1: \documentclass[12pt]{article}
2: \usepackage{epsf,latexsym,amsmath,amssymb}
3: % The preamble begins here.
4: \makeindex
5: \newcommand{\narrowmarg}{
6: \setlength{\oddsidemargin}{0in}
7: \setlength{\evensidemargin}{0in}
8: \setlength{\textwidth}{6.5in}
9: \setlength{\textheight}{9in}
10: \setlength{\topmargin}{0in}
11: \setlength{\headheight}{0in}
12: \setlength{\headsep}{0in}
13: }
14: % \narrowmarg
15:
16: % \tolerance=10000
17: \tolerance=10000
18:
19:
20: \newtheorem{theorem}{Theorem}[section]
21: \newtheorem{lemma}[theorem]{Lemma}
22: \newtheorem{proposition}[theorem]{Proposition}
23: \newtheorem{corollary}[theorem]{Corollary}
24: \newtheorem{conjecture}[theorem]{Conjecture}
25: \newtheorem{definition}[theorem]{Definition}
26: \newtheorem{sublemma}[theorem]{Sublemma}
27: \newtheorem{numberless}{Theorem}
28: \renewcommand{\thenumberless}{\Alph{numberless}}
29:
30: % These commands generate help generate a glossary. The pap.ind file,
31: % generated by makeindex, looks like
32: % [term:] definition, page number(s)
33: % and then the environment ``theindex'' is renewenviroment'ed to
34: % ``description''
35: \newcommand{\mytwoindex}[2]{\mythreeindex{#1}{#1}{#2}}
36: \newcommand{\mythreeindex}[3]{\index{#1@[#2:] {#3}}}
37: \newcommand{\myindexaj}{\mythreeindex{aj}{$a_j$}{the number of $\pi_j$ and
38: $\pi_j^{-1}$ appearing in a word or form}}
39: \newcommand{\myindexW}{\mythreeindex{WGamma}{$W_\Gamma,W_T,W_{T'}$}{the
40: number of potential walk classes or legal walks on a form, type, or
41: new type with various additional restrictions (such as irreducibility)}}
42: \newcommand{\myindexlambdaone}{\mythreeindex{lambda1}{$\lambda_1(G)$}{the
43: sup over all values $( c_G(v,v;k))^k$}}
44: \newcommand{\myindexEGamman}{\mythreeindex{EGamman}{E\char'133$\Gamma$\char'135$_n$}{The expected value of the number
45: of closed walks corresponding to a potential walk class (which depends
46: only the the form of the potential walk)}}
47:
48: \newcommand{\mob}{\mu}
49:
50: \newcommand{\newa}{P}
51:
52: \newcommand{\shriek}{{ATTENTION!!!! ATTENTION!!!! ATTENTION!!!! }}
53: \newcommand{\tohere}{{FINISHED CORRECTIONS UP TO HERE}}
54: \newcommand{\dtreelike}{{$d$-Ramanujan}}
55: % \newcommand{\dtreelike}{{$d$-tree-like}}
56: \newcommand{\opp}{{\rm opp}}
57: \newcommand{\dir}{{\rm dir}}
58: \newcommand{\sbd}{{\rm sbd}}
59: \newcommand{\walksum}[2]{{{\rm WalkSum}(#1,#2)}}
60: \newcommand{\supress}[2]{{#1\left[#2\right]_{\rm sup}}}
61: \newcommand{\realize}[2]{{{#1}|_{#2}}}
62: \newcommand{\ord}{{\rm ord}}
63: \newcommand{\from}{\colon}
64: \newcommand{\ignore}[1]{}
65:
66: \newcommand{\floor}[1]{\left\lfloor #1\right\rfloor}
67:
68: \newcommand{\El}{{E_{\rm long}}}
69: \newcommand{\Ef}{{E_{\rm fixed}}}
70: \newcommand{\kf}{k^{\rm fixed}}
71: \newcommand{\veckf}{{\vec k}^{\rm fixed}}
72:
73: \newcommand{\ck}{{\cal K}}
74: \newcommand{\cq}{{\cal Q}}
75: \newcommand{\ce}{{\cal E}}
76: \newcommand{\ct}{{\cal T}}
77: \newcommand{\cg}{{\cal G}}
78: \newcommand{\ch}{{\cal H}}
79: \newcommand{\cm}{{\cal M}}
80: \newcommand{\ci}{{\cal I}}
81: \newcommand{\cj}{{\cal J}}
82: \newcommand{\cjnd}{\cj_{n,d}}
83: \newcommand{\cind}{\ci_{n,d}}
84: \newcommand{\cinpd}{\ci_{n+1,d}}
85: \newcommand{\cknd}{\ck_{n,d}}
86: \newcommand{\cgnd}{\cg_{n,d}}
87: \newcommand{\chnd}{\ch_{n,d}}
88: \newcommand{\cw}{{\cal W}}
89: \newcommand{\cl}{{\cal L}}
90: \newcommand{\cf}{{\cal F}}
91:
92: \newcommand{\fund}{{\rm fund}}
93: \newcommand{\taufund}{{\tau_\fund}}
94: \newcommand{\tsetfund}{{\tset_\fund}}
95: \newcommand{\chfund}{{\ch_\fund}}
96:
97: \newcommand{\tang}{{\psi}} % variable for a tangle. ``tangle'' in use.
98: \newcommand{\tset}{{\Psi}}
99: \newcommand{\tseig}{{\tset_{\rm eig}}}
100: \newcommand{\tsmin}{{\tset_{\rm min}}}
101: \newcommand{\tsord}{{\tset_{\rm ord}}}
102:
103: \renewcommand{\complement}[1]{#1^{\rm c}}
104:
105: % \newcommand{\nv}{{\bf v}} % notated vertex
106: % \newcommand{\ned}{{\bf e}}
107: % \newcommand{\nt}{{\bf t}}
108: \newcommand{\nv}{{v}} % back to standard notion
109: \newcommand{\ned}{{e}}
110: % \newcommand{\nt}{{\bf t}}
111:
112: \newcommand{\oddf}{!_{\rm odd}}
113: \newcommand{\oddbinom}[2]{\binom{#1}{#2}_{\rm odd}}
114: \newcommand{\Esymm}{{\rm E}_{\rm symm}}
115:
116: \newcommand{\tree}{{\rm Tree}}
117: \newcommand{\tangle}{\ct}
118: \newcommand{\tanglefirst}{\tangle}
119: \newcommand{\selective}{{\rm SelTr}}
120: \newcommand{\irdsel}{{\rm IrSelTr}}
121: \newcommand{\sit}{{\rm SIT}}
122: \newcommand{\ssit}{{\rm SSIT}}
123:
124: \newcommand{\irdtr}[2]{{\rm IrredTr}\left(#1,#2\right)}
125: % \newcommand{\lamnonback}{{\lambda_{\rm Nb}}}
126: \newcommand{\intn}{{\{1,\ldots,n\}}}
127: \newcommand{\cestsl}{{\ce_{\rm STSL}}}
128: \newcommand{\cwstsl}{{\cw_{\rm STSL}}}
129: \newcommand{\pstsl}{{P_{\rm STSL}}}
130: \newcommand{\fstsl}{{f_{\rm STSL}}}
131: \newcommand{\fsl}{{f_{\rm SL}}}
132: \newcommand{\cwsl}{{\cw_{\rm SL}}}
133:
134:
135: \newcommand{\reals}{{\bf R}}
136: \newcommand{\integers}{{\bf Z}}
137: \newcommand{\complex}{{\bf C}}
138: \newcommand{\csphere}{\complex\cup\{\infty\}}
139:
140: \newcommand{\E}[1]{\mbox{E}\left[#1\right] }
141: \newcommand{\prob}[1]{{\rm Prob}\left\{ #1 \right\} }
142:
143: % \newcommand{\proof}{{\par\noindent{\bf Proof}\quad}}
144: \newcommand{\proof}{{\par\noindent {\bf Proof}\space\space}}
145: \newcommand{\proofbox}{\begin{flushright}$\Box$\end{flushright}}
146:
147: \newcommand{\ird}[1]{{\rm Irred}_{#1}}
148: \newcommand{\iwi}{i\stackrel{w}{\rightarrow}i}
149: \newcommand{\Eline}[1]{\mbox{E}\{#1\} }
150: \newcommand{\Gntd}{{\cal G}_{n,2d}}
151: \newcommand{\pr}[1]{{\rm Pr}\left( #1 \right) }
152: \newcommand{\coin}[1]{{\rm coin}\left( #1 \right) }
153: \newcommand{\doublescript}[2]{{\scriptstyle #1\atop{\atop\scriptstyle #2}}}
154:
155: \newcommand{\crud}{{(cdr)^{cr}}}
156: \newcommand{\descrud}{for some absolute constant $c$}
157: \newcommand{\crut}{{(cdr+m)^{crt}}}
158: \newcommand{\Pbar}{{\overline P}}
159: \newcommand{\Qbar}{{\overline Q}}
160: \newcommand{\Rbar}{{\overline R}}
161: \newcommand{\Lcalbar}{{\overline {\cal L}}}
162: \newcommand{\Qtilde}{{\widetilde Q}}
163: \newcommand{\ftilde}{{\tilde f}}
164: \newcommand{\mvect}{{\vec{m}}}
165: \newcommand{\kvect}{{\vec{k}}}
166: \newcommand{\Trace}[1]{{\rm Trace}\left( #1 \right) }
167:
168: \newcommand{\T}{{\rm T}}
169: \newcommand{\HH}{{\rm H}}
170:
171: \newcommand{\cat}{{\cal A}_{\rm t}}
172: \newcommand{\cah}{{\cal A}_{\rm h}}
173:
174:
175:
176:
177: \title{A Proof of Alon's Second Eigenvalue Conjecture and Related Problems}
178:
179: \author{Joel Friedman\thanks{
180: Departments of Computer Science and Mathematics,
181: University of British Columbia, Vancouver, BC\ \ V6T 1Z4
182: (V6T 1Z2 for Mathematics), CANADA.
183: {\tt jf@cs.ubc.ca}.
184: Research supported in part by an NSERC grant.}}
185:
186: \date{March 27, 2002 \\ Third Revision: May 3, 2004}
187:
188: \begin{document} % End of preamble and beginning of text.
189: \maketitle % Produces the title.
190: % \tableofcontents
191: \begin{abstract}
192: A $d$-regular graph has largest or first (adjacency
193: matrix) eigenvalue $\lambda_1=d$.
194: Consider for an even $d\ge 4$, a random $d$-regular
195: graph model
196: formed from $d/2$ uniform, independent permutations
197: on $\intn$. We shall show that for any $\epsilon>0$ we have
198: all eigenvalues aside from $\lambda_1=d$ are bounded by
199: $2\sqrt{d-1}\;+\epsilon$ with probability $1-O(n^{-\tau})$,
200: where $\tau=\lceil \bigl(\sqrt{d-1}\;+1\bigr)/2 \rceil-1$.
201: We also show that this probability is at most $1-c/n^{\tau'}$,
202: for a constant $c$ and a $\tau'$ that is either $\tau$ or
203: $\tau+1$ (``more often'' $\tau$ than $\tau+1$). We prove
204: related theorems for other models of random graphs, including
205: models with $d$ odd.
206:
207: These theorems resolve the conjecture of Alon, that says that
208: for any $\epsilon>0$ and $d$,
209: the second largest eigenvalue of ``most'' random $d$-regular graphs
210: are at most $2\sqrt{d-1}\;+\epsilon$ (Alon did not specify precisely
211: what ``most'' should mean or what model of random graph one should
212: take).
213:
214: %% In this paper we show the following conjecture of Alon.
215: %% Fix an integer $d>2$
216: %% and a real $\epsilon>0$. Then for sufficiently large $n$ we have that
217: %% ``most'' $d$-regular graphs on $n$ vertices
218: %% have all their eigenvalues except $\lambda_1=d$ bounded
219: %% by $2\sqrt{d-1}\;+\epsilon$ in absolute value. (Alon conjectured
220: %% this only for $\lambda_2$, but our methods, being trace methods,
221: %% also bound negative eigenvalues.)
222: % Here ``most'' is with respect to a number of different models
223: % of $d$-regular graphs on $n$ vertices, with the models here only requiring
224: % either $d$ or $n$ to be even.
225:
226: % Our method is a refinement
227: % of the trace method of the author, which
228: % built upon the Broder-Shamir trace method (which begins like
229: % Wigner's trace
230: % method).
231:
232: % Left open is the possibility of improving the above theorem by replacing
233: % a positive constant $\epsilon$ with
234: % $\epsilon=0$ or even some negative function $\epsilon=\epsilon(n)<0$.
235: \end{abstract}
236:
237: % \newpage
238:
239: \section{Introduction}
240:
241: The eigenvalues of the adjacency matrix of a finite undirected graph, $G$, are
242: real and hence can be ordered
243: \mythreeindex{lambda}{$\lambda_i(G)$}{$i$-th largest eigenvalue of the
244: adjacency matrix of a finite graph}
245: $$
246: \lambda_1(G)\ge\lambda_2(G)\ge\cdots\ge\lambda_n(G),
247: $$
248: where $n$ is the number of vertices in $G$. If $G$ is $d$-regular, i.e.,
249: each vertex is of degree $d$,
250: then $\lambda_1=d$.
251: In \cite{alon_eigenvalues}, Noga Alon conjectured that for any $d\ge 3$ and
252: $\epsilon>0$, $\lambda_2(G)\le 2\sqrt{d-1}+\epsilon$ for ``most'' $d$-regular
253: graphs on a sufficiently large number of vertices.
254: The Alon-Boppana bound shows that the constant
255: $2\sqrt{d-1}$ cannot be improved upon (see \cite{alon_eigenvalues,
256: nilli,friedman_geometric_aspects}).
257: The main goal of this paper is to prove this conjecture
258: for various models
259: of a ``random $d$-regular
260: graph.''
261:
262: Our methods actually show that for ``most'' $d$-regular graphs,
263: $|\lambda_i(G)|\le 2\sqrt{d-1}+\epsilon$
264: for all $i\ge 2$, since our methods are variants of the standard
265: ``trace method.''
266:
267: Our primary interest in Alon's conjecture, which was Alon's motivation,
268: is that fact that graphs with $|\lambda_i|$ small for $i\ge 2$ have various
269: nice properties, including being expanders or magnifiers
270: (see \cite{alon_eigenvalues}).
271:
272: For a fixed $n$ we can generate a random $d$-regular graph on
273: $n$ vertices
274: % added:
275: as follows, assuming $d$ is even (later we will give random graph models
276: that allow $d$ to be even or odd).
277: % end of added.
278: % by taking $d/2$ permutations on $V=\{1,\ldots,n\}$, $\pi_1,\ldots,\pi_{d/2}$,
279: Take $d/2$ permutations on $V=\{1,\ldots,n\}$, $\pi_1,\ldots,\pi_{d/2}$,
280: each $\pi_i$ chosen uniformly among all $n!$ permutations
281: with all the $\pi_i$ independent.
282: We then form
283: $$
284: E = \Bigl\{ \bigl(i,\pi_j(i)\bigr),\bigl(i,\pi_j^{-1}(i)\bigr) \bigm|
285: j=1,\ldots,
286: d/2, \quad i=1,\ldots,n \Bigr\} \;,
287: $$
288: yielding a directed graph $G=(V,E)$, which we may view as undirected.
289: We call this probability space of
290: random graphs $\cgnd$\mythreeindex{Gnd}{$\cgnd$}{space of random graphs
291: formed from $d/2$ permutations}.
292: $G$ can have multiple edges and self-loops, and
293: each self-loop contributes $2$ to the appropriate diagonal entry of $G$'s
294: adjacency matrix\footnote{Such a self-loop is a {\em whole-loop} in the
295: sense of \cite{friedman_geometric_aspects}; see also Section~2 of this
296: paper.}.
297:
298: The main goal of this paper is to prove theorems like the following,
299: which prove Alon's conjecture, for various models of a random $d$-regular
300: graph; we start with the model $\cgnd$.
301: \begin{theorem}\label{th:main}
302: Fix a real
303: $\epsilon>0$ and an even positive integer $d$. Then there is a constant,
304: $c$, such that for a random graph, $G$,
305: in $\cgnd$ we have that with probability at least
306: $1-c/n^\tau$ we have for all $i>1$
307: $$
308: |\lambda_i(G)|\le 2\sqrt{d-1}\;+\epsilon,
309: $$
310: where
311: $\tau=\taufund=\lceil \bigl(\sqrt{d-1}\;+1\bigr)/2 \rceil-1$
312: \mythreeindex{taufund}{$\taufund$}{smallest order of a supercritical
313: tangle}.
314: Furthermore, for some constant $c'>0$ we have that $\lambda_2(G)>2\sqrt{d-1}$
315: with probability at least
316: $c'/n^s$, where $s=\lfloor \bigl(\sqrt{d-1}\;+1\bigr)/2
317: \rfloor$. (So $s=\taufund$ unless $\bigl(\sqrt{d-1}\;+1\bigr)/2$ is an
318: integer, in which case $\taufund=s-1$.)
319: \end{theorem}
320:
321: Left open is the question of whether or not this theorem holds with
322: $\epsilon=0$ (which would yield ``Ramanujan graphs'')
323: or even some function $\epsilon=\epsilon(n)<0$.
324: Calculations such as those in
325: \cite{friedman_geometric_aspects} suggest that it does, even for some
326: negative function $\epsilon(n)$.
327: Examples of ``Ramanujan graphs,'' i.e., graphs where
328: $|\lambda_i(G)|\le 2\sqrt{d-1}$ except $i=1$ (and, at times, $i=n$
329: when $\lambda_n=-d$)
330: have been given in
331: \cite{lubotzky_phillips_sarnak,margulis,morgenstern} where $d$ is one
332: more than an odd prime or prime power. Theorem~\ref{th:main} demonstrates
333: the existence of ``nearly Ramanujan'' graphs of any even degree.
334: We shall soon address odd $d$, as well.
335:
336: Another interesting question arises in the gap between $\taufund$ and $s$ in
337: Theorem~\ref{th:main} in the case where
338: $\bigl(\sqrt{d-1}\;+1\bigr)/2$ is an integer;
339: it is almost certain that one of them can be
340: improved upon.
341: In the language of Section~4 of this paper,
342: $\taufund$\mythreeindex{taufund}{$\taufund$}{smallest order of a supercritical
343: tangle} is the smallest
344: order of a supercritical tangle, and $s$ that of a hypercritical tangle;
345: a gap between $\taufund$ and $s$ can only occur when there is a critical
346: tangle of order smaller than that of any hypercritical tangle.
347:
348: Previous bounds of the form $\lambda_2\le f(d)+\epsilon$
349: include $f(d)=(2d)^{1/2}(d-1)^{1/4}$ of the author
350: (see \cite{friedman_relative}),
351: which is slight improvement over the Broder-Shamir bound of
352: $f(d)=2^{1/2}d^{3/4}$ (see \cite{broder}). Asymptotically in $d$,
353: the bounds $f(d)=C\sqrt{d}$ of Kahn and Szemer\'edi (see
354: \cite{friedman_kahn_szemeredi}, here $C$ is some constant)
355: and $f(d)=2\sqrt{d-1}+2\log d + C$
356: of the author
357: (see \cite{friedman_kahn_szemeredi,friedman_random_graphs} and see
358: equation~(\ref{eq:rho}) for the more precise bound) are
359: improvements over the first two bounds.
360:
361: The value of $\taufund$ in Theorem~\ref{th:main} depends on the particular
362: model
363: of a random graph. Indeed, consider the model
364: $\chnd$\mythreeindex{Hnd}{$\chnd$}{random graph
365: space formed by $d/2$ permutations that are
366: cycles of length $n$}
367: of a random graph,
368: which is like $\cgnd$ except that we insist that each $\pi_i$ be one of the
369: $(n-1)!$ permutations whose cyclic decomposition consists of one cycle of
370: length $n$. The same methods used to prove Theorem~\ref{th:main} will show
371: the following variant.
372:
373: \begin{theorem}\label{th:mainh}
374: Theorem~\ref{th:main} holds with $\cgnd$ replaced by
375: $\chnd$ and $\taufund
376: =\lceil \sqrt{d-1}\;\rceil-1$ and $s=\lfloor \sqrt{d-1}\rfloor$,
377: except that when $d=4$ we take $s=2$.
378: \end{theorem}
379: Once again, $\taufund=s$, unless a certain expression, in this case
380: $\sqrt{d-1}$ (excepting $d=4$),
381: is an integer. Note that for $\chnd$, the value of
382: $\taufund$ is roughly twice as large as that for $\cgnd$ for $d$
383: large.
384:
385: Next consider two more models of random $d$-regular graphs; in these two
386: models $d$ may be even or odd.
387: Let $\cind$\mythreeindex{Ind}{$\cind$}{random graph model of $d$ perfect matchings ($n$ even)},
388: for positive
389: integers $n,d$ with $n$ even, be the model of a random $d$-regular graph
390: formed from $d$ random perfect matchings on $\intn$.
391:
392:
393: For an odd positive integer $n$, let a {\em near perfect matching} be a
394: matching of $n-1$ elements of $\intn$; such a matching becomes a $1$-regular
395: graph if it is complemented by a single half-loop\footnote{
396: Readers unfamiliar
397: with half-loops (i.e., self-loops contributing only $1$ to a
398: diagonal entry of the adjacency matrix)
399: can see Section~2 of this paper or
400: \cite{friedman_geometric_aspects}.
401: } at the unmatched
402: vertex. Taking $d$ independent such $1$-regular graphs gives a model,
403: $\cjnd$\mythreeindex{Jnd}{$\cjnd$}{random graph model formed from
404: $d$ permutations each with exactly one fixed point ($n$ odd)},
405: of a $d$-regular graph on $n$ vertices for $n$ odd.
406:
407: \begin{theorem}\label{th:maini}
408: Theorem~\ref{th:mainh} holds with $\chnd$ replaced by
409: $\cind$ and with no $d=4$ exception (i.e., $s=1$ for $d=4$).
410: Theorem~\ref{th:main} holds with $\cgnd$ replaced by
411: $\cjnd$.
412: \end{theorem}
413:
414:
415: We can assert the truth of the
416: Alon conjecture on more models of random graphs
417: by using results on continguity and related notions.
418: Consider
419: two families of probability spaces,
420: $(\Omega_n,{\cal F}_n,\mu_n)_{n=1,2,\ldots}$ and
421: $(\Omega_n,{\cal F}_n,\nu_n)_{n=1,2,\ldots}$ over the same sets
422: $\Omega_n$ and sigma-algebras ${\cal F}_n$; denote
423: $\mu=\{\mu_n\}$
424: and $\nu=\{\nu_n\}$. We say that $\mu$ {\em dominates} $\nu$ if
425: for any family of measurable events,
426: $\{E_n\}$ (i.e., $E_n\in{\cal F}_n$), we have $\mu_n(E_n)\to 0$ as
427: $n\to\infty$ implies $\nu_n(E_n)\to 0$ as $n\to\infty$.
428: We say that $\mu$ and $\nu$ are {\em contiguous} if $\mu$ dominates $\nu$
429: and $\nu$ dominates $\mu$.
430: \begin{corollary} Fix an $\epsilon>0$ and an integer
431: $d\ge 2$. Let ${\cal L}_{n}$
432: be any family of probability spaces of $d$-regular graphs on $n$ vertices
433: (possibly defined for only certain $n$)
434: that is dominated by $\cgnd$, $\chnd$, $\cind$, or $\cjnd$.
435: Then
436: for $G$ in ${\cal L}_n$ we have that with probability $1-o(1)$
437: (as $n\to\infty$) for
438: all $i$ with $2\le i\le n$ we have
439: $$
440: |\lambda_i(G)|\le 2\sqrt{d-1}+\epsilon.
441: $$
442: \end{corollary}
443: There are a lot of results regarding contiguity and (at least
444: implicitly) domination; see \cite{greenhill,kim,wormald} and the references
445: there. For example, if $\cgnd'$ is the restriction of $\cgnd$ to those
446: graphs without self-loops, then for $d\ge 4$ it is known that
447: (1) $\cgnd'$ and $\chnd$ are contiguous (by \cite{kim} and previous work), and
448: (2) $\cgnd$ dominates $\cgnd'$ (easy, since a self-loop occurs
449: in $\cgnd$ with
450: probability bounded away from $1$ for fixed $d$).
451: Thus the Alon conjecture for
452: $\cgnd$ implies the same for $\chnd$ (but this contiguity and/or domination
453: approach does not give as tight a bound on the probablity that
454: $\lambda_2\le 2\sqrt{d-1}\;+\epsilon$ fails to hold as is given
455: in Theorem~\ref{th:mainh}).
456: Also, $\cgnd$ is contiguous
457: with the ``pairing'' or ``configuration'' model of $d$-regular (pseudo)graphs
458: (see \cite{greenhill}); it follows that the Alon conjecture holds for the
459: latter model, and thus
460: (see \cite{wormald}, especially the beginning of Section~2 and
461: Corollary~4.17)
462: the conjecture holds for $n$ (and $d$) even
463: for $\cind$ or the uniform
464: measure on all $d$-regular (simple) graphs on $n$ vertices.
465:
466: Our method for proving Theorems~\ref{th:main}, \ref{th:mainh}, and
467: \ref{th:maini}
468: is a variant of the well-known ``trace method''
469: (see, for example
470: \cite{wigner,geman,komlos,broder,friedman_random_graphs}) originated
471: by Wigner,
472: especially the author's refinement in \cite{friedman_random_graphs}
473: of the beautiful Broder-Shamir style of analysis in
474: \cite{broder}.
475: The standard trace method involves taking the expected value
476: of the trace of a reasonably
477: high power\footnote{In
478: \cite{broder,friedman_random_graphs} this power
479: is roughly $c\log n$, where $c$ depends on $d$ and on aspects of the
480: method.}
481: of the adjacency matrix. In our situation we are
482: unable to analyze this trace accurately enough to prove
483: Theorem~\ref{th:main}, as certain infinite sums involved in our
484: analysis diverge
485: % (as they would in \cite{friedman_random_graphs} if
486: % we tried to continue the main expansion in $1/n$, equation~(3),
487: % there).
488: (for example, the infinite sum involving $W(T;\vec m)$ and $P_{i,T,\vec m}$
489: just above the middle
490: of page 351 in \cite{friedman_random_graphs}, for types of order $>d$).
491: This divergence is due to certain ``tangles'' that
492: can occur in a random graph and can adversely affect the eigenvalues
493: (see Sections~2 and 4).
494: To get around these ``tangles'' we introduce a {\em selective trace}.
495: We briefly sketch what a selective trace is in the next paragraph.
496:
497: Recall that a closed walk\mytwoindex{closed walk}{a walk beginning and
498: ending at the same vertex} about a vertex, $v$, is a walk in the graph
499: beginning and ending at $v$.
500: Recall that the trace of the $k$-th power of the adjacency matrix equals
501: the sum over all $v$ of the number of closed walks about $v$
502: of length $k$.
503: The {\em $k$-th irreducible trace} (used in both \cite{broder} and
504: \cite{friedman_random_graphs}) is the same sum as the $k$-th power
505: trace, except that we
506: require the closed walks to be {\em irreducible}\mytwoindex{irreducible}{
507: a walk (resp.\ word) that has no consecutive steps of an edge (resp.\ letter)
508: and its
509: inverse}, i.e., to have no edge
510: traversed and then immediately thereafter
511: traversed in the opposite direction.
512: A selective trace is a sum like an irreducible trace, but
513: where we further require that
514: the walk have no
515: small contiguous piece that ``traces out''
516: a ``supercritical
517: tangle'' (the notions of ``tracing out'' and ``supercritical
518: tangles'' will be
519: defined later;
520: roughly speaking, a ``supercritical
521: tangle'' is a small graph with many cycles).
522: Since these ``tangles'' occur with probability
523: at most proportional to $n^{-\tau}$, with $\tau=\taufund$
524: as in Theorem~\ref{th:main},
525: the selective trace usually agrees with the standard ``irreducible'' trace.
526:
527: Analyzing the selective traces involves a new concept of the
528: ``new type,'' which is a refinement of the ``type'' of
529: \cite{friedman_random_graphs}.
530:
531: We caution the reader about the notation used here. In this paper
532: we work with only $d$-regular graphs. In
533: \cite{broder} $2d$-regular graphs were studied; in
534: \cite{friedman_random_graphs} the graphs are usually $2d$-regular, although
535: for a part of Section~3 the graphs are $d$-regular.
536: We also caution the reader that here we use the term ``irreducible''
537: (as used in \cite{broder,friedman_random_graphs} and, for example,
538: in the text \cite{godsil}) as opposed to ``reduced'' (which is quite
539: common) or
540: ``non-backtracking'' (sometimes used in \cite{friedman_random_graphs})
541: in describing walks and related concepts.
542:
543:
544: We hope to generalize or
545: ``relativize'' the theorems here to
546: theorems about new eigenvalues of random covers
547: (see \cite{friedman_relative} for a relativized Broder-Shamir theorem).
548: In this paper we occasionally go out of our
549: way to use a technique that will easily generalize to this setting.
550:
551: The rest of this paper is organized as follows. In Section~2 we
552: review the trace method used in \cite{friedman_random_graphs} and
553: explain why it requres modification to prove Alon's conjecture; as a
554: byproduct we establish the part of Theorem~\ref{th:main} involving
555: $s$. In Section~3 we give some background needed for some technical
556: details in later sections. In Section~4 we formalize the notion of a
557: tangle, and discuss their properties; we prove the part
558: of Theorem~\ref{th:mainh} and \ref{th:maini}
559: involving $s$. In Section~5 we describe
560: ``types'' and ``new types,'' explaining how they help to estimate
561: ``walk sums;'' walk sums are generalizations of all notions of ``trace''
562: used here. In Section~6 we describe the ``selective trace'' used in
563: this paper; we give a crucial
564: lemma that counts certain types of selective closed walks in a graph.
565: In Section~7 we explain a little about ``{\dtreelike}'' functions, giving
566: a theorem to be used in Section~14 that also illustrates one of the main
567: technical points in Section~8.
568: In Section~8 we prove that certain selective traces have an asymptotic
569: expansion (in $1/n$) whose coefficients are ``\dtreelike.''
570: In Section~9 we show that the expansion in Section~8 still exists when
571: we count selective traces of graphs not containing any finite set of
572: tangles of order $\ge 1$.
573: In Section~10 we introduce strongly irreducible traces, that simplify
574: the proofs of the main theorems in this paper.
575: In Section~11 we prove a
576: crucial lemma that allows us to use the asymptotic expansion
577: to make conclusions about
578: certain eigenvalues; this lemma sidesteps the unresolved problem of
579: (even roughly) determining the coefficients of the asymptotic expansion
580: (in \cite{friedman_random_graphs} we actually roughly
581: determine the coefficients for the shorter expansion developed there).
582: In Section~12 we prove
583: the magnification (or ``expansion'') properties needed to apply the
584: sidestepping lemma of Section~11.
585: In Section~13 we complete the proof of Theorem~\ref{th:main}.
586: In Section~14 we complete the proof of Theorems~\ref{th:mainh} and
587: \ref{th:maini}, giving general conditions on a model of random graph
588: that are sufficient to imply the Alon conjecture.
589: In Section~15 we make some closing remarks.
590:
591: We mention that the reader interested only in the Alon conjecture
592: for only $\cgnd$
593: (i.e., the first part of Theorem~\ref{th:main}) need not read
594: Sections~2, 4 (assuming a willing to
595: believe Lemma~\ref{lm:order_increases}), and 14 and
596: Subsections~3.7, 3.8, 5.4, and 6.4.
597: Section~2 explains the problems with the trace method encountered
598: in \cite{friedman_random_graphs}. Subsections~3.7 and 3.8 and
599: Section~4 concern
600: themselves with the second part of Theorem~\ref{th:main} (the close
601: to matching bound on how many graphs fail the $2\sqrt{d-1}+\epsilon$
602: bound). Subsection~5.4 explains the new aspects in our approach to
603: the Alon conjecture; this subsection is not essential
604: to the exposition (but probably is helpful). Subsection~6.4 and
605: Section~14
606: involve the Alon conjecture for $\chnd,\cind,\cjnd$.
607:
608: Throughout the rest of this paper we will work with $\cgnd$ unless
609: we explicitly say otherwise, and we understand $d$ to be a fixed
610: integer at least
611: $3$. At times we insist that $d$ be even (for example, in
612: dealing with $\cgnd$ and $\chnd$).
613:
614:
615:
616:
617: \section{Problems with the Standard Trace Method}
618: \label{se:problem}
619:
620: In this section we summarize the trace method used in
621: \cite{friedman_random_graphs}, and why
622: this method cannot prove Alon's conjecture.
623: During this section we will review some of the ideas of
624: \cite{friedman_random_graphs}, involving asymptotic expansions of
625: various types of traces, which we modify in later sections
626: to complete our proof of Alon's
627: conjecture.
628:
629:
630: \subsection{The Trace Method}
631:
632: We begin by recalling the trace method as used in
633: \cite{friedman_random_graphs}, and why it did not yield the Alon
634: conjecture.
635:
636: The trace method (see \cite{wigner,geman,komlos,mckay,broder,
637: friedman_random_graphs}, for example) determines information on the
638: eigenvalues of a random graph in a certain probability space by computing
639: the expected value of a sufficiently high power of the adjacency matrix, $A$;
640: this expected value equals the expected value of the sum of that power of the
641: eigenvalues, since
642: $$
643: \Trace{A^k} = \lambda_1^k+\cdots+\lambda_n^k.
644: $$
645: Now $\Trace{A^k}$ may also be interpreted as the number of closed
646: walks (i.e., walks (see Section 3.1) in the
647: graph that start and end at the same vertex) of length $k$.
648: Now restrict our discussion to $\cgnd$.
649: A word, $w=\sigma_1\ldots\sigma_k$, of length $k$ over the alphabet
650: $$
651: \Pi=\{ \pi_1,\pi_1^{-1},\ldots,\pi_{d/2},\pi_{d/2}^{-1} \}
652: $$
653: (i.e. each $\sigma_i\in\Pi$), determines a random permutation, and the
654: $i,j$-th entry of $A^k$, is just the number of words, $w$, of length $k$,
655: taking $i$ to $j$.
656: But given a word, $w$, the probability, $P(w)$, that $w$
657: takes $i$ to $i$ is clearly independent of $i$.
658: Hence we have
659: $$
660: \E{\Trace{A^k}} = n \sum_{w\in\Pi^k} P(w)
661: $$
662:
663: In \cite{broder}, Broder and Shamir estimated the right-hand-side of the
664: above equality to obtain an estimate on $\lambda_2$. This analysis was
665: refined in \cite{friedman_random_graphs}. We review the ideas there.
666:
667:
668: First, a word, $w$, is
669: said to be {\em irreducible} if $w$ contains no consecutive occurence
670: of $\sigma,\sigma^{-1}$.
671: It is well-known that any word, $w$, has a unique
672: {\em reduction}
673: to an irreducible
674: word\footnote{
675: In fact, the irreducible word has length which is its distance
676: to the identity in the Cayley graph over the free group on
677: $d/2$ elements (see \cite{figa-talamanca_picardello}, Sections~1
678: and 7 of chapter 1).
679: Alternatively, see Proposition~2.5 of \cite{dicks} or
680: Theorem~1 of \cite{johnson} (this theorem says that a free
681: group on a set, $X$, is in one-to-one correspondence with the set
682: of reduced words, $X\cup X^{-1}$, which means that every word
683: over $X\cup X^{-1}$ has a unique corresponding reduced word;
684: here ``reduced'' is our ``irreducible'').
685: % To rigourously
686: % see that the reduction is unique, assume that $w=\sigma_1\ldots
687: % \sigma_k$ is a minimal
688: % length word with reductions to two distinct irreducible
689: % words $w_1=\sigma_{i_1}\ldots\sigma_{i_r}$ and
690: % $w_2=\sigma_{j_1}\ldots\sigma_{j_s}$. We easily establish the
691: % contraction that $v=\sigma_2\sigma_3\ldots\sigma_k$ also can
692: % be reduced to
693: % two distinct irreducible words. We do so by (1) arguing that
694: % the set of words to which $\sigma\sigma^{-1}u$ and $u$ can be
695: % reduced are the same (for any $\sigma\in\Pi$ and any $u\in\Pi^*$),
696: % and (2) then showing that $\sigma_1^{-1}w_i$ for $i=1,2$
697: % (which are both reductions of $\sigma_1^{-1}w$)
698: % reduce to two distinct words by
699: % considering the four cases of whether or not
700: % $\sigma_1=\sigma_{i_1}$ and/or $\sigma_1=\sigma_{j_1}$.
701: } (or reduced word), $w'$,
702: obtained from $w$ by repeatedly discarding any
703: consecutive occurences of $\sigma$ and $\sigma^{-1}$ in $w$, and
704: $P(w)=P(w')$.
705: Similarly a walk is said to be {\em irreducible} if it contains
706: no occurrence of a step along an edge immediately followed by the
707: reverse step along that edge\footnote{
708: In the case of an edge that is a half-loop (see Section 3), a half-loop
709: may not be traversed twice consecutively in an irreducible walk.
710: }.
711: Similarly, every irreducible walk has a unique reduction.
712: Let $\ird k$ be the set of irreducible words of
713: length $k$,
714: and let $\irdtr{A}{k}$\mythreeindex{IrredTr}{$\irdtr{A}{k}$}{the number of
715: closed irreducible walks of length $k$ in the graph underlying $A$}
716: be the number of irreducible closed walks of
717: length $k$ in $G$\footnote{We have admittedly defined $\irdtr{A}{k}$ in
718: terms of $G$, but we shall soon see (Lemma~\ref{lm:chebyshev}) that
719: $\irdtr{A}{k}$ can be defined as a polynomial in $A$ and $d$.
720: }.
721: We shall see that to
722: evaluate the expected value of $\Trace{A^m}$ it suffices, in a sense
723: (namely that of equation~(\ref{eq:other_way}) below),
724: to evaluate
725: $$
726: \E{\irdtr{A}{k}} = n\sum_{w\in\ird k} P(w),
727: $$
728: for $k=m,m-2,\ldots$.
729: It is easy to see that for any fixed word, $w$, we have a power series
730: expansion
731: $$
732: P(w)=\newa_0(w)+\frac{\newa_1(w)}{n}+\frac{\newa_2(w)}{n^2}+\cdots
733: $$
734: (see, for example, Theorem~\ref{th:exp_polys}).
735:
736: As examples, we note that for
737: a random permutation, $\pi$, on $\{1,\ldots,n\}$,
738: the probability that the sequence $1,\pi(1),\pi^2(1),\ldots$ first returns
739: to $1$ at $\pi^k(1)$ (i.e., the probability that $1$ lies on a cycle of
740: length exactly $k$) is $1/n$ for $k=1,\ldots,n$. It follows that
741: $P(\pi_1^m) = \phi(m)/n$, where $\phi(m)$ is the number of positive integral
742: divisors of $m$, assuming $m\le n$.
743: In this example $\newa_1(w)=\phi(m)$ involves number
744: theoretic properties of $m$.
745: For a second example, we first remark that $\pi^m$ maps a fixed vertex
746: to a different vertex with probability $1-\bigl(\phi(m)/n\bigr)$, and to
747: each of the $n-1$ different vertices with the same probability. It is
748: then easy to see that
749: $$
750: P(\pi_1^{m_1}\pi_2^{m_2}) = \frac{\phi(m_1)\phi(m_2)}{n^2} +
751: \frac{\bigl( n-\phi(m_1) \bigr) \bigl( n-\phi(m_2) \bigr) }{n^2(n-1)}
752: $$
753: provided that $m_1,m_2$ are at most $n$.
754: If $m_1,m_2$ are at least $2$, then the $\newa_i$ are non-zero for $i\ge 1$
755: and involve number theoretic functions of $m_1,m_2$ for $i\ge 2$.
756:
757: So set
758: \begin{equation}\label{eq:g_idef}
759: g_i(k)=\sum_{w\in\ird k} \newa_{i+1}(w)
760: \end{equation}
761: (we easily see $\newa_0(w)=0$ for $w\in\ird k$ and $k\ge 1$ and so
762: $g_{-1}(k)=0$ for $k\ge 1$).
763: \begin{definition}\label{de:ram}
764: A function, $f(k)$, on positive integers, $k$, is
765: said to be {\em \dtreelike} if there is a polynomial $p=p(k)$ and a
766: constant $c>0$ such
767: that
768: $$
769: |f(k)-(d-1)^kp(k)|\le ck^c(d-1)^{k/2}
770: $$
771: for all $k$. We call $(d-1)^kp(k)$ the {\em principal term} of $f$,
772: and $f(k)-(d-1)^kp(k)$ the {\em error term} (both terms are uniquely
773: determined if $d>2$).
774: \end{definition}
775: In \cite{friedman_random_graphs} it was shown (among other things) that
776: for all $i\le \sqrt{d-1}/2$ we have that $g_i$ as above is {\dtreelike}.
777: %, i.e.
778: %\begin{equation}\label{eq:g_iestimate}
779: %g_i(k)=d(d-1)^{k-1}\;f_i(k) + \; {\rm error}_i(k),
780: %\end{equation}
781: %where $f_i$ is some polynomial and
782: %\begin{equation}\label{eq:g_ierror}
783: %|{\rm error}_i(k)| \le (d-1)^{k/2}(ck)^c
784: %\end{equation}
785: %for some constant, $c$, depending on $i$ and $d$.
786: This, it turns out, gives a second eigenvalue bound of roughly
787: $2\sqrt{d-1}+2\log d+C+O(\log\log n/\log n)$ for a universal constant, $C$.
788: We now explain why.
789:
790: A standard counting and expansion argument is given in
791: \cite{friedman_random_graphs} (specifically Theorem~3.1 there)
792: to establish the following lemma.
793: \begin{lemma}\label{lm:counting}
794: For fixed even $d\ge 4$ there is an $\eta>0$ such that
795: with probability $1-n^{1-d}+O(n^{2-2d})$ we have that a
796: $G$ in $\cgnd$ has $max(\lambda_2,-\lambda_n)\le d-\eta$
797: (also with probability $n^{1-d}+O(n^{2-2d})$ we have that
798: $\lambda_2=d$).
799: \end{lemma}
800:
801: Next to $\lambda_1=d$, one (or both) of $\lambda_2,\lambda_n$ is
802: the next largest eigenvalue in absolute value; Lemma~\ref{lm:counting},
803: by bounding the eigenvalues other than $\lambda_1$, will eventually be
804: used to show that the $g_i$ of equation~(\ref{eq:g_idef}) are essentially
805: determined, for small $i$, by $\lambda_1$'s ``contribution'' to
806: $\irdtr{A}{k}$ (see below).
807:
808: Next we establish
809: the precise relationship between the traces of the $A^k$ and the
810: $\irdtr{A}{k}$. Let $A_k$ be the matrix whose $i,j$-th entry is the
811: number of irreducible walks of length $k$ from $i$ to $j$.
812: \begin{lemma}\label{lm:chebyshev}
813: The $A_k$ are given by
814: $$
815: A_k=q_k(A),
816: $$
817: where $q_k$ is the degree $k$ polynomial given via
818: \begin{equation}\label{eq:chebyshev}
819: q_k( 2\sqrt{d-1}\cos\theta) = \bigl( \sqrt{d-1}\bigr)^k
820: \biggl( \frac{2}{d-1} \cos k\theta + \frac{d-2}{d-1}\;\frac{\sin (k+1)\theta}
821: {\sin\theta}\biggr)
822: \end{equation}
823: (which is a type of Chebyshev polynomial);
824: alternatively we have $q_1(x)=x$, $q_2(x)=x^2-d$, and for $k\ge 3$ we have
825: $$
826: q_k(x) = x\;q_{k-1}(x)-(d-1)q_{k-2}(x).
827: $$
828: Also
829: $$
830: \irdtr{A}{k}=\Trace{A_k}=\sum_{i=1}^n q_k(\lambda_i).
831: $$
832: \end{lemma}
833: The proof is given in \cite{lubotzky} and
834: \cite{friedman_random_graphs} (specifically Lemma~3.3, page 356, in
835: \cite{friedman_random_graphs}; the $F_k$'s there are the $A_k$'s here).
836: In Section~10 we
837: shall use the fact that for fixed $\lambda$, $q_k=q_k(\lambda)$
838: satisfy the recurrence
839: $$
840: \bigl(\sigma_k^2-\lambda\sigma_k+(d-1)\bigr) q_k=0,
841: $$
842: where $\sigma_k$ is the ``shift in $k$'' operator (i.e.,
843: $\sigma_kq_k=q_{k+1}$)
844:
845: To go the other way
846: we note:
847: \begin{equation}\label{eq:other_way}
848: A^k = \sum_{i=k,k-2,k-4,\ldots} \;\; N_{k,i} A_i,
849: \end{equation}
850: where $N_{k,i}$ is the number of words of length $k$ that reduce to a given
851: irreducible word
852: of length $i$.
853: Thus
854: $$
855: \Trace{A^k} = \sum_{i=k,k-2,k-4,\ldots} \;\; N_{k,i} \;\irdtr{A}{i}.
856: $$
857: \begin{lemma}\label{lm:reducedwalks}
858: For $k,i$ even we have
859: $$
860: N_{k,i} \le \Bigl( 2\sqrt{d-1}\Bigr)^k (d-1)^{-i/2}\sqrt{(d-1)/d}
861: $$
862: if $i>0$ and
863: $$
864: N_{k,0} \le \Bigl( 2\sqrt{d-1}\Bigr)^k.
865: $$
866: \end{lemma}
867: An exact formula for $N_{k,i}$ is given in \cite{mckay}. A weaker
868: estimate than the above lemma
869: was used in \cite{friedman_random_graphs}. The proof of this
870: estimate is a simple spectral argument used by Buck (see \cite{buck,
871: friedman_relative}).
872: \proof
873: Consider
874: the adjacency matrix, $A_T$, of the infinite $d$-regular tree, $T$.
875: Our proof requires the following sublemma.
876: \begin{sublemma}\label{lm:tree_norm}
877: $A_T$ has norm $\le 2\sqrt{d-1}$.
878: \end{sublemma}
879: Actually, it is well-known that the norm of $A_T$ is exactly $2\sqrt{d-1}$
880: (see, for example, page 9 of
881: \cite{woess}, and the theorems on $\lambda_1$ in Section~3
882: here). However, the proof below is simple and generalizes to many other
883: situations (and can be used in many cases to determine the exact norm of
884: $A_T$).
885: \proof {\bf (of Sublemma~\ref{lm:tree_norm})}\space\space
886: Let $f$ be a function in $L^2(T)$, i.e., a function on the vertices
887: of $T$ whose sum of squares of values is finite. Fix a vertex, $v_0$, of $T$,
888: to be viewed as the root of $T$; the {\em children} of a vertex, $v$, are
889: defined to be those vertices adjacent to $v$ and of greater distance
890: than $v$ is to
891: %% from vs. to?
892: $v_0$.
893: We have
894: $$
895: (A_Tf,f) = \sum_v\sum_{w\in\;{\rm children}(v)} 2 f(v)f(w)
896: $$
897: which, by Cauchy-Schwarz, is
898: $$
899: \le \sum_v\sum_{w\in\;{\rm children}(v)} \left( f^2(w)\sqrt{d-1} +
900: \frac{f^2(v)}{\sqrt{d-1}} \right)
901: $$
902: $$
903: = f^2(v_0)d\Bigm/\sqrt{d-1}\;+\; \sum_{v\ne v_0} f^2(v)2\sqrt{d-1}.
904: $$
905: $$
906: \le \sum_v f^2(v) 2\sqrt{d-1} = 2\sqrt{d-1} \|f\|^2.
907: $$
908: Thus the norm of $A_T$ is $\le2\sqrt{d-1}$.
909: \proofbox
910: (To see that the norm of $A_T$ is exactly $2\sqrt{d-1}$ we find functions,
911: $f$, (of finite support)
912: for which the applications of Cauchy-Schwarz in the above proof are
913: ``usually'' tight.
914: Namely, we can take
915: $f(v)=(d-1)^{-{\rm dist}(v,v_0)/2}$ for ${\rm dist(v,v_0)}\le s$ and
916: $f(v)=0$ otherwise, where $s$ is a parameter which tends to $\infty$.
917: This technique works for some other graphs.)
918:
919: We resume the proof of Lemma~\ref{lm:reducedwalks}.
920: Let $v$ be a vertex of $T$, and let $S$ be the vertices
921: of distance $i$ to $s$. Then $|S|\,N_{k,i}$ is the dot product of
922: $A_T^k\chi_{\{v\}}$
923: with $\chi_S$, where $\chi_U$ denotes the characteristic function of $U$,
924: i.e. the function that is $1$ on $U$ and $0$ elsewhere.
925: So by Cauchy-Schwarz
926: $$
927: |S|\,N_{k,i}=(A_T^k\chi_{\{v\}},\chi_S)\le \|A_T\|^k |\chi_{\{v\}}| \;
928: |\chi_S| = \Bigl( 2\sqrt{d-1}\Bigr)^k \sqrt{|S|}.
929: $$
930: We finish with the fact that $|S|=1$ if $i=0$, and otherwise
931: $|S|=d(d-1)^{i-1}$.
932: \proofbox
933: Notice that clearly $N_{k,k}=1$, and so for $i=k$
934: Lemma~\ref{lm:reducedwalks}
935: is off by a multiplicative factor
936: of roughly
937: $2^k$; according to \cite{mckay,figa-talamanca_picardello}, the
938: Lemma~\ref{lm:reducedwalks}
939: estimate of $N_{k,0}$ is off by roughly a factor of $k^{3/2}$. The
940: roughness of Lemma~\ref{lm:reducedwalks} is unimportant for our purposes.
941:
942: Now notice that by Lemmas~\ref{lm:counting} and \ref{lm:chebyshev} we have
943: $$
944: \E{\irdtr{A}{k}} = q_k(d)\bigl( 1+n^{1-d}+O(n^{2-2d})\bigr) +
945: {\rm error},
946: $$
947: where
948: $$
949: |{\rm error}| \le
950: (n-1)\max_{|\lambda|\le d-\eta} |q_k(\lambda)|.
951: $$
952: It is easy to see (see \cite{friedman_random_graphs}) that
953: $q_k(d)=(d-1)^k$, and for some $\alpha>0$ we have
954: $$
955: \max_{|\lambda|\le d-\eta} |q_k(\lambda)| \le (d-1-\alpha)^k ck,
956: $$
957: for an absolute constant $c$ (with any $\eta$ as in
958: Lemma~\ref{lm:counting}). We wish to draw some conclusions about
959: the principal term of the $g_i$'s. We need the following
960: lemma:
961: \begin{lemma} For fixed $d,r$ there is a constant, $c$, such that
962: for $k\ge 1$ we have that in $\cgnd$
963: $$
964: \E{\irdtr{A}{k}} = g_0(k)+\frac{g_1(k)}{n} + \frac{g_2(k)}{n^2} + \cdots
965: + \frac{g_{r-1}(k)}{n^{r-1}} + {\rm error},
966: $$
967: where
968: $$
969: |{\rm error}| \le c (d-1)^{k-1}k^{4r+2}/n^{r}.
970: $$
971: \end{lemma}
972: \proof This follows from the calculations on page 352 of
973: \cite{friedman_random_graphs}; for each $i$, the $f_i$ in
974: \cite{friedman_random_graphs} is the polynomial in the principal
975: term of $g_i$ (and we mean $f_i$ corresponds precisely to $g_i$, not
976: $g_{i-1}$ or $g_{i+1}$). (Actually we shall later\footnote{
977: This stems from the fact that in \cite{friedman_random_graphs},
978: the $e^{(r+1)k/n}k^{2r+2}$ just above equation~(21) (page 352)
979: could have been replaced with $e^{rk/n}k^{2r}$.
980: } see that the
981: $4r+2$ in the error term estimate can be replaced by $4r$.)
982: \proofbox
983: We now take $k$ of order $\log^2n$ and use standard facts about
984: expansion and expansion's control on eigenvalues (namely our
985: Lemma~\ref{lm:counting})
986: to conclude, as in
987: \cite{friedman_random_graphs}, the following theorem.
988: \begin{theorem}\label{th:f_ivalues}
989: Let $g_0,g_1,\ldots,g_r$ be {\dtreelike}
990: for some $r\le d$. Then the principal term of $g_i$ vanishes for
991: $1\le i \le r$, and the principal term of $g_0$ is $d(d-1)^{k-1}$.
992: \end{theorem}
993: \proof See Theorem~3.5 of \cite{friedman_random_graphs}.
994: \proofbox
995:
996: We next apply Lemma~\ref{lm:reducedwalks} to estimate the expected value
997: of the trace of $A^k$ where $k$ is roughly
998: \begin{equation}\label{eq:hnrd}
999: h(n,r,d)=\frac{(r+1)\log n }{\log\Bigl( d/\bigl(2\sqrt{d-1}\,\bigr)\Bigr)} ,
1000: \end{equation}
1001: as in
1002: \cite{friedman_random_graphs}, in order to obtain the following theorem.
1003: \begin{theorem}\label{th:expansion_consequence}
1004: With the same hypotheses as Theorem~\ref{th:f_ivalues},
1005: we have (in $\cgnd$)
1006: \begin{equation}\label{eq:expansion_consequence}
1007: \E{ \sum_{i=2}^n \lambda_i^k} \le \rho^k
1008: \end{equation}
1009: for all $k\le h(n,r,d)$, with $h$ as above, where
1010: \begin{equation}\label{eq:rho}
1011: \rho = 2\sqrt{d-1}\; \Bigl( d/\bigl( 2\sqrt{d-1}\,\bigr) \Bigr)^{1/(r+1)}
1012: \left( 1 + \frac{c\log\log n}{\log n} \right)
1013: \end{equation}
1014: and $c$ depends only on $r,d$.
1015: \end{theorem}
1016: \proof First we take $k=\lfloor h(n,r,d) \rfloor$ and find that $\rho$
1017: can be taken as above. For
1018: smaller $k$ we appeal to Jensen's inequality. See
1019: \cite{friedman_random_graphs} for details.
1020: \proofbox
1021: From Theorem~2.18 of \cite{friedman_random_graphs} we see that we
1022: can take $r$ as large as $\lfloor \sqrt{d-1}\;/2\rfloor$. With this
1023: value of $r$, using equation~(\ref{eq:rho}), we conclude that
1024: equation~(\ref{eq:expansion_consequence}) holds with
1025: $\rho=2\sqrt{d-1}+2\log d + C+o(1)$ for an absolute constant, $C$, where
1026: $o(1)$ is a quantity that for fixed $d$ tends to $0$
1027: (proportional to $\log\log n/\log n$) as $n\to\infty$.
1028:
1029: Whenever equation~(\ref{eq:expansion_consequence}) holds,
1030: then the expected value
1031: of $\max(|\lambda_2|,|\lambda_n|)$ is bounded by $\rho$. The Alon conjecture
1032: would be implied if one could obtain $\rho\le 2\sqrt{d-1}\;+\epsilon$
1033: for any $\epsilon>0$.
1034:
1035:
1036:
1037: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1038:
1039: \subsection{Limitations of the Trace Expansion}
1040:
1041: In this subsection we will show that for some $i\le O(\sqrt{d}\log d)$,
1042: $g_i$ is not {\dtreelike}. We similarly
1043: show the part of Theorem~\ref{th:main} involving $s$. Both these facts are
1044: due to the possible occurrence of what we call {\em tangles}.
1045: Tangles and avoiding them are the main themes in this paper.
1046:
1047: We begin by describing an example of a {\em tangle}, and its effect
1048: on eigenvalues and traces.
1049: Consider $\cgnd$ for a fixed, even $d$ and a variable $n$ which we
1050: view as large.
1051: Fix an integer $m$ with $1\le m\le d/2$ (assume $d\ge 4$).
1052: Consider the event, $\tanglefirst$, that
1053: $$
1054: \mbox{$\pi_i(1)=1$ for $i=1,\ldots m$.}
1055: $$
1056: Clearly $\tanglefirst$ occurs with probability $1/n^m$.
1057:
1058: Assume $\tanglefirst$ occurs in a fixed $d$-regular graph, $G=(V,E)$.
1059: Let $W$ be the set of vertices of distance at least $2$ to the vertex $1$;
1060: $W$ is a random set of vertices, but always of size at least $n-d-1$.
1061: Consider the characteristic functions $\chi_{\{1\}},\chi_{W}$,
1062: where $\chi_U$ is the function that is $1$ on the vertices in $U$, and
1063: $0$ elsewhere. Let
1064: $$
1065: {\cal R}_A(v) = \frac{(Av,v)}{(v,v)}
1066: $$
1067: be the Rayleigh quotient associated to the adjacency matrix, $A$, of $G$.
1068: The following lemma is well-known.
1069: \begin{lemma}\label{lm:fix_bug}
1070: Let $A$ be a real, symmetric matrix.
1071: Let $u,v$ be nonzero vectors with $v$ orthogonal to $u$ and $Au$. Then
1072: $$
1073: \lambda_2 \ge \min\bigl( {\cal R}_A(u), {\cal R}_A(v) \bigr).
1074: $$
1075: \end{lemma}
1076: \proof Let $\mu$ denote the $\min$ on the right-hand-side of the above
1077: inequality.
1078: By the hypothesis of the lemma, $(Au,v)=0$; along with the symmetry of $A$,
1079: we have $(Av,u)=(v,Au)=0$.
1080: If $w=\alpha u+\beta v$ with $\alpha,\beta$ scalars,
1081: we have
1082: $$
1083: (Aw,w)= (Au,u)\alpha^2+(Av,v)\beta^2 \ge
1084: \mu(u,u)\alpha^2+\mu(v,v)\beta^2 = \mu(w,w).
1085: $$
1086: It follows that the Rayleigh quotient of any vector in the span of $u$ and
1087: $v$ is at least $\mu$. Since this span is a two-dimensional subspace, the
1088: max-min principle implies that $\lambda_2\ge \mu$.
1089: \proofbox
1090:
1091: We intend apply the above lemma with $u=\chi_{\{1\}}$ and $v=\chi_W$.
1092: $$
1093: {\cal R}_A(\chi_{\{1\}}) \ge 2m
1094: $$
1095: (since $(A\chi_U,\chi_U)$ counts twice the number of edges with both endpoints
1096: in $U$). Also
1097: $$
1098: (A\chi_{W},\chi_{W})
1099: = (A\chi_V,\chi_V)-2(A\chi_V,\chi_{V\setminus W})+(A\chi_{V\setminus W},
1100: \chi_{V\setminus W})
1101: $$
1102: $$
1103: \ge (A\chi_V,\chi_V)-2(A\chi_V,\chi_{V\setminus W}) \ge dn-2d(d+1)
1104: $$
1105: so
1106: $$
1107: {\cal R}_A(\chi_{W}) \ge \frac{dn -2d(d+1)}{n-d-1}=d-O(1/n)
1108: $$
1109: viewing $m,d$ as fixed.
1110: Since $\chi_{\{1\}}$ and $A\chi_{\{1\}}$ are supported in the neighbourhood
1111: of distance at most $1$ from the vertex $1$, $\chi_W$ is orthogonal to
1112: both of them.
1113: Lemma~\ref{lm:fix_bug} now implies
1114: $$
1115: \lambda_2 \ge \min\bigl(2m,d-O(1/n)\bigr).
1116: $$
1117: Next consider the probability that
1118: $$
1119: \mbox{$\pi_i(r)=r$ for $i=1,\ldots, m$}
1120: $$
1121: for at least one value of $r$. Inclusion/exclusion shows that the
1122: probability of this is at least
1123: $$
1124: \sum_r \prob{\mbox{$\pi_i(r)=r$ for $i=1,\ldots m$}} -
1125: $$
1126: $$
1127: \sum_{r,s} \prob{\mbox{$\pi_i(r)=r$ and $\pi_i(s)=s$ for $i=1,\ldots m$}}
1128: $$
1129: $$
1130: \ge n^{1-m}-\binom{n}{2} n^{-2m}.
1131: $$
1132: We summarize the above observations.
1133: \begin{theorem}\label{th:weak}
1134: For fixed integer $m$ with $1\le m\le d/2$,
1135: we have that $\lambda_2\ge 2m$
1136: for sufficiently large $n$ with probability at least
1137: $n^{1-m}-(1/2)n^{2-2m}$.
1138: \end{theorem}
1139:
1140: The proof of the above theorem did not exploit the fact that aside from
1141: having $m$ self-loops, the vertex $1$ is still adjacent to $d-2m$ other
1142: vertices of a $d$-regular graph.
1143: We seek a stronger theorem that exploits this fact.
1144:
1145: \begin{theorem}\label{th:improved_bound}
1146: For fixed integers $m\ge 1$ and $d\ge 4$, with $2m-1> \sqrt{d-1}$ and
1147: $m\le d/2$,
1148: we have that $\lambda_2>2\sqrt{d-1}$
1149: for sufficiently large $n$ with probability at least
1150: $n^{1-m}-(1/2)n^{2-2m}$.
1151: \end{theorem}
1152: (Notice that Theorem~\ref{th:weak} would require $m>\sqrt{d-1}$ for
1153: the same conclusion.)
1154: We are very interested to know if one can prove Theorem~\ref{th:improved_bound}
1155: when
1156: $2m-1=\sqrt{d-1}$ for integer $m$ and even integer $d$. We expect not.
1157: (The situation where $2m-1=\sqrt{d-1}$
1158: gives rise to what we will call a ``critical tangle,'' and
1159: $2m-1\ge\sqrt{d-1}$ to a ``supercritical tangle,'' in Section~4.)
1160: \proof
1161: Note: in Theorems~\ref{th:remarkable} and \ref{th:key_to_s} we give
1162: a proof of a generalization of this theorem requiring far less
1163: calculation (but requiring more machinery).
1164:
1165: Again, assume that $\pi_i(v_0)=v_0$ for $i=1,\ldots,m$ and some $v_0$.
1166: It suffices to show
1167: $\lambda_2>2\sqrt{d-1}$ for sufficiently large $n$, under the assumption
1168: that $2m-1>\sqrt{d-1}$.
1169:
1170: Let
1171: \begin{equation}\label{eq:alpha}
1172: \alpha(m) = (2m-1)+\frac{d-1}{2m-1}.
1173: \end{equation}
1174: By Cauchy-Schwarz we have $\alpha>2\sqrt{d-1}$ (equality does not hold,
1175: because $2m-1\ne (d-1)/(2m-1)$ since $2m-1>\sqrt{d-1}$).
1176:
1177: Let $\rho(v)$ denote $v$'s distance to $v_0$.
1178: For a fixed $r$, let
1179: $$
1180: f(v) = \left\{ \begin{array}{ll}
1181: (2m-1)^{-\rho} & \mbox{if $\rho\le r$,} \\
1182: 0 & \mbox{otherwise,} \end{array}\right.
1183: $$
1184: where $\rho=\rho(v)$. It is easy to check that $(Af)(v)\ge\alpha f(v)$
1185: provided that $\rho(v)<r$ (this includes the case $v=v_0$, since $\rho(v_0)=0$,
1186: but checking $v=v_0$ is a bit different from the other cases).
1187: It follows that
1188: \begin{equation}\label{eq:ray1}
1189: \frac{(Af,f)}{(f,f)} \ge \frac{\alpha(f,f)_{r-1}}{(f,f)_r},
1190: \end{equation}
1191: where
1192: $$
1193: (f,f)_t = \sum_{\rho(v)\le t} f^2(v).
1194: $$
1195: But $1=f^2(v_0)\le (f,f)_{r-1}$ if $r\ge 1$, and also
1196: $$
1197: (f,f)_r\le (f,f)_{r-1}+(d-2m)(d-1)^{r-1}(2m-1)^{-2r}
1198: $$
1199: (since, by induction, the number of vertices at distance $r$ from
1200: $v$ is at most $(d-2m)(d-1)^{r-1}$.)
1201: So
1202: \begin{equation}\label{eq:ray2}
1203: \frac{(f,f)_r}{(f,f)_{r-1}}
1204: \end{equation}
1205: can be made arbitrarily close to $1$ by taking $r$ sufficiently large
1206: (since $(d-1)(2m-1)^{-2}<1$).
1207:
1208: Let ${\cal R}$ be the Rayleigh quotient of $A$.
1209: The last paragraph, especially equations~(\ref{eq:ray1}) and
1210: (\ref{eq:ray2}) implies that
1211: for $\alpha'<\alpha$ there is an $r=r(\alpha')$ such that
1212: ${\cal R}(f)\ge \alpha'$.
1213:
1214: So let $N$ be those vertices of distance $1$ or $0$ to
1215: the support of $f$; the size, $|N|$, of $N$ is bounded as a function of
1216: $d$ and $r$.
1217: The function $f$ is
1218: orthogonal to $g=\chi_{V\setminus N}$ and $Ag$,
1219: and counting edges as before we see
1220: $$
1221: {\cal R}(g)\ge \frac{d|V|-2d|N|}{|V|-|N|}= d - O(|N|d/|V|).
1222: $$
1223: It follows that by taking $n$ sufficiently large, we can
1224: make $\lambda_2\ge \alpha'$; since $\alpha>2\sqrt{d-1}$, we can choose
1225: $\alpha'>2\sqrt{d-1}$, making
1226: $\lambda_2>2\sqrt{d-1}$.
1227: \proofbox
1228:
1229: Theorem~\ref{th:improved_bound} proves the part of Theorem~\ref{th:main}
1230: involving $s$, by taking $s=m-1$ with $m$ as small as possible
1231: (namely $m=\lfloor \bigl(\sqrt{d-1}\;+1\bigr)/2
1232: \rfloor+1$).
1233: The analogous parts of Theorems~\ref{th:mainh} and \ref{th:maini}
1234: are slightly trickier, since the ``tangle'' involved has automorphisms;
1235: we shall delay their proof (see Theorem~\ref{th:other_ss})
1236: until we give a more involved discussion of
1237: tangles in Section~4.
1238:
1239: Notice that our proof is really computing the norm of $A_H$ where $H$ is
1240: the $d$-regular graph with the vertex $1$ having $m$ self-loops, and which
1241: is a (an infinite) tree when these loops are removed. The function
1242: $f$ as above
1243: shows that $A_H$'s norm is at least $\alpha$.
1244: The statement and proof of Sublemma~\ref{lm:tree_norm}
1245: for the $d$-regular tree applies to the
1246: above tree (with $2m-1$ replacing $\sqrt{d-1}$, and with $\alpha$
1247: replacing $2\sqrt{d-1}$).
1248: In this way our proof of Theorem~\ref{th:improved_bound}
1249: is very much like one proof of
1250: the Alon-Boppana theorem (see \cite{nilli,friedman_geometric_aspects}).
1251:
1252:
1253:
1254: The discussion in this section leads to the following theorem.
1255: \begin{theorem}\label{th:nottreelike}
1256: There is an absolute constant (independent of $d$), $C$, such that the
1257: $g_i$ of equation~(\ref{eq:g_idef}) cannot be {\dtreelike}
1258: for all $i\le C\sqrt{d}\log d$.
1259: \end{theorem}
1260: \proof
1261: We fix an integer $s$ to be chosen later with
1262: $$
1263: \Bigl( \sqrt{d-1} \; -1 \Bigr)/2 < s < d/2.
1264: $$
1265: Set $s+1=m$ and apply Theorem~\ref{th:improved_bound}.
1266: Since $\alpha\ge 2s+1$ with $\alpha$ as
1267: in equation~(\ref{eq:alpha}), for $k$ even we have that $\lambda_2\ge 2s$
1268: with probability at least $n^{-s}+O(n^{-s-1})$. Thus
1269: $$
1270: \E{ \lambda_2^k}^{1/k} \ge \bigl(n^{-s}+O(n^{-s-1})\bigr)^{1/k} 2s.
1271: $$
1272: According to Theorem~\ref{th:expansion_consequence}, if $g_0,\ldots,g_r$
1273: are {\dtreelike} for some $r\le d$, then we have
1274: $$
1275: \E{ \lambda_2^k}^{1/k} \le \rho,
1276: $$
1277: with $\rho$ as in equation~(\ref{eq:rho}), provided that $k$ is even and
1278: bounded by $h(n,r,d)$ as in Theorem~\ref{th:expansion_consequence}.
1279: For some constant $c$ we have that for any $C$ and for
1280: $r=C\sqrt{d}\log d$, equation~(\ref{eq:rho}) gives
1281: $$
1282: \rho =2\sqrt{d-1} \bigl(1 + cC^{-1}d^{-1/2}+c(\log n)^{-1}\log\log n\bigr).
1283: $$
1284: In other words,
1285: \begin{equation}\label{eq:long}
1286: \bigl(n^{-s}+O(n^{-s-1})\bigr)^{1/k} 2s \le
1287: 2\sqrt{d-1} \bigl(1 + cC^{-1}d^{-1/2}+c(\log n)^{-1}\log\log n\bigr).
1288: \end{equation}
1289: Take $k$ even and as close to $h(n,r,d)$ as possible;
1290: note that by equation~(\ref{eq:hnrd}),
1291: $$
1292: \frac{\log n}{h}=
1293: \frac{ \log\Bigl( d/\bigl(2\sqrt{d-1}\,\bigr)\Bigr)}{r+1} \le
1294: \frac{\log d}{(C\sqrt{d}\log d)+1} \le 1/\Bigl(C\sqrt{d}\Bigr);
1295: $$
1296: hence, taking $n\to\infty$ in
1297: equation~(\ref{eq:long}) implies that
1298: for a universal constant, $c'$, we have
1299: $$
1300: e^{-sd^{-1/2}/C}2s\le 2\sqrt{d-1} \bigl(1 + cC^{-1}d^{-1/2}\bigr)
1301: \le 2\sqrt{d-1}\bigl(1 + cC^{-1}\bigr) .
1302: $$
1303: Choosing $s=C\sqrt{d}$ and dividing by $2$ yields
1304: $$
1305: C\sqrt{d}\;/e \le \sqrt{d-1} \bigl(1 + cC^{-1}\bigr).
1306: $$
1307: Choosing $C$ large enough so that $C/e> 1+(c/C)$ makes this impossible.
1308: \proofbox
1309:
1310:
1311: We have proven that not all $g_i$ are {\dtreelike} for $i\le r$ where
1312: $r=C\sqrt{d}\log d$.
1313: Notice that in our terminology, Theorem~2.18 of \cite{friedman_random_graphs}
1314: says that $g_i$ is {\dtreelike} for
1315: $i\le\lfloor \sqrt{d-1}\;\bigm/2 \rfloor-1$;
1316: again, for each $i$ the $f_i$ in \cite{friedman_random_graphs} is the
1317: polynomial in the principal part of our $g_i$.
1318: This leaves the question of
1319: whether Theorem~\ref{th:nottreelike} can be improved to an $r$ value closer
1320: to $\lfloor \sqrt{d-1}\;\bigm/2 \rfloor-1$; we conjecture that it can
1321: be improved to $r=\lfloor \bigl(\sqrt{d-1}\;+1\bigr)/2 \rfloor$, and
1322: that the tangle with $m=\lfloor \bigl(\sqrt{d-1}\;+1\bigr)/2
1323: \rfloor+1$ already ``causes'' this $g_r$ (or a lower one)
1324: to fail to be {\dtreelike}.
1325:
1326: This also leaves open the question of what can be said about the $g_i$
1327: that are not {\dtreelike}.
1328: Perhaps such $g_i=g_i(k)$ are a sum of $\nu^k p_\nu(k)$ over various $\nu$
1329: with some added error term. In this paper we avoid this issue,
1330: working with a modified trace (i.e., ``selective'' traces)
1331: for which the corresponding $g_i$ are {\dtreelike}.
1332:
1333: \section{Background and Terminology}
1334:
1335: In this section we review some ideas and techniques from the literature
1336: needed here. We also give some convenient terminology that is not
1337: completely standard.
1338:
1339:
1340: \subsection{Graph Terminology}
1341:
1342: We use some nonstandard notions in graph theory, and we carefully
1343: explain all our
1344: terminology and notions here.
1345:
1346: A directed graph, $G$, consists of a set of vertices, $V=V_G$, a set of
1347: edges, $E=E_G$, and an incidence map, $i=i_G\from E\to V\times V$; if
1348: $i(e)=(u,v)$ we will write $e\sim(u,v)$, say that $e$ is of {\em type}
1349: $(u,v)$, and say that $e$ originates in $u$ and terminates in $v$.
1350: (If $i$ is injective then it is usually safe to view $E$ as a subset of
1351: $V\times V$, and we say that $G$ has no multiple edges.) The adjacency
1352: matrix, $A=A_G$, is a square matrix indexed on $V$, where $A(u,v)$ counts
1353: the number of edges of type $(u,v)$. The outdegree at $v\in V$ is the
1354: row sum of $A$ at $v$, i.e., the number of edges originating in $v$; the
1355: indegree is the column sum or number of edges terminating in $v$.
1356:
1357:
1358: A graph\mytwoindex{graph}{a directed graph with a origin/terminal
1359: reversing pairing of its edges}, $G$,
1360: is a directed graph, $\widehat G$, such that each edge of type $(u,v)$
1361: is ``paired'' with an
1362: ``opposite edge'' of type $(v,u)$; in other words,
1363: we have a map
1364: $\opp=\opp_G\from E_{\widehat G}\to E_{\widehat G}$, such that
1365: $\opp(\opp)$ is the identity, and if $e\in E_{\widehat G}$ has
1366: $e\sim(u,v)$, then $\opp(e)\sim(v,u)$; in other words, the edges
1367: $E_{\widehat G}$ come in ``pairs,'' except that a self-loop, i.e., an
1368: $e\in E_{\widehat G}$ with $e\sim(v,v)$, can be paired with itself
1369: (which is a ``half-loop\mytwoindex{half-loop}{a self-loop paired with itself
1370: (in a graph)}'' in the terminology of
1371: \cite{friedman_geometric_aspects}) or paired with another self-loop at
1372: $v$ (which is a ``whole-loop\mytwoindex{whole-loop}{two self-loops (about the same vertex) paired with each other (in a graph)}''). Half-loops about $v$
1373: contribute $1$ to the adjacency matrix entry at $v,v$ (i.e., contribute
1374: $1$ to $A(v,v)$), and whole-loops
1375: contribute $2$. In this paper we primarily work with whole-loops,
1376: needing half-loops only in the model $\cjnd$.
1377: We refer to the {\em (undirected) edges}, $E_G$, of a graph, $G$,
1378: as the set of ``pairs'' of edges, $\{e,\opp(e)\}$. $G$'s vertex set
1379: and adjacency matrix are just those of the directed graph, $\widehat G$,
1380: i.e., $V_G=V_{\widehat G}$ and $A_G=A_{\widehat G}$.
1381:
1382: A {\em numbering} of a set, $S$, is a bijection $\iota\from S\to
1383: \{1,2,\ldots,s\}$, where $s=|S|$. A {\em partial numbering} of a set,
1384: $S$, is a numbering of some subset, $S'$, of $S$ (we allow $S'$ to be
1385: empty, in which case none of $S$ is numbered). We can speak of a graph,
1386: directed or not, as having numbered or partially numbered vertices and/or
1387: edges. A numbering can be viewed as a total ordering.
1388:
1389: Each letter $\pi\in\Pi=\{\pi_1,\pi_1^{-1},\ldots,\pi_{d/2}^{-1}\}$
1390: has its associated inverse, $\pi^{-1}\in\Pi$,
1391: and every word $w=\sigma_1\ldots\sigma_k$ over $\Pi$ has its associated
1392: inverse, $w^{-1}=\sigma_k^{-1}\ldots\sigma_1^{-1}$. If $\cw$ is any
1393: set of words over $\Pi$, then a {\em $\cw$-labelling} of an undirected
1394: graph, $G$, is a map or ``labelling'' $\cl\from E_{\widehat G}
1395: \to\cw$ such that
1396: $\cl\bigl( \opp(e)\bigr) = \bigl( \cl(e)\bigr)^{-1}$ for each $e\in
1397: E_{\widehat G}$.
1398: For example, any graph $G\in\cgnd$ automatically comes with a
1399: $\Pi$-labelling, namely $\bigl( i,\pi_j(i)\bigr)$ is labelled $\pi_j$, and
1400: $\bigl( i,\pi_j^{-1}(i)\bigr)$ is labelled $\pi_j^{-1}$.
1401:
1402: An {\em orientation} of an undirected graph, $G$, is the distinguishing
1403: for each $e\in E_G$ of one of the two directed edges corresponding to $e$.
1404:
1405: The following definition is special to this paper.
1406: \begin{definition}\label{de:structural}
1407: Fix sets $V,E$ and a set of words, $\cw$, over
1408: $\Pi$, with $\cw^{-1}=\cw$.
1409: A {\em structural map} is a map $s\from E\to\cw\times V\times V$.
1410: A structural map defines a unique $\cw$-labelled, oriented graph,
1411: $G$, with $V_G=V$ and $E_G=E$,
1412: as follows: for each $e\in E$ with $s(e)=(\sigma,u,v)$,
1413: we form a directed edge of type $(u,v)$ labelled $\sigma$ and declare
1414: it distinguished, and pair it with a directed edge of type $(v,u)$
1415: labelled $\sigma^{-1}$.
1416: \end{definition}
1417:
1418:
1419: \subsection{Variable-Length Graphs and Subdivisions}
1420:
1421: In this paper we will work with graphs that have large or infinite
1422: parts of them being paths or regular trees. In this case we can
1423: easily eliminate
1424: all the vertices in these parts
1425: by working with
1426: ``variable-length graphs.''
1427: This leads to simpler calculations (in Theorem~\ref{th:infinite_edge}, that is
1428: crucial to Lemma~\ref{lm:finiteness}, and in
1429: Theorem~\ref{th:crucial_cycle_count}). This also gives us a notion
1430: of regular tree of non-integral degree, in Theorem~\ref{th:remarkable}.
1431:
1432:
1433: Recall (see \cite{shannon,adler,heegard,friedman_geometric_aspects}) that a
1434: {\em VLG} or {\em variable-length graph}
1435: (respectively, {\em directed VLG} or {\em directed variable-length graph})\mytwoindex{variable-length graph (VLG)}{a graph, directed or undirected,
1436: with a positive integral ``length'' associated to each edge}
1437: \mythreeindex{vlg}{VLG}{variable-length graph}
1438: is a graph (respectively, directed graph)
1439: with an assignment of a positive
1440: integer to each edge called the edge's
1441: {\em length}\footnote{In Shannon's terminology of \cite{shannon}, Chapter~1,
1442: Section~1, the edges have various ``times'' (as opposed to
1443: ``lengths'') such as a dot versus a dash
1444: in Morse code.}.
1445: The length of a walk in a VLG is the sum of the lengths of its edges
1446: (each length is counted the number of times the edge appears in the walk).
1447:
1448: A graph can be regarded as a VLG with all edge lengths $1$.
1449: A VLG whose edge lengths are all $1$ can be identified with its underlying
1450: graph.
1451:
1452: A {\em bead}\mytwoindex{bead}{a vertex with indegree and outdegree $1$ (or, for undirected graphs, degree $2$) with no self-loops}
1453: in a directed graph (respectively graph)
1454: is a vertex with indegree and outdegree
1455: $1$ (respectively, degree $2$)
1456: and without a self-loop. A {\em beaded path} is a path where every
1457: vertex except possibly the endpoints are beads.
1458: \begin{definition}\label{de:subdivision}
1459: Let $G$ be a directed VLG. To {\em subdivide} an edge, $e$, from $u$ to $v$
1460: and of length $\ell$,
1461: in $G$ is
1462: to replace $e$ with a beaded path of length $\ell$ from $u$ to $v$ in
1463: $G$ (introducing $\ell-1$ new vertices).
1464: A {\em subdivided form of $G$} is a graph, $G_\sbd$,
1465: obtained by subdividing all edges
1466: of $G$.
1467: The same definition is made for VLG's, omitting the word ``directed''
1468: everywhere, provided no half-loops are of length $2$ or greater\footnote{This
1469: restriction will be explained just before
1470: Proposition~\ref{pr:ird_invar}. Actually, we can define a notion of
1471: subdivision for all half-loops of odd length, but in this paper we
1472: use half-loops only of length $1$.}.
1473: \end{definition}
1474: It should be clear that countings walks of certain types in a directed VLG,
1475: $G$, should translate to an appropriate similar counting in $G_\sbd$,
1476: and vice versa. We next define an opposite of subdivision, supression,
1477: and a vast generalization of supression, realization.
1478:
1479: \begin{definition} Let $G$ be a strongly connected directed graph,
1480: with $W\subset V_G$. The {\em realization of $G$ to with vertex set $W$}
1481: denoted $\realize{G}{W}$,
1482: is the directed VLG on vertices $W$ with the following set of edges.
1483: We create $\realize{G}{W}$
1484: one edge from $u$ to $v$ (for $u,v\in W$) of length
1485: $k$ for each walk from $u$ to $v$ in $G$ of length $k$ that
1486: contains no $W$ vertices except as the first and last
1487: vertices. (So self-loops or edges in $G$ involving $W$ vertices
1488: appear in $\realize{G}{W}$, since we regard self-loops or edges as
1489: walks with no vertices except the first and last vertices.)
1490: \end{definition}
1491: The notion of realization appears constrained coding theory
1492: (called ``fusion'' in \cite{heegard}, for example, and not given a name in
1493: \cite{adler}; see also
1494: \cite{friedman_geometric_aspects}).
1495: We remark that if
1496: $V_G\setminus W$ contains a cycle, then $\realize{G}{W}$ has
1497: infinitely many edges.
1498: \begin{definition}
1499:
1500:
1501: Let $G$ be a strongly connected directed graph,
1502: with $U\subset V_G$ a subset of beads in $G$ such that $U$ contains no
1503: cycle. The {\em supression of $U$
1504: (in $G$)}, denoted $\supress{G}{U}$, is the realization of $G$ with
1505: vertex set $V_G\setminus U$.
1506: \end{definition}
1507: The subdivision (by the supressed vertices)
1508: of a supression returns the original directed graph.
1509:
1510: We remark that if $G$ is a $\Pi$-labelled graph, then any supression
1511: in $G$ has a natural $\Pi^{+}$-labelling, where $\Pi^+$ is the set
1512: of words on $\Pi$ of length $\ge 1$.
1513:
1514: %% The same is true for undirected VLG's, provided that we count
1515: %% {\em irreducible walks} in the VLG and in its subdivision.
1516: %% We make some precise versions of the last two statement in what follows.
1517: %% We warn the reader that in the case of half-loops of positive
1518: %% even length, it is a problem to define a good notion on subdivision
1519: %% (that preserves the irreducible walk counts)
1520: %% without extending the notion of a VLG; fortunately, in our undirected
1521: %% VLG's, all subdivisions will involve only half-loops of length $1$.
1522:
1523:
1524: \subsection{$\lambda_1$ of a VLG}
1525:
1526: Let $G$ be a VLG (directed or undirected). For $u,v\in V_G$ and a
1527: non-negative integer, $k$, let
1528: $c_G(u,v;k)$\mythreeindex{cG}{$c_G(u,v;k)$}{the number of walks of
1529: length $k$ from $u$ to $v$ in $G$}
1530: denote the number of walks of length $k$ from $u$ to $v$ in
1531: $G$. We will use standard Perron-Frobenius theory
1532: (see Sections~1.3 and 7.1 in \cite{kitchens} or
1533: Chapters~1 and 6 in \cite{seneta}), which includes the rest of this
1534: paragraph.
1535: Assume that $G$ is strongly connected, i.e., for each $u,v\in V_G$ we
1536: have $c_G(u,v;k)>0$ for some $k$. Let $d=d_G$ be the period of $G$,
1537: i.e., the greatest common divisor of the lengths of all closed walks in $G$.
1538: Then all limits
1539: $$
1540: \limsup_{k\to\infty} \bigl( c_G(u,v;k) \bigr)^{1/k}, \quad
1541: \lim_{k\to\infty} \bigl( c_G(v,v;kd) \bigr)^{1/kd}
1542: $$
1543: exist and are all equal (so independent of $u,v$); we define this
1544: common limit to be $\lambda_1(G)$\myindexlambdaone,
1545: the {\em Perron value} of $G$.
1546: It is easy to see that $c_G(v,v;k)\le \lambda_1^k$, using that
1547: $c_G(v,v;k_1)c_G(v,v;k_2)\ge c_G(v,v;k_1+k_2)$. If $G$ is a finite graph,
1548: then $\lambda_1(G)$ is just the usual Perron-Frobenius (largest) eigenvalue
1549: of $A_G$.
1550:
1551: In directed graphs,
1552: supression, realization, and subdivision preserve walk counts
1553: (i.e., the $c(u,v;k)$'s) between appropriate vertices (those present
1554: in the two graphs in question). Therefore these operations also
1555: preserve $\lambda_1$.
1556:
1557: If $G$ is not strongly connected, we can define
1558: $\lambda_1(G)$\myindexlambdaone to be the supremum of $\lambda_1$ of
1559: all the strongly connected components of $G$, or equivalently as the
1560: supremum over all $\bigl( c_G(v,v;k)\bigr)^{1/k}$.
1561:
1562: One can equivalently define $\lambda_1(G)$ with $\tilde c_G$ replacing
1563: $c_G$, where $\tilde c_G(u,v;k)$ is the number of walks of length at
1564: most $k$. This is a sensible definition of $\lambda_1(G)$ when we
1565: allow non-integral edge lengths\footnote{Why should a ``dash'' in
1566: Morse code be precisely an integral
1567: multiple of a ``dot''?}.
1568: One can also extend all these defitions to graphs with positively
1569: weighted edges,
1570: where the weight of a walk becomes the product of its edge weights, and
1571: where $c_G$ or $\tilde c_G$ sums the weights of the walks.
1572:
1573:
1574:
1575: \subsection{Shannon's Algorithm and Formal Series}
1576:
1577: Shannon gives the following
1578: algorithm (see \cite{shannon}, Chaper~1, Section~1)
1579: for computing $\lambda_1(G)$ (or the ``valence'' or ``capacity'')
1580: of a finite graph:
1581: let $Z_G=Z_G(z)$ be the matrix whose $i,j$ entry is the sum of $z^\ell$ over
1582: all edge lengths, $\ell$, of edges from $i$ to $j$,
1583: with $z$ a formal
1584: parameter. Then $\lambda_1(G)$ is the reciprocal of the smallest real root
1585: in $z$ of
1586: \begin{equation}\label{eq:shannon}
1587: \det\bigl(I-Z_G(z)\bigr) = 0 .
1588: \end{equation}
1589: In this section we explain variants of this theorem/algorithm that hold
1590: for infinite VLG's. We first give some conventions that we will use
1591: with formal power series.
1592:
1593: By a {\em non-negative power series} we mean a series
1594: $f(z)=\sum_{k=0}^\infty a_k z^k$, with $a_k$ non-negative reals.
1595: We say that $f$ is the
1596: {\em generating function}\mytwoindex{generating function}{a power series, $\Sigma_k a_k z^k$, formed from coefficients $a_k$} of the $a_k$.
1597: At times we view $f$ as a formal power series, but we will also have cause
1598: to evaluate $f$ at non-negative reals (as a possibly diverging infinite
1599: sum); it is easy to see that if $f(z_0)$ converges for a positive $z_0$,
1600: then $f$ is continuous on $[0,z_0]$, and if $f(z_0)$ diverges then
1601: $f(y_0)$ gets arbitrarily large as $y_0$ approaches $z_0$.
1602:
1603: For such an $f$, $f$'s radius of convergence is
1604: $$
1605: \rho(f) = \limsup_{k\to\infty} a_k^{1/k}.
1606: $$
1607: The function
1608: $f$ has a singularity at $z=\rho$, and $f(z_0)=\sum a_kz_0^k$ diverges
1609: for $z_0>\rho$. If the singularity at $z=\rho$ is a pole (e.g., when
1610: $f(z)$ is a rational function), then $f(z)\to +\infty$ as $z\to\rho$
1611: from the left.
1612:
1613: N.B.: We do not identitfy a formal power series with any of its analytic
1614: extensions unless we specifically say so. For example,
1615: $f(z)=1+z+z^2+\cdots$ has only the value $+\infty$ for real $z> 1$
1616: unless we explicitly say to the contrary.
1617:
1618: For a VLG, $G$, we set
1619: $$
1620: M_G(z) = I + Z_G(z) + Z_G^2(z) + \cdots
1621: $$
1622: We have
1623: \begin{equation}\label{eq:M}
1624: \bigl( M_G(z) \bigr)_{u,v} = \sum_{k=0}^\infty c_G(u,v;k)z^k.
1625: \end{equation}
1626:
1627: We say that a non-negative power series, $f(z)=\sum a_k z^k$, {\em majorizes}
1628: another one, $g(z)=\sum b_k z^k$, if
1629: $$
1630: a_1+\cdots+a_j \ge b_1+\ldots+ b_j
1631: $$
1632: for all $j\ge 1$. If so, then $f(z_0)\ge g(z_0)$ for any $z_0\in[0,1]$
1633: (with appropriate convensions on the value $+\infty$). Given VLG's,
1634: $G$ and $H$, we say that $G$ {\em majorizes} $H$ if $Z_G(z)$ majorizes
1635: $Z_H(z)$ entry by entry;
1636: equivalently, there is an endpoint preserving
1637: injection from $E_H$ to $E_G$ that does not increase edge lengths.
1638: If so, clearly $M_G(z)$ majorizes
1639: $M_H(z)$ entry by entry.
1640:
1641: \begin{theorem}\label{th:shannon_infinite}
1642: Let $G$ be a strongly connected
1643: VLG, directed or undirected, on a countable number of
1644: vertices and edges. The following are equal:
1645: \begin{enumerate}
1646: \item $1/\lambda_1(G)$, and
1647: \item the radius of convergence of any entry
1648: of $M_G(z)$.
1649: \end{enumerate}
1650: If $G$ has a finite number of vertices, then the above two numbers also
1651: equal the following two:
1652: \begin{enumerate}\addtocounter{enumi}{2}
1653: \item the supremum of positive $z$ such that
1654: $Z_G(z)$ converges (in each entry)
1655: and has largest (i.e., Perron-Frobenius) eigenvalue
1656: less than $1$, and
1657: \item the supremum of positive $z$ such that
1658: $Z_G(z)$ converges and $\det\bigl(I-Z_G(y)\bigr)>0$
1659: for all $y$ with $0<y<z$.
1660: \end{enumerate}
1661: If $G$ has a finite number of vertices, and if the entries $Z_G(z)$
1662: are all rational functions of $z$, then the four above quantities equal
1663: \begin{enumerate}\addtocounter{enumi}{4}
1664: \item the first positive solution to $\det\bigl(I-Z_G(z)\bigr)=0$.
1665: \end{enumerate}
1666: Finally, all entries of $Z_G(z)$ will be rational functions of $z$ whenever
1667: $G$ is finite (i.e., has finitely many vertices and edges)
1668: or is a VLG realization of finite graph.
1669: \end{theorem}
1670: \proof (1)=(2): Clear from equation~(\ref{eq:M}).
1671:
1672: (2)=(3):
1673: Let $B$ be a non-negative matrix. According to Perron-Frobenius theory,
1674: all the eigenvalues of $B$ are of absolute value at most $\lambda_1(B)$,
1675: with equality only when the eigenvalue has the same algebraic and geometric
1676: multiplicity.
1677: It now follows from Jordan canonical
1678: form that
1679: a finite dimensional, non-negative matrix, $B$, has largest
1680: (Perron-Frobenius) eigenvalue less than $1$ iff $I+B+B^2+\cdots$ converges.
1681:
1682: (3)=(4): If $\lambda_1\bigl(Z_G(z_0)\bigr)<1$, then
1683: $\lambda_1\bigl(Z_G(y_0)\bigr)<1$ for all $y_0<z_0$, and hence
1684: $\det\bigl(I-Z_G(y_0)\bigr)>0$ for such $y_0$. Consider the first
1685: $z_0$ for which $\lambda_1\bigl(Z_G(z_0)\bigr)<1$ fails to hold
1686: (such a $z_0$ exists by the continuity of $\lambda_1$ as a function of
1687: its entries). Either $Z_G(z_0)$ does not converge, or else by continuity
1688: of $\lambda_1$ we have $\lambda_1\bigl(Z_G(z_0)\bigr)=1$ and hence
1689: $\det\bigl(I-Z_G(z_0)\bigr)=0$.
1690:
1691: (4)=(5): By the above paragraph, if suffices to show that
1692: $Z_G(z)$ cannot have a pole (in any of its entries)
1693: at the first $z_0$ for which
1694: $\lambda_1\bigl(Z_G(z_0)\bigr)<1$ fails to hold. But if $Z_G(z)$ has a
1695: pole in some entry at $z=z_0$, then for some $v\in V_G$ and positive
1696: integer, $k$, $Z^k_G(z)_{v,v}$ has a pole at $z=z_0$, by the strong
1697: connectivity of $G$. But $\lambda_1^k\bigl( Z_G(z) \bigr)\ge
1698: Z^k_G(z)_{v,v}$ for any $z$, and the latter tends to $+\infty$ as
1699: $z\to z_0$ from the left.
1700:
1701: Last part of the theorem:
1702: Let $G'$ be a realization of a finite VLG, $G$, on the set $U\subset V_G$.
1703: For $u_1,u_2\in U$, we shall calculate the $(u_1,u_2)$-entry of $Z_{G'}(z)$.
1704: We claim this entry is the $(u_1,u_2)$-entry of $Z_{G}(z)$, plus the
1705: $(u_1,u_2)$-entry
1706: $$
1707: v^{\rm T} (I+Z_{\bar G}(z)+Z_{\bar G}^2(z)+\cdots) w = v^{\rm T}
1708: \bigl(I-Z_{\bar G}(z)\bigr)^{-1}w,
1709: $$
1710: where $\bar U$ is the induced subgraph of $G$ on the vertex set
1711: $\bar U=V_G\setminus U$, where $v$ is the vector whose entries correspond
1712: to the edges from $u_1$ to the vertices of $\bar U$ and similarly for
1713: $w$. Since $v,w,Z_{\bar G}(z)$ all have polynomial entries,
1714: we conclude that the $(u_1,u_2)$-entry of $Z_{G'}(z)$ is a rational
1715: function.
1716: \proofbox
1717:
1718: Now we give two examples to show that Shannon's algorithm does not
1719: literally apply as is to infinite graphs. Let $G$ be a directed VLG with
1720: one vertex and $a_k=\lfloor 2^k / (k+1)^2 \rfloor$ edges of length $k$
1721: for all $k\ge 1$.
1722: Then $Z_G(z)$ is a $1\times 1$ matrix with sole entry is
1723: $f(z)=\sum_k a_k z^k$. In this case we have $f(1/2)\le (\pi^2/6)-1<1$
1724: and $f(z)$ diverges for any positive $z\ge 1/2$.
1725: It is not hard to see that $\lambda_1(G)=2$ (by
1726: Theorem~\ref{th:shannon_infinite},
1727: since $1+f+f^2+\cdots$ converges for $f<1$, which is the case
1728: when $z\le 1/2$, and clearly diverges when $z>1/2$ where the series even
1729: for $f$ diverges).
1730: But the expression $\det\bigl(I-Z_G(z)\bigr)$ fails to have a zero at
1731: $z=1/2$.
1732:
1733:
1734: Next consider an undirected VLG, $G$, whose vertex set is the integers,
1735: with each vertex having one self loop. Then $\det\bigl(I-Z_G(z)\bigr)$
1736: expanded in a power series has an infinite $z$ coefficient; if we
1737: simply multiply the diagonals together (since this is a diagonal matrix),
1738: we get the infinite product of $(1-z)$, which is $0$ for any $z>0$
1739: (yet $\lambda_1(G)=1$).
1740: While $G$ is not connected, we can add edges between $i$ and $i+1$ for
1741: all $i$; this yields a connected graph with similar problems.
1742:
1743: Finally we mention that Theorem~\ref{th:shannon_infinite} must be
1744: modified when $G$ is not strongly connected. Indeed, consider a directed
1745: VLG
1746: on three nodes, such that the first node has a single self-loop, and
1747: the only other edges are edges from the second node to the third,
1748: with $a_k$ such edges of length $k$. Then the most natural way to
1749: define $\lambda_1(G)$ is in terms of counting closed walks, so that
1750: $\lambda_1(G)=1$. But if $\sum a_k z^k$ diverges for any $z_0<1$, then
1751: Theorem~\ref{th:shannon_infinite} fails.
1752:
1753:
1754: \subsection{Limiting Graphs}
1755:
1756: Let $G_i$ be a sequence of finite VLG's on the same vertex set, $V$, and the
1757: same edge set, $E$. Let $E$ be partitioned into two sets, $E_1,E_2$,
1758: such that the following holds:
1759: \begin{enumerate}
1760: \item for each $e\in E_1$, the length of $e$ in $G_i$ is independent of $i$,
1761: and
1762: \item for each $e\in E_2$, the length of $e$ in $G_i$ tends to infinity
1763: as $i\to\infty$.
1764: \end{enumerate}
1765: The {\em limit of the $G_i$} is the graph, $G_\infty$, which is any
1766: $G_i$ with its $E_2$ edges discarded.
1767:
1768: This simple remark is crucial for an important finiteness lemma
1769: (Lemma~\ref{lm:finiteness}).
1770: \begin{theorem}\label{th:infinite_edge}
1771: With notation as above,
1772: $$
1773: \lim_{i\to\infty} \lambda_1(G_i) = \lambda_1(G_\infty).
1774: $$
1775: \end{theorem}
1776: \proof By counting closed walks we see that
1777: $\lambda_1(G_i)\ge\lambda_1(G_\infty)$; this establishes the theorem
1778: with ``$\ge$'' replacing ``$=$''. To see ``$\le$'' replacing ``$=$'',
1779: assume that
1780: $\lambda_1\bigl(A_{G_\infty}(z_0)\bigr)<1$ for some $z_0\in [0,1]$. Clearly
1781: $A_{G_i}(z_0)\to A_{G_\infty}(z_0)$ as $i\to\infty$. So the continuity
1782: of $\lambda_1$ on its entries implies $A_{G_i}(z_0)<1$ for $i$ sufficiently
1783: large, and so $1/z_0\ge \limsup\lambda_1(G_i)$. Now take a supremum over
1784: $z_0$ with $\lambda_1\bigl(A_{G_\infty}(z_0)\bigr)<1$.
1785: \proofbox
1786:
1787:
1788:
1789: \subsection{Irreducible Eigenvalues}
1790: \label{sb:lambda_irred}
1791: Let $G$ be an undirected graph with corresponding directed graph
1792: $\widehat G$.
1793: Let $G_{\ird{}}$\mythreeindex{GIrred}{$G_{\ird{}}$}{edge graph of $G$ with
1794: edges joined only when they form an irreducible path}
1795: be the graph with vertices $E_{\widehat G}$ and an edge
1796: from $e_1$ to $e_2$ iff $e_1e_2$ forms an irreducible path in $G$;
1797: i.e., $e_1$ and $e_2$ are not opposites
1798: (i.e., paired) in $G$, and $e_1$ terminates in
1799: the vertex where $e_2$ originates. (Therefore, if $e$ is a half-loop
1800: in $G$, then
1801: there is no edge from $e$ to itself in $G_{\ird{}}$.)
1802: Then walks in $G_{\ird{}}$ give ``irreducible''
1803: (or ``reduced'' or ``non-backtracking'') walks
1804: in $G$.
1805: A closed walk of length $k$ in $G_{\ird{}}$ gives a closed walk in $G$ (with specified
1806: starting vertex) that is {\em strongly irreducible}, meaning that the
1807: closed walk is irreducible and the last step in the closed walk is not the inverse of
1808: the first step.
1809: We define the {\em irreducible eigenvalues of $G$} to be those of
1810: $G_{\ird{}}$, and we define the largest or Perron-Frobenius eigenvalue
1811: of $G_{\ird{}}$ to be $\lambda_{\ird{}}=\lambda_{\ird{}}(G)$,\mythreeindex{lambdaIrred}{$\lambda_{\ird{}}(G)$}{the largest eigenvalue of $G_{\ird{}}$} the
1812: {\em largest irreducible eigenvalue of $G$}.
1813:
1814: For $G$ an undirected VLG, we may define
1815: $G_{\ird{}}$ as a VLG and hence define $\lambda_{\ird{}}(G)=
1816: \lambda_1(G_{\ird{}})$.
1817: In any graph,
1818: an irreducible walk that enters a beaded-path must directly traverse this
1819: path to its end (any backward step makes the walk reducible).
1820: It easily follows that if we subdivide edges in a VLG that are not
1821: half-loops
1822: (i.e., whole-loops or edges between distinct vertices),
1823: then counts of irreducible walks of a given length between $V_G$ vertices
1824: remain the same\footnote{We do not know any simple or very natural way
1825: to subdivide half-loops of even length
1826: in a VLG while keeping $\lambda_{\ird{}}$
1827: invariant. A half-loop of length $\ell$ should be traversable zero
1828: or one time in a row (but not twice or more in a row) in an irreducible
1829: walk. For odd lengths, $\ell$, this can be achieved by adding a path
1830: (of length $(\ell-1)/2$) with a
1831: half-loop at the end.
1832: Morally speaking, subdividing an edge in a VLG, $G$,
1833: can be
1834: viewed as subdividing the corresponding edge pair in
1835: underlying directed graph, $\widehat G$, and then gluing them
1836: together with an ``opposite'' pairing (that glues the newly introduced
1837: vertices together); for half-loops, we should glue a single directed
1838: self-loop of length $\ell$ to itself, via reflection about the middle);
1839: for even length half-loops, that fact that reflection leaves the middle
1840: vertex fixed creates problems if we wish to remain in the catergory of
1841: undirected graphs$\ldots$
1842: }.
1843: As a
1844: consequence we get the following simple but important proposition.
1845: \begin{proposition}
1846: \label{pr:ird_invar}
1847: We have
1848: $\lambda_{\ird{}}$ of a VLG
1849: is invariant under the subdivision of any set of edges devoid of
1850: half-loops; in particular, $\lambda_{\ird{}}$ is invariant under
1851: passing to a subdivided form of any VLG with all half-loops of length one.
1852: Similarly,
1853: $\lambda_{\ird{}}$ of a graph is invariant under any supression.
1854: \end{proposition}
1855:
1856: We now state a theorem for use later in this paper; the theorem requires
1857: a definition.
1858: \begin{definition} A connected graph, $G$, is {\em loopy} if
1859: $|E_G|\ge |V_G|$, or equivalently if $G$ contains an irreducible closed walk,
1860: or
1861: equivalently if $G$ is not a tree. $G$ is {\em $1$-loopy} if $G$ is
1862: connected and the removal of any edge from $G$ leaves a graph each of
1863: whose connected components are loopy.
1864: \end{definition}
1865:
1866: \begin{theorem}\label{th:loopy}
1867: For a connected graph, $G$, the following are equivalent:
1868: \begin{enumerate}
1869: \item $G$ is $1$-loopy,
1870: \item $G_{\ird{}}$ is strongly connected, and
1871: \item $G$ is not a cycle and all vertices in $G$ have degree at least $2$.
1872: \end{enumerate}
1873: %% \begin{theorem}\label{th:loopy} A connected graph, $G$, has $G_{\ird{}}$
1874: %% strongly connected iff $G$ is $1$-loopy.
1875: In particular, if $G$ is connected
1876: and $d$-regular for $d\ge 3$, then $G_{\ird{}}$ is strongly connected.
1877: \end{theorem}
1878: Condition~(3) in the above theorem was pointed out to us by a referee.
1879: \proof
1880: (2)$\Rightarrow$(1):
1881: Consider a directed edge, $e\sim(u,v)$, of $G$, and let $e'$ be
1882: $e$'s opposite (we permit $e=e'$, i.e., the case of a half-loop).
1883: If the removal
1884: of $e$ leaves the connected component of $v$ being a tree, then there is
1885: no irreducible walk from $e$ to $e'$. On the other hand, if this
1886: connected component is not a tree, then a cycle in this component
1887: about $v$ of minimum
1888: length is irreducible, which then extends to an irreducible walk from
1889: $e$ to $e'$. To summarize, if $G$ is not $1$-loopy, then $G_{\ird{}}$
1890: is not strongly connected; otherwise, each edge has an irreducible
1891: path to its opposite, i.e., each edge is connected to its opposite
1892: in $G_{\ird{}}$.
1893:
1894: (1)$\Rightarrow$(2):
1895: Assume $G$ is $1$-loopy.
1896: If $e_1,e_2$ are two distinct, unpaired
1897: directed edges originating in the same vertex,
1898: $v$, then an irreducible path from $e_1$ to its opposite followed by
1899: $e_2$ shows that $e_1$ and $e_2$ are connected by an irreducible path.
1900: Thus any two edges that share a vertex are strongly
1901: connected in $G_{\ird{}}$.
1902:
1903: Finally if $e_1,e_2$ are two undirected edges without a common vertex,
1904: then a shortest path connecting them gives an irreducible path
1905: from some orientation of $e_1$ to some of $e_2$. By the above we can
1906: follow them with irreducible paths to the edge of opposite orientation.
1907: Thus any two edges that do not share a vertex are strongly
1908: connected in $G_{\ird{}}$.
1909: So $G_{\ird{}}$ is strongly connected.
1910:
1911: (1)$\Rightarrow$(3)
1912: Since an isolated vertex is loopy, a $1$-loopy graph has all vertices
1913: of degree at least $2$. Also a $1$-loopy graph cannot be a single cycle,
1914: since removing any one edge would give a tree.
1915:
1916: (3)$\Rightarrow$(1)
1917: Any tree either (i) is an isolated vertex, or (ii) is a path with two
1918: vertices of degree one, or (iii) has at least three vertices of degree
1919: $1$ (this follows by considering two vertices, $u,v$, of maximum distance in
1920: a tree, and then taking a vertex of maximum distance to the unique
1921: path joining $u$ to $v$).
1922: Consider the tree, $T$, obtained as the
1923: non-loopy connected component by removing an edge, $e$, from a graph
1924: $G$ that is non-loopy. If $T$ is the only connected component, $e$
1925: has both its endpoints in $T$, and if not then $T$ has only one endpoint
1926: in $T$; in either situation, considering the cases (i)--(iii), it is
1927: easy to check that $G$ is a cycle or has at least
1928: one vertex (a vertex in $T$) of degree $1$.
1929: \proofbox
1930:
1931:
1932: \subsection{$\lambda_1$ and Closed Walks for Infinite
1933: Graphs}
1934:
1935: In this subsection we recall some facts about $\lambda_1$ and closed walks of
1936: graphs, either indicating the proofs or giving references. This section
1937: is geared to infinite graphs, since most of the facts are very easy when
1938: the graph is finite.
1939:
1940: Let $G$ be a graph of bounded degree, i.e. there is an $r$ such that
1941: the degree of each vertex is at most $r$.
1942: Then $A_G$, $G$'s adjacency matrix, is a bounded linear operator (bounded
1943: by $r$) on $L^2(G)$, the square summable functions on $G$'s vertices.
1944: The next theorem shows that $\|A_G\|$, the norm of the operator $\|A_G\|$,
1945: equals $\lambda_1(G)$ as was defined by counting closed walks about a
1946: vertex of at most some given length.
1947: \begin{theorem} Assume $G$ is connected. For any vertex, $v$, of $G$ and
1948: positive integer, $k$, recall that
1949: $c(v,k)$ is the number of closed walks of length $k$ in $G$ about $v$. Then
1950: $c(v,k)\le \|A_G\|^k$, and
1951: $$
1952: \lim_{r\to\infty} [c(v,2r)]^{1/(2r)}
1953: $$
1954: exists and equals $\|A_G\|$.
1955: \end{theorem}
1956: Since the limit above equals $\lambda_1(G)$, we see that
1957: $\lambda_1(G)=\|A_G\|$.
1958: \proof See \cite{buck}.
1959: \proofbox
1960:
1961: We also need the following simple fact.
1962: \begin{theorem} For every $\epsilon>0$ there is an $f\ne 0$ such that
1963: $\|Af\|\ge(\lambda_1-\epsilon)\|f\|$, where $f$ has finite support.
1964: \end{theorem}
1965: \proof By definition of norm, there is a $g\ne 0$ with
1966: $\|Ag\|\ge\bigl(\lambda_1-(\epsilon/2)\bigr)\|g\|$. For any $\nu>0$
1967: we may write $g=g_1+g_2$ where $g_1$ is of finite support and $\|g_2\|\le
1968: \nu$. It is easy to see (using the fact that $A$ is bounded) that
1969: if $\nu$ is sufficiently small we can take $f=g_1$ to satisfy the
1970: above theroem.
1971: \proofbox
1972:
1973:
1974: \subsection{A Curious Theorem}
1975:
1976: \begin{figure}
1977: \centering
1978: % \scalebox{0.6}{\includegraphics{switchdesign.eps}}
1979: \epsfysize=2.5in
1980: \epsfbox[0 0 745 370]{dtree.eps}
1981: \caption{A graph, $\tang$, and ${\rm Tree}_d(\tang)$
1982: with $d=4$.}
1983: \label{fg:dtree}
1984: \end{figure}
1985:
1986:
1987: \begin{definition} Let
1988: $\tang$ be a finite connected graph with each vertex of degree at
1989: most $d$ for
1990: an integer $d>2$.
1991: By $\tree_d(\tang)$ we mean the unique
1992: (up to isomorphism) undirected graph, $G$,
1993: that has an inclusion $\iota\from \tang\to G$
1994: such that $G$ is $d$-regular
1995: and such that $G$ becomes a forest when we remove (the
1996: image under $\iota$ of) $\tang$'s edges.
1997: \end{definition}
1998: We have seen an example of this construction in Section~2, where
1999: $\tang$ is one vertex with $m$ self-loops, in the proof of
2000: Theorem~\ref{th:improved_bound}. See Figure~\ref{fg:dtree} for
2001: another example.
2002:
2003: In the category of $d$-regular graphs with a $\tang$ inclusion,
2004: $\tree_d(\tang)$ is none other than the universal cover.
2005:
2006: The methods used to prove Theorems~\ref{th:main}, \ref{th:mainh},
2007: and \ref{th:maini}
2008: suggest the following curious theorem.
2009:
2010: \begin{theorem}\label{th:remarkable}
2011: Let $d\ge 3$, and let $\tang$ be a finite connected
2012: graph with each vertex of degree
2013: $\le d$. Then
2014: \begin{eqnarray*}
2015: \lambda_1\bigl( \tree_d(\tang) \bigr)=2\sqrt{d-1} & \Longleftrightarrow &
2016: \lambda_{\ird{}}(\tang)\le\sqrt{d-1}, \\
2017: \lambda_1\bigl( \tree_d(\tang) \bigr)>2\sqrt{d-1} & \Longleftrightarrow &
2018: \lambda_{\ird{}}(\tang)
2019: >\sqrt{d-1}.
2020: \end{eqnarray*}
2021: The same is true for any real $d>2$, provided that
2022: $\lambda_1\bigl( \tree_d(\tang) \bigr)$ is interpreted with an appropriate
2023: analytic continuation in $d$ (described below).
2024: \end{theorem}
2025: % Note that if $\tang$ has a vertex, $v$, of degree $r$, then
2026: % $\lambda_{\ird{},1}(\tang)\ge \sqrt{r-1}$ (consider walks starting at $v$
2027: % whose every other
2028: % vertex is $v$); so $\tree_d(\tang)$ exists, i.e. each vertex of $\tang$ is of
2029: % degree $\le d$, provided that $\lambda_{\ird{},1}(\tang)\le\sqrt{d-1}$.
2030:
2031: Before proving the theorem, we describe the analytic continuation to which
2032: we refer.
2033:
2034: Let $\widetilde T$ be an undirected rooted tree,
2035: where every vertex has $d-1$ children,
2036: with $d>1$ an integer (for now). ($\widetilde T$ has degree $d-1$ at the
2037: root and degree $d$ elsewhere, so $\widetilde T$ is not regular.)
2038: Let $a_n$ for $n=2,4,\ldots$ be the number of walks in $\widetilde T$
2039: from the root to
2040: itself that never pass through the root except at the beginning and end.
2041: It is easy to see that
2042: $$
2043: S = S_d(z)= \sum_{n=2}^\infty a_n z^n,
2044: $$
2045: satisfies the recurrence $S=z^2(d-1)(1+S+S^2+\cdots)$ (since the walks
2046: counted by $S$ take one step, in $d-1$ possible ways, to a child
2047: of the root, followed by some number (possibly zero)
2048: of $S$ walks, followed by the step back to the root); so
2049: $S(1-S)=z^2(d-1)$, so that near $z=0$
2050: \begin{equation}\label{eq:S_quadratic}
2051: S=S_d(z) = \frac{1-\sqrt{1-4(d-1)z^2}}{2}.
2052: \end{equation}
2053:
2054: One can alternatively say that if $H$ is the realization of $\widetilde T$
2055: with vertex set $\{v_0\}$, where $v_0$ is the root, then the $1\times 1$
2056: matrix $Z_H(z)$ has entry $S(z)$. Similarly, if $\widetilde T$ were
2057: modified to have $d-r$ children at the root (but every other vertex with
2058: $d-1$ children, as before), then $Z_H(z)$ would have entry $(d-r)S(z)/(d-1)$.
2059:
2060: By $\lambda_1\bigl( \tree_d(\tang) \bigr)$ for $d>1$ real
2061: (assuming each degree in $\tang$ is $\le d$)
2062: we mean $\lambda_1$
2063: of the VLG formed from $\tang$ with an additional self-loop about
2064: each vertex of degree $r$ that has ``formal weight''
2065: $(d-r)S_d(z)/(d-1)$ (this can be viewed as adding an infinite set of self-loops
2066: of given weights and lengths corresponding to this power series\footnote{
2067: In other words, we view a power series $\sum a_nz^n$ as the sum
2068: of terms representing
2069: $a_n$ edges of length $n$ (even when $a_n$ is not an integer).
2070: }, or can simply be viewed as a term to add
2071: to the diagonal of $Z_G(z)$).
2072: When $d$ in an integer, the realization of $\tree_d(\tang)$ with vertex
2073: set $V_\tang$ is exactly this VLG, and by Proposition~\ref{pr:ird_invar}
2074: we know $\lambda_{\ird{}}$ remains the same.
2075: \proof
2076: By \cite{godsil}, exercise 13 page 72, we know that
2077: $\lambda_{\ird{}}(A)$ is given by
2078: $1/y$ of the smallest root, $y$, of
2079: $$
2080: \det\bigl(I-yA+y^2(D-I)\bigr) = 0,
2081: $$
2082: where $D$ is the diagonal matrix whose entry at vertex $v$ is the degree of
2083: $v$, and $I$ is the identity matrix. Now
2084: $$
2085: Z_G(z) = zA + \frac{dI-D}{d-1}\; S(z).
2086: $$
2087: It is then easy to verify that
2088: $$
2089: I - Z_G(z) = \bigl( 1-S(z)\bigr) \bigl(I-yA+y^2(D-I)\bigr)
2090: $$
2091: for $y=y(z)=z/\bigl( 1-S(z)\bigr) $ (note that $y(z)$ is an increasing
2092: function in $z$ for as long as $z(1+S+S^2+\cdots)$ converges).
2093: But $S(z_1)\le 1/2$ for all
2094: positive real $z_1\le z_0$, where $z_0=1/\bigl( 2\sqrt{d-1}\bigr)$,
2095: and also $y(z_0)=1/\sqrt{d-1}$. So $I-y_1A+y_1^2(D-I)$ becomes non-invertible
2096: for a $y_1<y(z_0)$ precisely when $Z_G(z_1)$ has eigenvalue $1$ for a
2097: $z_1\in [0,z_0]$. Now apply the equality of quantities~(1) and (3)
2098: in Theorem~\ref{th:shannon_infinite}.
2099: \proofbox
2100: Note that the proof shows that if $1/y_1=\lambda_{\ird{}}(A)$
2101: and $1/z_1$ is $\tree_d(\tang)$, then $y_1=y(z_1)$ provided
2102: $1/z_1 > 2\sqrt{d-1}$.
2103:
2104:
2105:
2106: %% ************
2107:
2108: %% We give two proofs of the above theorem. The first is to consider
2109: %% $z\le z_0=1/\bigl( 2\sqrt{d-1}\bigr)$ in the series for
2110: %% $\lambda_1\bigl( \tree_d(\tang) \bigr)$. We have $S_d(z_0)= 1/2$
2111: %% (i.e., the formal power series defining $S_d(z)$ converges to $1/2$
2112: %% at $z=z_0$), since as $z\to z_0$ from the left, $S_d(z)$ tends to
2113: %% $1/2$ and the power series defining $S_d(z)$ consists of all positive terms.
2114: %% So
2115: %% a self-loop of $(d-r)S/(d-1)$ on a vertex of degree $r$ adds a
2116: %% $(d-r)/(2d-2)$
2117: %% to the diagonal of $Z_G(z_0)$. We get
2118: %% $$
2119: %% Z_G(z_0) = z_0 A + (dI-D)/(2d-2),
2120: %% $$
2121: %% where $D$ is the diagonal matrix whose entry at vertex $v$ is the degree of
2122: %% $v$, and $I$ is the identity matrx. We wish to know when
2123: %% $I-Z_G(z_0)$ has all positive
2124: %% eigenvalues (and $I-Z_G(z)$ for all $z<z_0$); but
2125: %% $$
2126: %% I-Z_G(z_0) = I - \frac{A}{2\sqrt{d-1}} - (dI-D)/(2d-2)
2127: %% = \bigl(I - y_0A+y_0^2(D-I) \bigl)/2,
2128: %% $$
2129: %% where $y_0=1/\sqrt{d-1}$.
2130: %% Our theorem now follows from the fact that $\lambda_{\ird{}}(A)$ is given by
2131: %% $1/y$ of the smallest root, $y$, of
2132: %% $$
2133: %% \det\bigl(I-yA+y^2(D-I)\bigr) = 0
2134: %% $$
2135: %% (see \cite{godsil}, exercise 13 page 72).
2136:
2137: %% We give another proof of the theorem; we first assume that $d$ is
2138: %% integral.
2139: %% Consider the universal
2140: %% cover, $T$, of $\tree_d(\tang)$. $T$ is a $d$-regular tree.
2141: %% Let $v$ be a vertex of $\tang$, which we may also view as a vertex of
2142: %% $\tree_d(\tang)$, and let $\bar v$ be a vertex of $T$ that maps to $v$.
2143: %% Let $b_k$ be the number of closed walks of length $k$ about $v$ in $\tree_d(\tang)$.
2144: %% Let $c_k$ be the number of closed walks of length $k$ in $T$ about $\bar v$
2145: %% (or about any vertex, since $T$ is a $d$-regular tree).
2146: %% Let $a_k$ be the number vertices in $T$ of
2147: %% distance $k$ to $\bar v$ that map to $v$, or equivalently the number of
2148: %% irreducible closed walks of length $k$ about $v$ in $\tang$ (or in $\tree_d(\tang)$).
2149: %% Set $f=f(z)$ to be the generating function for $a_k$, i.e.,
2150: %% $$
2151: %% f(z)=\sum_{k=0}^\infty a_k z^k,
2152: %% $$
2153: %% and set $g,h$ to be the generating functions for, respectively $b_k,c_k$.
2154:
2155: %% First of all, we claim that
2156: %% $$
2157: %% % h(z)=\bigl( 1-dz^2/(1-S_d) \bigr)^{-1};
2158: %% h(z)=\left( 1-\frac{dz^2}{1-S_d} \right)^{-1}
2159: %% = \left( 1-\frac{dS_d}{d-1} \right)^{-1};
2160: %% $$
2161: %% indeed, the generating function for the closed walks about a $\bar v$ that
2162: %% {\em do not} pass through $\bar v$ anywhere in the middle is
2163: %% $$
2164: %% R_d=dz^2/(1-S_d)= d S_d/(d-1),
2165: %% $$
2166: %% by the same argument used to derive the equation for
2167: %% $S_d$; allowing walks that pass through $\bar v$ any number of times
2168: %% is
2169: %% $$
2170: %% 1+R_d+R_d^2+\cdots = 1/(1-R_d) = h(z).
2171: %% $$
2172:
2173: %% We see that $h(z)$ has radius of convergence $1/\bigl( 2\sqrt{d-1}\bigr)$,
2174: %% given equation~(\ref{eq:S_quadratic}).
2175: %% Next, we claim that the generating function
2176: %% for the walks in $T$ from one vertex, $v_0$,
2177: %% to a neighbouring vertex, $v_1$, that don't
2178: %% pass through $v_1$ in the middle, is $z/(1-S_d)$;
2179: %% indeed, such a walk from $v_0$ to $v_1$ takes a walk from $v_0$ to itself,
2180: %% avoiding $v_1$ (which corresponds to a walk from the root of
2181: %% $\widetilde T$ to itself), followed by one step from $v_0$ to $v_1$.
2182: %% Now a closed walk about $v$
2183: %% in $\tree_d(\tang)$ has a unique lifting to a walk in $T$ beginning at $\bar v$
2184: %% and ending at a vertex, $\bar v'$, that maps to $v$;
2185: %% if the distance of $\bar v$ to $\bar v'$
2186: %% is $m$, then the generating function walks from $\bar v$ to $\bar v'$ is
2187: %% $$
2188: %% U_{m,d}(z)=\bigl( z/(1-S_d) \bigr)^m h(z),
2189: %% $$
2190: %% since such a walk first hits the vertex of distance $m-1$ to $\bar v'$ at
2191: %% some first time, then the vertex of distance $m-2$, etc., and then makes
2192: %% some closed walk about $\bar v'$. Since the number of $\bar v'$ mapping to $v$
2193: %% is counted by the $a_k$ and the
2194: %% generating function, $f$, and since each $\bar v'$
2195: %% contributes a $U_{m,d}$ to $g(z)$, we have
2196: %% \begin{equation}\label{eq:gen_fun_rel}
2197: %% g(z)=\sum_{k=0}^\infty a_k U_{k,d}(z) =
2198: %% h(z)\sum_{k=0}^\infty a_k \bigl( z/(1-S_d)\bigr)^k =
2199: %% h(z) f\bigl( z/(1-S_d)\bigr) .
2200: %% \end{equation}
2201: %% It follows that the radius of convergence of $g$ is less than
2202: %% $1/\bigl( 2\sqrt{d-1}\bigr)$ if and only if the same is true for
2203: %% $f\bigl( z/(1-S_d)\bigr)$. But at $z=1/\bigl( 2\sqrt{d-1}\bigr)$ we
2204: %% have $S_d=1/2$, so $z/(1-S_d)=(d-1)^{-1/2}$. So the question of $g$'s
2205: %% radius of convergence boils down to whether or not $f$'s radius of
2206: %% convergence is less than $(d-1)^{-1/2}$, and this proves the theorem
2207: %% (for $d$ integer); here our argument is simplest if we make use of the
2208: %% fact that a power series about $z=0$ with radius of convergence $\rho$
2209: %% and with non-negative coefficients always has a singularity at $\rho$
2210: %% (see the proof of Theorem~\ref{th:precise_remarkable}).
2211:
2212: %% If $d$ is real, the same proof carries over, although we treat
2213: %% self-loops of formal weight, representing root-children additions,
2214: %% as being collections of trees. So we take $T$ to be the universal
2215: %% cover of $H$, and then to each vertex of degree $r$ we add a
2216: %% self-loop of formal weight $(d-r)S/(d-1)$; in other words,
2217: %% the self-loops of formal
2218: %% weight in $\tree_d(\tang)$ are lifted to $T$
2219: %% but left intact in $T$ (since these self-loops are collections of trees).
2220: %% The rest of the proof goes through with the above minor adjustment.
2221: %% \proofbox
2222:
2223: %% The following consequence of the above proof is worth stating, but will
2224: %% not be used in the rest of this paper.
2225:
2226: %% \begin{theorem}\label{th:precise_remarkable}
2227: %% Let $z_0=1/\lambda_1(\tree_d(\tang))$ and
2228: %% $z_1=1/\lambda_{\ird{}}(\tang)$, with notation as in
2229: %% Theorem~\ref{th:remarkable} ($d>2$ is real, and each vertex of $\tang$
2230: %% has degree $\le d$). Then if $\lambda_1(\tree_d(\tang))>2\sqrt{d-1}$,
2231: %% we have
2232: %% $$
2233: %% z_1=\frac{z_0}{1-S(z_0)}.
2234: %% $$
2235: %% \end{theorem}
2236: %% \proof This follows from equation~(\ref{eq:gen_fun_rel}) and the
2237: %% well-known fact
2238: %% that a power series $\sum \alpha_n z^n$ with $\alpha_n\ge 0$ has a
2239: %% first (smallest) singularity on the positive real axis\footnote{
2240: %% If not, there is a series s(z)=$\sum\alpha_n z^n$ with $\alpha_n\ge 0$
2241: %% with radius of convergence $1$ such that
2242: %% $\sum \beta_m(z-1+\epsilon)^m$ agrees with the former power series
2243: %% near $z=1-\epsilon$, and $\sum \beta_m u^m$ has radius of
2244: %% convergence $3\epsilon$, for some $\epsilon>0$.
2245: %% But the $\beta_m$, arising from derivatives
2246: %% of $s$ at $z=1-\epsilon$, are also non-negative. Thus
2247: %% $\sum \beta_m u^m=\sum \alpha_n (u+1-\epsilon)^n$ can be
2248: %% expanded in a series of non-negative terms $\alpha_n u^m
2249: %% (1-\epsilon)^{n-m}\binom{n}{m}$. Since $\sum \beta_m (2\epsilon)^m$
2250: %% converges, so does $\sum \alpha_n (1+\epsilon)^n$ (by rearranging
2251: %% terms in a non-negative infinite sum); but then $s(z)$'s radius
2252: %% of convergence is $>1$.
2253: %% };
2254: %% indeed, if $g(z)$ has its first positive
2255: %% singularity at $z_0$, then $f(z)$ must have
2256: %% one at $z_1=z_0/\bigl(1-S_d(z_0)\bigr)$, since $h$ is analytic at $z=z_0$.
2257: %% Since $z/(1-S)=(S/z)/(d-1)$, $z/(1-S)$ can be viewed as a power series
2258: %% with non-negative coefficients; hence $z/(1-S)$ is increasing in real,
2259: %% positive $z$. If $f$ has a singularity before $z_1$, then $g$ would have
2260: %% a positive real singularity before $z_0$, which is impossible. Thus
2261: %% $f$'s first positive singularity is $z_1$.
2262: %% \proofbox
2263:
2264:
2265:
2266: \section{Tangles}
2267:
2268: In Section~2 we saw that a vertex with $m$ self-loops in a
2269: $\cgnd$ graph, with $m$ ``large,'' gives rise to a ``large'' second
2270: eigenvalue (i.e., larger than $2\sqrt{d-1}$ for sufficiently large $m$,
2271: as $n\to\infty$). Here we generalize this observation to what we call
2272: a ``tangle.'' Our proof of the Alon conjecture via a trace method must
2273: somehow overcome all ``hypercritical'' tangles.
2274:
2275:
2276: \begin{definition}
2277: Given two $\Pi$-labelled graphs, $G$ and $H$, we say $G$ {\em contains}
2278: $H$ (or $H$ occurs in $G$) if there is an
2279: inclusion\footnote{By an inclusion we mean a graph homomorphism that is
2280: a injection on the vertices and on
2281: the edges.} $\iota\from H\to G$ that preserves
2282: the labelling;
2283: the {\em number of times} a graph, $H$, {\em occurs} in $G$ is
2284: the number of distinct\footnote{
2285: For example, if $H=G$ consists of one edge joining two distinct
2286: vertices, $u,v$, then the identity is considered distinct from
2287: the morphism interchanging the vertices. By the same principle,
2288: if $H$ has exactly $k$ automorphisms, then the number of times
2289: $H$ occurs in a graph is always a multiple of $k$.
2290: } such $\iota$.
2291: A {\em tangle} (or {\em $\cgnd$-tangle}) is a $\Pi$-labelled
2292: connected graph, $\tang$, that is contained in some element of $\cgnd$.
2293: \end{definition}
2294: For example, in Section~2 we studied the
2295: tangle with one vertex and $m$ self-loops labelled $\pi_1,\ldots,\pi_m$.
2296: We define $\chnd$- and $\cind$- and $\cjnd$-tangles similarly, with
2297: the following modifications to the meaning of the $\pi_i$'s. For
2298: $\chnd$, each of the $d/2$ independent $\pi_i$'s is uniform over all
2299: permutations whose cyclic decomposition consists of a single cycle.
2300: For $\cind$, each of $d$ independent $\pi_i$'s are uniform over all
2301: perfect matchings, i.e., over all permutations that are involutions
2302: without fixed points.
2303: For $\cjnd$, each of $d$ independent $\pi_i$'s are uniform over all
2304: involutions with exactly one fixed point.
2305:
2306:
2307: \begin{theorem}\label{th:key_to_s}
2308: Fix a positive integer, $d\ge 3$, and a graph $\tang$. Any graph, $G$,
2309: on $n$ vertices,
2310: that contains $\tang$ has second eigenvalue at least
2311: $\rho-o(1)$, where $o(1)$ is a function of $n$ tending to $0$ as $n\to\infty$,
2312: and where $\rho$ is the norm of the adjacency matrix of $\tree_d(\tang)$.
2313: \end{theorem}
2314: This generalizes Theorem~\ref{th:improved_bound}.
2315: \proof Fix $\epsilon>0$; we will show that for $n$ sufficiently large any
2316: such $G$ has $\lambda_2(G)\ge\rho-\epsilon$. First, there is a finitely
2317: supported $f\ne 0$ on $\tree_d(\tang)$ with $\|Af\|\ge \|f\|(\rho-\epsilon)$,
2318: where $A$ is the adjacency matrix of $\tree_d(\tang)$.
2319: If $V_\tang$ is the set of vertices on which $f$ is non-zero, then replacing $f$
2320: with the non-negative first Dirichlet eigenfunction on $V_\tang$ (see
2321: \cite{friedman_geometric_aspects}) we may assume $f$ is non-negative,
2322: nonzero, and that $Af\ge f(\rho-\epsilon)$ (with equality everywhere except
2323: at the boundary of $V_\tang$).
2324: There is a covering map (see \cite{friedman_geometric_aspects} and
2325: \cite{friedman_relative}) $\pi\from\tree_d(\tang)\to G$; set $\pi_*f$ to be
2326: the function on $G$ defined by
2327: $$
2328: (\pi_* f)(v) = \sum_{\pi(w)=v} f(w).
2329: $$
2330: If $A_G$ is the adjacency matrix of $G$, then clearly
2331: $$
2332: A_G (\pi_* f) \ge (\rho-\epsilon)(\pi_* f).
2333: $$
2334: So, on the one hand,
2335: ${\cal R}(\pi_* f)\ge \rho-\epsilon$, where ${\cal R}$ is
2336: the Rayleigh quotient for $A_G$.
2337: On the other hand, the support, $N$, of $\pi_* f$ is of size no greater than
2338: that of $f$, and this size is bounded (independent of $G$ and $n$).
2339: So the same reasoning as in the proof of Theorem~\ref{th:improved_bound}
2340: shows that
2341: $$
2342: {\cal R}(g) \ge d-o(1), \qquad\mbox{where}\quad
2343: g=\chi_{V\setminus \widetilde N},
2344: $$
2345: where $\widetilde N$ is the set of vertices of distance $0$ or $1$ to $N$.
2346: Since $\pi_* f$ is orthogonal to both $g$ and $A_G g$,
2347: we are done (by Lemma~\ref{lm:fix_bug}).
2348: \proofbox
2349:
2350: \begin{definition} A tangle, $\tang$, is {\em critical}\mytwoindex{critical}{
2351: a tangle with $\lambda_{\ird{}}=\sqrt{d-1}$} (respectively,
2352: {\em supercritical}\mytwoindex{supercritical}{
2353: a tangle with $\lambda_{\ird{}}\ge\sqrt{d-1}$},
2354: {\em hypercritical}\mytwoindex{hypercritical}{
2355: a tangle with $\lambda_{\ird{}}>\sqrt{d-1}$}) if $\lambda_{\ird{}}(\tang)$
2356: equals (respectively, is at least, exceeds) $\sqrt{d-1}$.
2357: \end{definition}
2358: According Theorems~\ref{th:remarkable} and \ref{th:key_to_s},
2359: a fixed
2360: hypercritical tangle can only occur in a graph with sufficiently many vertices
2361: if the graph has $\lambda_2>2\sqrt{d-1}$.
2362:
2363:
2364: Now that we know how tangles affect eigenvalues, we want to know how often
2365: the tangles occur. This discussion, and the particular application to
2366: $\chnd$, $\cind$, and $\cjnd$, will take the rest of this section.
2367:
2368: \begin{definition} A {\em leaf} on a graph is a vertex of total degree $1$
2369: (whether the graph is directed or not).
2370: We say that a graph is {\rm pruned} if it has no leaves.
2371: A {\em simple pruning} is the act of removing one leaf and its incident
2372: edge from a graph; {\em pruning} is the repeated performance of some
2373: sequence of simple prunings; {\em complete pruning} is the act of pruning
2374: until no more pruning can be done.
2375: \end{definition}
2376: For example, completely pruning a tree results in a single vertex with
2377: no edges;
2378: completely pruning a cycle leaves the cycle unchanged.
2379: \begin{proposition}\label{pr:lies_on_irred}
2380: Given a graph, $G$, there is a unique pruned graph $H$
2381: obtainable from completely pruning $G$. Furthermore, $H$ is completely
2382: pruned iff each edge of $H$ lies on an irreducible cycle.
2383: \end{proposition}
2384: \proof Let $e_1,\ldots,e_t$ denote the edges pruned in
2385: one pruning of $G$, in the
2386: order in which they are pruned. Let $G'$ be a different complete pruning
2387: of $G$, which we assume does not contain all the $e_i$. Let $j$ be the
2388: smallest integer such that $e_j$ lies in $G'$. On the one hand,
2389: the removal of $e_1,\ldots,e_{j-1}$ from $G$, or any subgraph
2390: of $G$, allows $e_j$
2391: to be pruned from $G$, or any subgraph of $G$, including $G'$. On the
2392: other hand, the prunability of $e_j$ from $G'$ contradicts the completeness
2393: of the pruning that formed $G'$. It follows that any complete pruning
2394: contains all the edges of any other, and so any two are the same.
2395:
2396: We now address the last statement of the theorem.
2397: If $H$ has a leaf, then the edge incident upon this leaf does
2398: not lie in an irreducible cycle. Conversely, if $H$ has no leaves,
2399: and if $e$ is an edge with endpoints $u,v$, consider the graph, $H'$,
2400: obtained by removing $e$ from $H$. If $u$ and $v$ are connected in $H'$,
2401: then a minimal length path that joins them, along with $e$, gives an
2402: irreducible cycle containing $e$. Otherwise, since $u$'s connected
2403: component is not a tree (or else $H$ would have leaves), this
2404: component has an irreducible cycle, and a shortest walk from $u$ to
2405: this cycle, once around the cycle, and back to $u$ gives an irreducible
2406: cycle beginning and ending at $u$. Similarly there is such a cycle about
2407: $v$. The cycle about $u$, followed by $e$, followed by the $w$ cycle, and
2408: back through $e$ (in the other direction), gives an irreducible cycle
2409: containing $e$.
2410: \proofbox
2411:
2412: A morphism of tangles is a morphism of $\Pi$-labelled graphs, i.e., a
2413: graph morphism that preserves
2414: the edge labelling.
2415:
2416: \begin{definition}\label{de:tangle_order}
2417: The {\em order} of a tangle, $\tang$, is
2418: ${\rm ord}(\tang)=|E_\tang|-|V_\tang|$ (so while a whole-loop is counted as one
2419: edge, for the model $\cjnd$, to
2420: be considered soon, a half-loop is also counted as one edge\footnote{
2421: Rougly speaking,
2422: the reason for this is that a whole-loop and half-loop are both
2423: $1/n+O(1/n^2)$ probability events.}). More generally,
2424: for a graph, $G$, or any structure with an underlying graph, $G$ (such as
2425: a ``form'' or ``type'' to be defined in Section 5), its order is
2426: ${\rm ord}(G)=|E_G|-|V_G|$.
2427: \end{definition}
2428:
2429: \begin{theorem}\label{th:tangle_count}
2430: Let $\tang$ be a tangle of non-negative order.
2431: Then the expected number of occurrences
2432: of $\tang$ in an element, $G$, of $\cgnd$
2433: is $n^{-r}+O(n^{-r-1})$, where $r$ is the order of $\tang$.
2434: The probability that at least
2435: one occurrence occurs is at least
2436: $n^{-r}/c - n^{-2r}/(2c^2)+O(n^{-r-1})$, with $r$ as
2437: before and where
2438: $c$ is the number of automorphisms of the complete pruning of $\tang$.
2439: \end{theorem}
2440: Notice that the number of automorphisms of the complete pruning of a tangle,
2441: $\tang$, is at least as many as that of $\tang$, and it may be strictly
2442: greater\footnote{Indeed, a structural induction argument (i.e., by
2443: pruning one leaf) shows that any
2444: automorphism of the complete pruning of $\tang$
2445: has at most one extension to $\tang$.
2446: On the other hand, if $\tang$ is a cycle of length $q$ with all edges in one
2447: ``direction'' labelled $\pi_1$, then $\tang$ has $q$ automorphisms; yet if
2448: we add one edge labelled $\pi_2$
2449: to $\tang$ at any vertex, the new graph
2450: has only the trivial automorphism.}.
2451: The proof following will imply that the probability of $\tang$'s
2452: occurrence is also
2453: at most $n^{-r}/c+O(n^{-r-1})$, which matches the lower bound to first order,
2454: provided that $r\ge 1$.
2455: \proof
2456: Let $V_\tang=\{u_1,\ldots,u_s\}$, and for a tuple $\vec m=(m_1,\ldots,m_s)$
2457: of distinct integers between $1$ and $n$, let $\iota_{\vec m}$ denote the event
2458: that the map, $\iota$, mapping $u_i$ to $m_i$, is an occurrence of $\tang$
2459: in $G$.
2460: If $a_i$ is the number of $\tang$'s edges labelled $\pi_i$, then
2461: each event $\iota_{\vec m}$ involves
2462: setting $a_i$ values of
2463: $\pi_i$, all of which occur
2464: with probability
2465: \begin{equation}\label{eq:tangle_expecteds}
2466: \frac{(n-a_1)!}{n!}\cdots \frac{(n-a_{d/2})!}{n!}.
2467: \end{equation}
2468: Since the sum of the $a_i$ is $|E_\tang|$, this probability is
2469: $n^{-|E_\tang|}$.
2470: Since there are $n!/(n-|V_\tang|)!=n^{|V_\tang|}+O(n^{|V_\tang|-1})$ different
2471: $\iota_{\vec m}$'s,
2472: the expected number of occurrences is $n^{-r}+O(n^{-r-1})$, where
2473: $r={\rm ord}(\tang)$.
2474:
2475: Next notice that if $\tang'$ is a pruning of a tangle, $\tang$, then
2476: the probability that $\tang$ occurs is $1+O(n^{-1})$ times the probability
2477: that $\tang'$ occurs (adding each pruned edge adds a condition that occurs
2478: with probability between $1$ and $(n-c_1)/(n-c_2)$ with $c_1,c_2$
2479: constants). Hence we may assume $\tang$ is pruned.
2480:
2481: An automorphism of $\tang$ can be viewed as a permutation on
2482: $V_\tang$, which is the same as a permutation, $\sigma$, on
2483: $\{1,\ldots,s\}$ (identifying a $u_i\in V_\tang$ with $i$).
2484: Such a permutation, $\sigma$, acts by permuting the components
2485: of the $\vec m$'s. Say that $\iota_{\vec m}$ is equivalent to
2486: $\iota_{\vec k}$ if $\vec m$ and $\vec k$ differ by a permutation,
2487: $\sigma$, associated to an automorphism of $\tang$; i.e., if
2488: $\iota_{\vec m}$ and $\iota_{\vec k}$ correspond to the same subgraph
2489: of $G$. Let $R$ be a set
2490: of representatives in the equivalence classes of all $\vec m$'s.
2491:
2492: By inclusion/exclusion, the probability that $\tang$ occurs at least once is
2493: at least
2494: \begin{equation}\label{eq:incexc}
2495: \sum_{{\vec m}\in R} \prob{\iota_{\vec m}} -
2496: \frac{1}{2}\sum_{\substack{ {\vec k}\ne {\vec m} \\ {\vec k},{\vec m}\in R}}
2497: \prob{\iota_{\vec m}\cap\iota_{\vec k}}.
2498: \end{equation}
2499: The first summand is $(1/c)n^{-r}+O(n^{-r-1})$, by the argument given
2500: for the expected number. For the second summand, we may write
2501: $\iota_{\vec m}\cap\iota_{\vec k}$ as $\iota_{\vec q}(\tang')$, where
2502: $\vec q$ is a vector comprised of the
2503: distinct components of $\vec m$ and $\vec k$, and where $\tang'$ is the tangle
2504: obtained by gluing two copies of $\tang$ along certain vertices (corresponding
2505: to where the components of $\vec m$ and $\vec k$ coincide). If
2506: $\vec m$ is disjoint from (i.e., nowhere coincides with)
2507: $\vec k$, then $\tang'$ is two disjoint
2508: copies of $\tang$; for fixed $\vec k\in R$ we have
2509: $$
2510: \sum_{\substack{ {\vec m}\in R\\ {\vec m}\;{\rm disjoint\;from\;}\vec k}}
2511: \prob{\iota_{\vec m}\;|\;\iota_{\vec k}} = n^{-r}/c + O(n^{-r-1}),
2512: $$
2513: the summation being over conditional probabilities, since the conditioning
2514: of $\iota_{\vec k}$ and summing over $\vec m$ disjoint only affects
2515: equation~(\ref{eq:tangle_expecteds}) by changing $n!/(n-a_i)!$ terms
2516: into $(n-c_i)!/(n-c_i-a_i)!$ terms for constants $c_i$, which is a
2517: second order change.
2518: Hence
2519: $$
2520: \sum_{\substack{{\vec k},{\vec m}\in R \\ {\vec k},{\vec m}\;{\rm disjoint}}}
2521: \prob{\iota_{\vec m}\cap\iota_{\vec k}}
2522: =
2523: \sum_{ \vec k\in R} \prob{\iota_{\vec k}}
2524: \sum_{\substack{ {\vec m}\in R\\ {\vec m}\;{\rm disjoint\;from\;}\vec k}}
2525: \prob{\iota_{\vec m}\;|\;\iota_{\vec k}}
2526: $$
2527: $$
2528: = \sum_{ \vec k\in R} \prob{\iota_{\vec k}} \bigl( n^{-r}/c + O(n^{-r-1})
2529: \bigr)
2530: = n^{-2r}/c^2 + O(n^{-2r-1}).
2531: $$
2532: To understand the situation where $\vec m$ and $\vec k$ overlap somewhere,
2533: we pause for some lemmas.
2534:
2535: \begin{lemma}\label{lm:inclusion_order}
2536: Let $\iota\from \tang\to G$ be an inclusion of graphs, with
2537: $G$ connected.
2538: Then the order of $G$ is at least that of $\tang$.
2539: \end{lemma}
2540: \proof Let $G_\tang$ be $G$ with the vertices of $\iota(\tang)$ identified, and
2541: all $\iota(\tang)$ edges discarded. Then $G_\tang$ is connected, and so has
2542: order at least
2543: $-1$; on the other hand, clearly the order of $G_\tang$ is the
2544: order of $G$ minus that of $\tang$ minus $1$ (for the vertex that is the
2545: identification of all $\iota(\tang)$ vertices).
2546: Hence
2547: $$
2548: {\rm ord}(G) = {\rm ord}(G_\tang) + {\rm ord}(\tang) +1 \ge
2549: -1+{\rm ord}(\tang)+1={\rm ord}(\tang).
2550: $$
2551: \proofbox
2552:
2553: \begin{lemma}\label{lm:edge_removal}
2554: Let $G$ be a pruned graph and let $e\in E_G$. If $G\setminus\{e\}$
2555: (i.e., $G$ with $e$ removed) has two connected components, then each
2556: connected component has order at least $0$.
2557: \end{lemma}
2558: \proof
2559: Consider a connected component, $G'$, of $G\setminus\{e\}$. If $G'$ did not
2560: contain a cylce, then $G'$ would be a tree, and
2561: then $G$ would not be completely pruned.
2562: So $G'$ contains a cycle, and we may apply Lemma~\ref{lm:inclusion_order}
2563: to deduce that $G'$ has order at least $0$
2564: \proofbox
2565:
2566: \begin{lemma}
2567: \label{lm:order_increases}
2568: Let $\iota\from \tang\to G$ be an inclusion of pruned graphs.
2569: Then the order of $G$ is at least that of $\tang$, and $G$'s order is strictly
2570: greater than $\tang$'s if $\iota(\tang)$ is properly
2571: contained in $G$.
2572: \end{lemma}
2573: \proof A connected component of a graph that is pruned (and non-empty)
2574: has non-negative order. So we may assume $\iota(\tang)$ meets every connected
2575: component of $G$. It suffices to prove the case where $\iota(\tang)$ meets
2576: one connected component of $G$, i.e., the case where $G$ is connected.
2577: Choosing an edge, $e$, that is missed by $\iota(\tang)$, we have that
2578: $\iota$ includes $\tang$ into $G\setminus\{e\}$ (i.e., $G$ with
2579: $e$ removed); we apply Lemma~\ref{lm:inclusion_order} to those components
2580: of $G\setminus\{e\}$ containing part of $\iota(\tang)$, and to the
2581: possibly one other component of $G\setminus\{e\}$ we apply
2582: Lemma~\ref{lm:edge_removal}.
2583: It follows that the order of $\tang$ is at most that of
2584: $G\setminus\{e\}$; but the order of $G\setminus\{e\}$ is one less than
2585: that of $G$.
2586: \proofbox
2587:
2588: \begin{lemma}\label{lm:order_greater}
2589: Let $\iota_1,\iota_2\from \tang\to H$ be two inclusions
2590: of a tangle, $\tang$, in a connected (labelled) graph,
2591: $H$, such that $\iota_1(\tang)\cup\iota_2(\tang)=H$.
2592: Assume that $\iota_1(\tang)\ne H$. Then the order of $H$ is greater
2593: than that of $\tang$.
2594: \end{lemma}
2595: \proof Ignoring $\iota_2$, the preceeding lemma applies to $\iota_1$
2596: to immediately yield this lemma.
2597: \proofbox
2598:
2599: Lemma~\ref{lm:order_greater} shows that
2600: $$
2601: % \sum_{\scriptstyle {\vec k},{\vec m}\;{\rm not\;disjoint}
2602: % \atop {\vec k},{\vec m}\in R,\;\;{\vec k}\ne{\vec m}}
2603: \sum_{\substack{ {\vec k},{\vec m}\;{\rm not\;disjoint}
2604: \\ {\vec k},{\vec m}\in R,\;\;{\vec k}\ne{\vec m}}}
2605: \prob{\iota_{\vec m}\cap\iota_{\vec k}} = O(n^{-r-1}),
2606: $$
2607: since the summation can be broken down into a finite number of sums
2608: over tangles of order at least $r+1$. Thus
2609: $$
2610: \sum_{{\vec m}\in R} \prob{\iota_{\vec m}} -
2611: % \sum_{\scriptstyle {\vec k}\ne {\vec m} \atop {\vec k},{\vec m}\in R}
2612: \frac{1}{2}\sum_{\substack{ {\vec k}\ne {\vec m} \\ {\vec k},{\vec m}\in R}}
2613: \prob{\iota_{\vec m}\cap\iota_{\vec k}}.
2614: $$
2615: $$
2616: = n^{-r}/c+O(n^{-r-1})- n^{-2r}/(2c^2) + O(n^{-2r-1})+O(n^{-r-1}),
2617: $$
2618: which completes the proof of Theorem~\ref{th:tangle_count}.
2619: \proofbox
2620:
2621: It is easy to see that the above proof of Theorem~\ref{th:tangle_count}
2622: uses very little about the model of random graph, and therefore generalizes
2623: as follows.
2624: \begin{theorem} Let $\ck_n$ be a model of $d$-regular random graphs on
2625: $n$ vertices labelled $\intn$,
2626: defined for some values of $n$. Further assume that
2627: (1) $\ck_n$ is invariant under renumbering $\intn$, and (2) any
2628: tangle, $\tang$, has expected number of occurrences $n^{-r}+O(n^{-r-1})$
2629: where $r$ is the order of $\tang$. Then Theorem~\ref{th:tangle_count} holds
2630: for $\ck_n$.
2631: \end{theorem}
2632:
2633:
2634: \begin{theorem}\label{th:other_ss}
2635: For $G$ drawn from $\chnd$ or
2636: $\cind$, we have that $\lambda_2(G)>2\sqrt{d-1}$
2637: with probability at least $n^{-s}/2+O(n^{-s-1})$ where
2638: $s=\lfloor \sqrt{d-1}\rfloor$, except for when $d=4$ in $\chnd$, where
2639: we may take $s=2$. The same holds for $\cjnd$ with probability
2640: $n^{-s}+O(n^{-s-1})$, where
2641: $s=\lfloor \bigl(\sqrt{d-1}\;+1\bigr)/2
2642: \rfloor$ (and no exceptional values of $d$).
2643: \end{theorem}
2644: \proof
2645: For $\cjnd$, the tangle consisting of $1$ vertex with a number of self-loops
2646: (in this case half-loops), proves the theorem for $\cjnd$ just as it
2647: did for $\cgnd$ in Theorem~\ref{th:improved_bound}.
2648:
2649: Consider the tangle, $\tang$,
2650: with two vertices and $m$ edges joining the two vertices
2651: labelled $\pi_1,\pi_2,\ldots$; $\tang$ is an $\chnd$-tangle
2652: provided that $m\le d/2$, and an $\cind$-tangle if $m\le d$.
2653: $\lambda_{\ird{}}(\tang)$
2654: is clearly $m-1$, $s=\ord(\tang)=m-2$, and the automorphism group of
2655: $\tang$ is of order $2$.
2656: So if $m-1>\sqrt{d-1}$ and if $G$ contains
2657: this tangle, then we have
2658: $\lambda_2(G)>2\sqrt{d-1}$ for $n$ sufficiently large.
2659: If we take $s=\lfloor \sqrt{d-1}\rfloor$, then $m-1>\sqrt{d-1}$; we
2660: require $m\le d$ for $\cind$, amounting to
2661: $$
2662: \lfloor \sqrt{d-1}\rfloor + 2 \le d,
2663: $$
2664: which is satisfied for all $d\ge 3$. For $\chnd$ we require
2665: $$
2666: \lfloor \sqrt{d-1}\rfloor + 2 \le d/2,
2667: $$
2668: which is satisfied for all $d\ge 7$. Since $d$ is even and $\ge 4$ in
2669: $\chnd$, we finish by examining the cases $d=4$ and $d=6$.
2670:
2671: For $d=4$ consider the tangle, $\tang$, with vertices $v_1,v_2,v_3,v_4$,
2672: edges labelled $\pi_1,\pi_2$ from $v_1$ to $v_2$, from $v_2$ to $v_3$,
2673: and from $v_3$ to $v_4$.
2674: Then ${\rm ord}(\tang)=2$ and $\lambda_{\ird{}}(\tang)>
2675: \lambda_{\ird{}}(\tang')$ where $\tang'$ is the subgraph of $\tang$
2676: induced on $v_1,v_2,v_3$.
2677: But in the proof of Theorem~\ref{th:taufund_chnd} we compute
2678: $\lambda_{\ird{}}(\tang')=\sqrt{3}$. Hence $\tang$ is hypercritical
2679: of order $2$, so we may take $s=2$ in the theorem in the case of $d=4$
2680: and $\chnd$.
2681:
2682: For $d=6$, the proof of Theorem~\ref{th:taufund_chnd} gives a tangle
2683: of order $2$ with $\lambda_{\ird{}}>\sqrt{5}$
2684: (see equation~(\ref{eq:d6tangle}) and the discussion around it).
2685: So we may take
2686: $s=2$ in our theorem when $d=6$ in $\chnd$.
2687:
2688: \proofbox
2689:
2690:
2691:
2692:
2693:
2694: \section{Walk Sums and New Types}
2695: \label{se:newtypes}
2696:
2697: In this section we give some general techniques that are used in
2698: estimating the expected values of
2699: all the various traces that are used in this paper.
2700: The main idea, originated in \cite{broder} and strengthened in
2701: \cite{friedman_random_graphs}, is to group contributions to the trace
2702: in the following way.
2703: Consider the word, $w$, and vector, $t$
2704: $$
2705: w=(\sigma_1,\ldots,\sigma_{10})=
2706: (\pi_2^{-1},\pi_1,\pi_3,\pi_1^{-1},\pi_3,\pi_3,\pi_1^{-1},\pi_3,\pi_1^{-1},
2707: \pi_2),
2708: $$
2709: $$
2710: \vec t = (t_0,\ldots,t_{10})=(5,2,4,3,7,4,3,7,4,2,5)
2711: $$
2712: (see Figure~\ref{fg:form1}).
2713: This represents a possible or potential irreducible closed walk of length 10,
2714: from $5$ to $2$ along the edge
2715: labelled $\pi_2^{-1}$, from $2$ to $4$ along $\pi_1$, etc.
2716: Graphically we depict this potential walk by the subgraph it traces out, called
2717: its ``form'' (see Figure~\ref{fg:form1}).
2718: \begin{figure}
2719: \centering
2720: % \scalebox{0.6}{\includegraphics{switchdesign.eps}}
2721: \epsfysize=2.5in
2722: \epsfbox[0 0 389 426]{form1.eps}
2723: % \psfig{figure=switchdesign.eps}
2724: \caption{A form and its associated type.}
2725: \label{fg:form1}
2726: \end{figure}
2727: The pair $(w;\vec t\>)$ represents a possible contribution to
2728: $\irdtr{A}{10}$. For any word, $w$, of length $k$ and vector,
2729: $\vec t$, of length $k+1$, let $P(w;\vec t\>)$ be the probability that
2730: $\sigma_i(t_{i-1})=t_i$ for all $i$, i.e., that the cycle does occur.
2731: Then the expected value of $\irdtr{A}{10}$ is just a sum of appropriate
2732: $P(w;\vec t\>)$'s, what we will call a ``walk sum.''
2733:
2734: In the form of Figure~\ref{fg:form1}, we see that the numbers of the
2735: vertices $5,2,4,3,7$, are irrelevant, since our random graph model is
2736: ``symmetric,'' or invariant under renumbering the vertices. So we
2737: may replace the vertex $5$ by an abstract symbol $v_1$, $2$ by $v_2$,
2738: etc. (for reasons to made clear later, we do want to remember the order
2739: in which the vertices were traversed); this replacement
2740: is not necessary, but
2741: makes clear that there is no particular preference for any numbering
2742: of the vertices.
2743:
2744: In our example of Figure~\ref{fg:form1},
2745: $(w;\vec t\>)$, a loop about the vertex $4$ is traversed
2746: twice before we return to the starting vertex $2$.
2747: Broder and Shamir realized that the other vertices, those of degree $2$
2748: that are not the starting vertex (here the vertices $2,3,7$) are less
2749: interesting features of the ``form.''
2750: By supressing these ``less interesting'' vertices we get the ``type''
2751: of the form (see Figure~\ref{fg:form1}); in \cite{friedman_random_graphs},
2752: walk sums were grouped by their form, and the sums for each form were
2753: grouped by the type of the form.
2754: A type is a graph with certain
2755: features, but when a form gives rise to a type then the edges
2756: of the type inherit $\Pi^+$ labels from the form, and inherit ``lengths,''
2757: which are the lengths of the $\Pi^+$ labels (recall that $\Pi^+$ is the
2758: set of nonempty words over $\Pi$).
2759: For example, the edge $(5,4)$ of the type in Figure~\ref{fg:form1} inherits
2760: a length of $2$ and a label of $\pi_2^{-1}\pi_1$ from the form next to it;
2761: edge $(4,4)$ inherits a length of $3$ and a label of $\pi_3\pi_1^{-1}\pi_3$.
2762:
2763: This paper introduces a ``new type,'' used in analyzing walk sums
2764: corresponding to selective traces. A new type is a type with
2765: some additional information, primarily fixing the lengths of certain
2766: type edges and requiring all other lengths to be sufficiently large.
2767:
2768:
2769: \subsection{Walk sums}
2770:
2771: By a {\em weakly potential $(k,n)$-walk}, $(w;\vec t\>)$, we mean a pair
2772: consisting of
2773: a word $w=\sigma_1\ldots\sigma_k$ of length $k$
2774: over $\Pi$, and a vector,
2775: $\vec t=(t_0,\ldots,t_k)$, with each $t_i\in\intn$; we sometimes refer
2776: to $k$ as the {\em length} and $n$ as the {\em size} of the weakly potential
2777: walk. Given such a
2778: $(w;\vec t\>)$, let
2779: $\ce(w;\vec t\,)$ denote the event that the $\pi_i$ are chosen so that
2780: $\sigma_i(t_{i-1})=t_i$
2781: for all $i=1,\ldots,k$.
2782: Let $P(w;\vec t\>)$ denote the probability that $\ce(w;\vec t\>)$ occurs.
2783:
2784: A {\em potential $(k,n)$-walk} is a weakly potential $(k,n)$-walk that
2785: is {\em feasible}, meaning
2786: that $P(w;\vec t\>)\ne 0$, or equivalently that the
2787: following two feasibility
2788: conditions hold: (1) for any $i,j$ such that $\sigma_i=\sigma_j$, we have
2789: $$
2790: t_{i-1}=t_{j-1} \Leftrightarrow t_i=t_j,
2791: $$
2792: and for any $i,j$ such that $\sigma_i=\sigma_j^{-1}$, we have
2793: $$
2794: t_{i-1}=t_{j} \Leftrightarrow t_i=t_{j-1}.
2795: $$
2796: For example, if $\vec t=(1,1,2)$ and $w=\pi_1\pi_1$, then
2797: $(w,\vec t\>)$ is a weakly potential
2798: walk but not a potential walk.
2799: A {\em potential $(k,n)$-closed walk} is a potential walk, $(w,\vec t\>)$ as
2800: above with $t_k=t_0$.
2801:
2802: All our variants of traces, irreducible and not irreducible, selective
2803: and not, can be viewed as sums of $P(w;\vec t\>)$ over appropriate $w$'s
2804: and $\vec t\;$'s. In this subsection we formalize this notion and make
2805: preliminary remarks about
2806: such sums and asymptotic expansions.
2807:
2808: \begin{definition} A {\em walk collection, $\cw=\cw(k,n)$,} is a collection,
2809: for any two positive integers $k$ and $n$,
2810: of $(k,n)$-potential walks $(w;\vec t\>)$ as above,
2811: i.e. $w$ is a word over $\Pi$ of length
2812: $k$, and $\vec t=(t_0,\ldots,t_k)$ is a $k+1$ dimensional vector over $\intn$,
2813: and $(w;\vec t\>)$ is feasible.
2814: The {\em walk sum} associated to $\cw$ is
2815: $$
2816: \walksum{\cw}{k,n}=\sum_{(w;\vec t\>)\in\cw(k,n)} P(w;\vec t\>).
2817: $$
2818: \end{definition}
2819:
2820: The main goal of this paper is to organize the various $(w,\vec t\>)$ pairs
2821: into groups over which we can easily sum $P(w;\vec t\>)$. One simple
2822: organizational remark is that symmetries in the $(w;\vec t\>)$ pairs
2823: often simplify matters. Specifically, given an permutation, $\tau$,
2824: of the integers, and a vector $\vec t$ as above, let
2825: $$
2826: \tau(\vec t\>) = \bigl(\tau(t_0),\ldots,\tau(t_k)\bigr).
2827: $$
2828: Say that $\vec s$ and $\vec t$ {\em differ by a symmetry} if $\vec s=\tau
2829: (\vec t\>)$ for some $\tau$; in this case clearly $P(w,\vec s\>)=P(w,\vec t\;)$
2830: for any word $w$ of length $k$. We use $\vec s\sim\vec t$ to denote
2831: that $\vec s$ and $\vec t$ differ by a symmetry.
2832:
2833: \begin{definition} A walk collection, $\cw$, is {\em SSIIC} if it is
2834: \begin{enumerate}
2835: \item {\em symmetric}, i.e. $(w,\vec t\>)\in\cw(k,n)$ implies that
2836: $(w,\vec s\>)\in\cw(k,n)$ for all $\vec s\sim\vec t$ such that
2837: $s_i\le n$ for all $i$,
2838: \item {\em size invariant}, i.e. if $(w,\vec t\>)$ is a potential
2839: $(k,n)$-walk, then for any $n'>n$, $(w,\vec t\>)\in\cw(k,n)$ iff
2840: $(w,\vec t\>)\in\cw(k,n')$,
2841: \item {\em irreducible}, meaning that $(w,\vec t\>)\in\cw$ implies that
2842: $w$ is irreducible, and
2843: \item {\em closed}, meaning that $(w,\vec t\>)\in\cw$ implies that
2844: $t_0=t_k$.
2845: \end{enumerate}
2846: \end{definition}
2847: The walk sums of interest here, namely traces that are irreducible
2848: or strongly irreducible and possibly selective,
2849: will all be SSIIC.
2850: We now make a series of remarks about walk sums that apply to all
2851: SSIIC walk sums; some of the remarks apply more generally.
2852:
2853: \begin{definition}
2854: Let $\vec t$ be as above, i.e., a positive integer valued vector.
2855: Define the {\em equivalence class of $\vec t$}
2856: to be
2857: $$
2858: [\vec t\>] = \{ \vec s \mid \vec s\sim\vec t\},
2859: $$
2860: i.e., the set of all positive integer values vectors differing from
2861: $\vec t$ by a symmetry.
2862: Define the {\em $n$-th equivalence class of $\vec t$}
2863: $$
2864: [\vec t\>]_n = \{ \vec s \mid \vec s\sim\vec t \quad\mbox{and all}\quad
2865: s_i\le n\}
2866: $$
2867: (we may omit the $n$ subscript if $n$ is understood).
2868: \end{definition}
2869: Let $n$ be fixed, and set
2870: $$
2871: \Esymm(w;\vec t\>) = \Esymm(w;\vec t\>)_n
2872: = \sum_{\vec s\in[\vec t\>]_n} P(w;\vec s\>)
2873: $$
2874: $$
2875: = n(n-1)\cdots(n-v+1)P(w;\vec t\>),
2876: $$
2877: where $v$ is the number of distinct elements of $\vec t$. A symmetric
2878: walk sum is just the sum of certain $\Esymm(w,\vec t\>)$'s, and we can
2879: write
2880: $$
2881: \walksum{\cw}{k,n}=\sum_{(w;[\vec t\>])\in\cw(k,n)}\Esymm(w;\vec t\>),
2882: $$
2883: where
2884: the summation over $(w;[\vec t\>])\in\cw(k,n)$ means that
2885: we sum over one $\vec t$
2886: in each equivalence class and over all $w$.
2887:
2888: Each $\vec t=(t_0,\ldots,t_k)$
2889: has an $\vec s\sim\vec t$ where the size of $\vec s\;$'s
2890: entries are at most $k+1$. So for each $k$ there are a finite number
2891: of equivalence classes, $\cw(k)$, of $(w;\vec t\>)$ such that $w$ is
2892: of length $k$ and $\vec t$ is of some finite size. We refer to
2893: $\cw(k)$ as the set of {\em potential walk classes} of length $k$
2894: (or {\em potential closed walk classes} when we restrict to those $\vec t\;$'s
2895: with $t_k=t_0$). So if $\cw$ is
2896: size invariant we may write
2897: $$
2898: \walksum{\cw}{k,n}=\sum_{(w;[\vec t\>])\in\cw(k)}\Esymm(w;\vec t\>)_n,
2899: $$
2900: where the right-hand-side summation has a fixed, finite number of summands
2901: independent of $n$
2902: (for fixed $k$).
2903:
2904: Our next step is to comment about $\Esymm(w,\vec t\>)_n$. Notice that if
2905: the conditions $\sigma_i(t_{i-1})=t_i$ involve determining
2906: $a_j$\myindexaj values
2907: of $\pi_j$, then
2908: \begin{equation}\label{eq:probability}
2909: P(w;\vec t\>) = \prod_{i=1}^{d/2} \frac{1}{n(n-1)\cdots(n-a_j+1)}
2910: =\prod_{i=1}^{d/2}\frac{(n-a_j)!}{n!}.
2911: \end{equation}
2912: Let $e=a_1+\cdots+a_{d/2}$.
2913: We have
2914: $$
2915: \Esymm(w,\vec t\>)_n = n(n-1)\ldots (n-v+1) \prod_{i=1}^{d/2}
2916: \frac{1}{n(n-1)\cdots(n-a_j+1)}.
2917: $$
2918: Notice that in the power series expansions about $x=0$ of
2919: $$
2920: (1-x)(1-2x)\ldots(1-mx) \quad\mbox{and}\quad
2921: (1-x)^{-1}(1-2x)^{-1}\ldots(1-mx)^{-1},
2922: $$
2923: the $x^i$ coefficient
2924: is a polynomial (of degree at most $2i$) in $m$.
2925: It follows that
2926: there exist polynomials $p_0,p_1,\ldots$ in the
2927: variables $a_1,\ldots,a_{d/2},v$ such that
2928: \begin{equation}\label{eq:expansion_polynomials}
2929: \Esymm(w,\vec t\>)_n = n^{v-e} \sum_{i=0}^\infty n^{-i}
2930: p_i(a_1,\ldots,a_{d/2},v)
2931: \end{equation}
2932: for $n$ sufficiently large.
2933: \begin{definition}
2934: \label{de:expansion_polynomials}
2935: We define the {\em expansion polynomials}\mytwoindex{expansion polynomials}{The
2936: polynomials $p_i=p_i(a_1,\ldots,a_{d/2},v)$
2937: giving the $1/n$ expansion of $\Esymm(w,\vec t\>)$},
2938: $$
2939: p_i=p_i(a_1,\ldots,a_{d/2},v)
2940: \mythreeindex{pi}{$p_i=
2941: p_i(a_1,\ldots,a_{d/2},v)$}{the expansion polynomials},
2942: $$
2943: to be the polynomials that give the
2944: expansion in equation~(\ref{eq:expansion_polynomials}).
2945: Throughout the paper, $a_i$\myindexaj
2946: denotes the number of $\pi_i$ values determined
2947: by the relevant structure (in this case the potential walk,
2948: $(w,\vec t\>)$).
2949: \end{definition}
2950: We remark that since $e=a_1+\cdots+a_{d/2}$, we have that $v$ is determined
2951: from the $a_i$ if $v-e$ is fixed and known; in such situations, the
2952: expansion polynomials may be regarded as functions of the $a_i$ alone.
2953:
2954:
2955: \begin{theorem}\label{th:exp_polys}
2956: For any $w,\vec t$ and any integer $q\ge 0$ we have
2957: \begin{equation}\label{eq:simple_expansion}
2958: \Esymm(w,\vec t\>)_n = n^{v-e} \biggl( p_0+\frac{p_1}{n}+\cdots+\frac{p_q}{n^q}
2959: + \frac{{\rm error}}{n^{q+1}} \biggr)
2960: \end{equation}
2961: where
2962: $$
2963: |{\rm error}| \le \exp\bigl((q+1)k/(n-k)\bigr)\; k^{2q+2},
2964: $$
2965: and the $p_i$ are the expansion polynomials.
2966: \end{theorem}
2967: The proof is contained between
2968: Lemma~2.7 and Corollary~2.10 of \cite{friedman_random_graphs}, although
2969: the proof there has a minor error. We will correct it here and review
2970: the entire proof,
2971: since we will later
2972: need variants of this theorem for $\chnd,\cind,\cjnd$.
2973: If
2974: \begin{equation}\label{eq:g_rational}
2975: g(x)=(1-b_1x)\cdots(1-b_rx)(1-c_1x)^{-1}\cdots(1-c_sx)^{-1},
2976: \end{equation}
2977: where the $b_i$ and $c_j$ are positive constants,
2978: then $g$'s $i$-th derivative satisfies the bound
2979: $$
2980: |g^{(i)}(x)|/i!\le (1-xc_{\max})^{-i}\left(\sum b_j+\sum c_j\right)^i
2981: $$
2982: where $c_{\max}$ is the maximum of the $c_j$
2983: (by equation~(6) in \cite{friedman_random_graphs} on page~339).
2984: This estimate, using Taylor's theorem, expanding in $x=1/n$ about $x=0$,
2985: gives the error term for Theorem~\ref{th:exp_polys} of
2986: $$
2987: (1-\zeta c_{\max})^{-q-1} \left(\sum b_j+\sum c_j\right)^{q+1},
2988: $$
2989: for some $\zeta\in[0,1/n]$.
2990: Since
2991: $$
2992: -\log(1-\zeta c_{\max})=(\zeta c_{\max})+(\zeta c_{\max})^2/2+\cdots
2993: \le \zeta c_{\max}/(1-\zeta c_{\max})
2994: $$
2995: $$
2996: \le (k/n)/\bigl(1-(k/n)\bigr) = k/(n-k),
2997: $$
2998: we conclude that the error term is at most
2999: \begin{equation}\label{eq:error_term}
3000: \exp\bigl( (q+1)k/(n-k) \bigr) \left(\sum b_j+\sum c_j\right)^{q+1}.
3001: \end{equation}
3002: In the case of Theorem~\ref{th:exp_polys}, the $\sum b_j$ represents
3003: the sum of $0,1,\ldots,v-1$, which is at most $\binom{k}{2}$, and
3004: the $\sum c_j$ represents the sum over $j$ of all sums of $0,1,\ldots,a_j-1$,
3005: which is at most $\binom{k}{2}$. Since $2\binom{k}{2}\le k^2$, we
3006: get an error term at most $\exp\bigl((q+1)k/(n-k)\bigr)k^{2q+2}$.
3007: (See \cite{friedman_random_graphs} for more details.)
3008: \proofbox
3009:
3010: \begin{definition} Given a pair, $(w,\vec t\>)$, as above, its {\em order}
3011: is $e-v$, with $e,v$ as above.
3012: \end{definition}
3013:
3014: \begin{lemma}\label{lm:big_order}
3015: Given a word, $w$, of length $k$ over $\Pi$, we have
3016: $$
3017: \sum_{\vec t\;{\rm such\;that}\;(w,\vec t\>)\;{\rm is\;of\;order} \; \ge r}
3018: P(w,\vec t\>)
3019: \le n\binom{k}{r+1}\left(\frac{k}{n-k}\right)^{r+1},
3020: $$
3021: which for $k\le n/2$ is at most $ck^{2r+2}n^{-r}$ for some constant $c$
3022: depending only on $r$.
3023: \end{lemma}
3024: \proof
3025: This can be found in
3026: \cite{friedman_random_graphs}, second displayed equation and
3027: discussion before on page 352; this is the same idea used in the $r=1$
3028: case proven in \cite{broder}. (The extra factor of $n$ appears here
3029: but not in \cite{friedman_random_graphs}, since
3030: we do not fix the initial vertex of the walk. Also note that the ``order,''
3031: used here, is one less than the ``number of coincidences,''
3032: used in \cite{friedman_random_graphs}.) For the ease of reading, we shall
3033: discuss these ideas in our notation.
3034:
3035: We may evaluate $P(w,\vec t\>)$ by considering the steps of the walk
3036: $w=\sigma_1\ldots\sigma_k$ one by one; i.e., we fix a $v_0\in\intn$
3037: and consider the random walk $v_1=\sigma_1(v_0)$, $v_2=\sigma_2(v_1)$,
3038: etc. We inductively consider the probability that $(v_0,\ldots,v_s)$
3039: is equivalent to $(t_0,\ldots,t_s)$ for $s=1,\ldots,k$.
3040: Assuming $(v_0,\ldots,v_s)$
3041: is equivalent to $(t_0,\ldots,t_s)$, and assuming we wish to have
3042: $(v_0,\ldots,v_{s+1})$
3043: equivalent to $(t_0,\ldots,t_{s+1})$,
3044: one can divide the choice and outcome
3045: of the random variable $v_{s+1}=\sigma_{s+1}(v_s)$ into three cases:
3046: (1) $v_{s+1}$ has already been determined by previous information
3047: (i.e., the value $\sigma_{s+1}(v_s)$ has been determined in a previous
3048: step);
3049: (2) $v_{s+1}$ has not been determined and
3050: (2a) $v_{s+1}$ must occur as one of the $v_0,\ldots,v_s$
3051: (i.e., when $t_{s+1}$ occurs previously
3052: in $t_0,\ldots,t_s$), or (2b) $v_{s+1}$ must
3053: must be different from $v_0,\ldots,v_s$ (i.e., when $t_{s+1}$ does
3054: not occur previously).
3055: We call case (1) a forced choice, case (2)
3056: a free choice, with case (2a) a coincidence, and case (2b) a generic
3057: choice.
3058:
3059: For example, if $w=\pi_1\pi_1\pi_2\pi_3$ and $\vec t\> = (1,1,1,2,1)$,
3060: then $v_1$ is a coicidence (since $t_1$ occurs previously as $t_0$ and
3061: $\pi_1(v_0)$ has not been determined), $v_2$ is
3062: a forced choice (since $\pi_1(v_0)=v_0=v_1$ has been determined),
3063: $v_3$ is a generic choice ($t_3$ does not occur in $t_0,t_1,t_2$),
3064: and $v_4$ is a coincidence.
3065:
3066: A choice $v_{s+1}=\sigma_{s+1}(v_s)$ is deterministic if it is a forced
3067: choice; a choice occurs with probability $1 - O(s/n)$ if it is
3068: a generic choice, and with probability $1/n + O(s/n^2)$ if it is a
3069: coincidence.
3070: We see that a word of order $t$ has exactly $t+1$ coincidences.
3071: If follows that each word of order at least $r$ and length $k$ has some
3072: $r+1$ positions of $k$ that can be marked as coincidences. Fixing a
3073: marking, each coincidence of the marking occurs with probability
3074: at most $k/(n-k)$ (since $v_{s+1}$ must assume one of the at most
3075: $s+1\le k$ values
3076: $v_0,\ldots,v_s$, and at most $k$ values of the permutation $\sigma_{s+1}$
3077: could have been determined up to that point).
3078: Finally $v_0$ can be chosen in $n$ ways. We conclude
3079: the lemma.
3080: \proofbox
3081:
3082: \begin{lemma}\label{lm:equiv_classes}
3083: For any irreducible word, $w$, over $\Pi$, of length $k$,
3084: there are at most
3085: \begin{equation}\label{eq:equiv_classes}
3086: \sum_{j=0}^r \binom{k}{j} k^j \le c k^{2r}
3087: \end{equation}
3088: equivalence classes $[\vec t\>]$ whose order with $w$ is $\le r-1$.
3089: \end{lemma}
3090: \proof See the third displayed equation of page 352 of
3091: \cite{friedman_random_graphs} and the discussion preceding. For
3092: completeness and ease of reading, we repeat the argument here.
3093:
3094: If $w$ has order $j-1$, then it has $j$ coincidences occurring in $k$
3095: places. Again, in each coincidence, $v_{s+1}=\sigma_{s+1}(v_s)$,
3096: $v_{s+1}$ is being chosen from at most $k$ values. This gives the
3097: left-hand-side of equation~(\ref{eq:equiv_classes}). For the
3098: right-hand-side we notice that
3099: $$
3100: \sum_{j=0}^r \binom{k}{j} k^j \le \sum_{j=0}^r k^{2j}\le
3101: k^{2r}(1+k^{-2}+k^{-4}+\cdots) \le k^{2r} (4/3)
3102: $$
3103: for $k\ge 2$.
3104: \proofbox
3105: %% (The proofs of the two preceding lemmas, both based on page 352 of
3106: %% \cite{friedman_random_graphs}, are easy consequences of the idea of
3107: %% ``forced/fixed choices'' and ``coincidences'' of a walk of
3108: %% \cite{broder} and \cite{friedman_random_graphs}.)
3109:
3110: \begin{theorem}\label{th:SSIICexpansion}
3111: Let $\cw$ be SSIIC and let $r\ge 1$. Then for all $k\le n/2$ we have
3112: $$
3113: \walksum{\cw}{k,n} = f_0(k)+\frac{ f_1(k)}{n}+\cdots+
3114: \frac{ f_{r-1}(k)}{n^{r-1}}+\frac{{\rm error}}{n^r},
3115: $$
3116: where
3117: $$
3118: f_i(k) = \sum_{j=0}^{r-1}\;\; \sum_{(w;[\vec t\>]){\rm \;order\;}j,\in\cw(k)}
3119: p_{i-j}(w;[\vec t\>])
3120: $$
3121: (with $p_i$ the expansion polynomials, which can be viewed as a
3122: function of $(w;[\vec t\>])$ since $(w;[\vec t\>])$ determines the
3123: $a_i$ and $v$)
3124: and where for some $c$ depending only on $r$,
3125: $$
3126: |{\rm error}| \le ck^{4r}(d-1)^k.
3127: $$
3128: \end{theorem}
3129: \proof By Lemma~\ref{lm:big_order}, we introduce an error of at most
3130: $ck^{2r+2}n^{-r}$ per word by ignoring potential walks of order at least
3131: $r$.
3132: Each word, $w$, has at most $ck^{2r}$ associated potential walk classes
3133: of order at most $r-1$
3134: (by Lemma~\ref{lm:equiv_classes}), and truncating the associated
3135: asymptotic expansion, as in equation~(\ref{eq:simple_expansion}),
3136: of each associated potential walk class
3137: results in an error of at most $ck^{2r}$
3138: (by Theorem~\ref{th:exp_polys}).
3139: So each word, $w$, of length $k$ involved in $\cw$
3140: contributes an error of at most
3141: $ck^{4r}$, and there are at most $d(d-1)^{k-1}$ such words
3142: (for $k\ge 1$) since $\cw$ consists of only irreducible words.
3143: \proofbox
3144:
3145:
3146: \subsection{The Loop}
3147:
3148: Here we analyze walk sums associated with simple loops. This gives
3149: some ideas and a lemma to be used in Section 8.
3150:
3151: \subsubsection{The Singly Traversed Simple Loop}
3152: \label{sb:stsl}
3153: Let $w$ be an irreducible
3154: word over $\Pi$ of length $k$ and $\vec t$ a $k+1$ tuple
3155: over $\intn$ as in the previous subsection.
3156: If $t_0=t_k$ and the $t_i$'s are otherwise
3157: distinct, we say $(w;\vec t\>)$
3158: is a {\em singly traversed simple loop}
3159: or {\em STSL} for short; we define $\cwstsl$ to be the collection of all
3160: STSL's.
3161: %we write $\cestsl(w;\vec t\>)$ for the event that
3162: %$\ce(w;\vec t\>)$ occurs and is an STSL.
3163: When $\ce(w;\vec t\>)$ occurs and $(w;\vec t\>)\in\cwstsl$,
3164: the closed walk
3165: from $t_0$ following $w$ traces
3166: out a ``simple loop'' once,
3167: that begins and ends at $t_0$, moving through distinct edges
3168: and vertices throughout the closed walk.
3169:
3170: Clearly $\cwstsl$ is SSIIC, so according to Theorem~\ref{th:SSIICexpansion}
3171: we have an asymptotic expansion in $1/n$ with coefficients $f_i(k)$
3172: for the associated walk sum. We now briefly indicate why the
3173: $f_i$ are {\dtreelike}. This is a mildly tedious
3174: exercise, covered (in much greater generality)
3175: in \cite{friedman_random_graphs}. We quote the main points
3176: here.
3177:
3178: First, note that there is exactly one equivalence class $[\vec t\>]$
3179: of $\vec t\;$'s that appear in $\cwstsl$. So we may write $\Esymm(w)$
3180: for $\Esymm(w;\vec t)$ for any $\vec t$ of the equivalence class, and
3181: we may write $p_i(w)$ for the $p_i(w;\vec t\>)$ in
3182: Theorem~\ref{th:SSIICexpansion} or \ref{th:exp_polys}
3183: (note also that $v=k$ in the notation
3184: of Theorem~\ref{th:exp_polys}).
3185:
3186: %Let $\pstsl(w;\vec t\>)$ denote that probability that $\cestsl(w;\vec t\>)$ occurs.
3187: %We shall show that it is possible to show the existence of a simple expansion
3188: %$$
3189: %\fstsl(n,k)=\sum_{|w|=k,\;\vec t} \pstsl(w;\vec t\>)
3190: %$$
3191: %in $n$, with coefficients a certain type of function of $k$.
3192:
3193: %First note
3194: %that $\pstsl(w;\vec t\>)=0$ when
3195: %$\sigma_1=\sigma_k^{-1}$, for then $t_1=t_{k-1}$ if $t_0=t_k$.
3196: %Otherwise, $\cestsl(w;\vec t\>)$ requires one to fix $a_i=a_i(w)$ values of
3197: %the permutation $\pi_i$ for each $i$, where
3198: %$a_i(w)$ is the number of times $\pi_i$ and $\pi_i^{-1}$ occur
3199: %in $w$. The probability that $\pi_i$ is fixed as such is
3200: %$$
3201: %\frac{1}{n}\frac{1}{n-1}\cdots\frac{1}{n-a_i+1}=\frac{(n-a_i)!}{n!},
3202: %$$
3203: %and so
3204: %$$
3205: %\pstsl(w;\vec t\>) = \frac{\bigl((n-a_1(w)\bigr)!}{n!}\cdots
3206: %\frac{\bigl((n-a_{d/2}(w)\bigr)!}{n!},
3207: %$$
3208: %proved that $\sigma_1\ne\sigma_k^{-1}$.
3209: %Of course, $\pstsl(w)=\pstsl(w;\vec t\>)$ is independent of the particular
3210: %choice of $\vec t$ provided that $\cestsl(w;\vec t\>)$ is a STSL.
3211: %Since there are $n!/(n-k)!$ such choices of $\vec t$, we have
3212: %$$
3213: %\fstsl(n,k)=\frac{n}{(n-k)!}\sum_{w\;{\rm s.t.}\;\sigma_1\ne\sigma_k^{-1}}
3214: %\frac{\bigl((n-a_1(w)\bigr)!}{n!}\cdots
3215: %\frac{\bigl((n-a_{d/2}(w)\bigr)!}{n!}.
3216: %$$
3217: %The fact that $\fstsl(n,k)$ has a ``nice expansion'' is now a mildly tedious
3218: %exercise, done in \cite{friedman_random_graphs}. We give the main points
3219: %here, sticking to the notation in \cite{friedman_random_graphs} as closely
3220: %as possible.
3221: %\begin{lemma} We have
3222: %and summing, we have
3223: %$$
3224: %\frac{n}{(n-k)!}\frac{\bigl((n-a_1(w)\bigr)!}{n!}\cdots
3225: %\frac{\bigl((n-a_{d/2}(w)\bigr)!}{n!}=
3226: %p_0(w)+\frac{p_1(w)}{n}+\cdots,
3227: %$$
3228: %where each $p_i(w)$ is a polynomial of degree $\le i$ in the $(d/2)+1$
3229: %variables
3230: %$a_1(w),\ldots,a_{d/2}(w),k$. Also $p_0=1$ and
3231: %the error mumble.
3232: %\end{lemma}
3233: %\proof This follows from Lemmas 2.7--2.9 of
3234: %\cite{friedman_random_graphs} (as does Corollary 2.10 there).
3235: %\proofbox
3236: %Hence
3237: %$$
3238: %\fstsl(n,k)=\sum_{w\;{\rm s.t.}\;\sigma_1\ne\sigma_k^{-1}}
3239: %p_0(w)+\frac{p_1(w)}{n}+\cdots
3240: %$$
3241: %We apply this as follows.
3242:
3243: Let for $\sigma,\tau\in\Pi$, let
3244: $\ird{k,\sigma,\tau}$ denote the irreducible words of length $k$
3245: beginning with $\sigma$ and ending with $\tau$.
3246: For $w\in\Pi^k$ let $a_i(w)$\myindexaj denote the number of $\pi_i$ and
3247: $\pi_i^{-1}$ occurring in $w$. Since an STSL has no forced choices,
3248: the definition of $a_i$ here agrees with that in
3249: Definition~\ref{de:expansion_polynomials}.
3250:
3251: \begin{lemma}\label{lm:irdeigens}
3252: Let $p=p(a_1,\ldots,a_{d/2},k)$ be a polynomial.
3253: For every $\sigma,\tau$ there are polynomials $Q_1,Q_2,Q_3$ of $k$
3254: of degree at most the degree of $p$ such that
3255: $$
3256: \sum_{w\in\ird{k,\sigma,\tau}} p\bigl( a_1(w),\ldots,a_{d/2}(w),k\bigr)
3257: = (d-1)^k Q_1(k)+(-1)^k Q_2(k) + Q_3(k).
3258: $$
3259: \end{lemma}
3260: \proof This is immediate from Lemma 2.11 of \cite{friedman_random_graphs} or
3261: Corollary 2.12 (note that
3262: our $d$ is $2d$ in \cite{friedman_random_graphs}).
3263: \proofbox
3264: Note that the formula for the above lemma comes about from the fact that
3265: $d-1,-1,1$ are the eigenvalues of $G_{\ird{}}$ where $G$ is the graph with
3266: one vertex and $d/2$ whole-loops (with $G_{\ird{}}$ defined
3267: as in Subsection~\ref{sb:lambda_irred}).
3268:
3269: \begin{corollary}
3270: \label{cr:stsl_dtreelike}
3271: We have each $f_i(k)$ as in Theorem~\ref{th:SSIICexpansion}
3272: for ${\cal W}=\cwstsl$
3273: is {\dtreelike}.
3274: \end{corollary}
3275: \proof
3276: Recall that an $(w,\vec t\>)\in\cwstsl$ is of order $0$, and there is only
3277: one equivalence class of $\vec t$ for STSL's.
3278: So we have
3279: $$
3280: f_i(k)=\sum_{\sigma\ne\tau^{-1}} \sum_{\ird{k,\sigma,\tau}}
3281: p_i\bigl( a_1(w),\ldots,a_{d/2}(w),k\bigr),
3282: $$
3283: where the $p_i$ are the expansion polynomials.
3284: The result now follows from Lemma~\ref{lm:irdeigens}, with error term
3285: bounded by a polynomial in $k$.
3286:
3287:
3288: \subsubsection{Simple Loops}
3289:
3290: Consider any irreducible $(w,\vec t\>)$ that traces out a simple loop,
3291: i.e., the vertices and edges visited form one closed walk, but now we don't
3292: require that the loop is traversed only once.
3293: Corresponding to this geometric picture of a simple loop we can form
3294: the associated walk collection of {\em simple loop closed walks}; such closed walks,
3295: being irreducible, must traverse the loop traced out some number of
3296: times.
3297:
3298: So let $\cwsl$ be the set of $(w,\vec t\>)$ pairs with
3299: \begin{enumerate}
3300: \item $w=\sigma_1\raise1.6pt\hbox{\ldots}\sigma_k$ irreducible,
3301: % \item $w=\sigma_1\ldots\sigma_k$ irreducible,
3302: \item $\sigma_1\ne\sigma_k^{-1}$,
3303: \item $t_0,\ldots,t_{r-1}$
3304: distinct for some $r$ dividing $s$,
3305: \item $t_{i+r}=t_i$ for $0\le i\le k-r$, and
3306: \item $w=u^s$ for some word $u$
3307: with $rs=k$.
3308: \end{enumerate}
3309: $\cwsl$ is the walk collection of simple loop walks.
3310:
3311: Clearly we have
3312: \begin{equation}\label{eq:multiply_traversed}
3313: \walksum{\cwsl}{k,n}=\sum_{s|k} \walksum{\cwstsl}{s,n}.
3314: \end{equation}
3315: We easily conclude the following theorem.
3316: \begin{theorem} The $f_i(k)$ corresponding to $\cwsl$ are {\dtreelike},
3317: and have the same principal term as the $f_i(k)$ corresponding to
3318: $\cwstsl$.
3319: \end{theorem}
3320:
3321: \proof Consider the $f_i(k)$ corresponding to $\cwstsl$. By
3322: Corollary~\ref{cr:stsl_dtreelike}, all the $f_i$ are {\dtreelike}. By
3323: equation~(\ref{eq:multiply_traversed}), it suffices to show that
3324: $$
3325: \widetilde f_i(k) = \sum_{s|k} f_i(s)
3326: $$
3327: are also {\dtreelike}. Fixing an $i$ and setting
3328: $f_i(k)=p(k)(d-1)^k + r(k)$ as the decomposition of $f_i$ into principal
3329: and error terms, we see
3330: $$
3331: \widetilde f_i(k) = p(k)(d-1)^k + \widetilde r(k),
3332: $$
3333: where
3334: $$
3335: |\widetilde r(k)| \le \left| \sum_{s\le k/2} p(s)(d-1)^s \right| +
3336: \sum_{s\le k} cs^c (d-1)^{s/2}
3337: $$
3338: for some $c$ as in Definition~\ref{de:ram}. It is clear that
3339: the right-hand-side above is bounded by $(d-1)^{k/2}$ times a polynomial
3340: in $k$.
3341: \proofbox
3342:
3343:
3344: \subsection{Forms, Types, and New Types}
3345:
3346: In this subsection we will classify potential walks, $(w;\vec t\>)$,
3347: or more generally potential walk classes, according to some characteristics
3348: of the subgraph
3349: that the walk traces out.
3350:
3351: \begin{definition} A {\em form}\mytwoindex{form}{an oriented, $\Pi$-labelled
3352: graph with edges and vertices numbered that is meant to represent the
3353: graph traced out by a potential word, forgetting about the particular
3354: choices in $\intn$ of the vertices}, $\Gamma$, is an oriented, $\Pi$-labelled
3355: graph, $G_\Gamma=(V_\Gamma,E_\Gamma)$, with edges and vertices numbered.
3356: A {\em specialization} of a form, $\Gamma$, is an
3357: injection $\iota\from V_\Gamma\to \intn$.
3358: \end{definition}
3359:
3360: With each potential walk, $(w;\vec t\>)$, we can associate a form, $\Gamma=
3361: \Gamma(w;\vec t\>)$, with a
3362: specialization, $\iota$, as follows. (The form, $\Gamma$, is not unique, but
3363: is unique up to unique isomorphism, as described below; the
3364: specialization is unique given the form.)
3365: \begin{enumerate}
3366: \item Set $V_\Gamma=\{v_1,\ldots,v_r\}$ to be any numbered ($v_i$ numbered
3367: $i$)
3368: set of size $r$, where
3369: $r=|V_\Gamma|$ is the number of distinct elements among the $t_i$
3370: (where $\vec t=(t_0,\ldots,t_k)$)
3371: \item $\iota(v_i)$ is the $i$-th distinct element of the sequence
3372: $t_0,t_1,\ldots,t_k$,
3373: \item Set $E_\Gamma=\{e_1,\ldots,e_m\}$ to be any numbered set of size $m$
3374: ($e_i$ numbered $i$),
3375: where $m=|E_\Gamma|$ is the number of distinct triples
3376: $\{ (\sigma_i,t_{i-1},t_i) \}_{i=1,2,\ldots,k}$, where we identify a
3377: triple $(\sigma,s,t,)$ with $(\sigma^{-1},t,s)$,
3378: \item if $(\sigma_j,t_{j-1},t_j)$ is the $r$-th distinct tuple in
3379: $\{ (\sigma_i,t_{i-1},t_i) \}_{i=1,2,\ldots,k}$ (with the previous
3380: identification), then $s(\ned_r)=\bigl(\sigma_j,\iota^{-1}(t_{j-1}),
3381: \iota^{-1}(t_j)\bigr)$ defines the structural map
3382: (see Definition~\ref{de:structural})
3383: of $G_\Gamma$.
3384: \end{enumerate}
3385: (see the example in Figure~\ref{fg:form1} explained at the beginning
3386: of this section).
3387:
3388: In other words, the form is the subgraph traced out by $(w,\vec t\>)$,
3389: with some additional information (we remember the order in which
3390: the vertices and edges are visited, and the direction each edge is
3391: first traversed). We say that forms $\Gamma_1$ and $\Gamma_2$ are
3392: {\em isomorphic} if they are isomorphic as oriented, numbered,
3393: $\Pi$-labelled graphs. Because of the numbering, there is at most
3394: one isomorphism between any two forms (or a form and itself). We
3395: say that $(w,\vec t\>)$ is {\em of form $\Gamma$} or
3396: {\em associated to $\Gamma$}, written
3397: $(w,\vec t\>)\in\Gamma$, if one (or any) of the forms associated
3398: to $(w,\vec t\>)$ is
3399: isomorphic to $\Gamma$. Given $(w,\vec t\>)$, there is always an
3400: associated form, $\Gamma$, with $V_\Gamma=\{t_0,\ldots,t_k\}$ and
3401: associated specialization, $\iota$, being the identity; however, we
3402: usually view $V_\Gamma$ as any numbered set of the right size, since
3403: all of our random graph models are symmetric.
3404:
3405: If $(w,\vec t\>)\in\Gamma$, then define
3406: $$
3407: \E{\Gamma}_n = \Esymm(w,\vec t\>)_n
3408: $$\myindexEGamman
3409: which depends only on $\Gamma$, and not on the particular $(w,\vec t\>)$ to
3410: which $\Gamma$ is associated; indeed,
3411: \begin{equation}\label{eq:form_prob}
3412: \E{\Gamma}_n
3413: = \frac{n!}{(n-v)!}\prod_{i=1}^{d/2} \frac{(n-a_i)!}{n!},
3414: \end{equation}
3415: where $v=|V_\Gamma|$, and $a_i$\myindexaj
3416: is the number of edges in $\Gamma$ labelled
3417: with $\pi_i$ and $\pi_i^{-1}$ (this is exactly the $a_i(w)$ of
3418: equation~(\ref{eq:probability}) of any word, $w$, associated to $\Gamma$).
3419:
3420: Hence, if $\cw$ is symmetric and size invariant, we may write
3421: \begin{equation}\label{eq:fundamental}
3422: \walksum{\cw}{k,n} = \sum_\Gamma W_\Gamma(\cw,k)\E{\Gamma}_n,
3423: \end{equation}
3424: where $W_\Gamma(\cw,k)$\myindexW
3425: is the number of potential walk classes in $\cw(k)$
3426: associated to $\Gamma$, and we sum over one $\Gamma$ in each isomophism
3427: class of forms.
3428:
3429:
3430: \begin{definition} A {\em legal walk} in a form, $\Gamma$, is a walk
3431: starting in $\nv_1$ that visits all the vertices of $G_\Gamma$ in order
3432: (of their numbering),
3433: all the edges in order, and any edge is first traversed
3434: in the direction of its orientation. Each legal walk
3435: of length $k$ generates a walk class in the natural way.
3436: \end{definition}
3437:
3438: The following easy proposition is worth stating formally; it follows
3439: from the definitions.
3440: \begin{proposition} $W_\Gamma(\cw,k)$ is the number of legal walks on
3441: $\Gamma$ of length $k$.
3442: \end{proposition}
3443: %% \proof A potential walk $(w;\vec t\>)$, to which is associated
3444: %% a form, $\Gamma$, and a specialization, $\iota$,
3445: %% determines a legal
3446: %% walk, $W=W(w;\vec t\>)$,
3447: %% on $G_\Gamma$ by applying $\iota^{-1}$ to $\vec t$. Clearly
3448: %% $W=W(w;\vec t\>)$ depends only on hte equivalence class of $\vec t$.
3449: %% Conversely, any legal walk, $W$, traces out a word,
3450: %% $w=w(W)$, via the edge labelling, and any injection $\iota\from V_\Gamma\to
3451: %% \{1,2,\ldots\}$ determines a vector, $\vec t=\vec t\>(W,\iota)$;
3452: %% as $\iota$ varies, $\vec t\>(W,\iota)$ varies over an equivalence class.
3453: %% So we have an association, $(w,[\vec t\>])\mapsto W(w,[\vec t\>])$ and
3454: %% an association, $W\mapsto\bigl(w(W),[\vec t\>](W)\bigr)$, from
3455: %% a potential class to a legal walk and vice versa, that are clearly
3456: %% inverses of each other.
3457: %% \proofbox
3458:
3459: \begin{definition} The {\em order} of a form, $\Gamma$, is $\ord(\Gamma)
3460: =|E_\Gamma|-
3461: |V_\Gamma|$, subject to the Definition~\ref{de:tangle_order} convention
3462: that any self-loop, whole-loop or half-loop, counts as a single
3463: edge. (The order of a form
3464: equals the order of any potential walk to which it is
3465: associated.)
3466: \end{definition}
3467:
3468: The ``form'' allows us to group together potential walks that determine
3469: the same information on the graph; this can facilitate the task of studying
3470: a walk sum. A futher tool is the grouping of forms together by their
3471: ``type,'' which we now briefly motivate in rough terms to be made precise.
3472: A type arises from a form with all its maximal
3473: beaded paths collapsed to edges (by the appropriate supression).
3474: Each collapsed edge inherits an edge length and $\Pi^+$ labelling from
3475: the ``form,'' but in the type we forget this data. As long as we collect
3476: forms of a given type by
3477: the first and last letters of all the collapsed edge labellings, we can
3478: apply Lemma~\ref{lm:irdeigens} at each collapsed edge to study sums
3479: of the expansion polynomials in the $a_i$'s (and $v$)
3480: over forms in such collections.
3481:
3482: Before defining a ``type,''
3483: we recall that for a form, $\Gamma=\Gamma(w;\vec t\>)$,
3484: with specialization, $\iota$, the potential walk $(w;\vec t\>)$ pulls
3485: back under $\iota^{-1}$ to a walk on $G_\Gamma$;
3486: we remark that the vertex and edge numberings
3487: of $\Gamma$
3488: serve to remember in which order the vertices and edges were traversed
3489: in the walk.
3490:
3491: \begin{definition} A {\em type}\mytwoindex{type}{a graph representing a
3492: number of forms, where we forget certain features of the form, such as
3493: its $\Pi$-labelling and all or almost all its degree two vertices},
3494: $T$, is a connected,
3495: oriented graph
3496: $G_T=(V_T,E_T)$, with vertex and edge numberings
3497: such that all vertices except possibly the first one are of degree
3498: at least $3$.
3499: A {\em labelling} of a type means a $\Pi^+$-labelling (recall
3500: $\Pi^+$ is the
3501: set of words over $\Pi$ of length at least $1$).
3502: \end{definition}
3503:
3504: To each form, $\Gamma$, we associate a type, $T=T(\Gamma)$, as follows.
3505: Let $W$ be the set of beads of $G_\Gamma$ numbered greater than $1$.
3506: We claim $W$ cannot contain a cycle, for then this cycle would be
3507: disconnected from the vertex numbered $1$.
3508: So we may form the supression of $W$
3509: in $G_\Gamma$.
3510: This supression inherits a vertex and edge numbering from $G_\Gamma$
3511: (ordering a vertex before another in the supression if it is numbered
3512: less in $G_\Gamma$, and ordering edges in the supression by any
3513: associated edge in $E_\Gamma$).
3514: Of course, the $\Pi$-labelling of $\Gamma$ gives rise to a $\Pi^+$-labelling
3515: of $G_{T(\Gamma)}$.
3516:
3517: In other words, the beads of a form are ``less important'' features,
3518: and the type is just the form with these ``less important'' features
3519: supressed.
3520:
3521: A $\Pi^+$-labelled type uniquely determines a form, and vice versa. We
3522: wish to group together forms corresponding to one type (inducing different
3523: $\Pi^+$-labellings), the prototypical example being STSL's or SL's discussed
3524: earlier in this section. To do this it will be helpful to remember
3525: a small part of the labelling, namely the starting and ending letter of
3526: each $\Pi^+$-label.
3527:
3528: \begin{definition} A {\em lettering} of a type, $T$, is the assignment to
3529: each directed edge a {\em starting letter} in $\Pi$ and an {\em ending letter}
3530: in $\Pi$ (such that opposite directed edges are lettered with the starting
3531: letter of one being the inverse of the ending letter of the other).
3532: Given a form, $\Gamma$, with $T=T(\Gamma)$, or equivalently
3533: a $\Pi^+$-labelling of $T$, the associated lettering assigns to each edge the
3534: starting and ending letter of the $\Pi^+$-label assigned to it in its
3535: orientation.
3536: \end{definition}
3537:
3538: It turns out that the notion of a type is too coarse to attack the
3539: Alon conjecture (but sufficient for the results of
3540: \cite{friedman_random_graphs}). The problem is that some types, when
3541: their edges take on certain $\Pi^+$-labellings, contain supercritical
3542: tangles. When such tangles occur (which they
3543: can when their order is roughly $O(\sqrt{d})$), we must distinguish where
3544: there tangles occur.
3545: This is what a ``new type'' does, where the type edges are partitioned into
3546: ``long'' and ``fixed'' (in length)
3547: edges, and where supercritical tangles
3548: lie on type
3549: edges that are ``fixed.''
3550: We then modify our walk sums to be ``selective''
3551: (see Section~6), a notion which requires us to know where these tangles
3552: occur in the type.
3553:
3554: \begin{definition}
3555: \label{de:new_type}
3556: A {\em $B$-new type}\mytwoindex{new type}{a type
3557: with some additional information specified, such as a partition of
3558: the types edges into a ``long'' and a ``fixed'' edge set}
3559: is a collection, $\widetilde T=
3560: (T;\El,\Ef;{\vec k}^{\rm fixed})$,
3561: of (1) a lettered type,
3562: $T$, (2) a partition of $E_T$ into two sets, $\El,\Ef$, (3) for each
3563: $\ned_i\in \Ef$ an {\em edge length}, $\kf_i$, with $0<\kf_i<B$, and
3564: (4) a $\Pi^+$-labelling of $\Ef$ with each $\ned_i\in \Ef$
3565: labelled with a word
3566: of length $\kf_i$.
3567: A $\Pi^+$-labelling of $T$ (or, equivalently, a form, $\Gamma$), is said to be
3568: {\em of $B$-new type $\widetilde T$} if each $\El$ label is of length
3569: at least $B$, and each label corresponding to $\ned_i\in \Ef$ is of length
3570: $\kf_i$ and agrees with the label specified by $\widetilde T$.
3571: $\widetilde T$ is said to be {\em based on $T$}.
3572: \end{definition}
3573:
3574: \begin{theorem} For each $r>0$ there are finitely many types (up to
3575: isomorphism) of order
3576: at most $r$. For each type, $T$, and each $B>0$ there are finitely many
3577: $B$-new types based on $T$.
3578: \end{theorem}
3579: \proof The first statement is just Lemma~2.3 of \cite{friedman_random_graphs},
3580: except that ``coincidence'' is used instead of ``order'' (and the
3581: coincidence is the order plus one). The second statement is clear
3582: since there are finitely many (1) letterings, (2) partitions of $E_T$,
3583: (3) choices of $\kf_i$ with $e_i\in \Ef$, and (4) labellings of each $\Ef$
3584: edge, $e_i$, with a length $\kf_i<B$.
3585: \proofbox
3586:
3587:
3588: %% \input motivation
3589: %% new
3590: \subsection{Motivation of Types and New Types}
3591:
3592: So far we have defined walk sums; we have seen that symmetric, size
3593: invariant walk sums can be organized into forms, by
3594: equation~(\ref{eq:fundamental}); we have seen that forms can be
3595: grouped by type and new type. In this section we briefly explain
3596: how and why we use types and new types.
3597:
3598: Fix a symmetric, size invariant walk collection, $\cw$.
3599: Organizing forms by type, we may write equation~(\ref{eq:fundamental}) as
3600: $$
3601: \walksum{\cw}{k,n} =
3602: \sum_T
3603: \sum_{\Gamma\in T} W_\Gamma(\cw,k)\E{\Gamma}_n,
3604: $$
3605: where $\Gamma\in T$ means the form $\Gamma$ is of type $T$, and the
3606: summation in $T$ ranges over all types. Typically we need only sum over
3607: $T$ of at most some order, so the sum in $T$ will effectively be a finite
3608: sum.
3609:
3610: So fix a $T$ and define $\E{T}_{n,k}$ to mean
3611: $$
3612: \E{T}_{n,k} =
3613: \sum_{\Gamma\in T} W_\Gamma(\cw,k)\E{\Gamma}_n.
3614: $$
3615: Let $T$'s edges be $E_T=\{e_1,\ldots,e_b\}$.
3616: For a vector, $\vec k = (k_1,\ldots,k_{b})$, let $T(\vec k\>)$ denote
3617: the set of forms, $\Gamma$,
3618: of type $T$, such that for all $i$ the length of the beaded path in
3619: $\Gamma$ corresponding to the edge $e_i$ in $T$ has length $k_i$.
3620: For each $e_i\in E_T$ fix an integer $m_i\ge 1$.
3621: Let $W_\Gamma(\vec m)=W_\Gamma(m_1,\ldots,m_b)$
3622: denote the number of legal $\cw$ walks
3623: in $\Gamma$ that traverse edge $e_i$ exactly $m_i$ times.
3624: Clearly
3625: $$
3626: W_\Gamma(\vec m) = W_T(\vec m),
3627: $$
3628: i.e., $W_\Gamma(\vec m)$ depends only on $\vec m$ and the
3629: type, $T$, of $\Gamma$. This allows us to write
3630: $$
3631: \E{T}_{n,k} = \sum_{\vec m} W_T(\vec m)
3632: \sum_{\vec k \cdot \vec m = k} \ \
3633: \sum_{\Gamma\in T(\vec k\>)} \E{\Gamma}_n \;.
3634: $$
3635: Since each $\E{\Gamma}_n$ has a $1/n$ expansion, by adding expansions we
3636: get an asymptotic expansion
3637: $$
3638: \sum_{\Gamma\in T(\vec k\>)} \E{\Gamma}_n =
3639: n^{-{\rm ord}(T)} \bigl( q_0(\vec k\>) + q_1(\vec k\>)/n + \cdots \bigr),
3640: $$
3641: where we ignore error terms in this subsection;
3642: of course,
3643: $$
3644: q_i(\vec k\>) = \sum_{\Gamma\in T(\vec k\>)}
3645: p_i\bigl( a_1(\Gamma),\ldots,a_{d/2}(\Gamma) \bigr).
3646: $$
3647:
3648: Perhaps the main technical point of Section~8 (see Theorem~\ref{th:f_vecm_est}
3649: here, and Lemma~2.14 in \cite{friedman_random_graphs}) is that
3650: the $P_{\vec m\>,i}$ defined by
3651: $$
3652: P_{\vec m\>,i}(k)
3653: =\sum_{\vec k\cdot \vec m=k} q_i(\vec k\>)
3654: $$
3655: are {\dtreelike}, and
3656: roughly speaking $P_{\vec m\>,i}$'s principal part and error term
3657: decay at most like $(d-1)^{m/2}$ where
3658: $m=m_1+\cdots+m_k$ (actually, the principal part is shown
3659: in Lemma~2.14 in \cite{friedman_random_graphs} to decay at most like
3660: $(d-1)^{-m}$).
3661: The way this is done is very roughly as follows.
3662: First, we fix a lettering of $T$ and apply Lemma~\ref{lm:irdeigens}
3663: to each edge, $e_i$, of $T$, whose length is set to $k_i$.
3664: This shows that the $q_i(\vec k\>)$ is a sum of {\dtreelike} functions
3665: whose arguments are all sums of a subset of the $k_i$. As an example,
3666: consider
3667: a function, $g(k_1+\cdots+k_b)$ with $g$ being {\dtreelike}. The maximum
3668: value of $k_1+\cdots+k_b$, given $\vec k\cdot\vec m=k$ with fixed
3669: $\vec m$ and $k$, is $k-m+b$, which is achieved when and only when $k_i=1$
3670: whenever $m_i\ge 2$ (assuming at least one $m_i=1$; otherwise
3671: $g(k_1+\cdots+k_b)$ is bounded by
3672: $ck^c (d-1)^{k/2}$ since $k_1+\cdots+k_b\le k/2$).
3673: This $k-m+b$ is where the decay exponential in $m$ comes from
3674: (see Section~8 and/or Section~2 in \cite{friedman_random_graphs}
3675: for details).
3676:
3677: It follows that there is an asymptotic expansion
3678: $$
3679: \E{T}_{n,k} = n^{-{\rm ord}(T)} \bigl( Q_0(k) + Q_1(k)/n + \cdots \bigr),
3680: $$
3681: where
3682: \begin{equation}\label{eq:key_problem}
3683: Q_i(k) = \sum_{\vec m} W_T(\vec m\>) P_{\vec m\>,i}(k).
3684: \end{equation}
3685: Since the error term of the
3686: $P_{\vec m\>,i}$ decay like $(d-1)^{-m/2}$ (roughly speaking),
3687: the above sum for $Q_i(k)$ turns out to be {\dtreelike}
3688: provided
3689: that
3690: \begin{equation}\label{eq:bound_W_T}
3691: |W_T(\vec m\>)| \le c(d-1-\epsilon)^{m/2}
3692: \end{equation}
3693: for some $\epsilon>0$.
3694:
3695: If all $Q_i$ were {\dtreelike} for all $T$, then we'd have a
3696: fairly simple proof of the Alon conjecture. The results of
3697: \cite{friedman_random_graphs} are based on the fact that
3698: equation~(\ref{eq:bound_W_T}) for all types up to a certain order
3699: (of roughly $O(\sqrt{d})$); past order $O(\sqrt{d})$,
3700: equation~(\ref{eq:bound_W_T}) generally fails to hold.
3701: Furthermore, by Theorem~\ref{th:nottreelike} we see
3702: for certain that at least one $Q_i$ fails to be {\dtreelike} for
3703: an $i$ bounded by $O(\log d\sqrt{d})$.
3704:
3705: The new approach in this paper is as follows. Consider a
3706: constant positive integer $B$, and consider
3707: $$
3708: Q_i(k) = \sum_{\vec m} W_T(\vec m\>)
3709: \sum_{\vec k\cdot \vec m=k} q_i(\vec k\>).
3710: $$
3711: We can divide this sum by fixing some of the $k_i$'s at fixed values
3712: less than $B$ (call these $k_{t+1},\ldots,k_b$), and then summing
3713: over the remaining $k_i$'s (namely $k_1,\ldots,k_t$) subject to
3714: the remaining $k_i$'s being at least $B$ (or ``long'').
3715: This is where a ``new type'' comes from.
3716: Next we fix a constant
3717: $S<B$ and define a
3718: ``selective trace'' to be the sum of irreducible closed walks of a given
3719: length that have no subpath of length $S$ tracing out a graph of
3720: $\lambda_{\ird{}}$ at least $\sqrt{d-1}$. If
3721: $$
3722: M_1=m_1+\ldots+m_t,\quad M_2=m_{t+1}+\ldots+m_b,
3723: $$
3724: then the corresponding ``selective'' version of $W_T(\vec m\>)$
3725: (that depends on the new type, i.e., on knowing the particular $k_i$ that
3726: are fixed and their values) is bounded by roughly
3727: $(d-1-\epsilon)^{(BM_1+M_2)/2}$ (for appropriately large $S$ and $B$).
3728: But consider
3729: the $P_{\vec m\>,i}$ for the ``new type,'' i.e., the sum
3730: $$
3731: \sum_{\substack{\vec k\cdot \vec m=k \\ k_1,\ldots,k_t\ge B}} q_i(\vec k\>)
3732: $$
3733: with $k_{t+1},\ldots,k_b$ at their fixed values; it turns out they
3734: decay like $(d-1)^{-(BM_1+M_2)/2}$.
3735: Thus, after summing over all new types,
3736: the corresponding selective analogues of the $Q_i(k)$ are {\dtreelike}.
3737:
3738: Unfortunately, a selective trace does not generally equal the original
3739: trace unless the graph in question is free of certain tangles.
3740: Still, in Sections~9 and later we show how the asymptotic expansions
3741: with {\dtreelike} coefficients of
3742: selective traces can be used to control
3743: a random graph's eigenvalues.
3744:
3745: %% ***********
3746:
3747: %% Unfortunately equation~(\ref{eq:bound_W_T}) fails to hold some
3748: %% The problem is that this
3749: %% won't be the case when $T$'s order is large. Our strategy is then to
3750: %% modify our irreducible trace to make the modified $Q_i$ {\dtreelike}.
3751:
3752: %% In the case of a type, $T$, for which
3753: %% there is an $\epsilon>0$ and a constant $c$ for which
3754: %% $$
3755: %% |W_T(\vec m\>)| \le cm^c (d-1-\epsilon)^{m/2},
3756: %% $$
3757: %% then we claim all $Q_i$'s will be {\dtreelike}. This is because
3758: %% the sum in equation~(\ref{eq:key_problem}) is essentially (or, more
3759: %% correctly, behaves like) a
3760: %% convolution (the $P_{\vec m\>,i}$ can be viewed as independent of $\vec m$
3761: %% in some very rough sense); in
3762: %% convolutions the highest base wins, as witnessed by
3763: %% the simple example
3764: %% $$
3765: %% \sum_{m=1}^k \alpha^m \beta^{k-m},
3766: %% $$
3767: %% where the size of this sum, for fixed $\alpha,\beta>0$, is proportional
3768: %% to $\max(\alpha,\beta)^k$; for the $W_T$ with $P$, the $\sqrt{d-1-\epsilon}$
3769: %% base is competing with the $d-1$ base from the principal part of the
3770: %% $P$'s, and the $\sqrt{d-1}$ base from the error term.
3771: %% On the other hand, if $|W_T(\vec m\>)|$ grew as fast as
3772: %% $\sqrt{d-1+\epsilon}$ to the power $m$, then this term would win against
3773: %% the $\sqrt{d-1}$ base from the error term.
3774: %% (If $|W_T(\vec m\>)|$ grows as fast as $(d-1)^{m/2}$, then technicalities,
3775: %% such as the fact that the $P_{\vec m\>,i}$ are not really independent
3776: %% of $\vec m$, make it unclear whether or not the $Q_i$ are {\dtreelike}.)
3777:
3778: %% A type, $T$, with one vertex and more than $(\sqrt{d-1}+1)/2$ self-loops will
3779: %% have $|W_T(\vec m\>)|$ growing like $\beta^m$ for some $\beta>\sqrt{d-1}$.
3780: %% So we must find a way to modify the trace to better control
3781: %% $|W_T(\vec m\>)|$ in such situations. The genesis of the idea is as
3782: %% follows.
3783: %% In equation~(\ref{eq:key_problem}), the $k-m$ as argument of $P_{\vec m\>,i}$
3784: %% comes from the fact that the maximum sum of $k_1+\cdots+k_b$ is
3785: %% $k-m+b$ (and $b$ is ignored since it depends only on the type $T$).
3786: %% The maximality of this sum requires that $k_i=1$ whenever $m_i\ge 2$.
3787: %% Consider the situation $m_1=m_2=\cdots=m_s=1$, $k_1,\ldots,k_s$ arbitrary
3788: %% and $k_{s+1}=\cdots=k_b=1$, $m_{s+1},\ldots,m_b$ arbitrary; in this
3789: %% situation $|W_T(\vec m\>)|$ is bounded by $c \lambda_{\ird{}}(T')^m$,
3790: %% where $T'$ is the graph of $T$ with $e_1,\ldots,e_s$ discarded
3791: %% (since $e_1,\ldots,e_s$ can only be traversed once).
3792: %% If $\lambda_{\ird{}}(T')<\sqrt{d-1}$, then we can tolerate such
3793: %% contributions to $Q_i$ in the expression
3794: %% $$
3795: %% Q_i(k) = \sum_{\vec m} W_T(\vec m\>)
3796: %% \sum_{\vec k\cdot \vec m=k} q_i(\vec k\>).
3797: %% $$
3798: %% If not, we wish to modify the irreducible trace to discard irreducible
3799: %% closed walks that spend a large (constant) amount of time in $T'$.
3800: %% We do this simply by choosing a sufficiently large $S$ and forming
3801: %% a ``selective'' trace that counts all irreducible closed walks such that
3802: %% the subgraph traversed in any contiguous $S$ steps does not have
3803: %% $\lambda_{\ird{}}$ at least $\sqrt{d-1}$. This avoids the
3804: %% aforementioned problem when $\lambda_{\ird{}}(T')\ge\sqrt{d-1}$,
3805: %% because it turns out that a typical sufficiently long irreducible
3806: %% walk in $(T')_{\ird{}}$ will visit all of
3807: %% $T'$ in some $S$ steps at some point, and therefore be discarded.
3808: %%
3809: \section{The Selective Trace}
3810:
3811: In this section we define a {\em selective trace}, and discuss some of
3812: its properties.
3813:
3814:
3815: \subsection{The General Selective Trace}
3816:
3817: Fix a graph, $G=(V,E)$, coming from $\cgnd$, so that $V=\{1,\ldots,n\}$
3818: and $G$ is $\Pi$-labelled.
3819:
3820:
3821: By a {\em path\footnote{
3822: By a {\em path} one often means a sequence of vertices. In case there
3823: are multiple edges in the graph, one needs to note also which edge is
3824: traversed. Finally, in the case of whole-loops in an undirected graph,
3825: one needs to remember in which ``direction'' each whole-loop is being
3826: traversed. In the present situation, all the above information is contained
3827: simply in the
3828: initial vertex and the permutations, $\pi_i$ or $\pi_i^{-1}$, being taken
3829: on each step of the path.
3830: } of length $k$} in $G$ we shall mean a vertex,
3831: $v\in V$, and a word of length $k$,
3832: $w=\sigma_1\ldots\sigma_k$, over $\Pi$ (i.e., with each $\sigma_i\in\Pi$).
3833: Such a path determines a subgraph in $G$ of those vertices and edges
3834: traversed. We say a path {\em traverses} a tangle, $\tang$, if the subgraph
3835: traversed by the path contains the tangle, $\tang$.
3836:
3837: \begin{definition} Let $\tset=\{\tang_1,\ldots\}$ be a (finite or infinite)
3838: collection of
3839: tangles.
3840: For positive integer, $S$,
3841: the {\em set of $(S,\tset)$-selective closed walks (respectively, walks)}
3842: are those irreducible closed walks (respectively, walks)
3843: that have no subpath of length at most $S$ that traverses a
3844: tangle in $\tset$.
3845: The {\em $k$-th irreducible
3846: $(S,\tset)$-selective trace of $G$},
3847: $\irdsel_{S,\tset}(G;k)$\mythreeindex{IrSelTr}{$\irdsel_{S,\tset}(G;k)$}{
3848: $k$-th irreducible $(S,\tset)$-selective trace of $G$, i.e., the number of
3849: irreducible closed walks of length $k$ in $G$ such that no subpath of
3850: length at most $S$ traces out a tangle in $\tset$}, is the
3851: number of $(S,\tset)$-selective closed walks of length $k$.
3852: \end{definition}
3853: Intuitively, the selective trace modifies the standard irreducible
3854: trace on those graphs
3855: that have a tangle in $\tset$, and avoids those closed walks that
3856: in some
3857: short part trace out such a tangle.
3858:
3859:
3860: \subsection{A Lemma on Selective Walks}
3861:
3862: What is the point of the selective trace? We can answer this question
3863: in two ways. First, since hypercritical
3864: tangles give large eigenvalues, any trace
3865: with an arbitrarily long asymptotic expansion in $1/n$ with {\dtreelike}
3866: coefficients must avoid hypercritical tangles (according to
3867: Theorems~\ref{th:remarkable} and \ref{th:key_to_s});
3868: a trace must be selective or its asymptotic expansion coefficients
3869: will not all be {\dtreelike}.
3870: Second, there is a crucial technical theorem,
3871: Theorem~\ref{th:crucial_cycle_count}, about
3872: counting irreducible contributions to a selective trace.
3873: This lemma makes certain infinite sums converge for the selective trace
3874: that would have to diverge for the standard trace---
3875: for example, the infinite sum involving $W(T;\vec m)$ and $P_{i,T,\vec m}$
3876: just above the middle
3877: of page 351 in \cite{friedman_random_graphs}, for types of order $>d$;
3878: for the same reason, this crucial theorem makes the $1/n$ expansion for
3879: a selective trace have {\dtreelike} coefficients when they don't for a
3880: trace that is not selective--- indeed, the
3881: $(2d-1)^{k/2}$ bound in equation~(24) of \cite{friedman_random_graphs}
3882: depends critically on $2i+2\le\sqrt{2d-1}$, and this equation
3883: corresponds to the error in the $n^{-i}$ term in the expansion of the expected
3884: value of the irreducible
3885: trace (recall that $2d$ in \cite{friedman_random_graphs} corresponds to
3886: our $d$).
3887: We shall finish this subsection
3888: with the crucial technical theorem, Theorem~\ref{th:crucial_cycle_count},
3889: after setting up the necessary terminology.
3890:
3891: A {\em relabelling} of a tangle, $\tang$, is a tangle, $\tang'$, that differs
3892: from $\tang$ only in its edge labels.
3893:
3894: \begin{definition} A set, $\tset$, of tangles is called {\em closed under
3895: pruning (respectively, relabelling)}
3896: if $\tang\in\tset$ implies $\tang'\in\tset$ for any
3897: pruning (respectively, relabelling),
3898: $\tang'$, of $\tang$.
3899: \end{definition}
3900:
3901: Note: In the definition above, $\tang$ and $\tang'$ must be
3902: $\cgnd$-tangles (or tangles in whatever model is discussed)--- a vertex
3903: with two self-loops labelled both labelled $\pi_1$ is not a
3904: $\cgnd$-tangle and is therefore not considered a relabelling of the
3905: tangle where the self-loops are labelled $\pi_1$ and $\pi_2$.
3906:
3907: \begin{definition} For a positive integer, $\tau$, let $\tset_\ord(\tau)$ be
3908: the set of tangles whose order is at least $\tau$.
3909: For positive integers $\tau_1\le \tau_2$, let $\tset_\ord(\tau_1,\tau_2)$
3910: be the set of all tangles whose order is at least $\tau_1$ and at
3911: most $\tau_2$.
3912: We also write $\tset_\ord(\tau,\infty)$ for $\tset_\ord(\tau)$.
3913: \end{definition}
3914: Since pruning a tangle does not affect its order, $\tset_\ord(\tau_1,
3915: \tau_2)$ is
3916: closed under pruning; clearly $\tset_\ord(\tau_1,\tau_2)$
3917: is closed under relabelling.
3918:
3919: Consider a form, $\Gamma$,
3920: of type $T$, in which $T$'s edges,
3921: $e_i$, have length $k_i$ (as beaded paths arising from $\Gamma$).
3922: For each $e_i\in E_T$ fix an integer $m_i\ge 1$.
3923: Let $\tset$ be a set of tangles closed under relabelling.
3924: Let $W_\Gamma(\vec m;S,\tset)$ be the number of legal closed walks (in particular,
3925: beginning at the first vertex) in $\Gamma$ that traverse each $e_i$
3926: exactly $m_i$ times (in either direction) and that are
3927: $(S,\tset)$-selective. Since $\tset$ is closed under relabelling,
3928: $W_\Gamma$ depends only on the length, $k_i$, of $e_i$ in $\Gamma$, not
3929: on the particular $\Pi^+$ labels of length $k_i$. So we may write
3930: $$
3931: W_\Gamma(\vec m;S,\tset)=W_T(\vec m,\vec k;S,\tset)\myindexW .
3932: $$
3933: %% In other words, for any
3934: %% $\Pi^+$-labelling of $T$, the number of such walk classes
3935: %% with a given $k_i$ and $m_i$ for $T$, $W_T(\vec m,\vec k;S,\tset)$, that
3936: %% are irreducible cylces that are $(S,\tset)$ selective, is independent of
3937: %% the labelling.
3938:
3939: Now given the above setting, call an edge, $\ned_i$, of $T$ {\em long} if
3940: $k_i>S$, and {\em short} otherwise.
3941: If a walk contains some $\tang\in\tset$
3942: in any consecutive $S$ steps, then by possibly pruning these
3943: consecutive steps along long edges at the beginning and end,
3944: we get a consecutive walk over short edges that contains a pruning of $\tang$.
3945: In particular, if $B>S$ and $\widetilde T=(T;\El,\Ef;\vec k^{\rm fixed})$ is a
3946: $B$-new type based on $T$, then $W_T$ depends on only $\widetilde T$,
3947: $\vec m$, $S$, and $\tset$ provided that $k_i=\kf_i$; hence we may write
3948: $$
3949: W_T(\vec m,\vec k;S,\tset)=W_{\widetilde T}(\vec m;S,\tset).
3950: $$
3951:
3952: \begin{definition}
3953: We say that a collection of tangles, $\tset$, is {\em $r$-supercritical}
3954: if it contains all supercritical tangles of order at most $r$.
3955: \end{definition}
3956: Next we give two examples of very natural $r$-supercritical tangle sets.
3957: \begin{definition}
3958: Let $\taufund$ be the smallest order of a supercritical tangle,
3959: and let $\tset_{\rm fund}=\tsord(\taufund)$. Let $\tseig$ the be set of all
3960: supercritical tangles.
3961: \end{definition}
3962:
3963: $\tset_{\rm fund}$ and $\tseig$ are clearly $r$-supercritical for any $r$;
3964: $\tsord(\taufund,r)$ and $\tseig[r]=\tseig\cap\tsord(\taufund,r)$ are also
3965: clearly $r$-supercritical.
3966: We arrive at our crucial technical theorem, that is the key to the
3967: selective trace.
3968:
3969: \begin{theorem}\label{th:crucial_cycle_count}
3970: Let $T$ by any type, with specified edge set partition
3971: $\El$,$\Ef$, and a $\Pi$-lettering specified.
3972: Let the edges be indexed so that
3973: $$
3974: \El=\{e_1,\ldots,e_t\}, \qquad \Ef=\{e_{t+1},\ldots,e_b\}.
3975: $$
3976: Then there is a $c$, an
3977: $\epsilon>0$, and
3978: an $S_0$ such
3979: that the following is true for all $S\ge S_0$. Let
3980: $\tset$ be a set of tangles containing all supercritical tangles included
3981: in a form of type $T$; e.g., by Lemma~\ref{lm:inclusion_order} we may
3982: take $\tset$ to be any $r$-supercritical set for $r=\ord(T)$.
3983: Let
3984: $$
3985: W_{\widetilde T,S}(M_1,M_2) =
3986: % \sum_{\scriptstyle m_1+\ldots+m_t=M_1\atop
3987: % \scriptstyle m_{t+1}+\cdots+m_b=M_2}
3988: \sum_{\substack{ m_1+\ldots+m_t=M_1\\
3989: % \scriptstyle
3990: m_{t+1}+\cdots+m_b=M_2\\ m_i\ge 1}}
3991: W_{\widetilde T}(\vec m;S,\tset),
3992: $$
3993: for a $B$-new type, $\widetilde T$, with $B>S$ and with $\widetilde T$
3994: having edge set partition $\El,\Ef$. Then
3995: $$
3996: W_{\widetilde T,S}(M_1,M_2) \le c (\sqrt{d-1}-\epsilon)^{(S_0+1)M_1+M_2}
3997: \le c (\sqrt{d-1}-\epsilon)^{BM_1+M_2}.
3998: $$
3999: \end{theorem}
4000: \proof
4001: The general approach we take is based on the following crude estimate.
4002: Let $\{f_i(z)\}_{i\in I}$ be a collection of non-negative power series,
4003: and let $g(z)$ be a power series that majorizes each $f_i(z)$.
4004: Suppose that
4005: $g(z_0)$ converges for some $z_0\in(0,1)$. Then the $z^k$ coefficient
4006: of any of the $f_i(z)$'s is bounded by $g(z_0)z_0^{-k}$.
4007:
4008: Specifically, we shall show that there is an $S_0$ such that the
4009: following holds.
4010: Let $B>S_0$, and let $\widetilde T$ be a $B$-new type subject to the
4011: conditions of the theorem.
4012: Let $G$ be a VLG whose underlying graph is $T$, whose $\Ef$ edges take
4013: their lengths from $\widetilde T$, and whose $\El$ edges all have length
4014: $B$. Let $f(z)=\sum c_k z^k$, where $c_k$ is the number of
4015: $(S_0,\tset)$-selective irreducible walks of length $k$ in $G$.
4016: We shall show that
4017: there is a $z_0> (d-1)^{-1/2}$ and a $g=g(z)$ such that
4018: (1) $g$ majorizes $f$, (2) $g$ depends only on $T$, and (3) $g(z_0)$
4019: converges. In that case
4020: $$
4021: W_{\widetilde T,S}(M_1,M_2) \le g(z_0) z_0^{-BM_1-M_2},
4022: $$
4023: which completes the theorem.
4024:
4025: Let ${\cal G}_{\rm below}$ be the set of VLG's, $H$, whose underlying graph
4026: is a subgraph of $T$ containing only $\Ef$ edges, with the property that
4027: $\lambda_{\ird{}}(H)<\sqrt{d-1}$. Let ${\cal G}_{\rm extreme}$ be the set
4028: of elements
4029: of ${\cal G}_{\rm below}$ that are not majorized by a different member
4030: of ${\cal G}_{\rm below}$. We claim that ${\cal G}_{\rm extreme}$ is
4031: finite; indeed, if $H_1,H_2,\ldots$ were a distinct sequence of VLG's
4032: in ${\cal G}_{\rm extreme}$, then by passing to a subsequence we could
4033: assume that for every $e\in \Ef$ either the length of $e$ in $H_i$ is
4034: constant or the length tends to infinity; but then $H_1$ would majorize
4035: all $H_i$ with $i$ sufficiently large.
4036:
4037: Let ${\cal G}_{\rm extreme}=\{H_1,\ldots,H_m\}$, and let
4038: $h_i(z)=\sum c_{i,k} z^k$, where $c_{i,k}$ is the number
4039: of irreducible walks of length $k$ in $H_i$. These $c_{i,k}$ are
4040: given as the number of walks of length $k-1$ in $(H_i)_{\ird{}}$; it follows
4041: that $h_i(z)$ has radius of convergence greater than $(d-1)^{-1/2}$. Also,
4042: if $\widetilde c_{i,k}=c_{i,0}+\cdots+c_{i,k}$ is the number of irreducible
4043: walks in $H_i$ of length at most $k$, then
4044: $$
4045: \widetilde h_i(z) = \sum_k \widetilde c_{i,k} z^k = \frac{1}{1-z}h_i(z)
4046: $$
4047: also has radius of convergence greater than $(d-1)^{-1/2}$. Let
4048: $z_0>(d-1)^{-1/2}$ be a value at which all $\widetilde h_i$ converge.
4049:
4050: For any $S$ set
4051: $$
4052: \widetilde h_i^{S}(z) = \sum_{k>S} \widetilde c_{i,k} z^k,
4053: $$
4054: and set
4055: $$
4056: \widetilde h^{S}(z) = \sum_{i=1}^m \widetilde h_i^{S}(z) \quad\mbox{and}\quad
4057: h(z) = \sum_{i=1}^m h_i(z).
4058: $$
4059: Consider an
4060: $S_0$ sufficiently large so that
4061: $\widetilde h^{S_0}(z_0)< 1/d$ (later we impose other lower bounds on
4062: $S_0$'s value).
4063:
4064: Let $B>S_0$.
4065: Let $G$ be a VLG whose underlying graph is (the graph underlying) $T$, and
4066: whose $\El$ edges have length at least $B$. Let $d_k$ (respectively,
4067: $\widehat d_k$) be the number of irreducible walks in $G$ of length $k$
4068: that are $(S_0,\tset)$-selective (respectively, and never traverse an
4069: $\El$ edge). Let $f(z),\widehat f(z)$ be the generating functions
4070: of $d_k,\widehat d_k$, respectively.
4071:
4072: The functions $f(z),\widehat f(z)$ are clearly majorized by the
4073: $f(z),\widehat f(z)$ in the case where $B=S_0+1$ and all $\El$ edges
4074: have length $S_0+1$. We shall assume this to be the case.
4075:
4076: First, we claim that $f$ is majorized by
4077: $$
4078: \bigl(1 - \widehat f(z) dz^{S_0+1}\bigr)^{-1} \widehat f(z).
4079: $$
4080: This is because any walk in $G$ can be broken into alternating
4081: subwalks that remain in
4082: $\Ef$ and steps along $\El$ edges; each time an $\El$ edge is taken
4083: its length is at least $S_0+1$, and there are at most $d$ such
4084: $\El$ edges from which
4085: to choose after finishing the $\Ef$ walk\footnote{Of course,
4086: there are at most $d-1$ such $\El$ edges from which to choose except
4087: possibly at the very first step of a walk.}.
4088: To prove the theorem it therefore suffices to show that
4089: $\widehat f(z_0) dz_0^{S_0+1}$ is less than one for sufficiently large $S_0$.
4090:
4091: Next, we claim that $\widehat f(z)$ is majorized by
4092: $$
4093: \bigl(1 - d\widetilde h^{S_0}(z)\bigr)^{-1} h(z).
4094: $$
4095: Assuming this, it is clear that for sufficently large $S_0$ we have
4096: $\widehat f(z_0) dz_0^{S_0+1}<1$ and the theorem is proven.
4097: For $k\ge S_0$, let $b_k$ be the number of $(S_0,\tset)$-selective
4098: irreducible walks, $w=(v_0,e_1,v_1,\ldots,e_j,v_j)$,
4099: in $G$ of length $k$ through only $\Ef$ edges such that
4100: $w'=(v_0,e_1,v_1,\ldots,e_{j-1},v_{j-1})$ is of length less than $S_0$.
4101: The walk $w'$ is contained in a subgraph of $G$ that is majorized by
4102: one of the $H_i$.
4103: Clearly
4104: $$
4105: b_{S_0} \le \sum_{i=1}^m c_{i,S_0}.
4106: $$
4107: When $k>S_0$, then from $v_{j-1}$ the walk has at most $d$ possible
4108: directions to take (or at most $d-1$ possible directions if $j>1$),
4109: and so
4110: $$
4111: b_k \le \sum_{i=1}^m \widetilde c_{i,k}.
4112: $$
4113: Thus $\sum b_k z^k$ is majorized by $d\widetilde h^{S_0}(z)$. But
4114: any walk over $\Ef$ edges can be broken into a series of walks of
4115: the form $w$ as above, plus a final walk of length less than $S_0$.
4116: The generating function for such walks is clearly majorized by $h(z)$.
4117: \proofbox
4118:
4119: %% First consider, $\rho_1=\rho_1(k,v)$, the maximum number of
4120: %% $(S,\tset)$-selective irreducible
4121: %% walks of length $k$ in $T$
4122: %% there are from any given vertex through only $\Ef$ edges, for any $S\ge k$,
4123: %% where each $\Ef$ edge is, for now, taken to be of unit length.
4124: %% Such a walk traces out a tangle, $\tang$, and this tangle must have
4125: %% $\lambda_{\ird{}}<\sqrt{d-1}$. This tangle, $\tang$, is also a subgraph
4126: %% of $T$ (since each edge is of length one, for now).
4127: %% So such a walk is contained in one of a finite number of subgraphs,
4128: %% $T'$,
4129: %% of $T$, each with $\lambda_{\ird{}}(T')<\sqrt{d-1}$
4130: %% (since the walk is entirely in $\Ef$, and all $\Ef$'s edges are taken to
4131: %% be of length $1$).
4132: %% For each $T'$ there is a $c$ such that the number of $\Ef$ walks of length
4133: %% $k$ there is
4134: %% $$
4135: %% \le ck^c \bigl( \lambda_{\ird{}}(T') \bigr)^k,
4136: %% $$
4137: %% by considering the Jordan form of $A_{T'}$; for any $\eta>0$ this number is
4138: %% $$
4139: %% \le c'\bigl( \lambda_{\ird{}}(T')+\eta \bigr)^k,
4140: %% $$
4141: %% for all $k$, for sufficiently large $c'$.
4142: %% Hence taking $\epsilon>0$
4143: %% to be any number such that $\lambda_{\ird{}}<\sqrt{d-1}-\epsilon$ for these
4144: %% finitely many graphs, we conclude that there is a $c$ such that
4145: %% $$
4146: %% \rho_1(k) \le c(\sqrt{d-1}-\epsilon)^k.
4147: %% $$
4148:
4149: %% Next let $T'$ be $T$ with each $e_i\Ef$
4150: %% is subdivided into
4151: %% $\kf_i$ edges, and
4152: %% consider $\rho_2(k)$, the maximum number of $(S,\tset)$-selective
4153: %% irreducible walks, $\omega$, in $T'$,
4154: %% through only $\Ef$ edges.
4155: %% Each such walk, $\omega$, in $T'$,
4156: %% gives rise to a path, $\omega'$, in $T$,
4157: %% by possibly extending $\omega$ at the beginning and end to an
4158: %% irreducible $T'$ walk until the walk hits a $V_T$ vertex.
4159: %% The length of $\omega'$ (in $T$)
4160: %% is at most $k$. Each $\omega'$ walk can come
4161: %% from at most four $\omega$'s (the four arises from an
4162: %% ambiguity at the beginning and end of $\omega'$
4163: %% with subdivided self-loops in $\Ef$). Hence
4164: %% $$
4165: %% \rho_2(k) \le 4c \sum_{j=1}^k (\sqrt{d-1}-\epsilon)^j \le
4166: %% c' (\sqrt{d-1}-\epsilon)^k.
4167: %% $$
4168:
4169: %% Finally consider a $B$-new type, $\widetilde T$, under the assumption
4170: %% $B>S\ge k$. Let $G$ be a VLG with each $\El$ edge of length $\ge R+1$
4171: %% (for some parameter $R\ge S$
4172: %% to be specified shortly), and $\Ef$ edges subdivided
4173: %% according to their $\widetilde T$ lengths.
4174: %% (Any form, $\Gamma$, corresponding to $\widetilde T$ has each $\El$
4175: %% $\Pi^+$-label of length $\ge R+1$ provided that $R\le S$, since $S<B$.)
4176: %% The row sum of $Z^k_G(z)$
4177: %% represents how many walks there are originating from a given vertex of
4178: %% various lengths. Such a row sum consists of a sum over $j$ of
4179: %% representations of walks with $k-j$ steps in $\Ef$ and $j$ steps in $\El$;
4180: %% there are $\binom{k}{j}$ ways of choosing the $j$ steps in $\El$, and
4181: %% the contribution from these $j$ steps is $\le z^{Rj}$; the $k-j$ steps in
4182: %% $\Ef$ occur in at most $j+1$ contiguous $\Ef$ walks for length
4183: %% $q_1,\ldots,q_r$ with $r\le j+1$, and the $i$-th contiguous walk contributes
4184: %% at most
4185: %% $$
4186: %% c(\sqrt{d-1}-\epsilon)^{q_i}z^{q_i},
4187: %% $$
4188: %% for a total contribution, assuming $c\ge 1$, of at most
4189: %% $$
4190: %% c^{j+1}(\sqrt{d-1}-\epsilon)^{k-j}z^{k-j}.
4191: %% $$
4192: %% The total contribution is therefore at most
4193: %% \begin{equation}\label{eq:big_z_sum}
4194: %% c\sum_{j=0}^k \binom{k}{j}c^j(\sqrt{d-1}-\epsilon)^{k-j}z^{k+Rj}
4195: %% =c(\sqrt{d-1}-\epsilon+c z^R)^k z^k.
4196: %% \end{equation}
4197: %% Now choose $R$ large enough so that
4198: %% $$
4199: %% \sqrt{d-1}-\epsilon+c(\sqrt{d-1}-\epsilon)^{-R} \le \sqrt{d-1}-(\epsilon/2).
4200: %% $$
4201: %% Then for $z\le (\sqrt{d-1}-(\epsilon/4))^{-1}$ we have
4202: %% $$
4203: %% (\sqrt{d-1}-\epsilon+cz^R)^k z^k \le (1-\epsilon')^k
4204: %% $$
4205: %% for some $\epsilon'>0$. So choose $k$ large enough so that
4206: %% $c(1-\epsilon')^k<1$.
4207: %% From equation~(\ref{eq:big_z_sum}), each row sum of $Z_G^k(z)$ is then
4208: %% $<1$ for all $z\le (\sqrt{d-1}-(\epsilon/4))^{-1}$. Hence
4209: %% $\lambda_1(A_G)\le \sqrt{d-1}-(\epsilon/4)$.
4210:
4211: %% It follows that for $S_0=\max(k,R)$ we have that for any $S\ge S_0$ we have
4212: %% $$
4213: %% W_{\widetilde T,S}(M_1,M_2) \le \bigl(\sqrt{d-1}-(\epsilon/4)
4214: %% \bigr)^{(R+1)M_1+M_2}|V_G|;
4215: %% $$
4216: %% $|V_G|$, the number of vertices in the subdivided form of $G$,
4217: %% is at most the number of vertices
4218: %% in the type plus $B-2$ times the number of $\Ef$ edges plus $R$ times the
4219: %% number of $\El$ edges. In total, $|V_G|$ is at most a constant times $B$.
4220:
4221:
4222: \subsection{Determining $\taufund$ for $\cgnd$}
4223:
4224: In order to use the selective trace, we must determine $\taufund$.
4225: We begin by doing so for the model $\cgnd$; next subsection
4226: we use similar
4227: techniques for the models $\chnd$, $\cind$, and $\cjnd$.
4228:
4229: More generally, for a given $\tau$, consider the task of finding the
4230: tangle, $\tang$, in $\cgnd$, of order at most
4231: $\tau$ with $\lambda_{\ird{}}(\tang)$
4232: as large as possible. To simplify this task, notice that pruning leaves
4233: the order and $\lambda_{\ird{}}$ invariant (an irreducible closed
4234: walk can never visit a leaf, so pruning a leaf doesn't affect the number
4235: of irreducible closed walks); hence we may restrict our
4236: search to those $\tang$'s that are completely pruned.
4237:
4238:
4239: \begin{lemma}\label{lm:contract}
4240: Let $G$ be a graph with edge $e=\{u,v\}$ with $u\ne v$.
4241: Let $G_e$ be the {\em contraction} of $G$ along $e$, i.e. the graph
4242: obtained by discarding $e$ and identifying $u$ with $v$. Then
4243: $\lambda_{\ird{}}(G)\le \lambda_{\ird{}}(G_e)$.
4244: \end{lemma}
4245: \proof Consider an irreducible closed walk, $c$, about $u$ in $G$. Then we can
4246: associate to this closed walk one in $G_e$, $\iota(c)$,
4247: by discarding all occurrences
4248: of $e$. This association, $\iota$,
4249: is an injection, since given a $G_e$ irreducible
4250: closed walk about $u$ of the form $\iota(c)$,
4251: we can infer when $e$ was taken (since
4252: $e=\{u,v\}$ with $u\ne v$) in the $G$ closed walk, giving rise to (at most) a
4253: single $G$ closed walk. Since this injection does not increase the length of
4254: the closed walks, we conclude that the number of irreducible closed walks about $u$
4255: in $G$ of length $\le k$ is no more than the number in $G_e$. Hence
4256: the conclusion of the lemma.
4257: \proofbox
4258:
4259: Since edge contraction reduces the number of vertices and of edges by
4260: one each, edge contraction leaves the order invariant. So in looking
4261: for a $\lambda_{\ird{}}$ tangle of a given order, we may always assume
4262: the tangle has no edge contraction that leaves it a tangle\footnote{
4263: In $\chnd$, two vertices joined by between $2$ and $d/2$ edges is
4264: a tangle (with apropriate $\Pi$-labelling), but contracting any edge
4265: gives self-loops, which are not feasible in $\chnd$. Therefore
4266: edge contraction can take graphs that can be tangles to graphs that
4267: cannot, at least for certain random graph models.}.
4268:
4269: We now claim (by Lemma~\ref{lm:contract})
4270: that for $\cgnd$ and $\tau\le (d/2)-1$, a vertex with $\tau+1$
4271: whole-loops has the largest $\lambda_{\ird{}}$ of all tangles of order $\tau$
4272: (recall that each self-loop is counted as one edge, according to
4273: Definition~\ref{de:tangle_order}).
4274: For this graph we clearly have $\lambda_{\ird{}}=2\tau+1$; hence
4275: $\taufund$ is the smallest integer $\tau$ with $2\tau+1\ge\sqrt{d-1}$,
4276: provided that this $\tau$ is $\le (d/2)-1$.
4277: But we easily verify that this $\tau$,
4278: $$
4279: \taufund=\lceil (\sqrt{d-1}\;+1)/2 \rceil-1=\lceil (\sqrt{d-1}\;-1)/2 \rceil
4280: $$
4281: is indeed at most $(d/2)-1$ for all $d\ge 4$.
4282: We have just established
4283: the following theorem.
4284:
4285: \begin{theorem} For the model $\cgnd$, we have
4286: $\taufund=\lceil (\sqrt{d-1}\;+1)/2 \rceil-1$.
4287: \end{theorem}
4288: % We will ultimately see that this $\taufund$ is responsible for the $r$ of
4289: % Theorem~\ref{th:main} (which is why they agree).
4290:
4291: \subsection{Determining $\taufund$ for $\chnd$, $\cind$, and $\cjnd$}
4292:
4293: For $\chnd$ we have to remember that tangles can't have self-loops.
4294: Thus contractions can only be done along non-multiple edges,
4295: and $\taufund$ will not generally be the same for $\chnd$ and $\cgnd$.
4296:
4297: \begin{lemma}\label{lm:contract_distance_two}
4298: Let $u,v$ be vertices of distance two in a graph, $G$,
4299: i.e., there are no
4300: edges joining $u$ and $v$, but there is a $w$ with edges to each of
4301: $u,v$. Let $G'$ be the graph obtained by identifying $u$ and $v$
4302: and deleting one of the edges from $w$ to $u$ (or to $v$) (so that the order
4303: of $G'$ is the same as that of $G$).
4304: Then $\lambda_{\ird{}}(G)\le \lambda_{\ird{}}(G')$.
4305: \end{lemma}
4306: \proof Let $U$ be the vertex in $G'$ which is the identification of
4307: $v$ and $u$. Let the edges from $u$ to $w$ be enumerated
4308: $e_1,\ldots,e_s$, and those from $v$ to $w$ enumerated $f_1,\ldots,f_t$.
4309: The edges from $U$ to $w$ are $g_1,\ldots,g_r$, where $r=s+t-1$.
4310:
4311: First consider the case when $V_G=\{u,v,w\}$, and consider the irreducible
4312: closed walks about $w$ (which are necessary of even length). Such a closed walk begins
4313: in $w$ and takes two steps, visiting either $u$ or $v$, in, respectively,
4314: $s(s-1)$ or $t(t-1)$ ways. After coming back from a $u$ vertex, another
4315: step of length $2$ can either (1) visit a $u$ vertex, in $(s-1)^2$ ways,
4316: or (2) visit a $v$ vertex, in $t(t-1)$ ways; similarly for coming back from
4317: a $v$ vertex. Thus ``coming back from a $u$ vertex'' and
4318: ``coming back from a $v$ vertex'' forms a Markov chain, and the total number
4319: of irreducible closed walks of length $k$ about $w$ is
4320: \begin{equation}\label{eq:w_cycles}
4321: I_1(k)=
4322: \left[ \begin{array}{cc}s(s-1)&t(t-1)\end{array}\right]
4323: \left[ \begin{array}{cc}(s-1)^2&t(t-1)\\s(s-1)&(t-1)^2\end{array}\right]^{(k-2)/2}
4324: \left[ \begin{array}{c}1\\1\end{array}\right]
4325: \end{equation}
4326: We wish to compare this to the number of irreducible $G'$ closed walks about $w$,
4327: of which there are clearly
4328: $$
4329: I_2(k)=r(r-1)^{k-1} = (s+t-1)(s+t-2)^{k-1}.
4330: $$
4331: For starters, we see
4332: $$
4333: I_2(2)-I_1(2)=2(s-1)(t-1)
4334: $$
4335: which is non-negative, since both $s,t\ge 1$. Now since the maximum row sum
4336: in the $2\times 2$ matrix of equation~(\ref{eq:w_cycles}) is
4337: $$
4338: s^2+t^2-2(s+t)+1+\max(s,t),
4339: $$
4340: we have
4341: $$
4342: I_1(k+2)\le I_1(k)m_1,\quad\mbox{where}\quad
4343: m_1=s^2+t^2-2(s+t)+1+\max(s,t)
4344: $$
4345: for all $k$. But
4346: $$
4347: I_2(k+2) = I_2(k)m_2,\quad\mbox{where}\quad m_2=(s+t-2)^2,
4348: $$
4349: and
4350: $$
4351: m_2-m_1 = 2st-2(s+t)+3-\max(s,t) = 1+2(s-1)(t-1)-\max(s,t),
4352: $$
4353: which is positive unless $s$ or $t$ is $1$.
4354: Thus, provided that $s\ge 2$ and $t\ge 2$, we have
4355: $$
4356: \lambda_{\ird{}}(G)\le \sqrt{m_1}<\sqrt{m_2}=\lambda_{\ird{}}(G'),
4357: $$
4358: and
4359: \begin{equation}\label{eq:compare_I}
4360: I_1(k)\le I_1(2)m_1^{(k-2)/2}< I_2(2)m_2^{(k-2)/2} I_2(k) = I_2(k)
4361: \end{equation}
4362: for all even $k$. If $t=1$ we calculate
4363: \begin{equation}\label{eq:compare_I_t1}
4364: I_1(k)=s(s-1)^{k-1}=I_2(k),
4365: \end{equation}
4366: and similarly when $s=1$.
4367:
4368:
4369: We shall use the above calculation below. We can now assume that
4370: $V_G$ has a vertex, $x$, different from $u,v,w$.
4371:
4372: There is a natural bijection of edges, $\iota$ from $E_{G}
4373: \setminus(\{e_i\}\cup\{f_i\})$ to $E_{G'}\setminus\{g_i\}$.
4374: Extend $\iota$ to a map on all of $E_G$ by defining $\iota(e_i)$ and
4375: $\iota(f_i)$ to be a formal symbol $S$. For any irreducible $G$ closed walk
4376: about $x$ specified by its edges,
4377: $c=(c_1,\ldots, c_k)$ with $c_i\in E_G$, we associate a sequence
4378: $$
4379: \iota(c) = \bigl( \iota(c_1),\ldots,\iota(c_\ell) \bigr).
4380: $$
4381: We claim that the number of $c$ with a given image $\iota(c)$ is no
4382: more than the number of $E_{G'}$ closed walks corresponding to $\iota(c)$
4383: by changing all $g_i$ edges into $S$'s.
4384: Indeed, consider a block of consecutive $S$'s in $\iota(c)$, i.e.
4385: $\iota(c_a)=\iota(c_{a+1})=\cdots=\iota(c_b)=S$, and $\iota(c_{a-1})\ne S$
4386: and $\iota(c_{b+1})\ne S$; $\iota(c)$ cannot begin or end with an $S$,
4387: since the closed walk begins at $x$, and so we can assume $a\ge 2$ and $b\le \ell-1$.
4388: By looking at $\iota(c_{a-1})$ and $\iota(c_{b+1})$ we can determine
4389: whether or not the $S$-block begins in $u$, $v$, or $w$, and ends
4390: in $u$, $v$, or $w$. If the $S$-block begins in $w$ and ends in $w$, then
4391: equations~(\ref{eq:compare_I}) and (\ref{eq:compare_I_t1})
4392: show that there are no fewer $G'$ sequences
4393: for the corresponding $S$-block than $G$ sequences. Next compare those
4394: $S$-blocks
4395: that begin in a $u$ and end in a $w$. The number of such sequences
4396: in $G$ is
4397: $$
4398: \left[ \begin{array}{cc}s&0\end{array}\right]
4399: \left[ \begin{array}{cc}(s-1)^2&t(t-1)\\s(s-1)&(t-1)^2\end{array}\right]^{(b-a)/2}
4400: \left[ \begin{array}{c}1\\1\end{array}\right],
4401: $$
4402: whereas the number in $G'$ is $(s+t-1)(s+t-2)^{b-a}$ (since the non-$S$
4403: edge $\iota(c_{a-1})$ can be followed by any $U$ to $w$ edge in $G'$);
4404: so the $G'$ number is no less than the $G$ number for $b-a=0$ (since $t\ge 1$),
4405: and each time $b-a$ is increased by $2$, the former number
4406: gets multiplied by an
4407: $m_2$, the latter gets multiplied by no more than $m_1$, where $m_1<m_2$,
4408: provided that $s\ge 2$ and $t\ge 2$; the $s=1$ or $t=1$ case is easily
4409: checked to result in equality.
4410: The same arguement holds for $v$ to $w$ $S$-blocks. For an $S$-block starting
4411: and ending in $u$, we wish to compare
4412: $$
4413: \left[ \begin{array}{cc}s&0\end{array}\right]
4414: \left[ \begin{array}{cc}(s-1)^2&t(t-1)\\s(s-1)&(t-1)^2\end{array}\right]^{(b-a-1)/2}
4415: \left[ \begin{array}{c}s-1\\0\end{array}\right],
4416: $$
4417: with $(s+t-1)(s+t-2)^{b-a}$.
4418: Again, it suffices to compare when $b-a=1$, which is immediate, and to check
4419: $s=1$ or $t=1$ separately.
4420: We argue for $S$-blocks starting in either $u$ or $v$ and ending in either
4421: $u$ or $v$ similarly.
4422: \proofbox
4423:
4424:
4425:
4426: \begin{theorem}\label{th:taufund_chnd}
4427: For the model $\chnd$, we have
4428: $\taufund=\lceil \sqrt{d-1}\;\rceil-1$.
4429: \end{theorem}
4430: \proof As before, we consider a $\tau$ and search for those $\tang$ of order
4431: at most $\tau$ with $\lambda_{\ird{}}(\tang)$ as large as possible.
4432: By Lemma~\ref{lm:contract_distance_two},
4433: and by contractions (in Lemma~\ref{lm:contract}),
4434: we may restrict our search to those $\tang$ with
4435: two or more edges between every pair of nodes.
4436:
4437: First assume that $\tau+2\le d/2$.
4438: If $\tang$ has two vertices, then $\tang$ has $\tau+2$ edges joining the
4439: two vertices (since there are no self-loops in $\chnd$).
4440: In this case
4441: $\lambda_{\ird{}}(\tang)=\tau+1$.
4442: We claim that this is as large
4443: a $\lambda_{\ird{}}$ as possible (again, assuming $\tau+2\le d/2$).
4444: Indeed, if $\tang$ has $r>2$ vertices,
4445: then the maximum degree of a vertex is $|E|$ minus the edges not involved
4446: with that particular vertex, which is at least $2$ for each pair of the
4447: $r-1$ other vertices. So the maximum degree is at most
4448: $$
4449: |E|-\binom{r-1}{2} 2 \le (|V|+\tau)-\binom{r-1}{2} 2 = \tau+r-(r-1)(r-2).
4450: $$
4451: Since $\lambda_{\ird{}}$ is no greater than the maximum degree minus $1$,
4452: we have
4453: $$
4454: \lambda_{\ird{}} \le \tau+r-(r-1)(r-2)-1=\tau+1-(r-2)^2.
4455: $$
4456: It follows that if $r>2$, $\lambda_{\ird{}}$ is strictly less than
4457: $\tau+1$.
4458:
4459: To achieve $\lambda_{\ird{}}(\tang)=\tau+1$ with our $\tang$ having two vertices,
4460: we required $\tau+2\le d/2$. To get $\lambda_{\ird{}}(\tang)=\tau+1$ to equal or
4461: exceed $\sqrt{d-1}$, we require $\tau+1=\bigl\lceil \sqrt{d-1} \;\bigr\rceil$,
4462: for which we must have
4463: $$
4464: \bigl\lceil \sqrt{d-1} \;\bigr\rceil + 1 \le d/2.
4465: $$
4466: Since $d/2$ is an integer, this is equivalent to
4467: $$
4468: \sqrt{d-1}+1\le d/2,
4469: $$
4470: which we easily see holds for all even $d>2$ except $d=4,6$.
4471:
4472: We conclude that $\taufund=\bigl\lceil \sqrt{d-1} \;\bigr\rceil-1$
4473: for even $d\ge 8$.
4474: It suffices to analyze the cases $d=4,6$.
4475:
4476: For each order, $\tau$, and $d=4,6$, we must examine those tangles of
4477: order $\tau$ and determine the largest possible $\lambda_{\ird{}}(\tang)$.
4478: Let us note that if $\tang$ is a tangle of order $-1$, then it is a tree and
4479: has $\lambda_{\ird{}}(\tang)=0$. If $\tang$ is a completely pruned tangle of
4480: order $0$, then $\tang$ is a closed walk and has $\lambda_{\ird{}}(\tang)=1$.
4481:
4482: If $d=4$, then consider the tangle of order $1$ with three vertices,
4483: consisting of one ``middle''
4484: vertex joined by two edges to each of two vertices.
4485: (This is a tangle by labelling the left to middle edges and the middle to
4486: right edges $\pi_1,\pi_2$.)
4487: We easily compute $\lambda_{\ird{}}=\sqrt{3}$, as this graph is
4488: bipartite and the number of irreducible
4489: walks of length $2m$ from the middle vertex,
4490: all such walks being closed walks, is clearly $4\cdot 3^{m-1}$.
4491: So for $d=4$, $\taufund=1$.
4492:
4493: For $d=6$, consider the tangle, $\tang=(V,E)$ with $V=\{v_1,v_2,v_3\}$
4494: with three edges connecting $v_1$ to $v_2$ (labelled $\pi_1,\pi_2,\pi_3$)
4495: and two edges connecting
4496: $v_2$ to $v_3$ (labelled $\pi_1,\pi_2$).
4497: We claim that $\lambda_{\ird{}}(\tang)>\sqrt{5}$. Say that a
4498: closed walk about $v_2$ ends in ``state A'' if the last vertex before $v_2$
4499: was $v_1$, and otherwise in ``state B'' (i.e. the second to last vertex
4500: is $v_3$). From state A, taking two additional irreducible steps, there
4501: are 4 ways to reach another state A, and two ways to reach another state B.
4502: From state B, taking two irreducible steps, there is one way to reach
4503: another state B and six ways to reach another state A. It easy follows
4504: that $\lambda_{\ird{}}(\tang)$ is the square root of the largest eigenvalue
4505: of
4506: \begin{equation}\label{eq:d6tangle}
4507: \left[ \begin{array}{cc}4&2\\6&1\end{array}\right].
4508: \end{equation}
4509: But $\lambda_1$ of this matrix is $(5+\sqrt{57})/2$, and this $\lambda_1$
4510: is just
4511: $\lambda_{\ird{}}(\tang)$.
4512: It follows that $\lambda_{\ird{}}(\tang)>\sqrt{6}>\sqrt{5}$, and hence
4513: $\taufund\le 2$.
4514:
4515: We wish to rule out $\taufund=1$ when $d=6$. Since we are considering only
4516: completely pruned graphs, $\tang$,
4517: each vertex has degree $\ge 2$. Such a graph, $\tang$, of
4518: order $1$ has all vertices of degree $2$ except for one of degree $4$ or
4519: two of degree $3$. In the case where there are vertices, $u,v$,
4520: of degree $3$, therefore joined
4521: by three disjoint beaded paths, then $\lambda_{\ird{}}(\tang)$ is greatest when
4522: the beaded paths are each of length $1$ (by setting up an obvious map from
4523: irreducible closed walks about $u$ from the general graph to the one with
4524: beaded paths
4525: of length $1$); hence $\lambda_{\ird{}}(\tang)\le 2$ in this case,
4526: since the graph of two vertices joined by three edges has
4527: $\lambda_{\ird{}}=2$. Similarly,
4528: in the case with $u$ of degree $4$, therefore having two beaded closed walks
4529: from $u$, $\lambda_{\ird{}}(\tang)$ is greatest when the lengths of the two
4530: closed walks are two (they cannot be one since $\chnd$ does not permit self-loops);
4531: hence $\lambda_{\ird{}}(\tang)\le \sqrt{3}$ in this case.
4532: Hence $\taufund>1$ and therefore $\taufund=2$.
4533:
4534: We conclude that $\taufund=
4535: \bigl\lceil \sqrt{d-1} \;\bigr\rceil-1$ also when $d=4,6$.
4536: \proofbox
4537:
4538: \begin{theorem} For the model $\cind$, we have
4539: $\taufund=\bigl\lceil \sqrt{d-1}\;\bigr\rceil-1$ for all $d\ge 3$.
4540: \end{theorem}
4541: \proof We argue as with $\chnd$. The only difference is that in
4542: $\cind$, two vertices can have as many as $d$ edges between them
4543: in a tangle (as opposed to $d/2$ edges in an $\chnd$ tangle). So
4544: the argument in the previous thoerem shows that the two-vertex tangles
4545: give that $\tau=\bigl\lceil \sqrt{d-1} \;\bigr\rceil-1$ equals $\taufund$ provided
4546: that $\tau+2\le d$ (as opposed to $\tau+2\le d/2$ for $\chnd$).
4547: But we easily verify that
4548: $$
4549: \bigl\lceil \sqrt{d-1} \;\bigr\rceil+1 \le d
4550: $$
4551: for all $d\ge 3$ (indeed, we have equality for $d=3$, and
4552: each time we increase $d\ge 3$ by one, $\sqrt{d-1}$
4553: increases by less than one).
4554: \proofbox
4555:
4556:
4557: \begin{theorem} For the model $\cjnd$, we have
4558: $\taufund=\bigl\lceil \sqrt{d-1}\;\bigr\rceil-1$ for all $d\ge 3$.
4559: \end{theorem}
4560: \proof
4561: As in $\cind$, for any $\tau\le d-2$ there is a tangle $G_\tau$ that is
4562: a pair of vertices
4563: with $\tau+2$ edges
4564: joining them. $G_\tau$ has order $\tau$ and $\lambda_{\ird{}}=\tau+1$; since
4565: when $\tau+2=d$ we have $\lambda_{\ird{}}(G_\tau)
4566: =d-1\ge \sqrt{d-1}$ giving a supercritical
4567: tangle, we need worry only about whether or not there is a tangle
4568: of order $\tau\le d-2$ that can beat the $\lambda_{\ird{}}$ of $G_\tau$.
4569: Again, as with $\cind$,
4570: Lemma~\ref{lm:contract_distance_two} can be applied to graphs with
4571: half-loops, and so by the same argument as for $\cind$ we have that
4572: only graphs on one or two vertices can possibly beat $G_\tau$.
4573: So consider a graph on vertices $u,v$ with $a$ half-loops about $u$,
4574: $c$ half-loops about $v$, and $b$ edges from $u$ to $v$. An irreducible
4575: path traverses edges of four different states: (1) half-loops about $u$,
4576: (2) edges from $u$ to $v$, (3) edges from $v$ to $u$, and (4) half-loops
4577: about $v$. Now we write a transition matrix about the states: for example
4578: in state (1) we may either continue on one of $a-1$ half-loops in state (1)
4579: or continue on one of $b$ edges in state (2). We find the transition
4580: matrix
4581: $$
4582: \left[ \begin{array}{cccc}a-1&b&0&0\\0&0&b-1&c\\a&b-1&0&0\\0&0&b&c-1
4583: \end{array}\right],
4584: $$
4585: and $\lambda_{\ird{}}$ of our graph is this matrice's largest eigenvalue.
4586: The order of the graph is $a+b+c-2$
4587: (recall, each half-loop
4588: contributes one to the order of a graph).
4589: But the row sum is never greater than $a+b+c-1$ (and always less unless
4590: $a$ or $c$ vanishes), and so if this graph
4591: has order $\tau$ its $\lambda_{\ird{}}$ is no more than $\tau+1$.
4592: Hence
4593: no $\cjnd$ tangle of order $\tau$ beats $G_\tau$, provided that
4594: $\tau\le d-2$. Thus
4595: $\taufund$ is the smallest number with $\taufund+1\ge \sqrt{d-1}$.
4596: \proofbox
4597:
4598: \section{Ramanujan Functions}
4599:
4600: In this section we discuss Ramanujan functions in order to (1) explain
4601: their significance, and (2) give some intuition on some very technical
4602: issues surrounding the asymptotic expansion for irreducible traces
4603: (as in Section~\ref{se:expansion}).
4604:
4605: \begin{definition} A function, $f(k)$, on positive integers, $k$, is
4606: said to be {\em {\dtreelike} of order $\alpha>0$}
4607: if there is a polynomial $p=p(k)$ and a
4608: constant $c>0$ such
4609: that
4610: $$
4611: |f(k)-(d-1)^kp(k)|\le ck^c \alpha^k
4612: $$
4613: for all $k$. We call $(d-1)^kp(k)$ the {\em principal term} of $f$,
4614: and $f(k)-(d-1)^kp(k)$ the {\em error term} (both terms are uniquely
4615: determined if $\alpha<d-1$). A function is {\em super-{\dtreelike}}
4616: if it is {\dtreelike} of order $1$.
4617: \end{definition}
4618: A {\dtreelike} function as defined before, in Definition~\ref{de:ram},
4619: is just a
4620: {\dtreelike} of order $\sqrt{d-1}$.
4621:
4622: Let $N(k)$ be the number of irreducible cycles of length $k$
4623: in a $d$-regular graph. Then
4624: in \cite{lubotzky} it is shown that if $N(k)$ is {\dtreelike}, then
4625: any eigenvalue, $\lambda\ne \pm d$, of the graph satisfies
4626: $|\lambda|\le 2\sqrt{d-1}$. The discussion there also shows that
4627: in any case,
4628: if $\lambda$ is the eigenvalue of largest absolute value $<d$, then
4629: $N(k)$ is {\dtreelike} of order $\alpha$ with
4630: $$
4631: \alpha = \frac{ |\lambda|+\sqrt{\lambda^2-4(d-1)}}{2}
4632: $$
4633: (and not for any smaller an $\alpha$).
4634: Any discussion of irreducible traces and eigenvalues is bound to be tied to
4635: {\dtreelike} functions.
4636:
4637: One important property of {\dtreelike} functions of order $\alpha$
4638: is that they are closed under addition. Another very important
4639: property is that they are closed under {\em convolution}, which we
4640: now formally explain.
4641: This property
4642: will be used in Section~\ref{se:finish}, and
4643: refined versions of it will be used in Section~\ref{se:expansion}.
4644: \begin{theorem}\label{th:baby_convolute}
4645: Let $f_1,f_2$ be {\dtreelike} of order $\alpha$ with $\alpha<d-1$.
4646: Then their convolution,
4647: $$
4648: g(k)= (f_1*f_2)(k) = \sum_{j=1}^{k-1} f_1(j)f_2(k-j)
4649: $$
4650: is also {\dtreelike} of order $\alpha$.
4651: \end{theorem}
4652: The techniques in
4653: Section~\ref{se:expansion} prove a more precise version of this
4654: theorem (keeping track of the sizes of the
4655: the error term and the coefficients of the principal part); for this
4656: reason we keep the argument below concise.
4657: \proof
4658: For $i=1,2$ let
4659: $$
4660: f_i(k) = (d-1)^k p_i(k) + e_i(k)
4661: $$
4662: where $p_i$ are polynomials and the $|e_i(k)|$ are bounded
4663: by $ck^c\alpha^k$ for some $k$. We may also write
4664: $$
4665: f_i(k) = (d-1)^k\bigl(p_i(k)+\tilde e_i(k)\bigr), \qquad
4666: \mbox{where $\tilde e_i(k)=(d-1)^{-k}e_i(k)$.}
4667: $$
4668: Since convolution is bilinear, we easily see
4669: $$
4670: f*g= e_1*e_2+(d-1)^k(p_1*p_2 + p_1*\tilde e_2 + p_2*\tilde e_1).
4671: $$
4672: It suffices to show that
4673: $$
4674: e_1*e_2,\quad (d-1)^k(p_1*p_2)(k), \quad (d-1)^k(p_1*\tilde e_2)(k),
4675: \quad (d-1)^k(p_2*\tilde e_1)(k)
4676: $$
4677: are {\dtreelike} of order $\alpha$.
4678:
4679: According to Sublemma~2.15 of \cite{friedman_random_graphs},
4680: $p_1*p_2$ is a polynomial. Next
4681: $$
4682: (p_1*\tilde e_2)(k)=\sum_{j=1}^{k-1} (d-1)^{-j} p_1(k-j)e_2(j)
4683: = \Sigma_1-\Sigma_2
4684: $$
4685: where
4686: \begin{eqnarray*}
4687: \Sigma_1& =& \sum_{j=1}^\infty p_1(k-j)(d-1)^{-j}e_2(j),\\
4688: \Sigma_2 &=& \sum_{j=k}^\infty p_1(k-j)(d-1)^{-j}e_2(j).
4689: \end{eqnarray*}
4690: Writing
4691: $$
4692: p_1(k-j) = \sum a_{r,s} k^r j^s,
4693: $$
4694: we see that
4695: $\Sigma_1$ is a polynomial, and $\Sigma_2$ is bounded by
4696: $ck^c\alpha^k(d-1)^{-k}$ (see Section~8, especially Lemma~\ref{lm:useful_aux},
4697: for details).
4698: This shows $(d-1)^k(p_1*\tilde e_2)$ is {\dtreelike} of order $\alpha$.
4699: Similarly, so is $(d-1)^k(p_2*\tilde e_1)$; $e_1*e_2$ is easily also seen
4700: to be so (with zero principal term).
4701: \proofbox
4702:
4703:
4704: \section{An Expansion for Some Selective Traces}
4705: \label{se:expansion}
4706:
4707: In this section we prove the first crucial expansion theorem.
4708: Our second such theorem, Theorem~\ref{th:with}, will extend these ideas.
4709:
4710: \begin{theorem}\label{th:main_expansion}
4711: Let $r$ be a positive integer, and let $\tset$ be a set of
4712: tangles containing all supercritical tangles of order less than $r$.
4713: Then there is an $S_0=S_0(r)$
4714: such that for all $S\ge S_0$ the following holds.
4715: We have
4716: $$
4717: \E{\irdsel_{S,\tset}(G;k)} = f_0(k)+\frac{ f_1(k)}{n}+\cdots+
4718: \frac{ f_{r-1}(k)}{n^{r-1}}+\frac{{\rm error}}{n^r},
4719: $$
4720: where
4721: the $f_i$ are {\dtreelike} and the error term satisfies the bound given in
4722: Theorem~\ref{th:SSIICexpansion}.
4723: \end{theorem}
4724:
4725: We begin by explaining why this theorem is an easy consequence of
4726: Theorem~\ref{th:SSIICexpansion}
4727: and the following theorem.
4728:
4729: \begin{theorem}\label{th:reduce1}
4730: Fix a lettering, $\cl$, of type $T$, and fixed non-negative
4731: integers
4732: % $e_{ij}$ for $i=1,\ldots,e$ and $j=1,\ldots,d/2$.
4733: $\ell_1,\ldots,\ell_{d/2}$. Let
4734: $$
4735: R_{T,\cl}(k_1,\ldots,k_b) =
4736: \sum_{(w_1,\ldots,w_b)}
4737: % \prod_{i=1}^{e}
4738: \prod_{j=1}^{d/2}
4739: % a_j^{e_{ij}}(w_i),
4740: \bigl( a_j(w_1)+\ldots+a_j(w_b) \bigr)^{\ell_j}
4741: $$
4742: where the sum is over all tuples of words $(w_1,\ldots,w_b)$ such that each
4743: $w_i$ is irreducible and of length $k_i$ and is compatible with $\cl$.
4744: Let $\widetilde T$ be a $B$-new type based on $T$, and let
4745: \begin{equation}\label{eq:define_f}
4746: % f(k) = \sum_{\scriptstyle k_1m_1+\ldots+k_bm_b=k\atop \scriptstyle
4747: % k_i,m_i\ge 1} W_{\widetilde T}(\vec m;S,\tset)R_{T,\cl}(k_1,\ldots,k_b),
4748: f(k) = \sum_{m_i\ge 1}\;\;
4749: \sum_{\substack{ k_1m_1+\ldots+k_bm_b=k\\ k_i\ge B\;{\rm if}\;e_i\in
4750: \El \\ k_i=\kf_i\;{\rm if}\;e_i\in \Ef}}
4751: W_{\widetilde T}(\vec m;S,\tset)R_{T,\cl}(k_1,\ldots,k_b)
4752: \end{equation}
4753: (with $W$ as in Theorem~\ref{th:crucial_cycle_count}). Then
4754: $f$ is {\dtreelike} for all $B\ge B_0=B_0(T)$.
4755: \end{theorem}
4756:
4757:
4758: Assume Theorem~\ref{th:reduce1} for the moment. Let $\cw$
4759: (respectively, $\cw_T$ and $\cw_{\widetilde T}$) be the walk collections
4760: corresponding to irreducible, $(S,\tset)$-selective walks
4761: (respectively, and that are associated to the type $T$ and $\widetilde T$).
4762: These walk collections are all SSIIC. According to
4763: Theorem~\ref{th:SSIICexpansion}
4764: $$
4765: \walksum{\cw}{k,n} = f_0(k)+\frac{ f_1(k)}{n}+\cdots+
4766: \frac{ f_{r-1}(k)}{n^{r-1}}+\frac{{\rm error}}{n^r},
4767: $$
4768: where
4769: \begin{equation}\label{eq:want_to_be_ram}
4770: f_i(k) = \sum_{j=0}^{r-1}\;\; \sum_{(w;[\vec t\>]){\rm \;order\;}j,\in\cw(k)}
4771: p_{i-j}\bigl( a_1(w;[\vec t\>]),\ldots,a_d(w;[\vec t\>])\bigr).
4772: \end{equation}
4773: It suffices to show that these $f_i$ are {\dtreelike}.
4774: We know there are finitely many types of order at most $r-1$.
4775: Now fix
4776: $S_0$ to be the max over $B_0(T)$ over types, $T$, of order at most $r-1$
4777: (and $B_0$ as in Theorem~\ref{th:reduce1}). For any $S\ge S_0$,
4778: choose $B=S+1$; of course, $B\ge B_0(T)$ for any type, $T$, of order
4779: at most $r-1$.
4780: We know there are finitely many $B$-new types based on any type.
4781: So the sum involving $\cw$ in equation~(\ref{eq:want_to_be_ram}) decomposes
4782: as a finite sum over $\cw_T$'s or $\cw_{\widetilde T}$'s.
4783: Furthermore, each expansion polynomial $p_{i-j}$, involving a fixed new type,
4784: $\widetilde T$, is just
4785: a function of $a_1,\ldots,a_{d/2}$ over appropriate forms,
4786: and each $a_i$ of the form is just the sum of
4787: the $a_i$ along each edge of the form. Therefore
4788: Theorem~\ref{th:reduce1} just says that each $f_i$, when summing
4789: over a $\cw_{\widetilde T}$, is {\dtreelike}. Summing over all
4790: $\cw_{\widetilde T}$ shows that the $f_i$ corresponding to $\cw$ are
4791: also {\dtreelike}.
4792:
4793: \noindent{\bf Proof (of Theorem~\ref{th:reduce1})\ \ }
4794: Clearly it suffices to prove the following theorem with
4795: $R_{T,\cl}$ replaced by
4796: $$
4797: \sum_{(w_1,\ldots,w_b)}
4798: \prod_{i=1}^{b}
4799: \prod_{j=1}^{d/2}
4800: a_j^{\ell_{ij}}(w_i),
4801: $$
4802: with $\ell_{ij}$ any set of non-negative integers.
4803:
4804: Our Lemma~\ref{lm:irdeigens}
4805: reduces the above theorem to the following.
4806:
4807: \begin{theorem}\label{th:reduce3}
4808: With notation as in Theorem~\ref{th:reduce1}, let
4809: $K_1,K_2,K_3$ be a partition of $k_1,\ldots,k_b$, and let $|K_i|$ for
4810: $i=1,2,3$ denote the sum of the $k_j$ in $K_i$. Then for fixed non-negative
4811: integers $\ell_1,\ldots,\ell_b$, Theorem~\ref{th:reduce1} holds with
4812: $R_{T,\cl}$ replaced by
4813: \begin{equation}\label{eq:specific_R}
4814: R_{T,\cl}(k_1,\ldots,k_b) =
4815: (d-1)^{|K_1|}(-1)^{|K_2|}k_1^{\ell_1}\cdots
4816: k_b^{\ell_b}.
4817: \end{equation}
4818: More generally, Theorem~\ref{th:reduce1} holds with
4819: $R_{T,\cl}$ replaced by
4820: \begin{equation}\label{eq:general_R}
4821: R_{T,\cl}(k_1,\ldots,k_b) =
4822: (d-1)^{|K_1|} k_1^{\ell_1}\cdots
4823: k_u^{\ell_u} \beta(k_{u+1},\ldots,k_b),
4824: \end{equation}
4825: where the edges are ordered so that
4826: $$
4827: \{ i | \mbox{$e_i\in \El$ and $k_i\in K_1$} \} = \{1,\ldots,u\},
4828: $$
4829: and where $\beta$ is a function such
4830: that
4831: $$
4832: |\beta(k_{u+1},\ldots,k_b)| \le c(|k_{u+1}|+\cdots+|k_b|)^c
4833: $$
4834: for some constant $c$.
4835: \end{theorem}
4836: The $R$ of equation~(\ref{eq:specific_R}) is all that is needed for $\cgnd$;
4837: it will be convenient (if not necessary) to use the $R$ of
4838: equation~(\ref{eq:general_R}) for $\cjnd$ (see Section~14).
4839: \proof
4840: It suffices to deal with the $R$ of
4841: equation~(\ref{eq:general_R}).
4842: Let
4843: $$
4844: \widetilde K_i = \{ k_j\in K_i \mid \ned_j\in \El \} .
4845: $$
4846: We may assume $\El=\{\ned_1,\ldots,
4847: \ned_t\}$
4848: and $\widetilde K_1=\{k_1,\ldots,k_u\}$;
4849: set
4850: $$
4851: \begin{array}{lll}
4852: M_1=m_1+\cdots+m_t, & \qquad & M_2=m_{t+1}+\cdots+m_b, \\
4853: j_1=k_1m_1+\cdots+k_tm_t, & &
4854: \mbox{and $j_2=k_{t+1}m_{t+1}+\cdots+k_bm_b$.} \end{array}
4855: $$
4856: Clearly it suffices to prove the theorem for
4857: $$
4858: f(k)=\sum_{\vec m} W_{\widetilde T}(\vec m;S,\tset)
4859: \sum_{\substack{k_1m_1+\cdots+k_b m_b=k\\ k_i\ge B\;{\rm for}\; i\le t}}
4860: %\sum_{j_1+j_2=k} \sum_{\scriptstyle k_1m_1+\ldots+k_tm_t=j_1\atop
4861: %\scriptstyle k_{t+1}m_{t+1}+\cdots+k_bm_b=j_2}
4862: (d-1)^{|\widetilde K_1|}
4863: k_1^{\ell_1}\ldots k_u^{\ell_u}\beta(k_{u+1},\ldots,k_t),
4864: $$
4865: understanding that $k_{t+1},\ldots,k_t$ are fixed by $\veckf$.
4866:
4867: \begin{definition} The {\em (coefficient) norm}, $|p|$, of a polynomial, $p$
4868: (which is possibly multivariate), is the largest absolute value among
4869: its coefficients.
4870: \end{definition}
4871: Working with this notion of a norm is a bit ``weak,'' (i.e., sometimes
4872: much stronger statements would hold with other norms),
4873: but this notion is sufficient for our
4874: purposes.
4875:
4876: Let
4877: $$
4878: f_{\vec m}(k)=\sum_{\vec k\cdot\vec m=k}(d-1)^{|\widetilde K_1|}
4879: k_1^{\ell_1}\ldots k_u^{\ell_u}\beta(k_{u+1},\ldots,k_t).
4880: % (-1)^{|\widetilde K_2|} k_1^{\ell_1}\ldots k_t^{\ell_t}.
4881: $$
4882: \begin{theorem}\label{th:f_vecm_est}
4883: For any vector of positive integers, $\vec m$, we have
4884: $f_{\vec m}$ is {\dtreelike} with principal term $(d-1)^kp_{\vec m}(k)$
4885: and error term $e_{\vec m}(k)$ satisfying
4886: $$
4887: |p_{\vec m}| \le c(d-1)^{(-BM_1-M_2+c)/2},
4888: $$
4889: and
4890: $$
4891: |e_{\vec m}(k)| \le ck^c (d-1)^{(k-BM_1-M_2+Bc)/2},
4892: $$
4893: where $c$ depends only on the $\ell_i$ and $\beta$.
4894: \end{theorem}
4895: \proof Fix a value of $\vec m$. Without loss of generality we may assume
4896: $m_1=\cdots=m_s=1$ and $m_{s+1},\ldots,m_u\ge 2$. For now assume that
4897: $s\ge 1$; we will later indicate the minor changes needed for the situation
4898: $s=0$ (i.e. when there are no $m_i$ belonging to $\widetilde K_1$ that
4899: equal one).
4900: Let
4901: $$
4902: g_{\vec m}(r)= \sum_{\substack{ k_{s+1},\ldots,k_b{\rm \;s.t.}\\
4903: k_{s+1}m_{s+1}+\cdots+k_bm_b=r\\ k_i\ge B\;{\rm for}\; i\le t }}
4904: % g_{\vec m}(r)= \sum_{\scriptstyle k_{s+1},\ldots,k_b\atop
4905: % \scriptstyle k_{s+1}m_{s+1}+\cdots+k_bm_b=r}
4906: (d-1)^{k_{s+1}+\cdots+k_u}\beta(k_{u+1},\ldots,k_t).
4907: $$
4908: \begin{lemma}\label{lm:crucial_B_type_est}
4909: If $\widetilde T$ is a $B$-new type, then
4910: $$
4911: |g_{\vec m}(r)|\le cr^c (d-1)^{(r-BM_1-M_2+Bc)/2}
4912: $$
4913: for some constant $c$ depending only on $\widetilde T$.
4914: \end{lemma}
4915: \proof
4916: Since each $k_i$ is at most $r$,
4917: $$
4918: \beta(k_{u+1},\ldots,k_t)
4919: $$
4920: is bounded by $cr^c$, and
4921: it suffices to prove the
4922: estimate for $g_{\vec m}$ replaced with
4923: $$
4924: \sum_{\substack{k_{s+1}m_{s+1}+\cdots+k_bm_b=r\\
4925: k_i\ge B\;{\rm for}\; i\le t }}
4926: (d-1)^{k_{s+1}+\cdots+k_u}.
4927: $$
4928: But there are only $\binom{r+b-s-1}{b-s-1}$ ways of writing
4929: $r$ as the sum of $b-s$ positive integers. So it suffices to show
4930: $$
4931: (d-1)^{k_{s+1}+\cdots+k_u} \le (d-1)^{(r-BM_1-M_2-Bc)/2}.
4932: $$
4933: Now we have
4934: $$
4935: r = k_{s+1}m_{s+1}+\cdots+k_bm_b,
4936: $$
4937: so
4938: $$
4939: 2k_{s+1}+\cdots+2k_u =
4940: r - (k_{s+1}m_{s+1}+\cdots+k_bm_b)+(2k_{s+1}+\cdots+2k_u)
4941: $$
4942: $$
4943: =r - (m_{s+1}-2)k_{s+1} - \cdots - (m_u-2)k_u -
4944: m_{u+1}k_{u+1} - \cdots - m_bk_b.
4945: $$
4946: Since $m_i\ge 2$ for $i$ between $s+1$ and $u$, and since $k_i\ge B$ for
4947: $i\le t$ (and $k_i\ge 1$ for all $i$), we conclude
4948: $$
4949: 2k_{s+1}+\cdots+2k_u \le r - (m_{s+1}-2)B - \cdots - (m_u-2)B
4950: $$
4951: $$
4952: -m_{u+1}B - \cdots - m_tB-m_{t+1}-\cdots-m_b
4953: $$
4954: $$
4955: = r-(m_{s+1}+\cdots+m_t)B+2(u-s)B-(m_{t+1}+\cdots+m_b)
4956: $$
4957: $$
4958: = r - \bigl(M_1-s-2(u-s)\bigr)B-M_2.
4959: $$
4960: Hence
4961: $$
4962: k_{s+1}+\cdots+k_u\le (2k_{s+1}+\cdots+2k_u)/2
4963: \le \Bigl(r - \bigl(M_1-s-2(u-s)\bigr)B-M_2\Bigr)/2.
4964: $$
4965: But $u,s$ are bounded by the number of edges in $\widetilde T$.
4966: \proofbox
4967:
4968:
4969:
4970: We will need another lemma.
4971: \begin{lemma}\label{lm:Q}
4972: For any non-negative integers $\ell_1,\ldots,\ell_s$ there is
4973: a polynomial $Q$ such that for
4974: all $k\ge s$ we have
4975: $$
4976: \sum_{\substack{k_1+\cdots+k_s=k\\ {\rm integers}\;k_i\ge 1}}
4977: % \sum_{k_1+\cdots+k_s=k\atop {\rm integers}\;k_i\ge 1}
4978: k_1^{\ell_1}\ldots k_s^{\ell_s} = Q(k).
4979: $$
4980: \end{lemma}
4981: \proof This is a special case of Sublemma~2.15 of \cite{friedman_random_graphs}
4982: (proven in a straightforward induction on $s$).
4983: \proofbox
4984:
4985: Now let
4986: $$
4987: j_{11}=k_1m_1+\cdots+k_sm_s=k_1+\cdots+k_s, \qquad
4988: j_{12}=k_{s+1}m_{s+1}+\cdots+k_tm_t
4989: $$
4990: (so that $j_{11}+j_{12}=j_1$). In the notation of the above two lemmas,
4991: letting $j'=j_{12}+j_2$,
4992: we have
4993: $$
4994: f_{\vec m}(k) = \sum_{j_{11}+j'=k} (d-1)^{j_{11}} Q(j_{11})
4995: g_{\vec m}(j')
4996: $$
4997: \begin{equation}\label{eq:Qgsum}
4998: = \sum_{r=1}^{k-s} (d-1)^{k-r} Q(k-r) g_{\vec m}(r);
4999: \end{equation}
5000: here we sum until $r=k-s$ since Lemma~\ref{lm:Q} requires $k\ge s$ (and
5001: $Q(k-r)$ vanishes for $k<s$), and we
5002: sum from $r=1$ to simplify the expression,
5003: despite the fact that $g_{\vec m}(r)$ clearly vanishes
5004: for
5005: $$
5006: r<B(m_{s+1}+\cdots+m_t)+m_{t+1}k_{t+1}+\cdots+m_bk_b.
5007: $$
5008: The
5009: sum in equation~(\ref{eq:Qgsum}) is clearly $\Sigma_1(k)-\Sigma_2(k)$, where
5010: \begin{eqnarray*}
5011: \Sigma_1(x) =& \sum_{r=1}^{\infty} &(d-1)^{k-r} Q(x-r) g_{\vec m}(r), \\
5012: \Sigma_2(x) =& \sum_{r=k-s+1}^{\infty}& (d-1)^{k-r} Q(x-r) g_{\vec m}(r),
5013: \end{eqnarray*}
5014: assuming these sums converge.
5015:
5016: We claim $\Sigma_1(k)$ will be the principal part of $f_{\vec m}(k)$,
5017: and $\Sigma_2(k)$ will be the error term. First we need the following
5018: lemma.
5019:
5020: \begin{lemma}\label{lm:useful_aux}
5021: For any positive integer, $D$, there is a $C_2$ such that the following
5022: holds.
5023: Let $g(r)$ be a function defined on non-negative integers, $r$,
5024: such that $|g(r)|\le C_1r^D\rho^r$, with $\rho<1$. Let $Q=Q(x)$
5025: be any polynomial
5026: of degree at most $D$.
5027: Then (1) the infinite sum
5028: $$
5029: h(x) = \sum_{r=1}^\infty Q(x-r) g(r)
5030: $$
5031: is convergent (in coefficient norm), (2) the degree of $h$ is that of $Q$, and
5032: (3) we have
5033: $$
5034: |h|\le C_1C_2(1-\rho)^{-2D} |Q|.
5035: $$
5036: The same is true for the
5037: sum
5038: $$
5039: h_u(x) = \sum_{r=u+1}^\infty Q(x-r) g(r),
5040: $$
5041: for any positive integer $u$, except that we replace the last claim with
5042: the estimate
5043: $$
5044: |h_u|\le C_1C_2u^{2D}(1-\rho)^{-2D}\rho^u|Q|
5045: $$
5046: (however, the $C_2$ in this equation might need to be larger than that
5047: in the estimate for $h$).
5048: \end{lemma}
5049: \proof
5050: First we observe that since
5051: $$
5052: \sum_{r=0}^\infty \binom{r}{j}\rho^r = \frac{\rho^{j+1}}{(1-\rho)^j},
5053: $$
5054: we have
5055: $$
5056: \sum_{r=0}^\infty r^j\rho^r = \frac{q_j(\rho)}{(1-\rho)^j},
5057: $$
5058: where $q_j$ is some polynomial of degree at most $j+1$.
5059:
5060: We first prove the claims on $h$.
5061: By the binomial theorem, and the fact that $Q$'s degree is bounded, it
5062: suffices to examine only the cases where
5063: $Q(x-r)$ is replaced by $x^ir^j$ for $i+j\le D$.
5064: In this case $h(x)$ becomes
5065: $$
5066: \sum_{r=1}^\infty x^ir^jg(r) =
5067: x^i\sum_{r=1}^\infty r^jg(r),
5068: $$
5069: and we have
5070: $$
5071: \sum_{r=1}^\infty r^j |g(r)| \le
5072: \sum_{r=0}^\infty r^j C_1r^D\rho^r = \frac{C_1\rho^{j+D+1}}{(1-\rho)^{j+D}}.
5073: $$
5074: This establishes the claim on $h$. The claim on
5075: $h_u$ is reduced to $h$ via
5076: $$
5077: h_u(x) = \sum_{r=1}^\infty \tilde Q(x-r) \tilde g(r),
5078: $$
5079: where $\tilde Q(x)=Q(x-u)$ and $\tilde g(r)=g(r+u)$. So $\tilde g$
5080: satisfies the same estimate as does $g$, except with an extra factor
5081: of $(r+u)^Dr^{-D}\rho^u\le Cu^D\rho^u$;
5082: the binomial theorem implies that $|\tilde Q|$ is at most
5083: $|Q|u^D$ times a constant depending on $D$.
5084: \proofbox
5085:
5086: We continue with the proof of Theorem~\ref{th:f_vecm_est}.
5087: We have
5088: $$
5089: (d-1)^{-k}\Sigma_1(k) = \sum_{r=1}^\infty Q(k-r) [(d-1)^{-r}g_{\vec m}(r)]
5090: = \sum_{r=1}^\infty Q(k-r) \tilde g(r),
5091: $$
5092: where $\tilde g(r)=(d-1)^{-r}g_{\vec m}(r)$. Now $Q$ is fixed in the
5093: theorem, so $|Q|$ can be regarded as a constant. Also, since
5094: $$
5095: |g_{\vec m}(r)| \le cr^c(d-1)^{(r-BM_1-M_2+Bc)/2},
5096: $$
5097: according to Lemma~\ref{lm:crucial_B_type_est}, we have
5098: $$
5099: |\tilde g(r)| \le cr^c(d-1)^{(-BM_1-M_2+Bc)/2}.
5100: $$
5101: It follows that $(d-1)^{-k}\Sigma_1(k)=h(k)$, where $h$ is a polynomial with
5102: $$
5103: |h|\le c (d-1)^{(-BM_1-M_2+Bc)/2},
5104: $$
5105: assuming that $d>2$ (so that $1-\rho$ with $\rho=(d-1)^{-1/2}$ is strictly
5106: positive).
5107:
5108: Furthermore, Lemma~\ref{lm:useful_aux} also implies that
5109: $$
5110: |\Sigma_2|\le c(d-1)^k (k-s+1)^{2D}(d-1)^{-(k-s)/2}(d-1)^{(-BM_1-M_2+Bc)/2}
5111: $$
5112: $$
5113: \le c'k^{2D}(d-1)^{(-BM_1-M_2+Bc+k)/2}.
5114: $$
5115: Now we see that $\Sigma_1(k)+\Sigma_2(k)$ is the decomposition of
5116: $f_{\vec m}(k)$ into principal and error terms, as claimed before, and
5117: that these terms satisfy the bounds stated in Theorem~\ref{th:f_vecm_est}.
5118:
5119: Finally we indicate the minor changes when $s=0$. In this case we take
5120: $Q$ to be the function $Q(0)=1$ and $Q$ vanishing elsewhere.
5121: Then $f_{\vec m}(k)=g_{\vec m}(k)$, so
5122: Lemma~\ref{lm:crucial_B_type_est} shows that $f_{\vec m}$ is {\dtreelike}
5123: with zero principal part.
5124: \proofbox
5125:
5126: We continue with the proof of Theorem~\ref{th:reduce3}.
5127: We are studying
5128: $$
5129: f(k) = \sum_{\vec m} W_{\widetilde T}(\vec m;S,\tset)f_{\vec m}(k).
5130: $$
5131: Set
5132: $$
5133: F(M_1,M_2;k) = \sum_{\substack{m_1+\ldots+m_t=M_1\\
5134: \scriptstyle m_{t+1}+\cdots+m_b=M_2}}
5135: % F_(M_1,M_2;k) = \sum_{\scriptstyle m_1+\ldots+m_t=M_1\atop
5136: % \scriptstyle m_{t+1}+\cdots+m_b=M_2}
5137: W_{\widetilde T}(\vec m;S,\tset)f_{\vec m}(k),
5138: $$
5139: so that
5140: $$
5141: % f(k)=\sum_{\scriptstyle M_1,M_2>0\atop\scriptstyle M_1+M_2\le k}
5142: f(k)=\sum_{M_1,M_2>0}
5143: F(M_1,M_2;k).
5144: $$
5145: Theorem~\ref{th:crucial_cycle_count} combined with
5146: Theorem~\ref{th:f_vecm_est} gives that for fixed $M_1,M_2$ we have
5147: that $F(M_1,M_2;k)$ is {\dtreelike} with principle term
5148: $(d-1)^k P_{M_1,M_2}(k)$ and error term $E_{M_1,M_2}(k)$ where
5149: $$
5150: |P_{M_1,M_2}| \le (d-1)^{(-BM_1-M_2+c)/2}
5151: c (\sqrt{d-1}-\epsilon)^{BM_1+M_2},
5152: $$
5153: $$
5154: |E_{M_1,M_2}(k)| \le
5155: ck^c (d-1)^{(k-BM_1-M_2+Bc)/2}
5156: c (\sqrt{d-1}-\epsilon)^{BM_1+M_2},
5157: $$
5158: and the degree of $P_{M_1,M_2}$ is bounded independent of $M_1,M_2$.
5159: So we sum over all $M_1,M_2$
5160: to conclude that
5161: $f$ is {\dtreelike}.
5162: \proofbox
5163: \section{Selective Traces In Graphs With (Without) Tangles}
5164:
5165: Let us review our general approach to the Alon conjecture.
5166: We are interested in expansions in $1/n$
5167: of the expected value of the $k$-th
5168: irreducible trace. Unfortunately these expansions have some coefficients
5169: that fail to be {\dtreelike}, and, as explained in Section~2, this
5170: prevents us from proving the Alon conjecture.
5171: Replacing irreducible traces with
5172: selective traces gives
5173: {\dtreelike} coefficients up to any desired power of $1/n$. However,
5174: selectivity, in the presence of appropriate tangles, modifies the
5175: irreducible trace in a way that seems hard to control.
5176: Thus we don't know how to use the results of the last section to
5177: conclude the Alon conjecture.
5178:
5179: The last main
5180: idea of the proof is to get an expansion for the expected value
5181: of selective traces counted only when appropriate tangles are present
5182: (i.e., the selective trace multiplied by a characteristic function
5183: over those graphs in $\cgnd$ with appropriate tangles).
5184: The methods of the last section generalize, rather tediously, to
5185: such expansions. These expansions will also have
5186: {\dtreelike} coefficients up to any desired power of $1/n$.
5187: It follows that we also get such expansions for the expected value
5188: of the selective trace counted only when appropriate tangles are
5189: {\em not} present; for this count, the selective trace and irreducible
5190: trace are the same. This information turns out to be enough to prove
5191: the Alon conjecture (with an auxilliary lemma proven in
5192: Section~11).
5193:
5194: %% We shall give an overview of the techniques of this section.
5195: %% Consider the tangle, $\tang$, of two vertices, $\{u_1,u_2\}$
5196: %% with three edges between
5197: %% them labelled $\pi_1,\pi_2,\pi_3$ in the direction from $u_1$ to $u_2$.
5198: %% Let $A$ be the event that $\tang$ appears in a graph in $\cgnd$.
5199: %% Let $C$ be the number of times a singly traversed simple loop
5200: %% (STSL, as in Subsection~\ref{sb:stsl}) of length $k$ labelled
5201: %% $\pi_1^5$ is traversed. Consider the task of computing
5202: %% the $\cgnd$ expectation
5203: %% $$
5204: %% \E{ C \chi_A},
5205: %% $$
5206: %% where $\chi_A$ is $1$ on $A$ and $0$ outside of $A$.
5207:
5208: %% Before doing this, consider the task of simply computing $\E{\chi_A}$,
5209: %% i.e., the probability that $\tang$ occurs. It is true that $u_1,u_2$
5210: %% can occur in $\intn$ in $n(n-1)/2$ ways, and for each way there is
5211: %% a $1/n^3$ chance that $\pi_1,\pi_2,\pi_3$ take $u_1$ to $u_2$.
5212: %% Of course, the expression $n(n-1)/2$ times $1/n^3$ counts the expected
5213: %% number of times $\tang$ occurs. The first inclusion/exclusion type
5214: %% correction would be to subtract an expression counting how many times
5215: %% two distinct tangles occur. This gets complicated, as $\tang$ may
5216: %% occur twice with distinct vertices or with shared vertices. In
5217: %% the distinct case we anticipate four distinct vertices, $u_1,\ldots,u_4$,
5218: %% with $\binom{n}{4}$ possible choices, and $\pi_i(u_{2j-1})=v_{2j}$
5219: %% for $i=1,2,3$, $j=1,2$, occurring with probability $(n^2-n)^{-3}$.
5220: %% However, after multiplying $\binom{n}{4}$ with $(n^2-n)^{-3}$, we
5221: %% must remember that this situation has a two-fold symmetry; given two
5222: %% disjoint $\tang$'s in the graph, they will be counted twice by
5223: %% such an expression. Thus we include a factor of $1/2$; this gives
5224: %% the right count when there is only one tangle or when there are two
5225: %% disjoint tangles. Next we consider two tangles with shared vertices,
5226: %% namely $u_1,u_2,u_3$ with edges $\pi_i(u_{j-1})=u_j$ for $i=1,2,3$,
5227: %% $j=2,3$; this time we multiply $\binom{n}{3}$ by $(n^2-n)^{-3}$ and
5228: %% stop, for there are no symmetries or automorphisms to this
5229: %% configuration.
5230:
5231: %% So far we have seen that calculating $\E{\chi_A}$ involves
5232: %% inclusion/exclusion type counting over configuarations that look like
5233: %% forms. The counting is made more complicated by (1) automorphism, and
5234: %% (2) the way tangles can meet. The above procedure turns out to be
5235: %% finite if we only need an expansion up to some power in $1/n$.
5236:
5237: %% To calculate $\E{ C \chi_A}$, we can now add the geometry of the STSL
5238: %% into the picture.
5239: %% Below we define a {\em formoid}, which consists of
5240: %% the form of a potential walk, with appropriate tangle
5241: %% configuration data; it is much like a form, but, for example, need not
5242: %% be connected. We will also define a {\em typoid}, which contains the
5243: %% type data plus tangle configuration data; the typoid will have some
5244: %% edges of fixed length one (with a label) from the tangle, and some edges
5245: %% of undertermined length (as do types). Complications arise
5246: %% because a potential walk can share
5247: %% vertices and edges with tangles; for example, the numbering of typoid
5248: %% edges (indicating in what order they first appear in the closed walk)
5249: %% can include edges of fixed length one.
5250:
5251:
5252:
5253: %% In this section we
5254: %% take any of the selective traces discussed in this paper and to
5255: %% remove the contribution from graphs with tangles. What is left
5256: %% comes from graphs without tangles, where the selective
5257: %% trace (based on those tangles)
5258: %% will always agree with the trace without selectivity.
5259: %% This eliminates the need to analyze the effect of selectivity.
5260:
5261: %% The expansions of Section 8 carry over to traces where we exclude
5262: %% the contribution from tangled graphs.
5263: %% This is a tedious but fairly straightforward consequence of
5264: %% the methods already developed, along with the following very important
5265: %% lemma.
5266:
5267: Before doing the above,
5268: it is crucial to know that a certain set of tangles
5269: is finite.
5270:
5271: \begin{definition}
5272: Recall that $\tseig$ is the set of supercritical tangles, i.e.,
5273: whose $\lambda_{\ird{}}$ is at least
5274: $\sqrt{d-1}$. Recall that $\tseig[r]$ is the subset of elements of
5275: $\tseig$ of
5276: order at most $r$. Let $\tsmin[r]$ be the set of tangles of
5277: $\tseig[r]$ that are minimal with respect to inclusions, i.e., that
5278: don't have another element of $\tseig[r]$ properly included in it.
5279: \end{definition}
5280:
5281: \begin{lemma}\label{lm:finiteness}
5282: The set $\tsmin[r]$ is finite.
5283: \end{lemma}
5284: This means that containing a supercritical tangle of order at most $r$
5285: is equivalent to containing
5286: one of a finite set of tangles.
5287: \proof
5288: Assume that $\tsmin[r]$ is not finite. With each tangle we associate
5289: a type which is the labelled graph obtained by supressing the degree
5290: two vertices. Since there are finitely many types of order at most $r$,
5291: there must be an infinite number of $\tsmin[r]$ tangles of some type,
5292: $T$. By passing to a subsequence we may assume there is an infinite
5293: sequence of $\tsmin[r]$ elements, $\{\tang_i\}$, such that for each edge
5294: of $T$ the associated labelling is either constant or has length
5295: tending to infinity; furthermore, the length must tend to infinity
5296: along at least one $T$ edge. Let $\tang_\infty$ be the limiting tangle,
5297: where we discard all edges with length tending to infinity.
5298:
5299: We claim $\lambda_{\ird{}}(\tang_i)=\lambda_1(T_{\ird{}}^i)$, where
5300: $T_{\ird{}}^i$ is the VLG with underlying directed graph $T_{\ird{}}$,
5301: and where $e=(v_1,v_2)\in E_{T_{\ird{}}}$ has length equal to the length,
5302: $\ell(v_1)$,
5303: of $v_1$ in $\tang_i$ (recall $v_1$ can be viewed as a directed edge of
5304: $T$); indeed, with a vertex path $v_1,\ldots,v_r$ in $T_{\ird{}}$ with
5305: $v_j=(u_j,u_{j+1})$ for $u_j\in V_T$, we associate the walk
5306: $u_1,\ldots,u_{r+1}$, which has $\tang_i$ length equal to the sum of
5307: the $\ell(v_i)$.
5308: For a closed walk, where $u_{r+1}=u_1$, its length (in $T_{\ird{}}^i$)
5309: $\ell(v_1)+\cdots+\ell(v_r)$, which corresponds to a unique
5310: $(\tang_i)_{\ird{}}$ closed walk of the same length, arising from
5311: the subdivided $v_j$. This correspondence is clearly a length preserving
5312: bijection between $T_{\ird{}}^i$ closed walks and $(\tang_i)_{\ird{}}$ closed walks.
5313: Hence $\lambda_{\ird{}}(\tang_i)=\lambda_1(T_{\ird{}}^i)$.
5314:
5315:
5316: Similarly $\lambda_{\ird{}}(\tang_\infty)=\lambda_1(T_{\ird{}}^\infty)$
5317: with $T_{\ird{}}^\infty$ defined similarly. Now by
5318: Theorem~\ref{th:infinite_edge}, $\lambda_1(T_{\ird{}}^i)\to
5319: \lambda_1(T_{\ird{}}^\infty)$, and so
5320: $\lambda_{\ird{}}(\tang_\infty)\ge \sqrt{d-1}$. Also $\tang_\infty$'s
5321: order is less than that of the $\tang_i$ (because of the edge removal(s)).
5322: Hence
5323: $\tang_\infty$ is again a $\tseig[r]$ tangle.
5324: But $\tang_\infty$ properly
5325: contains (all) $\tang_i$, which contradicts the supposed
5326: minimality of
5327: the $\tang_i$.
5328: Hence $\tsmin[r]$ is finite.
5329: \proofbox
5330: We illustrate the above lemma with an example. Let $\tang_i$ be a
5331: sequence of tangles whose underlying graph is the same except for one
5332: beaded cycle of length $i$ about some vertex.
5333: This infinite collection of tangles would prove troublesome
5334: to the methods of this section.
5335: However either (1) $\lambda_{\ird{}}(\tang_i)<\sqrt{d-1}$ for some $i$,
5336: at which point only finitely many of the $\tang_i$ are relevant,
5337: or (2) the limiting tangle, $\tang_\infty$, has
5338: $\lambda_{\ird{}}(\tang_\infty)\ge\sqrt{d-1}$, in which case a $\tang_i$
5339: inclusion implies a $\tang_\infty$ inclusion.
5340:
5341: \begin{theorem}\label{th:with}
5342: Let $\tset$ be a finite set of pruned (nonempty)
5343: tangles of order at least $1$. Let $\chi_\tset$ be the
5344: indicator function of the event that $G\in\cgnd$ contains a (i.e., at
5345: least one) tangle
5346: from $\tset$, i.e.,
5347: $$
5348: \chi_\tset(G)=\left\{ \begin{array}{ll} 1 & \mbox{if $G$ contains a tangle
5349: from $\tset$,} \\ 0 & \mbox{if not.}\end{array} \right.
5350: $$
5351: Let $\tset'$ be a set of tangles including all
5352: supercritical tangles of order less than $r$.
5353: Then for any $r$ there is an $S_0=S_0(r,\tset,\tset')$ such
5354: that for all $S\ge S_0$ we have an expansion
5355: \begin{equation}\label{eq:with_expansion}
5356: \E{\chi_\tset\irdsel_{S,\tset'}(G;k)} = f_0(k)+\frac{ f_1(k)}{n}+\cdots+
5357: \frac{ f_{r-1}(k)}{n^{r-1}}+\frac{{\rm error}}{n^r},
5358: \end{equation}
5359: where
5360: the $f_i$ are {\dtreelike} and the error term satisfies
5361: % the bound given in Theorem~\ref{th:SSIICexpansion}.
5362: $$
5363: |{\rm error}| \le c k^{\widetilde r}(d-1)^k
5364: $$
5365: with $c$ and $\widetilde r$ depending only on $r$, $\tset$, and $\tset'$.
5366: \end{theorem}
5367:
5368: We believe this theorem is true even if $\tset$ contains cycles, i.e.,
5369: tangles of order $0$. But to prove this would be more difficult,
5370: since the last part of Lemma~\ref{lm:truncate} would not be true
5371: (see the remark that follows this lemma).
5372:
5373: Before giving the proof of Theorem~\ref{th:with}, we give an important
5374: corollary of it and Theorem~\ref{th:main_expansion}.
5375: \begin{corollary} With notation and conditions
5376: as in Theorem~\ref{th:with}, we have
5377: that
5378: $$
5379: \E{(1-\chi_\tset)\irdsel_{S,\tset'}(G;k)}
5380: $$
5381: also has an expansion of the form given by the right-hand-side of
5382: equation~(\ref{eq:with_expansion}).
5383: \end{corollary}
5384:
5385: \proof {\bf (of Theorem~\ref{th:with})}\ \
5386: We want to generalize walk sums, forms, types, etc. into structures
5387: that incorporate the presence of a $\tset$ tangle. We shall first
5388: explain why we run into inclusion/exclusion and tangle automophisms
5389: (as in Theorem~\ref{th:tangle_count}).
5390:
5391: Define a
5392: {\em potential tangle specialization} as a pair, $(\Omega,\sigma)$,
5393: of a tangle, $\Omega$, and an inclusion $\sigma\from V_\Omega\to\intn$.
5394: Let $\chi_{\Omega,\sigma}$ denote the indicator function of the event
5395: that $\Omega$ is contained in $G\in\cgnd$ via the map $\sigma$.
5396: Let $\chi_{(w;\vec t\;)}$ be the indicator function of the event,
5397: ${\cal E}(w;\vec t\;)$, of the
5398: potential
5399: walk $(w;\vec t\;)$.
5400: We plan to study sums involving various
5401: $\chi_{(w;\vec t\;)}\chi_{\Omega,\sigma}$.
5402:
5403: For now consider $\chi_{\Omega,\sigma}$ alone. Since the word ``form''
5404: was used earlier to mean ``look at the graph traced out and
5405: forget the specific values of the vertices,''
5406: it makes sense to view $\Omega$ itself as the ``form'' of $(\Omega,\sigma)$,
5407: and let
5408: $$
5409: \E{\Omega}_n = \sum_\sigma \E{\chi_{\Omega,\sigma}}
5410: =\frac{n!}{(n-v)!}\prod_{i=1}^{d/2} \frac{(n-a_i)!}{n!},
5411: $$
5412: where, as usual, the $a_i$ are the number of $\Omega$ edges labelled $\pi_i$
5413: and $v$ is the number of vertices in $\Omega$. The problem is that
5414: $\E{\Omega}_n$ gives the expected number of times $\Omega$ is included
5415: into a random graph; when computing traces and walk sums, we do wish
5416: to count; but when looking at tangle inclusions, when don't wish to count---
5417: we want to compute only the function $\chi_\tset$, i.e., we want a $1$ when
5418: a $\tset$ tangle is present, and a $0$ otherwise.
5419:
5420: Given two tangles, $\tang,\tang'$, let $N(\tang,\tang')$ denote the
5421: number of inclusions of $\tang$ into $\tang'$. Of course,
5422: $N(\tang,\tang)$ is the number of automorphisms of $\tang$. Also,
5423: $\E{\Omega}_n$ is just the expected value of $N(\Omega,G)$ for a
5424: $G\in\cgnd$ (viewing $G$ as a tangle). Let $\tang\le\tang'$
5425: (respectively, $\tang<\tang'$) denote that $\tang$ has an inclusion
5426: (respectively, proper inclusion) into $\tang'$.
5427: By a $\tset$-tangle we mean any tangle isomorphic to an element of
5428: $\tset$.
5429: A {\em derived tangle of $\tset$} is a tangle that is the
5430: nonempty union of $\tset$-tangles. Let
5431: $\tset^+$ be a collection of one derived tangle of $\tset$ in
5432: every tangle isomorphism class.
5433:
5434: \begin{proposition} There exist reals numbers $\{\mob_\Omega\}_{\Omega\in
5435: \tset^+}$ such that for any $\Omega'\in\tset^+$ we have
5436: $$
5437: \sum_{\Omega\le\Omega'} N(\Omega,\Omega')\mob_\Omega = 1.
5438: $$
5439: \end{proposition}
5440: \proof This is just generalized M\"obius inversion: $\tset^+$ is a
5441: partially ordered set, and given $\Omega'$
5442: there are only finitely many $\Omega$
5443: with $\Omega<\Omega'$. Furthermore $N(\Omega',\Omega')$ is positive for
5444: all $\Omega'$. So we can inductively solve for $\mob_\Omega$.
5445: \proofbox
5446:
5447: It follows that
5448: \begin{eqnarray}\label{eq:mother1}
5449: \E{\chi_\tset\irdsel_{S,\tset'}(G;k)}
5450: &=& \sum_{(w,\vec t\>)\in\cw} \E{\chi_{(w;\vec t\;)}\chi_\tset} \\
5451: \label{eq:mother2}
5452: &=& \sum_{(w,\vec t\>)\in\cw}
5453: \quad\sum_{(\Omega,\sigma),\Omega\in\tset^+}
5454: \E{\chi_{(w;\vec t\;)}\chi_{\Omega,\sigma}} \mob_\Omega,
5455: \end{eqnarray}
5456: where $\cw$ is the walk collection corresponding to the irreducible
5457: $(S,\tset')$ selective walks.
5458:
5459: \begin{lemma}\label{lm:truncate}
5460: Let $\cw_{<r}$ be those
5461: elements of $\cw$ of order less than $r$,
5462: and similarly for $\tset^+_{<r}$.
5463: In equation~(\ref{eq:mother2}), by replacing the summation over $\cw$
5464: and $\tset^+$ by summation over $\cw_{<r}$ and $\tset^+_{<r}$, the
5465: difference is at most $C(d-1)^k n^{-r}k^{2r}$.
5466: Furthermore, the set $\tset^+_{<r}$ is finite for each $r$.
5467: \end{lemma}
5468: We remark that if $\tset$ could contain a tangle of order $0$, i.e.,
5469: one whose underlying graph is a cycle, then $\tset^+_{<1}$ would be
5470: infinite, containing arbitrarily large disjoint unions of tangles
5471: whose underlying graph is a cycle.
5472: \proof
5473: We know that we can ignore $(w;\vec t\;)$'s of order at least $r$ by
5474: using
5475: Theorem~\ref{th:SSIICexpansion} and the fact that $|\chi_\tset|\le 1$
5476: in equation~(\ref{eq:mother1}).
5477: This leaves us with a sum over $\cw_{<r}$ and $\tset^+$, and it is left
5478: to see what happens when $\tset^+$ is replaced by $\tset^+_{<r}$.
5479:
5480: For the $(\Omega,\sigma)$'s, notice that every tangle in $\tset^+$
5481: of order at least $r$ must contain a tangle of order between
5482: $r$ and $r+s-1$, where $s$ is the maximum number of edges in a $\tset$-tangle;
5483: this follows because taking the union of any graph with a $\tset$-tangle
5484: increases the number of edges by at most $s$.
5485: Lemma~\ref{lm:order_increases} shows that there are only finitely many
5486: elements of $\tset^+$ of order at most $r+s-1$ (and that any such element
5487: is the union of at most $r+s-1$ elements of $\tset$).
5488: Theorem~\ref{th:tangle_count} now shows that there is a $C=C(r)$
5489: such that a graph contains a $\tset^+$-tangle of order at least $r$
5490: with probability at most $Cn^{-r}$. Let $\tset^+_{<r}$ be the subset
5491: of tangles of order less than $r$ in $\tset^+$. Then
5492: $$
5493: \chi_\tset(G) = h(G) + \sum_{\Omega\in \tset^+_{<r}} N(\Omega,G) \mob_\Omega,
5494: $$
5495: where $h(G)$ is a function bounded by a $1$ plus the finite sum
5496: of all $|\mob_\Omega|$ over $\Omega\in \tset^+_{<r}$.
5497: Truncating the sum
5498: in equation~(\ref{eq:mother2}) to $\Omega\in \tset^+_{<r}$ therefore
5499: introduces an error of at most
5500: $$
5501: \sum_{(w,\vec t\>)\in\cw} \E{\chi_{(w;\vec t\;)}h(G)}\le
5502: d(d-1)^{k-1} C n^{-r} \max(h).
5503: $$
5504: \proofbox
5505:
5506: %%
5507: %%
5508:
5509: According to Lemma~\ref{lm:truncate}, it suffices to fix an
5510: $\Omega\in\tset^+_{<r}$ and to show that
5511: \begin{equation}\label{eq:Omega_aim}
5512: \sum_{(w,\vec t\>)\in\cw_{<r}}\sum_\sigma
5513: \E{\chi_{(w;\vec t\;)}\chi_{\Omega,\sigma}}
5514: \end{equation}
5515: has an expansion like the right-hand-side of
5516: equation~(\ref{eq:with_expansion}).
5517: So fix an $\Omega\in\tset^+_{<r}$; we now define the ``form'' of
5518: $(w;\vec t\>)$, but we incorporate into the form
5519: the information of how $\Omega,\sigma$
5520: overlaps with $(w;\vec t\>)$ by allowing $\Omega$ and $\Gamma$ to share
5521: vertices and edges.
5522:
5523: \begin{definition}
5524: \label{de:Omega_isom}
5525: An $\Omega$-specialization of a form, $\Gamma$,
5526: is an inclusion $\iota\from V_\Gamma\cup V_\Omega\to\intn$. An
5527: $\Omega$-isomorphism of forms is an isomophism of forms $\Gamma_1\to\Gamma_2$
5528: which is the identity from $V_{\Gamma_1}\cap V_\Omega$ to
5529: $V_{\Gamma_2}\cap V_\Omega$. We introduce the notation
5530: $$
5531: \E{\Gamma;\Omega}_n
5532: = \frac{n!}{(n-v)!}\prod_{i=1}^{d/2} \frac{(n-a_i)!}{n!},
5533: $$
5534: where $v=|V_\Gamma\cup V_\Omega|$
5535: and $a_i$ is the number of $\pi_i$ labels occurring
5536: in $\Gamma\cup\Omega$.
5537: \end{definition}
5538: Given $(w;\vec t\;)$ and $\Omega,\sigma$ as above, we set
5539: $[\vec t;\sigma]$ to be the set of all pairs,
5540: $(\vec s,\tau)$, such that there is a permutation of
5541: the integers taking $\vec t$ to $\vec s$ and $\sigma$ to $\tau$;
5542: we set $[\vec t;\sigma]_n$ to the same, with the additional requirement that
5543: the components of $\vec s$ and the image of $\tau$
5544: lie in $\intn$.
5545: If we let $\Gamma$ be the form of $w$, sharing vertices and edges
5546: with $\Omega$
5547: where $\sigma$ and $\vec t$ ``overlap,'' then
5548: $$
5549: \E{\Gamma;\Omega}_n = \sum_{ (\vec s,\tau)\in [\vec t;\sigma]_n}
5550: \E{ \chi_{(w;\vec s)}
5551: \chi_{\Omega;\tau} }.
5552: $$
5553: %% $\E{\Gamma;\Omega}_n$ is clearly the expected value of
5554: %% $\chi_{(w;\vec t\;)}\chi_{\Omega,\sigma}$ for a word, $w$, of form $\Gamma$,
5555: %% summed over all $\vec t$ and $\sigma$ that together include
5556: %% $V_\Gamma\cup V_\Omega$ into $\intn$.
5557: It follows that the expression in
5558: equation~(\ref{eq:Omega_aim}) is just
5559: \begin{equation}\label{eq:Omega_fundamental}
5560: \sum_{\Gamma\in \cw_{<r}} W_\Gamma(\cw,k)\E{\Gamma;\Omega}_n,
5561: \end{equation}
5562: summed over one $\Gamma$ from each $\Omega$-isomophism class
5563: (compare equation~(\ref{eq:fundamental})).
5564:
5565: We now go through the rest of Sections~5, 6, and 8, indicating how to
5566: modify our results to allow deal with sums in
5567: equation~(\ref{eq:Omega_fundamental}).
5568: Let us begin by remarking that the expansion of Theorem~\ref{th:exp_polys}
5569: generalizes easily here, in that
5570: $$
5571: \E{\Gamma;\Omega}_n = n^{v-e} \biggl( p_0+\frac{p_1}{n}+\cdots+\frac{p_q}{n^q}
5572: + \frac{{\rm error}}{n^{q+1}} \biggr)
5573: $$
5574: where
5575: the $p_i=p_i(a_1,\ldots,a_{d/2},v)$ are the expansion polynomials, and
5576: $$
5577: |{\rm error}| \le e^{qk/(n-k)}\bigl( v(v-1)/2 + a_1(a_1-1)/2 + \cdots
5578: + a_{d/2}(a_{d/2}-1)/2 \bigr)^q
5579: $$
5580: (by equation~(\ref{eq:error_term}) ). Since our potential walk is of
5581: length $k$, we have
5582: $$
5583: v \le k + |V_\Omega|,
5584: $$
5585: and
5586: $$
5587: \sum a_i \le k + |E_\Omega|.
5588: $$
5589: Since $\Omega$ is fixed, it follows that the error is bounded
5590: by $ck^{2r-2}$ in the range $k\le n/2$,
5591: since the order of $\Gamma\cup\Omega$ is at least $1$
5592: (since $\Gamma\cup\Omega$ is pruned and contains an element of $\tset$).
5593:
5594: We next claim that given $w$ (and our fixed
5595: $\Omega$), the number of equivalence
5596: classes $[\vec t;\sigma]$ corresponding to $w$ and $\Omega$ is
5597: bounded by $Ck^{2r+|V_\Omega|}$;
5598: indeed, Lemma~\ref{lm:equiv_classes} shows there at most $Ck^{2r}$
5599: $\vec t$ classes. The additional $|V_\Omega|$ choices of $\sigma$
5600: values can be chosen from at most $k+|V_\Omega|-1$ ``old'' values plus
5601: one ``new'' value each time, for a total number of equivalence classes
5602: of at most $Ck^{2r}$ times $(k+|V_\Omega|)^{|V_\Omega|}$, which is
5603: the claimed bound.
5604:
5605: We obtain that the sum in equation~(\ref{eq:Omega_fundamental}) is
5606: \begin{equation}\label{eq:Omega_Gamma_exp}
5607: \sum_{\Gamma\in \cw_{<r}} W_\Gamma(\cw,k)\biggl[
5608: \frac{p_0(\Gamma;\Omega)}{n^{\ord(\Gamma;\Omega)}} + \cdots +
5609: \frac{p_{r-1-\ord(\Gamma;\Omega)}(\Gamma;\Omega)}{n^{r-1}}
5610: + \frac{{\rm error}(\Gamma;\Omega)}{n^r} \biggr]
5611: \end{equation}
5612: with $p_i(\Gamma)=p_i\bigl(a_1(\Gamma;\Omega),\ldots,v(\Gamma;\Omega)\bigr)$ the
5613: expansion polynomials,
5614: with
5615: $$
5616: \sum_{\Gamma\in \cw_{<r}} W_\Gamma(\cw,k) {\rm error}(\Gamma;\Omega) =
5617: O(k^{4r-2+|V_\Omega|})(d-1)^k.
5618: $$
5619:
5620: To attack the expansion polynomial sums in equation~(\ref{eq:Omega_Gamma_exp}),
5621: we introduce straightforward generalizations of types and new types.
5622: \begin{definition} An $\Omega$-type\mythreeindex{Omega-type}{$\Omega$-type}{
5623: a generalization of type that incorporates information of an included tangle,
5624: $\Omega$} is an oriented graph
5625: $G_T=(V_T,E_T)$, with vertex and edge partial numberings
5626: such that (1) $\Omega$ is a subgraph of $G_T$,
5627: (2) all vertices of $G_T$
5628: except possibly the first one and possibly $V_\Omega$ vertices
5629: are of degree
5630: at least $3$, and (3) all vertices and edges not in $\Omega$ are numbered.
5631: \end{definition}
5632: To a form, $\Gamma$, we associate its $\Omega$-type, $T$, by taking
5633: $\Gamma\cup\Omega$ and supressing the degree $2$ vertices in
5634: $V_\Gamma\setminus V_\Omega$. $T$'s edges are partially numbered since
5635: we number the edges in they order they must be traversed by a
5636: corresponding walk, and we don't number an edge if the walk doesn't
5637: traverse the edge (such edges are the edges of $E_\Omega\setminus E_\Gamma$).
5638:
5639: \begin{definition} A $B$-new $\Omega$-type is a collection
5640: $\widetilde T=
5641: (T;\El,\Ef;{\vec k}^{\rm fixed})$,
5642: (1) a lettered $\Omega$-type,
5643: $T$, (2) a partition of $E_T$ into two sets, $\El,\Ef$, (3) for each
5644: $\ned_i\in \Ef$ an {\em edge length}, $\kf_i$, with $0<\kf_i<B$, and
5645: (4) a $\Pi^+$-labelling of $\Ef$ with each $\ned_i\in \Ef$
5646: labelled with a word
5647: of length $\kf_i$. Furthermore, we require that $E_\Omega$ is contained
5648: in $\Ef$, and that each $E_\Omega$ edge has length $1$ in $\widetilde T$
5649: and is labelled with its $\Omega$ label in $\widetilde T$.
5650: \end{definition}
5651: So conditions (2)--(4) are just as for a new type
5652: (See Definition~\ref{de:new_type}), and condition (1) here involves an
5653: $\Omega$-type instead of a type.
5654:
5655: Again, it is easily seen that there are finitely many (isomorphism
5656: classes) of $\Omega$-types of a given order, and finitely many
5657: $B$-new $\Omega$-types belonging to a type, $T$,
5658: for a given $B$ (and $T$). It suffices to show that for $B$ and $S$
5659: sufficiently large with $B>S$ we have that for any $B$-new $\Omega$-type,
5660: $\widetilde T$, and any polynomial, $p$, in the $a_i$'s we have
5661: $$
5662: \sum_{\Gamma\in \widetilde T} W_\Gamma(\cw,k) % p(\Gamma;\Omega)
5663: p\bigl( a_1(\Gamma;\Omega),\ldots,a_{d/2}(\Gamma;\Omega) \bigr)
5664: $$
5665: is {\dtreelike}.
5666:
5667: For integers $\{m_i\}$ indexed over $e_i$ edges of $E_T$,
5668: set
5669: $$
5670: W_{\widetilde T}(\vec m;S,\tset')
5671: $$
5672: to the number of walk classes of new type $\widetilde T$,
5673: traversing the edge $e_i$ $m_i$ times, that are
5674: irreducible $(S,\tset')$ selective cycles and that respect the
5675: $T$ partial numbering and orientation; we assume $m_i\ge 1$ or $m_i=0$
5676: according to whether or not $e_i$ is numbered (i.e., $e_i$ is to be
5677: traversed on our walk); assuming $\widetilde T$
5678: is a $B$-new type with $B>S$, this number does not depend on the
5679: ``long'' (i.e., $\El$) edge lengths.
5680: Index the edges so that $\El=\{e_1,\ldots,e_t\}$
5681: and $\Ef=\{e_{t+1},\ldots,e_{b'}\}$, with $\{e_{t+1},\ldots,e_b\}$ the
5682: numbered edges (so $m_{b+1}=\cdots=m_{b'}=0$).
5683: We claim that the proof of Theorem~\ref{th:crucial_cycle_count} shows that
5684: $$
5685: W_{\widetilde T,S}(M_1,M_2) =
5686: \sum_{\substack{ m_1+\ldots+m_t=M_1\\
5687: \scriptstyle m_{t+1}+\cdots+m_b=M_2} }
5688: W_{\widetilde T}(\vec m;S,\tset'),
5689: % \sum \biggl\{ W_{\widetilde T}(\vec m;S,\tset') \biggm|
5690: % \sum_{e\in \El} m_e = M_1,\quad
5691: % \sum_{e\in \Ef} m_e = M_2 \biggr\}
5692: $$
5693: satisfies the bound
5694: $$
5695: W_{\widetilde T,S}(M_1,M_2) \le c B (\sqrt{d-1}-\epsilon)^{cM_1+M_2}.
5696: $$
5697: for $S\ge S_0$
5698: for some $c$, $S_0$, and $\epsilon$ depending only on $\widetilde T$.
5699: This is because the argument of Theorem~\ref{th:crucial_cycle_count}
5700: is unaffected by the two essential new features that
5701: $T$ has over types, which are the possible presence of
5702: (1) some more degree $2$ vertices (other than just the vertex numbered
5703: $1$), and (2) some edges whose lengths are
5704: fixed at $1$ (namely $E_\Omega$ edges).
5705: %% But then Theorem~\ref{th:reduce1} holds
5706: %% for $\widetilde T$ being a $B$-new $\Omega$-type.
5707: But each $a_i=a_i(\Gamma;\Omega)$,
5708: which is the number of edges labelled $\pi_i$ in $\Gamma\cup\Omega$, has
5709: $$
5710: a_i(\Gamma;\Omega)=a_i(\Gamma)+a_i(\Omega\setminus\Gamma),
5711: $$
5712: and $a_i(\Omega\setminus\Gamma)$ is fixed in
5713: a new $\Omega$-type, $\widetilde T$ (by the edge partial numbering).
5714: Thus for any polynomial $p$ and new
5715: typoid, $\widetilde T$, there is a polynomial $\widetilde p$ such that
5716: \begin{equation}\label{eq:almost_done}
5717: p\bigl( a_1(\Gamma;\Omega),\ldots,a_{d/2}(\Gamma;\Omega) \bigr) =
5718: \widetilde p\bigl(
5719: a_1(\Gamma),\ldots,a_{d/2}(\Gamma) \bigr).
5720: \end{equation}
5721: But it is easy to see that Theorem~\ref{th:reduce1} holds if we extend
5722: the definition of type and new type to allow any fixed number of vertices
5723: of degree $2$. Furthermore
5724: the condition that $\Gamma$ belongs to $\widetilde T$ is the same as
5725: $\Gamma$ belonging to this extended notion of new type.
5726: We conclude that either side of equation~(\ref{eq:almost_done}) is
5727: {\dtreelike}.
5728: \proofbox
5729:
5730:
5731:
5732:
5733:
5734:
5735:
5736:
5737: \section{Strongly Irreducible Traces}
5738:
5739: We wish to use Theorem~\ref{th:with} to estimate eigenvalues. However,
5740: it is easier to use {\em strongly irreducible traces} rather than
5741: irreducible traces. We explain this and develop the properties of
5742: the strongly irreducible trace in this section.
5743:
5744:
5745: \begin{definition} A word $w\in\Pi^*$ is {\em strongly irreducible} if
5746: $w$ is irreducible and $w=\sigma_1\ldots\sigma_k$ with $\sigma_1\ne
5747: \sigma_k^{-1}$.
5748: \end{definition}
5749: For any of our irreducible traces, selective irreducible traces,
5750: irreducible walk sums, etc., we can form its ``strongly irreducible''
5751: version where we discard contributions from words that are not
5752: strongly irreducible.
5753: In any graph, labelled or not, one can speak of strongly
5754: irreducible closed walks as those closed walks that are irreducible and whose last
5755: step is not the opposite of its first step.
5756: \begin{definition} The {$k$-th strongly irreducible trace} of a graph,
5757: $G$, is the number of strongly irreducible closed walks of length $k$ for a
5758: positive integer $k$; we
5759: denote it $\sit(G,k)$ or $\sit(A,k)$ if $A$ is the adjacency matrix of $G$.
5760: If $G$ has half-loops, we consider each half-loop to be a strongly
5761: irreducible closed walk (of length 1) and include it in our count for
5762: $\sit(G,1)$ or $\sit(A,1)$.
5763: \end{definition}
5764: Half-loops are only a concern for us in the model $\cjnd$; the reason
5765: that we count half-loops as strongly irreducible is to make
5766: Lemma~\ref{lm:main_sit} hold.
5767:
5768: With a closed walk of length $k$ in $G_{\ird{}}$ about a directed edge, $e$, of
5769: $G$, we may associate the strongly irreducible closed walk in $G$ about the vertex
5770: in which $e$ originates. It follows that if $\mu_i$ are the eigenvalues
5771: of $G_{\ird{}}$, we have
5772: \begin{equation}\label{eq:SIT_sum}
5773: \sit(G,k) = \sum_{i=1}^{nd} \mu_i^k
5774: \end{equation}
5775: for all $k\ge 2$; for $k=1$ we must add the number of half-loops to the
5776: right-hand-side of the above equation, since there is no edge in
5777: $G_{\ird{}}$ from a vertex to itself when the vertex corresponds to a
5778: half-loop.
5779:
5780: We will study the relationship between $\irdtr{G}{k}$ and $\sit(G,k)$
5781: and its consequences. The most important consequence is the following
5782: theorem.
5783: \begin{theorem}\label{th:strongtrace}
5784: For $|\lambda|\le d$, let
5785: $$
5786: \mu_{1,2}(\lambda) = \frac{\lambda\pm\sqrt{\lambda^2-4(d-1)}}{2},
5787: $$
5788: and set
5789: \begin{equation}\label{eq:tilde_q}
5790: \widetilde q_k(\lambda)=\mu_1^k(\lambda)+\mu_2^k(\lambda)+
5791: \bigl( 1+(-1)^k \bigr)(d-2)/2.
5792: \end{equation}
5793: If $G$ is a $d$-regular graph with no half-loops and
5794: adjacency matrix eigenvalues $\lambda_1\ge\cdots\ge\lambda_n$,
5795: then
5796: \begin{equation}\label{eq:q_k_sum}
5797: \sit(G,k) = \sum_{i=1}^n \widetilde q_k(\lambda_i).
5798: \end{equation}
5799: Furthermore, if instead $G$ has no whole-loops, then the same is
5800: true with $\widetilde q_k$ replaced by $\widehat q_k$ where
5801: $$
5802: \widehat q_k(x) = \left\{ \begin{array}{ll}\widetilde q_k(x)&
5803: \mbox{if $k$ is even or $k=1$,} \\ \widetilde q_k(x)-x&
5804: \mbox{if $k\ge 3$ is odd.}\end{array}\right.
5805: $$
5806: \end{theorem}
5807: Furthermore, we shall see that $\widetilde q_k$, like the $q_k$
5808: of Lemma~\ref{lm:chebyshev}, are polynomials of degree $k$ that may
5809: alternatively be expressed as a simple linear combination of
5810: Chebyshev polynomials (plus the $\pm 1$ eigenvalue contribution for
5811: the $\widetilde q_k$).
5812: (In a sense, equation~(\ref{eq:q_k_sum}) says that
5813: to each eigenvalue, $\lambda$, of $G$, there correspond
5814: eigenvalues $\mu_{1,2}(\lambda)$ of multiplicity one each and eigenvalues
5815: $1$ and $-1$ of multiplicity $(d-2)/2$ each in $G_{\ird{}}$.)
5816:
5817:
5818: \begin{lemma}\label{lm:main_sit}
5819: Let $G$ be a $d$-regular graph on $n$ vertices with $h$ half-loops.
5820: Then for all
5821: integers $k\ge 2$ we have\footnote{
5822: For $k=2$ the summation in the formula to follow is ignored, since it
5823: ranges from $i=1$ to $i=\lfloor(k-1)/2\rfloor=0$.
5824: }
5825: $$
5826: \irdtr{G}{k} = \sit(G,k)+(d-2)\sum_{i=1}^{\lfloor(k-1)/2\rfloor}(d-1)^{i-1}
5827: \sit(G,k-2i)
5828: $$
5829: $$
5830: + \left\{\begin{array}{ll} 0 & \mbox{if $k$ is even,}\\ (d-1)^{(k-3)/2}h &
5831: \mbox{if $k$ is odd.} \end{array}\right.
5832: $$
5833: Therefore for all $k\ge 4$ (and for $k=3$ when $h=0$) we have
5834: \begin{equation}\label{eq:sit_recur}
5835: \irdtr{G}{k}-(d-1)\irdtr{G}{k-2} = \sit(G,k)-\sit(G,k-2)
5836: \end{equation}
5837: \end{lemma}
5838: \proof
5839: The last equation follows from the previous one, with a little care
5840: when $k=3$; indeed, for $k=3$ we have $\sit(G,k-2)=\sit(G,1)=2w+h$,
5841: where $w,h$ are the number of whole- and half-loops.
5842: So
5843: $$
5844: \irdtr{G}{3}=\sit(G,3)+(d-2)\sit(G,1)+h,
5845: $$
5846: and
5847: $$
5848: \irdtr{G}{1}=\sit(G,1)=\Trace{A} = 2w+h.
5849: $$
5850: Hence
5851: $$
5852: \irdtr{G}{3}-(d-1)\irdtr{G}{1}-\sit(G,3)+\sit(G,1)=h.
5853: $$
5854: Thus equation~(\ref{eq:sit_recur}) holds with $k=3$ if $h=0$.
5855: We similarly show that this equation holds regardless of $h$ for $k\ge 4$.
5856:
5857: So it suffices to prove the first equation of the lemma.
5858: Each irreducible closed walk about a vertex, $v$, begins by traversing a
5859: path, $p$, to a vertex, $w$,
5860: then follows a strongly irreducible (nonempty)
5861: closed walk about $w$, and then backtracks
5862: over $p$ (and this statement is only true
5863: if we count half-loops as strongly irreducible); each irreducible closed walk
5864: has a uniquely determined such $p$ and $w$.
5865: So we may count irreducible closed walks of length $k$ in $G$ by counting how
5866: many paths of length $i$ there are from $w$ that when combined with
5867: a strongly irreducible closed walk about $w$
5868: of length $k-2i$ yield an irreducible closed walk, $C$, of length $k$.
5869: The strongly irreducible closed walk's length, $k-2i$, must be positive, or
5870: else $C$ isn't irreducible.
5871: If $C$
5872: is of length at least $2$ or is a whole-loop, then there are
5873: $d-2$ possibilities for the first edge of the path (since two edges are
5874: ruled out by $C$ in either case), and $d-1$ possibilities for all edge
5875: choices thereafter. For a half-loop there are $d-1$ possibilities for
5876: the first edge (since only the single half-loop edge is ruled out).
5877: So the contribution per strongly irreducible closed walk of length $k-2i$
5878: is $1$ for $i=0$, $(d-2)(d-1)^{i-1}$ for $i\ge 1$, except in a half-loop,
5879: where the contribution per half-loop is $(d-1)^i$, i.e., an additional
5880: $(d-1)^{i-1}$ beyond the standard contribution. Since $k$ is odd and
5881: $i=(k-1)/2$ in the case of a half-loop, the additional amount beyond
5882: the standard contribution is $(d-1)^{(k-3)/2}$ per half-loop.
5883: \proofbox
5884:
5885:
5886: We return to the proof of the theorem.
5887: First assume that $G$ has no half-loops.
5888: We will prove by induction on $k\ge1$
5889: that
5890: \begin{equation}\label{eq:tilde_trace}
5891: \sit(G,k)=\sum_{i=1}^n \widetilde q_k(\lambda_i),
5892: \end{equation}
5893: where $\widetilde q_k$ are polynomials of degree $k$.
5894: Clearly
5895: $$
5896: \sit(G,1)={\rm Trace}(A),
5897: $$
5898: and so $\widetilde q_1$ exists as desired with
5899: $\widetilde q_1(\lambda)=\lambda$. Of the closed walks of length $2$, all
5900: irreducible closed walks are strongly irreducible, so $\widetilde q_2(\lambda)
5901: =\lambda^2-d$.
5902: Lemmas~\ref{lm:main_sit} and \ref{lm:chebyshev} now imply (by induction on
5903: $k$) that
5904: polynomials $\widetilde q_k(\lambda)$ exist of degree $k$ satisfying
5905: equation~(\ref{eq:tilde_trace}), and that the $\widetilde q_k$, for
5906: $k\ge 2$, are
5907: annihilated by
5908: \begin{equation}\label{eq:recurrence}
5909: (\sigma_k^2-1)\bigl(\sigma_k^2-\lambda\sigma_k+(d-1)\bigr),
5910: \end{equation}
5911: where $\sigma_k$ is the ``shift in $k$'' operator, i.e.,
5912: $\sigma_k\bigl( f(k) \bigr)=f(k+1)$ (here we use the fact that the $q_k$
5913: are annihilated by $\sigma_k^2-\lambda\sigma_k+(d-1)$, mentioned below
5914: Lemma~\ref{lm:chebyshev}).
5915: % Since the $q_k$ depend only on $d$ and $\lambda$, it suffices to
5916: % verify equation~(\ref{eq:tilde_q}) for a countable number of $\lambda$'s.
5917: Since $\mu_{1,2}(\lambda),\pm 1$ are the four roots in $\sigma_k$
5918: of equation~(\ref{eq:recurrence}), we have
5919: $$
5920: \widetilde q_k(\lambda) = c_1\mu_1^k+c_2\mu_2^k+c_3+c_4(-1)^k,
5921: $$
5922: where the $c_i=c_i(\lambda)$ assuming that the four roots are distinct.
5923: There are now two ways to finish the theorem.
5924:
5925: The first method to finish the theorem is to calculate
5926: $\widetilde q_3,\widetilde q_4$ and verify equation~(\ref{eq:tilde_q})
5927: holds for $k=2,3,4,5$, i.e.,
5928: that $c_1=c_2=1$ and $c_3=c_4=(d-2)/2$ work in those cases. Then by
5929: uniqueness (i.e., the nonvanishing of a Vandermonde determinant),
5930: those $c_i$'s must be the unique $c_i$'s that work for all $k$.
5931:
5932: Another way to finish the theorem is to use
5933: the fact that $c_1(d)=1$ (see the remark after
5934: Lemma~\ref{lm:irdeigens}), and then argue that $c_1(\lambda)=c_2(\lambda)=1$
5935: by analytic continuation. First, the $c_i$'s are the unique solutions to a
5936: $4\times 4$ system of equations with coefficient analytic in $\lambda$;
5937: hence the $c_i$ are, indeed, analytic in $\lambda$. Next,
5938: notice that $\mu_1(\lambda)$ at
5939: $\lambda=d$ analytically continues to $\mu_2(\lambda)$ at $\lambda=d$
5940: by one loop about $2\sqrt{d-1}$; thus is suffices to prove that $c_1=1$
5941: near $d$.
5942: Next notice that different $\lambda$'s give different $\mu$'s
5943: (indeed, $r^2-\lambda_1r
5944: +(d-1)=0$ and $r^2-\lambda_2 r+(d-1)=0$ for the same $r$ implies
5945: $r(\lambda_2-\lambda_1)=0$), so if $\lambda\ne d$ is an eigenvalue of
5946: multiplicity $k$ in a graph, then $c_1(\lambda)$ times $k$ must be
5947: an integer.
5948: But there is a sequence, $z_n\to 1$, with $c_1(z_n)$ an integer
5949: or half-integer (namely a cycle of length $n$
5950: where each edge has multiplicity $d/2$
5951: has $z_n=(d/2)\cos(2\pi/n)$ as an eigenvalue of multiplicty two).
5952: So by continuity $c_1(z_n)=1$
5953: for sufficiently large $n$, and thus $c_1$ is identically $1$.
5954:
5955: It suffices to determine
5956: $c_3,c_4$, which from $\widetilde q_1,\widetilde q_2$ we find are
5957: (the constant functions) $c_3=c_4=(d-2)/2$.
5958:
5959: Finally, if $G$ has only half-loops (no whole-loops),
5960: then the $k$ even formula and $k=1$ formula
5961: are unchanged.
5962: We easily see by induction on odd $k\ge 3$ that the polynomial $\widehat q_k
5963: =\widetilde q_k(x)-x$ works (we use the fact that $h$, the number of
5964: half-loops, is the trace of $A_G$, i.e., $h=\sum\lambda_i$).
5965: \proofbox
5966:
5967: \begin{theorem} Fix an integer $d>2$ and a real
5968: $\epsilon>0$. There
5969: is an $\eta>0$ such that if $G$ is a $d$-regular graph with
5970: $|\lambda_i|\le d-\epsilon$ for all $i>1$, then the $\mu_i$ of
5971: equation~(\ref{eq:SIT_sum}) satisfy $\mu_1=d-1$ and
5972: $|\mu_i|\le d-\eta$ for all $i>1$.
5973: \end{theorem}
5974: \proof The $\mu_i$ must be $\pm 1$ or roots of the equation in $\mu$
5975: $$
5976: \mu^2-\mu\lambda_i+(d-1)=0,
5977: $$
5978: or
5979: $$
5980: \mu = \frac{\lambda_i \pm \sqrt{\lambda_i^2-4(d-1)}}{2}.
5981: $$
5982: For $i=1$ we have $\lambda_1=d$ and the corresponding $\mu$'s are
5983: $\mu=d-1,1$.
5984: % Since $G_{\ird{}}$ is $(d-1)$-regular and strongly connected
5985: % (by Theorem~\ref{th:loopy}), we have $\mu=d-1$ occurs with multiplicity
5986: % $1$.
5987: The other $\mu$'s are either $1$ or come from $\lambda_i$ with
5988: $i>1$. But for $|\lambda_i|\le d-\epsilon$ it is easy to see that
5989: the corresponding $\mu$'s are bounded away from $d-1$.
5990: \proofbox
5991:
5992: We can form a selective, strongly irreducible trace by taking
5993: $$
5994: \ssit_{S,\tset'}(G;k)
5995: $$
5996: to be the number of strongly irreducible closed walks of length $k$ that
5997: are $(S,\tset')$-selective. Notice that there are no more strongly
5998: irreducible closed walks than irreducible closed walks in any lettered type,
5999: and the strong irreducibility of a potential walk can
6000: be determined from its image in the corresponding lettered type.
6001: Hence the expansion theorems of Sections 6--9, especially
6002: Theorem~\ref{th:with}, carry over to $\ssit$ replacing $\irdsel$,
6003: by simply replacing
6004: $$
6005: W_{\widetilde T,S}(M_1,M_2)
6006: $$
6007: by the same number of walk classes of the new $\Omega$-type, $\widetilde T$,
6008: except requiring that the walks are strongly irreducible.
6009: % \input sectseparate % eigenvalue separation theorems
6010: \section{A Sidestepping Lemma}
6011:
6012: \begin{lemma}\label{lm:ugly}
6013: Fix integers $r,\widetilde r,d$ with $d>2$,
6014: polynomials $p_0,\ldots,p_r$, a constant, $c$, and
6015: an integer $D$.
6016: Assume that for each $n$ we have complex-valued random variables
6017: $\theta_1,\ldots,\theta_m$ such that $m=Dn$.
6018: Assume that $1-\theta_i$ is of absolute value at most $1$, and
6019: is purely real if its absolute value is greater than $(d-1)^{-1/2}$.
6020: Furthermore assume
6021: that for all integers $k\ge 1$ we have
6022: \begin{equation}\label{eq:ugly_start}
6023: \E{ \sum_i (1-\theta_i)^k } = \sum_{j=0}^{r-1} p_j(k)n^{-j} \; + \;
6024: O\bigl(k^{\widetilde r}n^{-r} + k^c (d-1)^{-k/2}\bigr).
6025: \end{equation}
6026: Then for sufficiently large $n$ we have
6027: \begin{equation}\label{eq:ugly_finish}
6028: \E{ \sum_i \chi_{\{|\theta_i|>\log^{-2}n\}} (1-\theta_i)^k }
6029: = O(Dn^{1-(r/3)}+k^c(d-1)^{-k/2})
6030: \end{equation}
6031: for all $k$ with $1\le k\le n^\gamma$ for some constant $\gamma>0$, where the
6032: constant $\gamma$ and the constant in the $O(\,\cdot\,)$ notation depends
6033: only on $r,\widetilde r,d$ and the maximum degree of the $p_i$.
6034: \end{lemma}
6035:
6036: After this section we will apply this lemma with the $(d-1)(1-\theta_i)$
6037: being the eigenvalues of $G_{\ird{}}$,
6038: with $k$ proportional (for fixed $r$) to $\log n$.
6039:
6040:
6041: If $\sigma_k$ denotes the ``shift with respect to $k$,'' i.e.,
6042: $\sigma_k\bigl( f(k) \bigr) = f(k+1)$, then some fixed power of
6043: $\sigma_k-1$ annihilates the $p_j(k)$, and also
6044: $$
6045: (\sigma_k-1)^i (1-\theta)^k = (-\theta)^i (1-\theta)^k.
6046: $$
6047: This allows us to say a lot about the $\theta_i$ by applying some power
6048: of $\sigma_k-1$ to equation~(\ref{eq:ugly_start}).
6049: For the irreducible trace, the $(1-\theta_i)^k$ are replaced by certain
6050: Chebyshev polynomials of $\theta_i$, and applying powers of $\sigma_k-1$
6051: to them seems more awkward; this is why we have introduced strongly
6052: irreducible traces.
6053:
6054: In \cite{friedman_random_graphs}, we worked with irreducible traces, not
6055: strongly irreducible traces. There we had the $(1-\theta_i)^k$ replaced
6056: by Chebyshev polynomials of $1-\theta_i$; however (1) we knew that
6057: $\theta_1=0$ and $\theta_i$ was bounded away from $0$ with probability
6058: $1-O(n^{1-d})$, and (2) we could only prove the asymptotic expansion up to
6059: $r$ which was roughly propotional to $d^{1/2}$. So we could directly
6060: apply the analogue of equation~(\ref{eq:ugly_start}) with $k$ roughly
6061: $\log^2 n$ to determine that $p_0(k)=1$ and the higher $p_j$ vanish
6062: (up to $j$ roughly proportional to $d^{1/2}$). In this paper the
6063: arbitrary length of the asymptotic expansion for a type of trace comes
6064: at the cost of having far less control over the $\theta_i$, and we have
6065: no ability to determine the $p_j$ exactly. Fortunately
6066: Lemma~\ref{lm:ugly} allows us to control the $\theta_i$ bounded away
6067: from $0$, and fortunately we will see that
6068: the $\theta_i$ for $i>1$ are bounded away
6069: from $0$ with probability $1-O(n^{-\taufund})$.
6070:
6071: We wish to comment that one expects polynomials $p_j=p_j(k)$ to arise from
6072: the binomial expansion. Namely (by Taylor's theorem),
6073: \begin{equation}\label{binom_expansion}
6074: (1-\theta)^k = \sum_{i=0}^{s-1} \binom{k}{i}(-\theta)^i \; + \;
6075: \binom{k}{s}(-\theta)^s(1-\xi)^{k-s},
6076: \end{equation}
6077: for some $\xi\in[0,\theta]$. So for those $\theta\le n^{-\beta}$ for some
6078: constant $\beta>0$, we may take $s$ roughly $r/\beta$ and get an error term
6079: in $\xi$ bounded by roughly $O(n^{-r})$; we get a similarly bounded error
6080: term when taking expected
6081: values of $(1-\theta)^k\chi_E$ where $E$ is
6082: the event that $\theta\le n^{-\beta}$.
6083: In this way, equation~(\ref{binom_expansion}) could give rise to the terms
6084: of an asymptotic expansion.
6085: \proof (of Lemma~\ref{lm:ugly}).
6086: Let $s$ be a fixed even integer such that
6087: the maximum degree of the $p_j$ is at most $s-1$. We apply $(\sigma_k-1)^s$ to
6088: equation~(\ref{eq:ugly_start}) to conclude that
6089: $$
6090: \E{ \sum_i (-\theta_i)^s(1-\theta_i)^k } =
6091: O\bigl(k^{\widetilde r}n^{-r} + k^c (d-1)^{-k/2}\bigr)
6092: $$
6093: for any $k$, where the constant in the $O(\,\cdot\,)$ notation depends
6094: on $d$ and $s$. We conclude that for $\log^2 n\le k\le n^{r/(2\widetilde r)}$
6095: we have
6096: $$
6097: \E{ \sum_i \theta_i^s(1-\theta_i)^k } = O(n^{-r/2}).
6098: $$
6099: Applying this for $k=\lfloor n^\gamma\rfloor$ and
6100: $k=\lfloor\log^2 n\rfloor$, where
6101: $\gamma=r/(2\widetilde r)$, and subtracting
6102: we conclude
6103: $$
6104: \E{ \sum_i \theta_i^s \bigl( (1-\theta_i)^{\lfloor\log^2n\rfloor} - (1-\theta_i)^{\lfloor n^\gamma\rfloor}
6105: \bigr)}
6106: = O(n^{-r/2}).
6107: $$
6108: Since $\theta_i$ is real unless $1-\theta_i=(d-1)^{-1/2}$, we conclude
6109: that
6110: $$
6111: \E{ \sum_i |\theta_i|^s \bigl( |1-\theta_i|^{\lfloor\log^2n\rfloor} - |1-\theta_i|^{\lfloor n^\gamma\rfloor}
6112: \bigr)}
6113: $$
6114: $$
6115: = O(n^{-r/2})+O\bigl( n(d-1)^{-(\log^2 n)/2} \bigr)= O(n^{-r/2}).
6116: $$
6117: Now for any $\theta_i\le \log^{-2}n$ we have $(1-\theta_i)^{\log^2n}$
6118: is at least roughly $1/e$ for large $n$; also for $\theta_i\ge n^{-\alpha}$
6119: for a constant $\alpha>0$ we have $(1-\theta_i)^{n^\gamma}$ is near $0$
6120: for large $n$ provided $\alpha<\gamma$, and also $\theta_i^s\ge n^{-s\alpha}$.
6121: We conclude that
6122: $$
6123: \E{ \sum_i (\chi_{\{n^{-\alpha}\le\theta_i\le\log^{-2}n\}}) n^{-s\alpha} }
6124: =O(n^{-r/2}),
6125: $$
6126: and hence, since $\theta_i$ is real for $|\theta_i|<\log^{-2}n$ for
6127: $n$ large,
6128: \begin{equation}\label{eq:midrange}
6129: \E{ \chi_{n^{-\alpha}\le\theta_i\le\log^{-2}n} } = O(n^{s\alpha-(r/2)}).
6130: \end{equation}
6131:
6132: Let $\alpha>0$ be fixed with $s\alpha<r/6$, and set
6133: \begin{eqnarray*}
6134: A_1[i,n] &=& \;\mbox{The event that $\theta_i<n^{-\alpha}$},\\
6135: A_2[i,n] &=& \;\mbox{The event that $n^{-\alpha}\le \theta_i\le\log^{-2}n$},\\
6136: A_3[i,n] &=& \;\mbox{The event that $\log^{-2}n<|\theta_i|$}.
6137: \end{eqnarray*}
6138: Equation~(\ref{eq:midrange})
6139: implies that
6140: $$
6141: \prob{ A_2[i,n]} = O(n^{-r/3}).
6142: $$
6143: Since the number of
6144: $\theta_i$ is linear in $n$, we conclude that
6145: \begin{equation}\label{eq:big_expected}
6146: \E{ \sum_i \chi_{A_2[i,n]} (1-\theta_i)^k } = O(n^{1-(r/3)}).
6147: \end{equation}
6148:
6149: By the comment just before the proof, we have (using Taylor's theorem)
6150: $$
6151: \E{ \chi_{A_1[i,n]} (1-\theta_i)^k } =
6152: \sum_{j=0}^{j\le 2r/\alpha} \binom{k}{j} \E{ \chi_{A_1[i,n]}(-\theta_i)^j}
6153: \;+\;O(k^{1+(2r/\alpha)}n^{-2r}).
6154: $$
6155: Summing over $i$ in the above involves summing over $i$ in the expected
6156: values and multiplying the error term by a number linear in $n$.
6157: So let
6158: $$
6159: q(k,n) = \sum_i\sum_{j=0}^{j\le r/\alpha} \binom{k}{j}
6160: \E{ \chi_{A_1[i,n]}(-\theta_i)^j},
6161: $$
6162: which is a polynomial of fixed degree in $k$ whose coefficients depend on
6163: $n$ (and the $\theta_i$ which are given for each value of $n$).
6164: Fix a $\gamma$ for which
6165: $$
6166: O(k^{1+(2r/\alpha)}n^{1-2r}) = O(n^{-r/3})
6167: $$
6168: for all $k\le n^\gamma$, i.e., fix a $\gamma$ with
6169: $$
6170: \gamma\bigl( 1 + (r/\alpha) \bigr) \le 5r/3-1.
6171: $$
6172: Then for all $k\le n^\gamma$ we have
6173: \begin{equation}\label{eq:A1}
6174: \E{ \sum_i \chi_{A_1[i,n]} (1-\theta_i)^k } = q(k,n) + O(n^{-r/3}).
6175: \end{equation}
6176:
6177: Now combine
6178: $$
6179: \E{ \sum_i (1-\theta_i)^k } = \sum_{j=1}^3 \E{ \sum_i \chi_{A_j[i,n]}
6180: (1-\theta_i)^k }
6181: $$
6182: with equations~(\ref{eq:A1}) and (\ref{eq:big_expected}) to conclude
6183: that for $k\le n^\gamma$ we have
6184: $$
6185: \E{ \sum_i (1-\theta_i)^k } = q(k,n) + \E{ \sum_i \chi_{A_3[i,n]}
6186: (1-\theta_i)^k } + O(n^{1-(r/3)}).
6187: $$
6188: On the other hand, equation~(\ref{eq:ugly_start}) just says
6189: $$
6190: \E{ \sum_i (1-\theta_i)^k } = p(k,n) +
6191: O\bigl(k^{\widetilde r}n^{-r} + k^c (d-1)^{-k/2}\bigr),
6192: $$
6193: where $p(k,n)$ is the polynomial in $k$ given as the sum of the $p_j(k)/n^j$.
6194: Therefore
6195: $$
6196: \E{ \sum_i \chi_{A_3[i,n]}(1-\theta_i)^k }
6197: $$
6198: \begin{equation}\label{eq:A_three_sum}
6199: = p(k,n) -q(k,n)+
6200: O(n^{1-(r/3)}) + O\bigl(k^{\widetilde r}n^{-r} + k^c (d-1)^{-k/2}\bigr)
6201: \end{equation}
6202: for all $k\le n^\gamma$.
6203: Since $A_3[i,n]$ implies $|1-\theta_i|\le (1-\log^{-2}n)$ and thus
6204: $|1-\theta_i|^k\le e^{O(\log^{-2}n)}$ for $k\ge \log^4n$, we have
6205: \begin{equation}\label{eq:for_the_end}
6206: \E{ \sum_i \chi_{A_3[i,n]}
6207: |1-\theta_i|^k } = O(n^{-r/3})
6208: \end{equation}
6209: for $k\ge\log^4n$ for $n$ sufficiently large. We conclude that
6210: \begin{equation}\label{eq:pq_close}
6211: p(k,n)-q(k,n) = O(n^{1-r/3})
6212: \end{equation}
6213: for all $k$ with $\log^4 n\le k\le n^\gamma$.
6214:
6215: \begin{sublemma} Let $g(k)$ be a polynomial in $k$ of degree $\le s-1$
6216: such that $|g(i)|\le 1$ for integers $i=a,a+1,\ldots,b$ for some
6217: integers $a,b$ with $a\le b$. Then $|g(i)|\le 2^s-1$ for integers $i$
6218: with
6219: $$
6220: a-\frac{b-a}{s-1} \le i \le a.
6221: $$
6222: \end{sublemma}
6223: \proof We have $(\sigma_k-1)^s g =0$, and therefore
6224: $$
6225: g(x) = \sum_{i=1}^d \binom{s}{i}(-1)^{i-1} g(x+hi)
6226: $$
6227: for any $x$ and $h$. Given $i<a$, let $h=a-i$ and $x=i$ in the above;
6228: $x+h,x+2h,\ldots,x+sh$ are integers between $a$ and $b$ provided that
6229: $$
6230: i+(a-i)s \le b,
6231: $$
6232: so $i\ge (as-b)/(s-1)$ or $i\ge a-(b-a)/(s-1)$. If so, then
6233: $$
6234: |g(i)| \le \sum_{i=1}^d \binom{s}{i} |g(x+hi)| \le
6235: \sum_{i=1}^d \binom{s}{i} = 2^s-1.
6236: $$
6237: \proofbox
6238:
6239: Recall equation~(\ref{eq:pq_close}) and the fact that $p$ and $q$ are
6240: polynomials in $k$ (for fixed $n$) of bounded degree. So applying the
6241: above sublemma for $a=2\lceil(\log^4 n)/2\rceil$
6242: and $b=2\lfloor n^\gamma/2\rfloor$
6243: implies
6244: that $p(k,n)-q(k,n)=O(n^{1-(r/3)})$ for $1\le k\le\log^4n$.
6245: Equation~(\ref{eq:pq_close}) now holds for all $1\le k\le n^\gamma$.
6246: We conlude that equation~(\ref{eq:for_the_end}) holds
6247: for all $1\le k\le n^\gamma$.
6248: Adding this to
6249: equation~(\ref{eq:big_expected}) yields the desired
6250: equation~(\ref{eq:ugly_finish}).
6251: \proofbox
6252: \section{Magnification Theorems}
6253:
6254: In this section we use standard counting arguments to prove theorems
6255: implying
6256: ``magnification'' or ``expansion'' for ``most'' random graphs;
6257: here
6258: ``most'' means all graphs excepting a set of probability $O(n^{-\taufund})$.
6259: These theorems will then be used with Lemma~\ref{lm:ugly} to
6260: prove Theorems~\ref{th:main}, \ref{th:mainh}, and
6261: \ref{th:maini}.
6262:
6263: A graph, $G$, with $n$ vertices is said to be a $\gamma$-magnifier
6264: if for all subsets of vertices, $A$, of size at most $n/2$ we have
6265: $$
6266: |\Gamma(A)-A| \ge \gamma|A|,
6267: $$
6268: where $\Gamma(A)$ denotes those vertices connected to some member of $A$
6269: by an edge. Alon has shown that any $d$ regular $\gamma$-magnifier has
6270: $$
6271: \lambda_2(G) \le d-\frac{\gamma^2}{4+2\gamma^2}
6272: $$
6273: (see \cite{alon_eigenvalues}; see \cite{dodziuk,jerrum1,jerrum2} for related
6274: ``edge magnification'' results).
6275:
6276: \begin{definition}
6277: Say that a $d$-regular graph on $n$ vertices is a {\em $\gamma$-spreader}
6278: if for every subset,
6279: $A$, of at most $n/2$ vertices we have
6280: $$
6281: |\Gamma(A)|\ge (1+\gamma)|A|.
6282: $$
6283: \end{definition}
6284: \begin{theorem}\label{th:separation}
6285: Let $G$ be a $d$-regular $\gamma$-spreader.
6286: Then for all $i>1$ we have
6287: $$
6288: \lambda_i^2(G) \le d^2 -\frac{\gamma^2}{4+2\gamma^2}.
6289: $$
6290: \end{theorem}
6291: \proof
6292: Since the graph is $d$-regular, we have $|\Gamma(B)|\ge |B|$ for all subsets
6293: of vertices, $B$. Taking $B=\Gamma(A)$ yields
6294: $$
6295: |\Gamma^2(A)|\ge |\Gamma(A)| \ge (1+\gamma)|A|.
6296: $$
6297: Hence $G^2$, the graph on $V_G$ whose edges are paths in $G$ of length $2$
6298: (and whose adjacency matrix is $A_G^2$), is a $d^2$-regular $\gamma$-magnifier.
6299: Now apply Alon's result on magnification and eigenvalues to $G^2$, whose
6300: eigenvalues are $\lambda_i^2(G)$.
6301: \proofbox
6302:
6303: We now establish that for all our models, a graph will be a
6304: $\gamma$-spreader for some fixed $\gamma=\gamma(d)>0$ with probability
6305: $1-O(n^{-\taufund})$.
6306:
6307: \begin{theorem}\label{th:magnify_cgnd}
6308: For any $\epsilon>0$ and even $d\ge 4$ there is a $\gamma>0$
6309: such that $G\in\cgnd$ is a $\gamma$-spreader
6310: with probability
6311: $1-O(n^{-\taufund})$.
6312: \end{theorem}
6313: Later we shall prove this theorem for other models of random graphs, by
6314: very similar calculations. This theorem is easy for $d$ sufficiently
6315: large; but when $d=4$ (or later possibly $d=3$) one has to calculate
6316: fairly carefully.
6317: \proof
6318: Fix $A,B\subset\intn$, and consider the event that $\Gamma(A)\subset B$.
6319: We will impose the condition that $a=|A|\le n/2$ and
6320: $|B|=a+\lfloor \gamma a\rfloor$.
6321:
6322: Fix constants, $\gamma,C$ with $0<\gamma<1/C$. Consider the situation
6323: where
6324: $|A|<C$. In this case
6325: $|B|=|A|=a$.
6326: But since $G$ is $d$-regular and we have $d|A|$ edges
6327: leaving $A$,
6328: these edges comprise all edges incident upon $B$ (since $|B|=|A|$).
6329: Thus $A\cup B$ is a union of connected components of $G$. But this
6330: cannot occur if $G$ has no supercritical
6331: tangles of size at most $2C$ (since each connected component of $G$ has
6332: $\lambda_{\ird{}}=d-1$). For a constant
6333: $C$ there are only a constant number of tangles of size at most $2C$.
6334: Thus, by forsaking a probability of $O(n^{-\taufund})$, we may assume
6335: that $a=|A|\ge C$ for any fixed constant, $C$ (provided that we then
6336: take $\gamma<1/C$) for sufficiently large $n$.
6337:
6338: So consider a random
6339: permutation, $\pi=\pi_i$, and consider the event that $\pi$
6340: and $\pi^{-1}$ map $A$ to $B$.
6341: Let $C_1=A\cap B$, $C_2=A\setminus B$, $C_3=B\setminus A$, and let
6342: $c_i=|C_i|$.
6343: We view $\pi$ as determined by a perfect matching of a bipartite graph
6344: on inputs, $I$, and outputs, $O$, with $I,O$ being copies of $\intn$
6345: (and $i\in I$ mapped to $\pi(i)\in O$).
6346: Viewing $\pi$ as a bipartite matching, it consists of
6347: (1) $r$ edges from $C_1$ to $C_1$ (i.e., the $I$ vertices corresponding
6348: to $C_1$ to those $O$ vertices corresponding to $C_1$),
6349: (2) $c_1-r$ edges from $C_1$ to $C_3$, (3) $c_1-r$ edges from $C_3$ to
6350: $C_1$, (4) $c_2$ edges from $C_2$ to $C_3$, and (5) $c_2$ edges from
6351: $C_3$ to $C_2$. (This is true since a $C_2$ vertex, either input or output,
6352: must be paired with a $C_3$ vertex, and a $C_1$ vertex must be paired with
6353: either a $C_1$ or $C_3$ vertex.) So the event that $\pi$ and $\pi^{-1}$
6354: map $A$ to $B$ with $c_1,c_2,c_3,r,A,B$ all held fixed has probability
6355: $$
6356: p(c_1,c_2,c_3,r)=\left[ \binom{c_1}{r}^2 r! \right]
6357: \left[ \binom{c_3}{c_1-r} (c_1-r)! \right]^2 \times
6358: $$
6359: $$
6360: \left[ \binom{c_3-c_1+r}{c_2} c_2! \right]^2
6361: \left[ n(n-1)\cdots(n-2c_1-2c_2+r+1)\right]^{-1}
6362: $$
6363: (The first expression in square brackets corresponds to choosing $r$
6364: $C_1$ to $C_1$ edges; the second expression corresponds
6365: to choosing $c_1-r$ $C_1$ to
6366: $C_3$ edges, and is squared to include choosing the $C_3$ to $C_1$ edges;
6367: etc.)
6368: The probability taken over all $A,B$ of a given $c_1,c_2,c_3$ (and with
6369: $r$ fixed) is therefore at most
6370: \begin{equation}\label{eq:bimon_and_p}
6371: \binom{n}{c_1,c_2,c_3,n-c_1-c_2-c_3} p^{d/2}(c_1,c_2,c_3,r).
6372: \end{equation}
6373: It suffices to show that this expression is $O(n^{-s})$ with
6374: $s=\taufund+4$,
6375: provided that $a$ is sufficiently large (and at most $n/2$), since then
6376: we can sum equation~(\ref{eq:bimon_and_p}) over the at most $n^4$ relevant
6377: values of $c_1,c_2,c_3,r$.
6378: We should remind
6379: ourselves that $c_1,c_2,c_3,r$ range over integers with
6380: $$
6381: c_1+c_2=a,\quad
6382: c_1+c_3=a+\lfloor \gamma a\rfloor,\quad
6383: r \le c_1.
6384: $$
6385: Furthermore, considering the expression defining $p$, we have
6386: $c_3-c_1+r\ge c_2$.
6387:
6388: We now write
6389: \begin{equation}\label{eq:b_factorials}
6390: b=b(c_1,c_2,c_3,r,n) = \binom{n}{c_1,c_2,c_3,n-c_1-c_2-c_3}
6391: \end{equation}
6392: $$
6393: = \frac{n!}{c_1!\;c_2!\;c_3!\;(n-c_1-c_2-c_3)!}
6394: $$
6395: and
6396: \begin{equation}\label{eq:p_factorials}
6397: p=p(c_1,c_2,c_3,r,n) = \frac{(c_1!\;c_3!)^2\;(n-2c_1-2c_2+r)!}{
6398: \bigl( (c_1-r)!\;(c_3-c_1-c_2+r)!\bigr)^2\; r!\;n!}.
6399: \end{equation}
6400: We make some general remarks about analyzing the factorials in the
6401: above two equations:
6402: \begin{enumerate}
6403: \item All factorials in the above equations are of the form $(\mu n)!$
6404: for some $\mu\in[0,1]$. Stirling's formula $m!\sim (m/e)^m\sqrt{2\pi m}$
6405: implies that
6406: \begin{equation}\label{eq:Stirling}
6407: \frac{1}{n}\log[ (\mu n)! ] = \mu\log(n/e) + \mu\log\mu + O\left(\frac{\log n}{
6408: n}\right),
6409: \end{equation}
6410: where the constant in the $O(\,\cdot\,)$ is independent of $n$ and
6411: $\mu\in[0,1]$.
6412: \item In analyzing $b$ and $p$ above, we may ignore the $\mu\log(n/e)$ term
6413: in equation~(\ref{eq:Stirling}). This is because $b,p$ are {\em balanced}
6414: in that the sum of the numbers to which factorials are applied is the same
6415: in the numerator and denominator; in other words, the $\mu\log(n/e)$ terms
6416: in the numerator will exactly cancel those in the denominator.
6417: \item Let $f(\theta)=-\theta\log\theta$.
6418: We claim that for $\theta_1,\theta_2\in[0,1]$ we have
6419: $$
6420: |f(\theta_1)-f(\theta_2)| \le \max\bigl( f(|\Delta\theta|),f(1-|\Delta\theta|)
6421: \bigr),\qquad\mbox{with}\quad \Delta\theta=\theta_2-\theta_1.
6422: $$
6423: Indeed, since $f''(\theta)=-1/\theta<0$ for $\theta>0$, $f$ is concave
6424: in $[0,1]$, and so $g(\theta)=f(\theta+\Delta\theta)-f(\theta)$ is decreasing
6425: in $\theta$ for $\Delta\theta$ fixed; so $|g|$'s maximum over an interval is
6426: taken at its endpoints, and since $f(0)=f(1)=0$, the above claim is
6427: established.
6428:
6429: Next, a Taylor expansion shows that $-\epsilon\log\epsilon\ge
6430: -(1-\epsilon)\log(1-\epsilon)$ for sufficiently small $\epsilon>0$. Hence
6431: there is an $\epsilon_0$ such that
6432: \begin{equation}\label{eq:difference}
6433: |f(\theta_1)-f(\theta_2)| \le f(|\theta_1-\theta_2|)
6434: \end{equation}
6435: for all $\theta_1,\theta_2\in[0,1]$ with $|\theta_1-\theta_2|\le \epsilon_0$.
6436: \end{enumerate}
6437:
6438: Let $\nu_i,\rho,\alpha,\delta$ ($i=1,2,3$) be the non-negative reals given by
6439: $$
6440: c_i=\nu_i n,\quad r=\rho n,\quad a=\alpha n,\quad \lfloor \gamma a\rfloor
6441: =\delta n.
6442: $$
6443:
6444: We have that
6445: $$
6446: \nu_1+\nu_2=\alpha,\quad \nu_1+\nu_3=\alpha+\delta,
6447: \quad \rho\le\nu_1,\quad \rho\ge \nu_1+\nu_2-\nu_3.
6448: $$
6449: We conclude that
6450: $$
6451: |\nu_2-\nu_3| \le \delta,\qquad |\nu_1-\rho|\le \delta.
6452: $$
6453: It follows
6454: from equation~(\ref{eq:Stirling}), remark~(2) below it, and
6455: equation~(\ref{eq:difference}),
6456: that we may replace $\nu_3$ with $\mu_2$ and $\rho$ with
6457: $\nu_1$ in calculating $(\log b)/n$ and incur an additive error term
6458: of at most $O(\delta\log\delta)$. Thus we get
6459: $$
6460: \frac{\log b}{n} = h(\nu_1,\nu_2)+
6461: O\left(|\delta\log\delta|+\frac{\log n}{n}\right),
6462: $$
6463: where
6464: \begin{equation}\label{eq:ubiqu_h}
6465: h(\nu_1,\nu_2) = -\nu_1\log\nu_1-2\nu_2\log\nu_2
6466: -(1-\nu_1-2\nu_2)
6467: \log(1-\nu_1-2\nu_2).
6468: \end{equation}
6469: Similarly we calculate
6470: $$
6471: \frac{-\log p}{n} = h(\nu_1,\nu_2)+
6472: O\left(|\delta\log\delta|+\frac{\log n}{n}\right),
6473: $$
6474: i.e., we have
6475: the exact same equation (!)
6476: for $\log b$ replaced by $-\log p$
6477: (this ``coincidence'' happens for the other models as well).
6478: Hence
6479: $$
6480: \frac{\log(bp^2)}{n} = -h(\nu_1,\nu_2)
6481: + O\left(|\delta\log\delta|+\frac{\log n}{n}\right).
6482: $$
6483: Since $\nu_1+\nu_2=\alpha$, we have either (or both) $\nu_i$ are
6484: $\ge \alpha/2$. Hence
6485: $$
6486: h(\nu_1,\nu_2) \ge -(\alpha/2)\log(\alpha/2).
6487: $$
6488:
6489: Now we claim that for any constant $C>0$ there is a constant $\gamma>0$
6490: such that for all $\alpha\in[0,1/2]$ we have
6491: \begin{equation}\label{eq:alphalog}
6492: -\alpha\log\alpha \ge -C (\gamma\alpha)\log(\gamma\alpha).
6493: \end{equation}
6494: Indeed, for $\gamma<1$ fixed we have
6495: \begin{equation}\label{eq:g_alpha}
6496: g(\alpha) = \frac{(\gamma\alpha)\log(\gamma\alpha)}{\alpha\log\alpha}
6497: =\gamma+\frac{\gamma\log\gamma}{\log\alpha}
6498: \end{equation}
6499: is increasing for $\alpha\in[0,1/2]$. Hence it suffices to choose a
6500: $\gamma>0$ sufficiently small so that
6501: $$
6502: g(1/2)=\frac{(1/2)\log(1/2)}{(\gamma/2)\log(\gamma/2)} \ge C,
6503: $$
6504: so that
6505: $$
6506: g(\alpha)\ge g(1/2)\ge C,
6507: $$
6508: which along with equation~(\ref{eq:g_alpha}) yields
6509: equation~(\ref{eq:alphalog}).
6510:
6511: It follows that for sufficiently small $\gamma>0$ we have
6512: $$
6513: \frac{\log(bp^2)}{n} \ge -(\alpha/2)\log(\alpha/2) +
6514: O\left(|\delta\log\delta|+\frac{\log n}{n}\right),
6515: $$
6516: and, since $\delta\le\gamma a/n=\gamma\alpha$, this expression is
6517: $$
6518: \ge -(\alpha/4)\log(\alpha/2) +
6519: O\left(\frac{\log n}{n}\right).
6520: $$
6521: Hence for any constant, $C_1$, there is a $C_2$ such that if $a\ge C_2$
6522: (i.e., $\alpha\ge C_2/n$) then
6523: $$
6524: \frac{\log(bp^2)}{n} \le \frac{-C_1\log n}{n}
6525: $$
6526: for all $n$ sufficiently large. In other words $bp^{d/2}$,
6527: i.e., the expression in equation~(\ref{eq:bimon_and_p}),
6528: is at most $n^{-C_1}$; this, by the discussion after
6529: equation~(\ref{eq:bimon_and_p}),
6530: completes the proof.
6531: \proofbox
6532:
6533:
6534: \begin{theorem} Theorem~\ref{th:magnify_cgnd} holds in the models
6535: $\chnd$, $\cind$, and $\cjnd$.
6536: \end{theorem}
6537: \proof
6538: In $\chnd$ each permutation occurs with probability at most $n$ times
6539: its probability in $\cgnd$. Therefore the same analysis goes through,
6540: except that $p$ is multiplied by at most a factor of $n$. This changes
6541: the expression for $n^{-1}\log(bp^2)$ by an $O(n^{-1}\log n)$ factor,
6542: so the same proof carries over.
6543:
6544: For $\cind$ we again set $C_i$ and $c_i$ as before. A perfect matching
6545: in $\intn$ will have
6546: (1) $r$ vertices of $C_1$ paired amongst themselves,
6547: (2) $c_1-r$ vertices of $C_1$ paired with $C_3$ vertices, and
6548: (3) $c_2$ vertices of $C_2$ paired with $C_3$ vertices.
6549: This data determines the pairing for $r+2(c_1-r)+2c_2$ vertices.
6550: The expression for $b$, representing the number of ways the $C_i$ can
6551: be chosen, is the same as before. We now derive an
6552: expression for $p$, the probability that a single perfect matching
6553: matches all $A$ vertices to those in $B$.
6554:
6555:
6556: For an even integer, $m$, let {\em $m$ odd factorial} be
6557: $$
6558: m\oddf = (m-1)(m-3)\cdots 3 = \frac{m!}{2^{m/2}(m/2)!},
6559: $$
6560: which is just
6561: the number of perfect matchings of $m$ elements.
6562: Stirling's formula yields
6563: $$
6564: m\oddf \sim \sqrt{2}\;(m/e)^{m/2}
6565: $$
6566: (so that for our purposes $m\oddf$ can be regarded as replacable by
6567: the square root of $m!$).
6568:
6569: We have
6570: $$
6571: p=p(\{c_i\},r,n)=
6572: \left[ \binom{c_1}{r} r\oddf \right]\;
6573: \left[ \binom{c_3}{c_1-r} (c_1-r)! \right]\;
6574: $$
6575: $$
6576: \times
6577: \left[ \binom{c_3-c_1+r}{c_2} c_2! \right]\;
6578: \frac{ (n-2c_1-2c_2-r)\oddf }{n\oddf}.
6579: $$
6580: We get that $-\log p$ is
6581: $$
6582: \frac{h(\nu_1,\nu_2)}{2} +
6583: O\left(|\delta\log\delta|+\frac{\log n}{n}\right),
6584: $$
6585: with $h$ as in equation~(\ref{eq:ubiqu_h}).
6586: Since $b$ is unchanged,
6587: by analyzing as before we see that there is a fixed $\gamma>0$ such that
6588: for any constant $C_1$ there is a $C_2$ such that
6589: $bp^3=O(n^{-C_1})$ provided that $a\ge C_2$.
6590:
6591: Next we consider $\cjnd$. Let $G$ be a random graph in $\cjnd$,
6592: so $V_G=\intn$.
6593: Consider the graph $G'$ formed by adding one new
6594: vertex, $w=n+1$, to $G$ and replacing
6595: each half-loop about a vertex, $v$, in $G$ by an edge from $v$ to $w$.
6596: Then $G'$ is precisely distributed as an element of $\cinpd$; indeed,
6597: a perfect matching on $V_{G'}$ matches $w$ to some element of $V_G=\intn$
6598: and then randomly matches the remaining $n-1$ elements of $V_G$.
6599:
6600: Now we know that $G'$ is a $\gamma$-spreader with probability
6601: $1-O(n^{-\taufund})$. But for any $A\subset V_G$, $\Gamma_{G'}(A)$
6602: consists of at most one more vertex than $\Gamma_G(A)$. Hence
6603: for $|A|\le |V_G|/2$ and $G'$ being a $\gamma$-spreader, we have
6604: $$
6605: |\Gamma_G(A)\setminus A| \ge \gamma|A|-1 \ge \gamma'|A|,
6606: $$
6607: where $\gamma'=\gamma-(1/c')$, provided that $|A|\ge c'$. Hence $G$
6608: is a $\gamma$-spreader on sets, $A$, of size $\max(c,c')\le |A|\le n/2$.
6609: On smaller sets, $A$, we have $G$ is a $\gamma'$-spreader with probability
6610: $1-O(n^{-\taufund})$, assuming $\gamma'\le 1/\max(c,c')$, by the
6611: argument given before for $\cgnd$. (Notice that the $\taufund$
6612: for $\cind$ and $\cjnd$ are the same.) Hence a random graph in $\cjnd$ is a
6613: $\gamma'$-spreader with probability $1-O(n^{-\taufund})$.
6614: \proofbox
6615:
6616: % \input oldsectmagnify % magnification results go here
6617: \section{Finishing the $\cgnd$ Proof}
6618:
6619: Here we quickly finish the proof of Theorem~\ref{th:main}, which
6620: proves Alon's conjecture for $\cgnd$.
6621:
6622: Fix a value of $r$ to
6623: be specified later.
6624: Let $\tset'=\tseig[r-1]$ be the set of supercritical tangles of order less
6625: than $r$,
6626: and let
6627: $\tset=\tsmin[r-1]$ (which we recall is the set of minimal $\tset'=
6628: \tseig[r-1]$ elements with respect to inclusion);
6629: we know that $\tset$ is finite by
6630: Lemma~\ref{lm:finiteness}, and we recall that if $G$ contains a
6631: $\tset'$ tangle then it contains an element of $\tset$.
6632: The probability that $\chi_\tset(G)=1$, i.e., that
6633: $G$ contains at least
6634: one element of $\tset$, is at most $O(n^{-\taufund})$, since
6635: every one of the finitely many tangles in $\tset$ occurs with probability
6636: proportional to $1/n$ to the order of tangle
6637: (Theorem~\ref{th:tangle_count}), and each tangle order is
6638: at least $\taufund$. Given that $\chi_\tset(G)=0$, we have that $G$
6639: contains no supercritical tangle of order less than
6640: $r$, and hence no irreducible
6641: closed walk
6642: can fail to
6643: be $(S,\tset')$-selective for any $S$. Hence for all $S$ and $k$ we have
6644: $$
6645: \chi_\tset(G)=0 \quad\mbox{implies} \quad\ssit_{S,\tset'}(G;k)=\sit(G;k).
6646: $$
6647: Thus
6648: $$
6649: \E{ (1-\chi_\tset)\ssit_{S,\tset'}(G;k)} =
6650: \E{ (1-\chi_\tset) \sit(G;k)}.
6651: $$
6652:
6653: Now according to Theorem~\ref{th:strongtrace} we have
6654: $$
6655: \sit(G;k) = \sum_{i=1}^n \mu_1^k(\lambda_i)+\mu_2^k(\lambda_i)+
6656: \bigl( 1+(-1)^k \bigr)(d-2)/2,
6657: $$
6658: for even $k\ge 2$, where
6659: $$
6660: \mu_{1,2}(\lambda) = \frac{\lambda\pm\sqrt{\lambda^2-4(d-1)}}{2}.
6661: $$
6662: In other words, there are $nd$ numbers $\nu_i$,
6663: such that $\sit(G;k)$ is the sum of
6664: the $k$-th powers of these numbers.
6665: Also for each $i$ we have $\nu_i$ is not real only if
6666: it is of absolute value $\sqrt{d-1}$.
6667: Combining this and Theorem~\ref{th:with} we see that
6668: $$
6669: \theta_i=1-(1-\chi_\tset)\nu_i/(d-1)
6670: $$
6671: are random variables that
6672: satisfy
6673: the conditions of Lemma~\ref{lm:ugly} for each $i$ (and $G$).
6674: It follows that
6675: \begin{equation}\label{eq:verybig}
6676: \E{ (1-\chi_\tset)\sum_{i=1}^n
6677: \sum_{\substack{j\;\rm such\;that\cr |\mu_j(\lambda_i)|\le (d-1)(1-\log^{-2}n)}}
6678: \mu_j(\lambda_i)^k }
6679: = O(Dn^{1-(r/3)}+k^c(d-1)^{-k/2})
6680: \end{equation}
6681: for all $k$ with $1\le k\le n^\gamma$ for some constant $\gamma>0$
6682: depending only on $r$.
6683:
6684: According to Theorems~\ref{th:separation} and \ref{th:magnify_cgnd}
6685: there is an $\epsilon>0$ such that with probability
6686: $1-O(n^{-\taufund})$ we have $|\lambda_i|\le d-\epsilon$ for all
6687: $i\ne 1$; in this case there is an $\epsilon'=\epsilon'(\epsilon)>0$
6688: such that $|\mu_j(\lambda_i)|\le (d-1)-\epsilon'$
6689: for all $j=1,2$ and $i\ne 1$.
6690:
6691: We claim that for any $G$ and an even integer $k$ we have
6692: $$
6693: \sum_{i\;{\rm s.t.}\;\mu_{1,2}(\lambda_i)\;{\rm not\;real}}\quad \sum_{j=1}^2
6694: \mu_j(\lambda_i)^k \ge - 2(n-1)(d-1)^{k/2}
6695: $$
6696: indeed, if $\mu_j(\lambda_i)$ is not real, it
6697: is of absolute value $\sqrt{d-1}$; if $\mu_j(\lambda_i)$ is real then
6698: its $k$-th power is non-negative.
6699:
6700: Now let $A$ be the event that $\chi_\tset=0$ and that
6701: $|\mu_j(\lambda_i)|\le (d-1)-\epsilon'$
6702: for all $j=1,2$ and $i\ne 1$.
6703: Let $B=B(\eta)$ be the event that
6704: for some $j$ and some $i\ne 1$ we have
6705: $|\mu_j(\lambda_i)|\ge e^\eta\sqrt{d-1}$ for
6706: an arbitrary fixed $\eta>0$.
6707: $A\cap B$ implies that for even integer $k$ we have
6708: $$
6709: \sum_{i=2}^n\sum_{j=1}^2
6710: \mu_j(\lambda_i)^k \ge \left( e^\eta\sqrt{d-1}\right)^k - 2(n-2)(d-1)^{k/2}.
6711: $$
6712: It follows, using equation~(\ref{eq:verybig}), that for even $k$,
6713: $$
6714: \prob{ A\cap B } \left( e^\eta\sqrt{d-1}\right)^k \le
6715: \E{ \sum_{i\;{\rm s.t.}\;\mu_{1,2}(\lambda_i)\;{\rm real}} \quad\sum_{j=1}^2
6716: \mu_j(\lambda_i)^k}
6717: $$
6718: $$
6719: = \E{ \sum_{i=2}^n \sum_{j=1}^2
6720: \mu_j(\lambda_i)^k} -
6721: \E{ \sum_{i\;{\rm s.t.}\;\mu_{1,2}(\lambda_i)\;{\rm not\;real}}\quad
6722: \sum_{j=1}^2
6723: \mu_j(\lambda_i)^k}
6724: $$
6725: $$
6726: \le O(Dn^{1-(r/3)}(d-1)^k+k^c(d-1)^{k/2}) + 2(n-1)(d-1)^{k/2}.
6727: $$
6728: We now take
6729: $$
6730: k = 2 \left\lceil \frac{r\log n}{3\log(d-1)} \right\rceil.
6731: $$
6732: We have
6733: $$
6734: (k/2)-1 \le \frac{r\log n}{3\log(d-1)} \le k/2.
6735: $$
6736: Hence
6737: $$
6738: n^{-r/3}\le (d-1)^{-(k/2)+1},
6739: $$
6740: and so
6741: $$
6742: \prob{ A\cap B } \le
6743: c\max(k^c,n) e^{-k\eta}
6744: $$
6745: $$
6746: \le cn e^{-k\eta}
6747: =cn n^{-\alpha r},
6748: $$
6749: where $\alpha=(2/3)\eta/\log(d-1)$, i.e. $\alpha$ is a positive constant
6750: (depending only on $\eta$ and $d$). Choosing $r$ so that
6751: $\alpha r -1 > \taufund$, we have
6752: $$
6753: \prob{ A\cap B } = O(n^{-\taufund}).
6754: $$
6755: But we have already seen
6756: (Theorems~\ref{th:separation} and \ref{th:magnify_cgnd})
6757: that
6758: $$
6759: \prob{ \complement A } = O(n^{-\taufund}),
6760: $$
6761: where $\complement A$ is the complement of $A$.
6762: Hence
6763: $$
6764: \prob{ B} = \prob{B\cap A} + \prob{B\cap\complement A} = O(n^{-\taufund}).
6765: $$
6766: For any $\epsilon>0$ there is an $\eta>0$ such that $|\lambda|\ge
6767: 2\sqrt{d-1}\;+\epsilon$ implies $|\mu_i(\lambda)|\ge e^\eta\sqrt{d-1}$
6768: for at least one $i$, which is the event
6769: $B=B(\eta)$ above. It follows that for any $\epsilon>0$ we have
6770: $$
6771: \prob{ \mbox{$|\lambda_i|\ge 2\sqrt{d-1}\;+\epsilon$ for some $i>1$}}
6772: = O(n^{-\taufund}).
6773: $$
6774: This (and Theorem~\ref{th:improved_bound}) proves Theorem~\ref{th:main}.
6775: \section{Finishing the Proofs of the Main Theorems}
6776: \label{se:finish}
6777:
6778: We now complete the proofs of Theorems~\ref{th:mainh}
6779: and \ref{th:maini}, i.e., we establish the Alon
6780: conjecture for $\chnd$, $\cind$, and $\cjnd$.
6781:
6782: The proofs of the theorems are as the proof for $\cgnd$.
6783: We only need to establish the following results for the different models
6784: of random graph:
6785: \begin{enumerate}
6786: \item Labelling:
6787: The model comes with edges labelled from a set $\Pi$ such that
6788: to each $\pi\in\Pi$ we associate a $\pi^{-1}\in\Pi$ such that
6789: $(\pi^{-1})^{-1}=\pi$ (in other words,
6790: the elements of $\Pi$ are paired, with the possibility that an element
6791: is paired with itself).
6792: \item Coincidence: If $k$ of the random edges have been determined,
6793: and if we fix any two vertices, $v,w$,
6794: in the graph, then the probability that an
6795: edge of a given label takes $v$ to $w$ is at most $c/(n-ck)$ for some
6796: constant $c$. We have only briefly mentioned coincidences in this paper, but
6797: our Lemmas~\ref{lm:big_order} and \ref{lm:equiv_classes}, proven in
6798: \cite{friedman_random_graphs}, require a property like this.
6799: \item Expansion with Error:
6800: Consider a $\Pi$-labelled graph, $H$, with vertices a subset of $\intn$,
6801: that can occur as a subgraph of a graph in our model.
6802: The probability that $H$ occurs must
6803: depend only on the
6804: number of edges, $a_\pi$, of each label, $\pi$ (of course,
6805: $a_\pi=a_{\pi^{-1}}$). Furthermore, this probability
6806: times the number subsets of $\intn$ of size $V_H$ is, for every positive
6807: integer $r$,
6808: $$
6809: \Esymm(H)_n = \left(\sum_{i=0}^{r-1} \frac{p_i(\vec a)}{n^i}\right)
6810: + \frac{{\rm error}}{n^r},
6811: $$
6812: where $p_i$ are polynomials in $\vec a$ (where $\vec a$ is the collection
6813: of all $a_\pi$) and where
6814: $$
6815: |{\rm error}| \le c k^{r'}
6816: $$
6817: for all $k\le n/c$, where $c_1,r'$ depend only on $r$.
6818: Furthermore, $p_i=0$
6819: if $i$ is less than the order of $H$.
6820: \item Simple Word Sum:
6821: Let $\ird{k,\sigma,\tau}$ be those words that begin with $\sigma$, end
6822: in $\tau$, and are {\em irreducible} (meaning no consecutive occurrence of
6823: $\pi$ and $\pi^{-1}$). Then for any polynomial, $p=p(\vec a)$ (with
6824: $\vec a$ as above), we require
6825: \begin{equation}\label{eq:simple_word}
6826: \sum_{w\in\ird{k,\sigma,\tau}} p\bigl( a_1(w),\ldots,a_{d/2}(w),k\bigr)
6827: = (d-1)^k Q_1(k)+ E(k)
6828: \end{equation}
6829: for a polynomial, $Q_1$, and a function $E$ with $|E(k)|\le ck^c$ for
6830: some constant $c$ (i.e., the above sum is super-{\dtreelike}).
6831: \item $\taufund$ determination:
6832: We must determine $\taufund$ for the model.
6833: \item Spreading:
6834: There is a constant $\gamma>0$ such that the probability that a
6835: random graph has $|\lambda_i|\ge d-\gamma$ for some $i\ne 1$ is of order
6836: at most $n^{-\taufund}$.
6837: \end{enumerate}
6838:
6839: We have already shown spreading and determined $\taufund$ for all
6840: three models. The labelling of the models is quite simple:
6841: $\chnd$ is labelled like $\cgnd$; $\cind$ is labelled with
6842: $\Sigma=\{\sigma_1,\ldots,\sigma_d\}$ where $\sigma_i^{-1}=\sigma_i$
6843: (each $\sigma_i$ represents a perfect matching);
6844: $\cjnd$ is labelled with $\Sigma\cap T$ with $\Sigma$ as before and
6845: $T=\{\tau_1,\ldots,\tau_d\}$ with $\tau_i^{-1}=\tau_i$, and where
6846: the $\sigma_i$ represent the near perfect matching and the $\tau_i$
6847: represents the single completing half-loop for $\sigma_i$.
6848:
6849: Coincidence is easily checked for all three models.
6850:
6851: We address the issue of Simple Word Sum. The word sum for $\chnd$
6852: is the same as for $\cgnd$. For $\cind$, the technique of Lemma~2.11
6853: of \cite{friedman_random_graphs} reduces the matter to the irreducible
6854: eigenvalues
6855: of a vertex with $d$ half-loops; since these eigenvalues are the
6856: eigenvalues of a $d\times d$ matrix which is $0$ on the diagonal and
6857: $1$'s elsewhere, the eigenvalues are $d-1$ with multiplicity $1$ and
6858: $-1$ with multiplicity $d-1$. Hence the simple word sum of
6859: equation~(\ref{eq:simple_word}) is given by
6860: \begin{equation}\label{eq:matching_walk}
6861: (d-1)^k Q_1(k)+(-1)^k Q_2(k)
6862: \end{equation}
6863: where $Q_i$ are polynomials.
6864: For $\cjnd$ we can break the sum by how many half-loops are involved.
6865: For a fixed set of half-loops involved in the irreducible word,
6866: the sum is a convolution of functions of the form in
6867: equation~(\ref{eq:matching_walk}), which by Theorem~\ref{th:baby_convolute}
6868: is again super-{\dtreelike}.
6869:
6870:
6871: We now establish Expansion with Error for the three models.
6872: Equation~(\ref{eq:probability}) has the $\chnd$ analogue
6873: $$
6874: P(w;\vec t\>) = \prod_{i=1}^{d/2} \frac{(n-a_i-1)!}{(n-1)!}.
6875: $$
6876: Now recall the proof of Theorem~\ref{th:exp_polys}, especially
6877: equations~(\ref{eq:g_rational}) and (\ref{eq:error_term}).
6878: For $\chnd$, we have
6879: \begin{equation}\label{eq:chnd_form}
6880: \Esymm(H)_n= n(n-1)\cdots(n-v+1)
6881: \prod_{i=1}^{d/2} \frac{(n-a_i-1)!}{(n-1)!}.
6882: \end{equation}
6883: $$
6884: = n^{v-e} g(1/n),
6885: $$
6886: with $g$ as in equation~(\ref{eq:g_rational})
6887: with $b_1,\ldots,b_v$ being $0,1,\ldots,v-1$
6888: and $c_1,\ldots,c_e$ being the collection of the sequences
6889: $1,2,\ldots,a_i$. Hence for a walk of length at most $k$ we have
6890: $$
6891: \sum b_j+\sum c_j \le \binom{k}{2}+\binom{k+1}{2}=k^2.
6892: $$
6893: Accordingly Expansion with Error holds for $\chnd$ with expansion polynomials
6894: determined by equation~(\ref{eq:chnd_form}), and with
6895: error term bounded by
6896: $$
6897: e^{rk/(n-k)}k^{2r};
6898: $$
6899: this bound is $\le ck^{r'}$ for all $k\le n$ with $r'=2r$ and $c=e^r$.
6900:
6901: Similarly for $\cind$ we have the analogue
6902: $$
6903: P(w;\vec t\>) = \prod_{i=1}^{d} \frac{(n-a_i)\oddf}{n\oddf}.
6904: $$
6905: The analysis goes through essentially as before; in the error bound
6906: we have $\sum b_j$ is again $\binom{k}{2}$, but this time the
6907: $\sum c_j$ is as large as
6908: $$
6909: 1+3+5+\cdots+(2k-1) = k^2
6910: $$
6911: (taking one $a_i=2k$ and the rest $0$).
6912: So Expansion with Error holds for $\cind$ with
6913: error term bounded by
6914: $$
6915: e^{rk/n}\left(k^2+\binom{k}{2}\right)^{r}\le
6916: e^{rk/n}(2k)^{2r}.
6917: $$
6918:
6919:
6920: For $\cjnd$, consider a random $1$-regular graph, $G'$, consisting of a near
6921: perfect matching plus one complementing half-loop on the vertex set
6922: $\intn$. Notice that the
6923: number of such graphs is $n(n-1)\oddf$. Hence the probability of
6924: occurrence of a specified half-loop and $a$ other matchings in $G'$ is
6925: $$
6926: \frac{(n-1-2a)\oddf}{n(n-1)\oddf} = \frac{1}{n(n-2)\cdots(n-2a)},
6927: $$
6928: and the probability of $a$ specified matchings (with no specified half-loop)
6929: is
6930: $$
6931: \frac{(n-2a)(n-1-2a)\oddf}{n(n-1)\oddf} = \frac{1}{n(n-2)\cdots(n-2a-2)}.
6932: $$
6933: So for any specification of half-loops in $H$, i.e., any fixing of each
6934: $a_{\tau_i}$ to $0$ or $1$,
6935: $\Esymm(H)_n$ is a polynomial in the $a_{\sigma_i}$'s; this makes
6936: $\Esymm(H)_n$ a polynomial
6937: in the $\vec a$, namely
6938: $$
6939: \sum_{I\subset\{1,\ldots,d\}} \biggl(p_I(a_{\sigma_1},\ldots,a_{\sigma_d})
6940: \prod_{i\in I} a_{\tau_i}
6941: \prod_{i\notin I} (1-a_{\tau_i})\biggr).
6942: $$
6943: We also see that, in the terminology above, $\sum b_j = \binom{k}{2}$
6944: and $\sum c_j \le 2\binom{k-1}{2}$. Hence Expansion with Error holds
6945: for $\cjnd$ as well.
6946:
6947: This establishes the six required results mentioned at the beginning of
6948: this section for the models $\chnd$, $\cind$, and $\cjnd$.
6949: Theorems~\ref{th:mainh} and \ref{th:maini} follow.
6950:
6951:
6952: \section{Closing Remarks}
6953:
6954: We make a number of final remarks.
6955:
6956: \paragraph{Stronger conjectures:}
6957: As mentioned before, numerical experiments indicate that the average
6958: (and median) $\lambda_2$ for a random graph is $2\sqrt{d-1}+\epsilon(n)$,
6959: where $\epsilon(n)$ is a negative function (tending to $0$ as $n\to\infty$).
6960: By the results of Friedman and Kahale (extending the Alon-Boppana result),
6961: $-\epsilon(n)\le O(\log^{-2}n)$ (see \cite{friedman_geometric_aspects}).
6962: However, the trace method, even with selective traces, seems to require
6963: some fundamental new idea in order to have any hope of achieving
6964: $\epsilon(n)$ that is zero or negative.
6965:
6966:
6967: \paragraph{Critical $d$:}
6968: As mentioned before, when there is a critical tangle of order
6969: strictly less than that of any hypercritical tangle, then our techniques
6970: leave a gap in that we can only prove $\lambda_2>2\sqrt{d-1}$ with
6971: probability at least $c/n^s$ where $s>\taufund$. This case is extremely
6972: interesting, since it seems that there should be a theorem that closes
6973: this gap, and such a theorem would either get around a poorly bounded
6974: $W_{\widetilde T,S}(M_1,M_2)$ or improve the very interesting
6975: Theorem~\ref{th:remarkable} (or do something else).
6976:
6977:
6978: \paragraph{Relative Alon Conjecture:}
6979: Following \cite{friedman_relative}, it seems quite possible to
6980: relativize the main theorems in this paper. Namely, fix a
6981: ``base'' graph, $B$, (or, more generally, a ``base''
6982: pregraph, in the sense of
6983: \cite{friedman_geometric_aspects}). Fix an $\epsilon>0$. Then we
6984: believe that most
6985: random coverings of $B$ of degree $n$ have all ``new'' eigenvalue
6986: $\le \epsilon+\rho$, where $\rho$ is the spectral radius of the
6987: universal cover of $B$. Similarly, we can ask for $\epsilon$
6988: to be zero or even a negative function of $n$. See
6989: \cite{friedman_relative,friedman_tillich_generalized} for further
6990: discussion and a result in this direction.
6991:
6992:
6993: \paragraph{Alternate Proof with Trace (see the end of Section~2):}
6994: It may be possible to analyze the expected irreducible trace over all
6995: of $\cgnd$. As remarked in Theorem~\ref{th:nottreelike} and the
6996: discussion thereafter, the coefficients of $g_i(k)$ there could no
6997: longer be {\dtreelike}.
6998: It may be possible to analyze selective traces without discarding
6999: contributions from tangled graphs. In other words, if we better
7000: understood how selectivity affected irreducible traces, we might
7001: not need Section~9 (and certain parts of our understanding of
7002: these traces might improve). Clearly selectivity in $G$ can be
7003: expressed in terms of walks in an induced subgraph of a ``higher block
7004: presentation'' of $G$ (see \cite{marcus,kitchens}).
7005: However, it is not clear what can be said about the eigenvalues
7006: of induced subgraphs of a higher block presentation; the author has
7007: only some weak results in this directions
7008: (see \cite{friedman_SVD}).
7009:
7010:
7011:
7012: \section*{Glossary}
7013: {\em This glossary contains a term or a piece of notation, followed
7014: by a colon (:), followed by a brief description, followed by the page
7015: number(s) where the term/notation is explained.}
7016: \medskip
7017: \renewenvironment{theindex}{\begin{description}}{\end{description}}
7018: \input all.ind % this index file is set up to look like a glossary,
7019: % and appears like a glossary using the environment
7020: % ``description''
7021:
7022: %\appendix
7023: %\input appendixa
7024:
7025: \begin{thebibliography}{GJKW02}
7026:
7027: \bibitem[AFKM86]{adler}
7028: R.~Adler, J.~Friedman, B.~Kitchens, and B.~Marcus.
7029: \newblock State splitting for variable length graphs.
7030: \newblock {\em IEEE Transactions on Information Theory}, 32(1):108--115,
7031: January 1986.
7032:
7033: \bibitem[Alo86]{alon_eigenvalues}
7034: N.~Alon.
7035: \newblock Eigenvalues and expanders.
7036: \newblock {\em Combinatorica}, 6(2):83--96, 1986.
7037: \newblock Theory of computing (Singer Island, Fla., 1984).
7038:
7039: \bibitem[BS87]{broder}
7040: Andrei Broder and Eli Shamir.
7041: \newblock On the second eigenvalue of random regular graphs.
7042: \newblock In {\em 28th Annual Symposium on Foundations of Computer Science},
7043: pages 286--294, 1987.
7044:
7045: \bibitem[Buc86]{buck}
7046: Marshall~W. Buck.
7047: \newblock Expanders and diffusers.
7048: \newblock {\em SIAM J. Algebraic Discrete Methods}, 7(2):282--304, 1986.
7049:
7050: \bibitem[DD89]{dicks}
7051: Warren Dicks and M.~J. Dunwoody.
7052: \newblock {\em Groups acting on graphs}.
7053: \newblock Cambridge University Press, Cambridge, 1989.
7054:
7055: \bibitem[Dod84]{dodziuk}
7056: Jozef Dodziuk.
7057: \newblock Difference equations, isoperimetric inequality and transience of
7058: certain random walks.
7059: \newblock {\em Trans. Amer. Math. Soc.}, 284(2):787--794, 1984.
7060:
7061: \bibitem[FK81]{komlos}
7062: Z.~{F\"uredi} and J.~{Koml\'os}.
7063: \newblock The eigenvalues of random symmetric matrices.
7064: \newblock {\em Combinatorica}, 1(3):233--241, 1981.
7065:
7066: \bibitem[FKS89]{friedman_kahn_szemeredi}
7067: J.~Friedman, J.~Kahn, and E.~Szemer{\'e}di.
7068: \newblock On the second eigenvalue of random regular graphs.
7069: \newblock In {\em 21st Annual ACM Symposium on Theory of Computing}, pages
7070: 587--598, 1989.
7071:
7072: \bibitem[Fri]{friedman_SVD}
7073: Joel Friedman.
7074: \newblock {SVD} misalignment and eigenvalue separation.
7075: \newblock {I}n preparation.
7076:
7077: \bibitem[Fri91]{friedman_random_graphs}
7078: Joel Friedman.
7079: \newblock On the second eigenvalue and random walks in random $d$-regular
7080: graphs.
7081: \newblock {\em Combinatorica}, 11(4):331--362, 1991.
7082:
7083: \bibitem[Fri93]{friedman_geometric_aspects}
7084: Joel Friedman.
7085: \newblock Some geometric aspects of graphs and their eigenfunctions.
7086: \newblock {\em Duke Math. J.}, 69(3):487--525, 1993.
7087:
7088: \bibitem[Fri03]{friedman_relative}
7089: Joel Friedman.
7090: \newblock Relative expanders or weakly relatively {R}amanujan graphs.
7091: \newblock {\em Duke Math. J.}, 118(1):19--35, 2003.
7092:
7093: \bibitem[FT02]{friedman_tillich_generalized}
7094: Joel Friedman and Jean-Pierre Tillich.
7095: \newblock Generalized {A}lon-{B}oppana theorems and error-correcting codes.
7096: \newblock May 2002.
7097: \newblock preprint.
7098:
7099: \bibitem[FTP83]{figa-talamanca_picardello}
7100: Alessandro Fig{\`a}-Talamanca and Massimo~A. Picardello.
7101: \newblock {\em Harmonic analysis on free groups}.
7102: \newblock Marcel Dekker Inc., New York, 1983.
7103:
7104: \bibitem[Gem80]{geman}
7105: S.~Geman.
7106: \newblock A limit theorem for the norm of random matrices.
7107: \newblock {\em Ann. of Prob.}, 8(2):252--261, 1980.
7108:
7109: \bibitem[GJKW02]{greenhill}
7110: Catherine Greenhill, Svante Janson, Jeong~Han Kim, and Nicholas~C. Wormald.
7111: \newblock Permutation pseudographs and contiguity.
7112: \newblock {\em Combin. Probab. Comput.}, 11(3):273--298, 2002.
7113:
7114: \bibitem[God93]{godsil}
7115: C.~D. Godsil.
7116: \newblock {\em Algebraic combinatorics}.
7117: \newblock Chapman \& Hall, New York, 1993.
7118:
7119: \bibitem[HMS91]{heegard}
7120: Chris~D. Heegard, Brian~H. Marcus, and Paul~H. Siegel.
7121: \newblock Variable-length state splitting with applications to average
7122: runlength-constrained ({ARC}) codes.
7123: \newblock {\em IEEE Trans. Inform. Theory}, 37(3, part 2):759--777, 1991.
7124:
7125: \bibitem[Joh90]{johnson}
7126: D.~L. Johnson.
7127: \newblock {\em Presentations of groups}.
7128: \newblock Cambridge University Press, Cambridge, 1990.
7129:
7130: \bibitem[JS89]{jerrum2}
7131: Mark Jerrum and Alistair Sinclair.
7132: \newblock Approximating the permanent.
7133: \newblock {\em SIAM J. Comput.}, 18(6):1149--1178, 1989.
7134:
7135: \bibitem[Kit98]{kitchens}
7136: Bruce~P. Kitchens.
7137: \newblock {\em Symbolic dynamics}.
7138: \newblock Springer-Verlag, Berlin, 1998.
7139:
7140: \bibitem[KW01]{kim}
7141: Jeong~Han Kim and Nicholas~C. Wormald.
7142: \newblock Random matchings which induce {H}amilton cycles and {H}amiltonian
7143: decompositions of random regular graphs.
7144: \newblock {\em J. Combin. Theory Ser. B}, 81(1):20--44, 2001.
7145:
7146: \bibitem[LM95]{marcus}
7147: Douglas Lind and Brian Marcus.
7148: \newblock {\em An introduction to symbolic dynamics and coding}.
7149: \newblock Cambridge University Press, Cambridge, 1995.
7150:
7151: \bibitem[LPS86]{lubotzky}
7152: A.~Lubotzky, R.~Phillips, and P.~Sarnak.
7153: \newblock Explicit expanders and the ramanujan conjectures.
7154: \newblock In {\em 18th Annual ACM Symposium on Theory of Computing}, pages
7155: 240--246, 1986.
7156:
7157: \bibitem[LPS88]{lubotzky_phillips_sarnak}
7158: A.~Lubotzky, R.~Phillips, and P.~Sarnak.
7159: \newblock Ramanujan graphs.
7160: \newblock {\em Combinatorica}, 8(3):261--277, 1988.
7161:
7162: \bibitem[Mar88]{margulis}
7163: G.~A. Margulis.
7164: \newblock Explicit group-theoretic constructions of combinatorial schemes and
7165: their applications in the construction of expanders and concentrators.
7166: \newblock {\em Problemy Peredachi Informatsii}, 24(1):51--60, 1988.
7167:
7168: \bibitem[McK81]{mckay}
7169: B.~McKay.
7170: \newblock The expected eigenvalue distribution of a large regular graph.
7171: \newblock {\em Lin. Alg. Appl.}, 40:203--216, 1981.
7172:
7173: \bibitem[Mor94]{morgenstern}
7174: Moshe Morgenstern.
7175: \newblock Existence and explicit constructions of $q+1$ regular {R}amanujan
7176: graphs for every prime power $q$.
7177: \newblock {\em J. Combin. Theory Ser. B}, 62(1):44--62, 1994.
7178:
7179: \bibitem[Nil91]{nilli}
7180: A.~Nilli.
7181: \newblock On the second eigenvalue of a graph.
7182: \newblock {\em Discrete Math.}, 91(2):207--210, 1991.
7183:
7184: \bibitem[Sen81]{seneta}
7185: E.~Seneta.
7186: \newblock {\em Nonnegative matrices and {M}arkov chains}.
7187: \newblock Springer Series in Statistics. Springer-Verlag, New York, second
7188: edition, 1981.
7189:
7190: \bibitem[SJ89]{jerrum1}
7191: Alistair Sinclair and Mark Jerrum.
7192: \newblock Approximate counting, uniform generation and rapidly mixing {M}arkov
7193: chains.
7194: \newblock {\em Inform. and Comput.}, 82(1):93--133, 1989.
7195:
7196: \bibitem[SW49]{shannon}
7197: Claude~E. Shannon and Warren Weaver.
7198: \newblock {\em The Mathematical Theory of Communication}.
7199: \newblock University of Illinois Press, 1971/1949.
7200:
7201: \bibitem[Wig55]{wigner}
7202: E.~Wigner.
7203: \newblock Characteristic vectors of bordered matrices with infinite dimensions.
7204: \newblock {\em Annals of Math.}, 63(3):548--564, 1955.
7205:
7206: \bibitem[Woe00]{woess}
7207: Wolfgang Woess.
7208: \newblock {\em Random walks on infinite graphs and groups}.
7209: \newblock Cambridge University Press, Cambridge, 2000.
7210:
7211: \bibitem[Wor99]{wormald}
7212: N.~C. Wormald.
7213: \newblock Models of random regular graphs.
7214: \newblock In {\em Surveys in combinatorics, 1999 (Canterbury)}, pages 239--298.
7215: Cambridge Univ. Press, Cambridge, 1999.
7216:
7217: \end{thebibliography}
7218:
7219: \end{document}
7220:
7221: