
1123: of $v_0$, and write $Z^N$ for the number of hyperedges identifiable from
1124: $v_0$. Set $\bar{T}^N \deq N^{-1}T^N$ and $\bar{Z}^N \deq N^{-1}Z^N$. Both
1125: the \emph{microscopic variables} $(T^N, Z^N)$, and the
1126: \emph{macroscopic variables}
1127: $(\bar{T}^N, \bar{Z}^N)$ have non-trivial limits as $N \rightarrow
1128: \infty$, which we now describe. The coefficient $\beta_2$ plays a distinguished
1129: role.
1130:
1131: \begin{lem}
1132: Let $\{\xi_n\}_{n \in \N}$ be a random walk on the integers, started
1133: at $\xi_0 = 1$, whose increments are of the form
1134: $\xi_n - \xi_{n - 1} = -1 + {\rm Poisson}(2\beta_2)$. Let $\phi$ be
1135: the largest root in $[0,1]$ of $2 \beta_2 x +
1136: \log(1 - x) = 0$, so $\phi = 0$ for $2 \beta_2 \leq 1$, and
1137: $0 < \phi < 1$ otherwise. Then the first passage time to $0$,
1138: \begin{equation} \label{eq:M_definition}
1139: M \deq \inf\{n \geq 0 \,:\, \xi_n = 0 \} \,,
1140: \end{equation}
1141: has the following distribution:
1142: \begin{equation}
1143: \begin{split} \label{eq:bp_population}
1144: \Pb{M = n} & = e^{-2 \beta_2 n} \left(
1145: 2 \beta_2 n \right)^{n - 1} /
1146: n!\,, \quad n \in \N \,; \\
1147: \Pb{M = \infty} & = \phi \,,
1148: \end{split}
1149: \end{equation}
1150: \end{lem}
1151: \begin{rmk*}
1152: $M$ is distributed as the total number of individuals in a
1153: branching process with one ancestor, and Poisson($2\beta_2$) offspring
1154: distribution. This distribution describes the sizes of small
1155: components in an Erd\H{o}s-R\'enyi random graph; see
1156: \ocite{B:RG}.
1157: \end{rmk*}
1158:
1159: \begin{proof}
1160: The fact that $\Pb{M = \infty} = \phi$ is an elementary fact from the
1161: theory of branching processes. The formula for $\Pb{M=n}$ is a special
1162: case of a formula of \ocite{D:TP}, which is proved in detail on p.\ 300
1163: of \ocite{D:BP}.
1164: \end{proof}
1165:
1166: Assume that $\beta'(1) < \infty$. Then the set
1167: \eqref{eq:18} is non-empty, and
1168: its infimum is $g \deq g(t)$, as defined in
1169: \eqref{eq:gsdefinition}. Assume further that
1170: $\beta'(x) + \log(1 - x) > 0$ for all $x \in (0,g)$. If either of these
1171: assumptions fail, then the techniques of \ocite{DN:HG},
1172: combined with some arguments given below, still establish the desired
1173: asymptotics. We omit the details.
1174:
1175: Set
1176: \begin{equation} \begin{split}
1177: \bar{T} & \deq g \one_{\{M = \infty\}} \,; \\
1178: \bar{Z} & \deq \left[ \beta(g) - (1 - g)\log(1 - g) \right]
1179: \one_{\{M = \infty\}}\,.
1180: \end{split} \end{equation}
1181:
1182: \begin{thm} \label{thm:identif_sans_patches}
1183: Consider a Poisson random hypergraph without patches, and fix a distinguished
1184: vertex $v_0$. The number of vertices in the domain of $v_0$, and number
1185: of hyperedges identifiable from $v_0$, obey the following limits in
1186: distribution as $N \rightarrow \infty$:
1187: \begin{equation}
1188: (T^N, Z^N) \weakconv (M, M)\,; \quad
1189: (\bar{T}^N, \bar{Z}^N) \weakconv (\bar{T}, \bar{Z})\,.
1190: \end{equation}
1191: Here $M$ is considered as a random variable taking values in the
1192: one-point compactification $\N \cup \{\infty\}$ of $\N$.
1193: \end{thm}
1194:
1195: \begin{proof}
1196: \emph{Step I}. Set $\Lambda_0 \deq \Lambda + \one_{\{v_0\}}$, and let
1197: $\{\Lambda_n\}_{n \in \N}$ be a sequence of hypergraphs obtained
1198: by randomized collapse. Denote by $Y^N_n$ and $Z^N_n$ the numbers of
1199: patches and debris, respectively, in $\Lambda_n$. Then
1200: \[
1201: T^N \deq \inf\{n \geq 0 \,:\, Y^N_n = 0\}\,;
1202: \quad Z^N \deq Z^N_{T^N} \,.
1203: \]
1204:
1205: We know that $\{(Y^N_n, Z^N_n)\}_{n \geq 0}$ is a Markov chain,
1206: starting from $(1,0)$: the increments, conditional on
1207: $Y^N_n = m \geq 1$ and $Z^N_n = k$, are as given in
1208: \eqref{eq:y_and_z} and \eqref{eq:u_and_w}.
1209:
1210: For fixed $n \geq 0$ and $m \geq 1$, the random variable $W_{n+1}$
1211: defined in \eqref{eq:u_and_w} converges to $0$ in
1212: distribution as $N \rightarrow \infty$. Also
1213: \begin{equation} \label{eq:lambda2_limit}
1214: (N - n - 1) \lambda_2(N, n) \rightarrow 2 \rho_2 \,.
1215: \end{equation}
1216: so the random variable $U_{n+1}$ defined in
1217: \eqref{eq:u_and_w}
1218: converges to Poisson($2\beta_2$) in distribution as
1219: $N \rightarrow \infty$. Hence, for all $n \geq 0$,
1220: \[
1221: \{(Y^N_j, Z^N_j)\}_{0 \leq j \leq n}
1222: \weakconv \{(\xi_j, j)\}_{0 \leq j \leq n}
1223: \]
1224: which implies $(T^N, Z^N) \weakconv (M, M)$ as $N \rightarrow \infty$.
1225: If $2\beta_2 \leq 1$, then $\Pb{M = \infty} = 0$,
1226: so the proof is complete. It only remains to prove the second
1227: convergence assertion in the case where $2\beta_2 > 1$, and
1228: $0 < \phi < 1$.
1229:
1230: \medskip
1231: \emph{Step II}. Introduce an auxiliary time variable $t$, and let
1232: $\{\nu_t\}_{t \geq 0}$ be a Poisson process of rate $N$. Set
1233: \begin{align*}
1234: \bar{Y}^N_t & \deq N^{-1}Y^N_{\nu_t} \,, \\
1235: \bar{Z}^N_t & \deq N^{-1}Z^N_{\nu_t} \,, \\
1236: \bar{\nu}^N_t & \deq N^{-1}\nu_t \,, \\
1237: \tau^N & \deq \inf\{t \geq 0 \,:\,
1238: \bar{Y}^N_t = 0 \} \,.
1239: \end{align*}
1240: With reference to \ocite{DN:HG}, set
1241: \begin{align*}
1242: y(t) & \deq (1 - t)(\beta'(t) + \log(1 - t)) \,;\\
1243: z(t) & \deq \beta(t) - (1-t)\log(1 - t) \,.
1244: \end{align*}
1245: By Theorem 6.1 and Remark 6.2 of \ocite{DN:HG},
1246: for all $\delta > 0$,
1247: \begin{equation} \label{eq:old_fluid_limit_result}
1248: \limsup_{N \rightarrow \infty} \frac{1}{N}
1249: \log\left( \Pb{ \sup_{t \leq \tau^N}
1250: \left\| (\bar{\nu}^N_t,\bar{Y}^N_t, \bar{Z}^N_t) -
1251: (t, y(t), z(t)) \right\| > \delta} \right) < 0 \,.
1252: \end{equation}
1253: Observe that $\bar{\nu}^N_{\tau^N} = \bar{T}^N$, which will have the
1254: same limit in probability as does $\tau^N$. We will show that, for
1255: all $\theta \in (\log(1-\phi)],0)$, there exists $\delta > 0$
1256: and $N_0$ such that
1257: \begin{equation} \label{eq:unlikely_to_die_before_delta}
1258: \Pb{\bar{T}^N \leq \delta} \leq e^{\theta} \,,
1259: \quad \text{for all } N \geq N_0 \,.
1260: \end{equation}
1261: By \eqref{eq:bp_population} and the fact
1262: that $T^N \weakconv M$, we know that, for all
1263: $\delta > 0$ and all $\phi' > \phi$:
1264: \begin{equation*}
1265: \Pb{\bar{T}^N \leq \delta} \geq 1 - \phi'
1266: \end{equation*}
1267: for all sufficiently large $N$. Also from
1268: \eqref{eq:old_fluid_limit_result} we obtain, for all
1269: $\delta > 0$,
1270: \begin{equation*}
1271: \Pb{\bar{T}^N \in (\delta, g - \delta) \cup (g + \delta, \infty)}
1272: \rightarrow 0
1273: \end{equation*}
1274: as $N \rightarrow \infty$. Hence the claim that
1275: $(\bar{T}^N, \bar{Z}^N) \weakconv (\bar{T}, \bar{Z})$ will follow
1276: as soon as we have proved \eqref{eq:unlikely_to_die_before_delta};
1277: then \eqref{eq:old_fluid_limit_result} will strengthen this
1278: to show $(\bar{T}^N,\bar{Z}^N) \weakconv (\bar{T},\bar{Z})$.
1279:
1280: \medskip
1281: \emph{Step III}. The remainder of the proof is to establish
1282: \eqref{eq:unlikely_to_die_before_delta}. Given $Y^N_n = m \geq 1$, set
1283: \begin{align*}
1284: \Phi^N(m, n) & \deq
1285: \E \exp\left\{\theta(-1 - W_{n + 1} + U_{n + 1}) \right\} \\
1286: & = \exp\left\{-\theta + F\left(m - 1, \frac{1}{N - n},
1287: -\theta\right) + G((N - n - 1)\lambda_2(N, n), \theta) \right\} \,.
1288: \end{align*}
1289: where
1290: \begin{equation*}
1291: F(k, p, \theta) \deq k \log\left(1 - p + p e^{\theta} \right) \,;
1292: \quad G(\mu, \theta) \deq \mu(e^{\theta} - 1) \,.
1293: \end{equation*}
1294: Lemma 6.1 of \ocite{DN:HG} implies that
1295: \begin{equation*}
1296: \sup_{n \leq N/2} \left|(N - n - 1)\lambda_2(N, n) -
1297: \left(1 - \frac{n}{N} \right) \beta''(n/N) \right|
1298: \rightarrow 0 \,,
1299: \end{equation*}
1300: as $N \rightarrow \infty$. Since $\theta > \log(1 - \phi)$, there is
1301: $\bar{\phi} < \phi$ such that $\theta > \bar{\theta}
1302: \deq \log(1 - \bar{\phi})$; by construction of $\phi$,
1303: $2\beta_2\bar{\phi} + \log(1 - \bar{\phi}) > 0$, so
1304: $2 \beta_2 (1 - e^{\bar{\theta}}) + \bar{\theta} > 0$;
1305: in other words,
1306: \begin{equation*}
1307: \exp\{ - \bar{\theta} + G(2\beta_2, \bar{\theta}) \} < 1 \,.
1308: \end{equation*}
1309: We can therefore find $\delta > 0$ and $N_0$ such that
1310: \begin{equation} \label{eq:phi_bound}
1311: \Phi^N(m, n) \leq 1\,, \quad \text{for all } m,\ n \leq N\delta,
1312: \quad \text{for all } N \geq N_0 \,.
1313: \end{equation}
1314: Consider the martingale
1315: \begin{equation*}
1316: M_n \deq e^{\bar{\theta} Y^N_n}
1317: \left(\prod_{k = 0}^{n-1} \Phi^N( Y^N_k,k ) \right)^{-1} \,,
1318: \end{equation*}
1319: and set $R^N \deq \inf\{n \geq 0 \,:\, Y^N_n \geq N\delta \}$. It
1320: follows from \eqref{eq:phi_bound} that, on
1321: the event $\{ T^N \leq R^N \wedge N\delta\}$,
1322: \[
1323: M_{T^N} \geq 1\,, \quad \text{for all } N \geq N_0 \,.
1324: \]
1325: Hence for $N \geq N_0$,
1326: \[
1327: e^{\theta} > \E M_0 = e^{\bar{\theta}} =
1328: \E M_{T^N \wedge R^N \wedge N\delta}
1329: \geq \Pb{T^N \leq R^N \wedge N\delta} \,.
1330: \]
1331: However \eqref{eq:old_fluid_limit_result} implies that, for
1332: $\delta < g/2$, $\Pb{R^N < T^N \leq N\delta} \rightarrow 0$,
1333: and \eqref{eq:unlikely_to_die_before_delta} follows.
1334: \end{proof}
1335:
1336: %
1337: %
1338: \section{Identifiability In Patch-Free Processes}
1339: \label{sec:patch-free_process}
1340: %
1341: %
1342: We now focus on the case of patch-free hypergraph processes,
1343: proving in this section Theorem \ref{thm:intro_no_patch_process}.
1344:
1345: \subsection{A Coupled Family of Random Walks}
1346: Let $\{P_t(n)\}_{t \geq 0}$, $n \in \N$, be a family of independent Poisson
1347: processes, all of rate $2 \rho_2 > 0$, and consider the coupled family
1348: of random walks $\{\xi_t(n)\}_{n \geq 0}$, for
1349: $t \in \R_+$, where $\xi_t(0) = 1$ for all $n$, and
1350: \begin{gather}
1351: \xi_t(n + 1) = \xi_t(n) + (P_t(n + 1) - 1) \one_{\{n < M_t\}} \,; \\
1352: M_t \deq \inf\{n \geq 0 \,:\, \xi_t(n) = 0\}
1353: \in \N \cup \{ \infty \} \,. \label{eq:rw_extinction_time}
1354: \end{gather}
1355:
1356: The marginal law of $M_t$ is given by \eqref{eq:bp_population}
1357: with $\beta_2 \deq t\rho_2$. There is a relation between
1358: $\{\xi_t(n)\}_{n \geq 0}$ and the multigraph structure function:
1359: since $g_2(t)$ is the largest root in $[0,1]$ of $2 t \rho_2 x +
1360: \log(1 - x) = 0$, we have as a special case of
1361: \eqref{eq:bp_population}:
1362:
1363: %\addtocounter{subsection}{1}
1364: \begin{lem}
1365: The first time $t$ at which $\{\xi_t(n)\}_{n \geq 0}$ escapes to
1366: infinity is related to the multigraph lower envelope \eqref{eq:g2s}
1367: as follows:
1368: \begin{equation}
1369: \Pb{M_t = \infty} = g_2(t) \,.
1370: \end{equation}
1371: Moreover $t \mapsto M_t$ is an increasing process by the coupling,
1372: so $\chi \deq \inf \{t \geq 0 \,:\, M_t = \infty\}$ is a continuous
1373: random variable with distribution function $g_2(t)$.
1374: \end{lem}
1375:
1376: \subsection{Notation}
1377: \label{sec:no_patch}
1378: We finally turn to the case of a Poisson($\rho$) hypergraph process
1379: $\{\Lambda_t\}_{t \geq 0}$ without patches, i.e.\ such that
1380: \[
1381: \rho_0 \deq \rho_1 \deq 0 < \rho_2\,,
1382: \quad \rho(x) \deq \sum_{k \geq 2} \rho_k x^k \,,
1383: x \in [0, 1] \,.
1384: \]
1385: Write $T^N_t$ for the number of vertices in the domain of $v_0$ in
1386: $\Lambda_t$, and write $Z^N_t$ for the number of hyperedges
1387: identifiable from $v_0$ in $\Lambda_t$. Set
1388: $\bar{T}^N_t \deq N^{-1}T^N_t$ and
1389: $\bar{Z}^N_t \deq N^{-1}Z^N_t$. Using
1390: \eqref{eq:rw_extinction_time}, we define what will turn out to be the
1391: macroscopic limits for Theorem \ref{thm:intro_no_patch_process}.
1392: \begin{align*}
1393: \bar{T}_t & \deq g(t) \one_{\{M_t = \infty\}} \,; \\
1394: \bar{Z}_t & \deq \left\{ t\rho(g(t)) - [1 - g(t)]\log(1 - g(t))
1395: \right\} \one_{\{M_t = \infty\}} \,.
1396: \end{align*}
1397:
1398: \begin{proof}[Proof of Theorem \ref{thm:intro_no_patch_process}]
1399: \hspace{0.1in}
1400:
1401: \emph{Step I}. Extending the notation of Theorem
1402: \ref{thm:identif_sans_patches} %\ref{thm:no_patch_process}
1403: let $\Lambda_t(n)$ denote the hypergraph that
1404: results from applying $n$ steps of randomized collapse to
1405: $\Lambda_t + \one_{\{v_0\}}$; $Y^N_t(n)$ and $Z^N_t(n)$ count the number
1406: of patches, and the amount of debris, respectively in
1407: $\Lambda_t(n)$, and $n$ is assumed to satisfy:
1408: \[
1409: n \leq T^N_t \deq
1410: \inf\{n \geq 0 \,:\, Y^N_t(n) = 0 \} \,.
1411: \]
1412: Consider a finite set of time points
1413: $0 < t_1 < \ldots < t_r$.
1414: The hypergraph collapses of
1415: $\Lambda_{t_1} + \one_{\{v_0\}}, \ldots,
1416: \Lambda_{t_r} + \one_{v_0}$ are coupled together as follows: perform
1417: the $(n+1)$st step of randomized collapse by choosing a patch
1418: uniformly at random from the smallest unstable hypergraph. Poisson
1419: symmetries imply that this amounts to randomized collapse for each
1420: of the unstable hypergraphs. Condition on the event:
1421: \begin{equation} \label{eq:condition_on_m_and_k}
1422: \bigcap_{i = 1}^{r} \{ Y^N_{t_i}(n) = m_i\,,
1423: \ Z^N_{t_i}(n) = k_i \} \,.
1424: \end{equation}
1425: For $i$ such that $m_i = 0$, evidently $Y^N_{t_i}(n + 1) = 0$
1426: and $Z^N_{t_i}(n + 1) = k_i$.
1427: For those $i$ such that $m_i \geq 1$, we may write:
1428: \begin{align*}
1429: Y^N_{t_i}(n + 1) & = m_i - 1 - W^N_{t_i}(n + 1) + U^N_{t_i}(n + 1)\,; \\
1430: Z^N_{t_i}(n + 1) & = k_i + 1 + W^N_{t_i}(n + 1)\,,
1431: \end{align*}
1432: where the random increments are distributed as follows.
1433: Take $q$ to be the least $i \in \{1, 2, \ldots, r\}$ for which
1434: $m_i \geq 1$, and take $W^N_{t_q}(n + 1)$ and $U^N_{t_q}(n + 1)$
1435: independent such that
1436: \begin{equation} \begin{split}
1437: W^N_{t_q}(n + 1) & \sim {\rm Binomial}\left( m_q - 1,
1438: \frac{1}{N-n}\right) \,; \\
1439: U^N_{t_q}(n + 1) & \sim {\rm Poisson} \left(
1440: (N - n - 1) t_q \lambda_2(N-n) \right) \,,
1441: \end{split} \end{equation}
1442: where $\lambda_2(N, n)$ is as in
1443: \eqref{eq:lambda2}. Because of the coupling, we may take subsequent
1444: increments (for $i = q, \ldots, r - 1$) to be independent and of the form:
1445: \begin{align*}
1446: W^N_{t_{i+1}}(n + 1) - W^N_{t_i}(n + 1) & \sim
1447: {\rm Binomial} \left(m_{i+1} - m_i, \frac{1}{N-n} \right) \,; \\
1448: U^N_{t_{i+1}}(n + 1) - U^N_{t_i}(n + 1) & \sim
1449: {\rm Poisson} \left( (N - n - 1)(t_{i+1} - t_i) \lambda_2(N,n)
1450: \right) \,.
1451: \end{align*}
1452:
1453: \emph{Step II}. Observe that the behavior of $\lambda_2(N, n)$ depends
1454: on whether $n \deq O(1)$, or $n \deq O(N)$. It follows from
1455: \eqref{eq:lambda2_limit} and the calculations in Step I that,
1456: conditional on \eqref{eq:condition_on_m_and_k}, the joint law of
1457: \[
1458: \left( (Y^N_{t_1}(n + 1), Z^N_{t_1}(n + 1)), \ldots,
1459: (Y^N_{t_r}(n + 1), Z^N_{t_r}(n + 1)) \right)
1460: \]
1461: converges as $N \rightarrow \infty$ to the conditional law of
1462: \[
1463: \left( (\xi_{t_1}(n + 1), k_1 + 1), \ldots,
1464: (\xi_{t_r}(n + 1), k_r + 1) \right)
1465: \]
1466: given that
1467: $\xi_{t_1}(n) = m_1, \ldots, \xi_{t_r}(n) = m_r$.
1468: Evidently $Z_{t_i}^N(0) = 0$ for all $i$.
1469: Since $n$ was arbitrary, and since for each $t$ both
1470: $\{\xi_t(n)\}_{n \geq 0}$ and
1471: $\{ (Y^N_t(n), Z^N_t(n)) \}_{n \geq 0}$
1472: are Markov, we have now proved convergence in distribution as
1473: $N \rightarrow \infty$:
1474: \begin{multline*}
1475: \{ (Y^N_{t_1}(n), Z^N_{t_1}(n)), \ldots,
1476: (Y^N_{t_r}(n), Z^N_{t_r}(n)) \}_{n \geq 0} \\
1477: \weakconv
1478: \{ (\xi_{t_1}(n), n \wedge M_{t_1}), \ldots,
1479: (\xi_{t_r}(n), n \wedge M_{t_r})\}_{n \geq 0} \,.
1480: \end{multline*}
1481: In particular, in the notation of
1482: \eqref{eq:rw_extinction_time} and Section
1483: \ref{sec:no_patch},
1484: \begin{equation} \label{eq:fd_distribs}
1485: \left( (T^N_{t_1}, Z^N_{t_1}), \ldots,
1486: (T^N_{t_r}, Z^N_{t_r}) \right)
1487: \weakconv
1488: \left( (M_{t_1}, M_{t_1}), \ldots,
1489: (M_{t_r}, M_{t_r}) \right) \,.
1490: \end{equation}
1491:
1492: \emph{Step III}. To prove
1493: \eqref{eq:micro_no_patch_fluid_limit}
1494: it suffices, in the light of
1495: \eqref{eq:fd_distribs},
1496: to prove tightness of
1497: $\{(T^N_t, Z^N_t)\}_{t \geq 0}$ with
1498: respect to the Skorohod topology of
1499: $D([0, \infty), (\N \cup \{\infty\})^2)$.
1500: On $(\N \cup \{\infty\})^2$, we shall use the metric
1501: \[
1502: d\left( (m,n),(p,q) \right) \deq
1503: \max\left\{\left|\frac{1}{p} - \frac{1}{m} \right|,
1504: \left|\frac{1}{q} - \frac{1}{n} \right|\right\} \,
1505: \]
1506: understanding that $1/\infty = 0$. We shall
1507: verify the condition of Aldous for tightness of
1508: $\{(T^N_t, Z^N_t)\}_{t \geq 0}$, as stated
1509: in \ocite{B:CPM2}, p.\ 176, or \ocite{K:FMP},
1510: p.\ 314, with respect to this metric.
1511: Since $s \mapsto T^N_s$ and $s \mapsto Z^N_s$ are
1512: non-decreasing processes, the condition takes a slightly simpler form
1513: than usual: it suffices to show that, for each $\epsilon > 0$ and
1514: $\eta > 0$, there exist $h$ and $N_0$ such that for every bounded sequence
1515: of optional times $\sigma^N$ with respect to
1516: $\{(T^N_t, Z^N_t)\}_{t \geq 0}$,
1517: and for every $N \geq N_0$,
1518: \begin{equation} \label{eq:aldous_crit_1}
1519: \Pb{\max\left\{\left|\frac{1}{T^N_{\sigma + h}} -
1520: \frac{1}{T^N_{\sigma}}\right|,
1521: \left|\frac{1}{Z^N_{\sigma + h}}
1522: - \frac{1}{Z^N_{\sigma}}\right| \right\} \geq
1523: \ep } < \eta \,,
1524: \end{equation}
1525: where $\sigma$ is short for $\sigma^N$ in the subscripts.
1526:
1527: Proposition \ref{prop:markov_property}
1528: established
1529: that $\{ (T^N_t, Z^N_t) \}_{t \geq 0}$ is a Markov process. By the strong
1530: Markov property, the conditional law of
1531: $T^N_{\sigma + h} - m^N$, given
1532: that $T^N_\sigma = m \deq m^N$, and $Z^N_\sigma = q^N$, is that
1533: same as that of the number of identifiable vertices in a
1534: Poisson($\hat{\beta}$) random hypergraph $\hat{\Lambda}^N$
1535: on $\hat{N} \deq N - m$
1536: vertices, where by the reasoning of Lemma
1537: \ref{lem:stopset} %\ref{sec:residual_hg},
1538: and the fact that $\rho_1 = 0$,
1539: \[
1540: \hat{\beta}_1 \deq \frac{hN}{N - m}
1541: \sum_{k \geq 2} \rho_k \binom{m}{k - 1}
1542: \binom{N - m}{1} / \binom{N}{k} \,.
1543: \]
1544: Suppose $\epsilon > 0$ and $\eta > 0$ are given. In the case where
1545: $\min\{m^N, q^N\} > 1/\epsilon$, it follows that
1546: \begin{equation}
1547: \max\left\{ \left| \frac{1}{T^N_{\sigma + h}}
1548: - \frac{1}{T^N_\sigma} \right|, \
1549: \left| \frac{1}{Z^N_{\sigma + h}} -
1550: \frac{1}{Z^N_\sigma} \right| \right\} < \epsilon \,.
1551: \end{equation}
1552:
1553: On the other hand, if $m^N \leq 1/\epsilon$, then
1554: \[
1555: \hat{\beta}_1 \hat{N} \leq
1556: h N^2 \sum_{k \geq 2} \rho_k \binom{2/ \ep}{k - 1}
1557: / \binom{N}{k} = \frac{2 h \rho_2}{\ep} +O(N^{-1}) \,.
1558: \]
1559: Choose $N_0$ so large that, for $N \geq N_0$, the right side is not more than
1560: $3 h \rho_2/\epsilon$; now it is true that, for any
1561: \[
1562: h \leq \frac{-\epsilon \log(1 - \eta)}{3\rho_2} \,,
1563: \]
1564: and for any $N \geq N_0$, the probability that $\hat{\Lambda}^N$ has no
1565: patches, and hence no identifiable vertices nor identifiable hyperedges,
1566: is at least $1-\eta$; in that case, $T^N_{\sigma + h}
1567: = T^N_\sigma$ and
1568: and $Z^N_{\sigma + h} = Z^N_{\sigma}$. In summary, for such $N$ and
1569: $h$,
1570: \eqref{eq:aldous_crit_1} holds. Hence
1571: $\{(T^N_t, Z^N_t)\}_{t \geq 0}$ is tight,
1572: and \eqref{eq:micro_no_patch_fluid_limit}
1573: follows.
1574:
1575: \emph{Step IV}. As for \eqref{eq:macro_no_patch_fluid_limit}
1576: %\eqref{eq:micro_no_patch_fluid_limit},
1577: we need only check the convergence of finite-dimensional
1578: distributions, i.e.\ that
1579: \begin{equation} \label{eq:macro_fd_distribs}
1580: \left( (\bar{T}^N_{t_1}, \bar{Z}^N_{t_1}), \ldots,
1581: (\bar{T}^N_{t_r}, \bar{Z}^N_{t_r}) \right)
1582: \weakconv
1583: \left( (\bar{T}_{t_1}, \bar{Z}_{t_1}), \ldots,
1584: (\bar{T}_{t_r}, \bar{Z}_{t_r}) \right) \,.
1585: \end{equation}
1586: for every finite set of time points $0 < t_1 < \ldots < t_r$. For the
1587: case $r = 1$, the validity of \eqref{eq:macro_fd_distribs}
1588: follows from Theorem
1589: \ref{thm:identif_sans_patches}.
1590: For the sake of brevity, restrict our discussion of the
1591: case $r > 1$ to the $\bar{T}$ component; the argument for
1592: the $\bar{Z}$ component is similar. It suffices to show,
1593: for all $q = 2,\ldots,r$, and all
1594: $\epsilon > 0$, that
1595: \begin{equation} \label{eq:jump_to_infinity}
1596: \P\!\left(\bigcap_{\substack{ i,j \\ 1 \leq i < q \leq j \leq r}}
1597: \{ \bar{T}^N_{t_i} < \ep \} \cap
1598: \{ | \bar{T}^N_{t_j} - g(t_j)| < \ep \} \right)
1599: \rightarrow
1600: \Pb{M_{t_{q-1}} < \infty = M_{t_q}} \,.
1601: \end{equation}
1602:
1603:
1604: By our knowledge of the finite dimensional distributions from Theorem
1605: \ref{thm:identif_sans_patches},
1606: the left side of
1607: is well approximated by
1608: \[
1609: 1 - \Pb{\bar{T}^N_{t_{q - 1}} \geq \epsilon}
1610: - \Pb{\bar{T}^N_{t_q} \leq g(t_q) - \epsilon} \,,
1611: \]
1612: and for $\epsilon$ sufficiently small, this converges to the right
1613: side of \eqref{eq:jump_to_infinity}.
1614: \end{proof}
1615:
1616: \section{Future Directions} \label{sec:future}
1617:
1618: We have not explained here the role of the upper envelope
1619: \eqref{eq:g_star_definition}, even though it was
1620: included in the classification of structure functions. It is related
1621: to dual hypergraph collapse and the size of the core, as in
1622: \ocite{C:core}.
1623: We shall give the corresponding asymptotic results in a
1624: future paper.
1625:
1626: \section*{Acknowledgments} We thank Peter Matthews and the
1627: referees for suggesting various expository improvements.
1628:
1629: \begin{bibdiv}
1630: \begin{biblist}
1631:
1632: \bib{A:PCH}{book}{
1633: author = {Aldous, D.},
1634: title = {Probability Approximations via the Poisson Clumping
1635: Heuristic},
1636: publisher = {Springer},
1637: year = {1989},
1638: }
1639:
1640: \bib{B:CPM2}{book}{
1641: author = {Billingsley, P.},
1642: title = {Convergence of Probability Measures},
1643: publisher = {John Wiley},
1644: year = {1999},
1645: edition = {Second edition},
1646: }
1647:
1648: \bib{B:RG}{book}{
1649: author = {Bollob{\'a}s, B.},
1650: title = {Random Graphs},
1651: publisher = {Cambridge University Press},
1652: year = {2001},
1653: edition = {Second edition},
1654: }
1655:
1656: \bib{C:core}{article}{
1657: author = {Cooper, C.},
1658: title = {The size of the cores of random graphs with a
1659: given degree sequence},
1660: note = {Preprint},
1661: address = {Goldsmiths College, University of London},
1662: year = {2002},
1663: }
1664:
1665: \bib{DN:HG}{article}{
1666: author = {Darling, R.W.R.},
1667: author = {Norris, J.R.},
1668: title = {Structure of large random hypergraphs},
1669: journal = {Ann. App. Probab.},
1670: note = {to appear},
1671: year = {2004},
1672: eprint = {arXiv:math.PR/0109020},
1673: }
1674:
1675: \bib{D:BP}{article}{
1676: author = {Devroye, L.},
1677: title = {Branching processes and their applications in the
1678: analysis of tree structures and tree algorithms},
1679: booktitle = {Probabilistic Methods for Algorithmic Discrete
1680: Mathematics},
1681: publisher = {Springer},
1682: year = {1998},
1683: editor = {Habib, M},
1684: editor = {McDiarmid, C},
1685: editor = {Ramirez-Alfonsin, J},
1686: editor = {Reed, B},
1687: }
1688:
1689: \bib{D:HG}{article}{
1690: author = {Duchet, P.},
1691: title = {Hypergraphs},
1692: booktitle = {Handbook of Combinatorics},
1693: publisher = {Elsevier Science B.V.},
1694: year = {1995},
1695: editor = {Graham, R},
1696: editor = {Gr{\"o}tschel, M},
1697: editor = {Lov{\'a}sz, L},
1698: }
1699:
1700: \bib{D:TP}{article}{
1701: author = {Dwass, M.},
1702: title = {The total progeny in a branching process},
1703: journal = {Journal of Applied Probability},
1704: year = {1969},
1705: volume = {6},
1706: pages = {682\ndash 686},
1707: }
1708:
1709: \bib{F:core}{article}{
1710: author = {Fountoulakis, N.},
1711: title = {On the structure of the core of sparse random
1712: graphs},
1713: note = {Preprint},
1714: address = {Mathematics Institute, University of Oxford},
1715: year = {2002},
1716: }
1717:
1718: \bib{GN:EE}{article}{
1719: author = {Goldschmidt, C.A.},
1720: author = {Norris, J.R.},
1721: title = {Essential edges of {P}oisson random hypergraphs},
1722: journal = {Random Structures \& Algorithms},
1723: note = {to appear},
1724: year = {2002},
1725: }
1726:
1727: \bib{K:FMP}{book}{
1728: author = {Kallenberg, O.},
1729: title = {Foundations of Modern Probability},
1730: publisher = {Springer},
1731: year = {2002},
1732: edition = {Second edition},
1733: }
1734:
1735: \bib{JS:LT1}{book}{
1736: author = {Jacod, J.},
1737: author = {Shiryaev, A.N.},
1738: title = {Limit Theorems for Stochastic Processes},
1739: publisher = {Springer},
1740: year = {1987},
1741: }
1742:
1743: \bib{PSW:core}{article}{
1744: author = {Pittel, B.},
1745: author = {Spencer, J.},
1746: author = {Wormald, N.},
1747: title = {Sudden emergence of a giant {$k$}-core in a random graph},
1748: journal = {Journal of Combinatorial Theory, Series B},
1749: year = {1996},
1750: volume = {67},
1751: pages = {111\ndash 151},
1752: }
1753: \end{biblist}
1754: \end{bibdiv}
1755:
1756:
1757: \end{document}
1758:
1759: