1: \documentclass[10pt,a4paper,reqno]{amsart} 

2: %\pagenumbering{arabic}

3: %\usepackage{showkeys}

4: 

5: %\usepackage{fullpage}

6: \usepackage{latexsym,floatflt}

7: \usepackage{epsfig}

8: \usepackage{rotating}

9: \usepackage{amssymb}

10: \usepackage[T1]{fontenc}% pour les accents français

11: \usepackage{afterpage}

12: \usepackage{color}

13: 

14: \usepackage{colordvi}

15: 

16: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

17: %     New style for sections and subsections

18: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

19: \catcode`\@=11

20: \def\section{\@startsection{section}{1}%

21:  \z@{.7\linespacing\@plus\linespacing}{.5\linespacing}%

22:  {\normalfont\bfseries\scshape\centering}}

23: %\def\section{\@startsection{section}{1}%

24: %  \z@{.7\linespacing\@plus\linespacing}{-.5em}%

25: %  {\normalfont\bfseries}}

26: 

27: %\def\subsection{\@startsection{subsection}{2}%

28: %  \z@{.5\linespacing\@plus.7\linespacing}{-.5em}%

29: %  {\normalfont\scshape}}

30: \def\subsection{\@startsection{subsection}{2}%

31:   \z@{.5\linespacing\@plus\linespacing}{.5\linespacing}%

32:   {\normalfont\bfseries\scshape}}

33: 

34: 

35: \def\subsubsection{\@startsection{subsubsection}{3}%

36: %  \z@{.5\linespacing\@plus.7\linespacing}{-.5em}%

37:  \z@{.5\linespacing\@plus\linespacing}{-.5em}%{.5\linespacing}%

38:   {\normalfont\bfseries\itshape}}

39: \catcode`\@=12

40: 

41: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

42: \addtolength{\textheight}{-4mm} \topmargin6mm

43: \addtolength{\textwidth}{14mm} 

44: \hoffset -8mm

45: %\voffset -6mm

46: 

47: \newcommand\mps[1]{\marginpar{\small\sf#1}}

48: 

49: \newcommand{\epsf}[1]{\epsfbox{#1}}

50: \def\cqfd{$\hfill{\vrule height 3pt width 5pt depth 2pt}$}

51: \def\qee{$\hfill{\Box}$}

52: 

53: \newcommand{\ov}{\overline}

54: \newcommand{\un}{\underline}

55: \newcommand{\bwx}{\phi}

56: \newcommand{\BWX}{\Phi}

57: \newcommand{\f}{{\hbox{\small {fr}}}}

58: \newcommand{\GK}{\mathbb{K}}

59: \newcommand{\GL}{\mathbb{L}}

60: \newcommand{\ns}{\mathbb{N}}

61: \newcommand{\zs}{\mathbb{Z}}

62: 

63: \newcommand{\qs}{\mathbb{Q}}

64: \newcommand{\rs}{\mathbb{R}}

65: \newcommand{\bx}{\bar x}

66: \newcommand{\by}{\bar y}

67: \newcommand{\bv}{\bar v}

68: \newcommand{\bz}{\bar z}

69: \newcommand{\bu}{\bar u}

70: \newcommand{\cs}{\mathbb{C}}

71: \newcommand{\cf}{\mathbb{F}}

72: 

73: \newcommand{\bm}[1]{\mbox{\boldmath \ensuremath{#1}}}

74: \newcommand{\bs}[1]{\mbox{\boldmath \ensuremath{\scriptstyle #1}}}

75: 

76: \newcommand{\GE}{\mathbb{E}}

77: \newcommand{\R}{\mathcal R}

78: \newcommand{\A}{\mathcal A}

79: \newcommand{\I}{\mathcal I}

80: \newcommand{\D}{\mathcal D}

81: \newcommand{\E}{\mathcal E}

82: \newcommand{\C}{\mathcal C}

83: \newcommand{\T}{\mathcal T}

84: \newcommand{\J}{\mathcal J}

85: \renewcommand{\S}{\mathcal S}

86: \newcommand{\Sn}{\ensuremath{\mathcal S}} 

87: 

88: \newcommand{\Ha}{\mathcal H}

89: \newcommand{\GA}{\mathbb A}

90: \newcommand{\GB}{\mathbb B}

91: \newcommand{\EE}{\mathbb{E}}

92: \newcommand{\PP}{\mathbb{P}}

93: 

94: \newcommand{\Ref}[1]{(\ref{#1})}

95: \newcommand{\beq}{\begin{equation}}

96: \newcommand{\eeq}{\end{equation}}

97: \newcommand{\gf}{generating function}

98: \newcommand{\gfs}{generating functions}

99: \newcommand{\al}{\alpha}

100: \newcommand{\be}{\beta}

101: \newcommand{\cd}{\cdot}

102: \newcommand{\eps}{\epsilon}

103: \newcommand{\bin}[2]{\genfrac{(}{)}{0pt}{}{#1}{#2}}

104: 

105: \newcommand{\fps}{formal power series}

106: \def\cqfd{\par\nopagebreak\rightline{\vrule height 4pt width 5pt depth 1pt}

107: \medbreak}

108: \def\emm#1,{{\em #1}}

109: \newcommand{\deq}{\stackrel{\mathrm{d}}{=}}

110: \newcommand{\dconv}{\stackrel{\mathrm{d}}{\rightarrow}}

111: \font\dsrom=dsrom10 scaled 1400

112: \def \1{\textrm{\dsrom{1}}}

113: 

114: \tolerance=10000

115: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

116: 

117: \newtheorem{Theorem}{Theorem}

118: \newtheorem{Propo}[Theorem]{Proposition}

119: \newtheorem{Coro}[Theorem]{Corollary}

120: \newtheorem{Lemma}[Theorem]{Lemma}

121: \newtheorem{Definition}[Theorem]{Definition}

122: \newtheorem{Conj}[Theorem]{Conjecture}

123: \newtheorem{Example}{Example}

124: \begin{document}

125: 

126: 

127: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

128: \title[Limit laws for embedded trees]{Limit laws for embedded trees.\\

129:  Applications to the integrated

130:   superBrownian excursion} 

131: 

132: 

133: 

134: 

135: 

136: \author{Mireille Bousquet-M\'elou}

137: \address{CNRS, LaBRI, Universit\'e Bordeaux 1, 351 cours de la Lib\'eration,

138:   33405 Talence Cedex, France}

139: \email{mireille.bousquet@labri.fr}

140: \thanks{MBM  was partially supported by the European Commission's IHRP

141:   Programme, grant HPRN-CT-2001-00272, ``Algebraic Combinatorics in

142:   Europe''} 

143: \keywords{}

144: %\subjclass{Primary: subject; Secondary: subject}

145: \date{January 17, 2005}

146: 

147: 

148: 

149: 

150: 

151: \begin{abstract}

152: We study three families of labelled  plane

153: trees. In all these trees, the root is labelled $0$, and the labels of

154: two adjacent nodes differ by $0, 1$ or $-1$. 

155: 

156: One part of the paper is devoted to enumerative results. For each

157: family, and for all $j\in \ns$, we obtain closed form expressions for the

158: following three generating functions: the generating function of trees having no label larger than

159: $j$; the (bivariate) generating function  of trees, counted by the  number of

160: edges and the number of nodes labelled $j$; and finally the

161: (bivariate) generating function  of trees, counted by the  number of edges

162:  and the number of nodes labelled \emm at least, $j$. Strangely

163: enough, all these series turn out to be algebraic, but we have no

164: combinatorial intuition for this algebraicity.

165: 

166: The other part of the paper is devoted to deriving limit laws from

167:  these enumerative results. In each of our families of

168: trees, we endow the trees of size $n$ with the uniform

169: distribution, and study the following random variables: $M_n$, the

170: largest label occurring in a (random) tree; $X_n(j)$, the number of

171: nodes labelled $j$; and $X_n^+(j)$, the number of nodes labelled $j$

172: or more. We obtain limit laws for scaled versions of these random

173: variables.

174: 

175: Finally, we translate the above limit results into statements dealing

176: with  the

177: integrated superBrownian excursion (ISE). In particular, we describe

178: the law of the supremum of its support (thus recovering some earlier

179: results obtained by Delmas), and the law of its

180: distribution function at a given point. We also conjecture the law of

181: its density (at a given point).

182: \end{abstract}

183: 

184: \maketitle

185: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

186: \section{Introduction}

187: \label{section-intro}

188: We study in this paper three families of labelled  plane

189: trees.  In all these trees, the root is labelled $0$, and the labels of

190: two adjacent nodes differ by $0, 1$ or $-1$. 

191: 

192:  More precisely, the first family we consider is the set of plane

193:  trees, and the increments of 

194: the labels along edges are constrained to be $\pm 1$. In the closely

195: related second family, these increments can be $0, \pm 1$. The third

196: family is a bit different. It is simply the set of (incomplete) binary

197: trees, in 

198: which the nodes are labelled in a deterministic way: the label

199: of a node is the difference between the number of right steps and the

200: number of left steps occurring in the path that yields from the root

201:  to the node under consideration. See

202: Figure~\ref{fig-three-families} for an illustration. We call this

203: labelling the \emm natural labelling, of the binary

204: tree. Note that the label of each node is simply its abscissa, if we

205: draw the tree in the plane in such a way the right (resp.~left) son of

206: a node lies one unit to the right (resp.~left) of its father. For this

207: reason, we will sometimes call these labelled binary trees \emm

208: naturally embedded binary trees., More generally, for \emm any,\/

209: plane labelled tree,  we may consider that the label of each node tells

210: where to embed it  in  $\zs$; hence the title of the paper.

211: 

212: \begin{figure}[hbt]

213: \begin{center}

214: \input{tree-examples.pstex_t}

215: \end{center}

216: \caption{A labelled plane tree with increments $\pm1$. --- A labelled tree

217:   with increments $0, \pm 1$. --- A naturally embedded binary tree.}

218: \label{fig-three-families}

219: %\hrule

220: \end{figure}

221: 

222: In each of these three families, we endow the set of trees having a

223: given size (say, $n$ edges) with the uniform distribution.  We address

224: (via \gfs) 

225: the following three questions:

226: \begin{enumerate}

227: \item

228: What is the maximal label that occurs in the tree? This label is in

229: fact a random variable $M_n$. We prove that $M_n/n^{1/4}$ converges in

230: distribution to a random variable $N$ having a density. We give this

231: density explicitly. We also compute the moments of $N$ and prove the

232: convergence of the moments of $M_n/n^{1/4}$  

233:  to those  of $N$.

234: \item

235: How many nodes of the tree have label $j$? Let $X_n(j)$ denote the

236: corresponding random variable. If $j$ is fixed, and $n$ goes to

237: infinity, then the answer to this question is independent of $j$. We

238: prove that for any $j \in \zs$, the variable  $X_n(j)/n^{3/4}$

239: converges in distribution to  $cT^{-1/2}$, where 

240: $c$ is a constant depending on which family of trees we consider, and

241: $T$ follows a unilateral stable law of parameter $2/3$. 

242: 

243: Given that the maximal label grows like $n^{1/4}$, we get a better

244: insight on the label distribution by asking how many nodes in a tree

245: of size $n$ have label $\lfloor  \lambda n^{1/4} \rfloor$. We prove

246: that, for any $\lambda \in \rs$, the random variable $X_n(\lfloor

247: \lambda n^{1/4} \rfloor)/n^{3/4}$ converges in distribution to a limit

248: variable $Y(\lambda)$. This variable  admits a Laplace transform,

249: which we give explicitly. The convergence of the Laplace transform, and

250: of the moments,  hold as well. We say we have obtained a \emm

251: local limit law, for embedded trees, because we look at \emm one, value of

252: the labels only.

253: \item

254: Finally, we also obtain a \emm global limit law, by studying the

255: variable  $X_n^+(j)$ that gives the number of nodes having label $j$

256: at least. Remarkably, we prove that $X_n^+(0)/n$, the (normalized)

257: number of nodes having a non-negative label, converges to the uniform

258: distribution on $[0,1]$. More generally, for $\lambda \in \rs$, the

259: variable $X_n^+(\lambda n^{1/4})/n$ converges in distribution to a

260: variable $Y^+(\lambda)$.  This variable  admits a Laplace transform,

261: which we give explicitly. Once again, the convergence of the Laplace

262: transform, and of the moments,  hold as well. 

263: \end{enumerate}

264: The laws of $N$, $Y(\lambda)$ and $Y^+(\lambda)$ naturally depend on

265: which family of trees we consider, but only by a simple normalization

266: factor. 

267: 

268: \subsection{Embedded trees and the integrated superBrownian excursion}

269: Why should one study such labelled trees? 

270: 

271: The first two classes of

272: trees we consider have a close connection with certain families of

273: planar

274: maps~\cite{bdg-statistics,chassaing-schaeffer,cori-vauquelin}. In

275: particular, the 

276: diameter of a random quadrangulation having $n$ faces is distributed

277: like the largest 

278: label in \emm non-negative, random trees of our second

279: family. Moreover, once scaled by $n^{1/4}$, this diameter has the same limit law

280: as $(M_n-m_n)n^{-1/4}$, where $M_n$ (resp.~$m_n$) is the largest

281: (resp.~smallest)  label occurring in

282: a random tree of our second family~\cite{chassaing-schaeffer}. 

283: 

284: 

285: The third class we study

286: is the good old family of binary trees, and this may suffice to motivate

287: its study! More seriously, the three questions addressed above

288: have, for binary trees, a natural geometric formulation. 

289: The random variable $M_n$ (the 

290: maximum label) tells us about the ``true width'' of a binary tree (as

291: opposed to the maximal number of nodes lying at the same

292: level, which is known to grow like $\sqrt n$). More generally, the

293: variables $X_n(j)$ 

294: tell us about the \emm vertical 

295: profile, of the tree (as opposed to the \emm horizontal profile, which

296: describes the repartition of nodes by

297: level~\cite{drmota-profile}). See Figure~\ref{fig-profils}. 

298: 

299: \begin{figure}[t]

300: \begin{center}

301: \input{profils.pstex_t}

302: \end{center}

303: \caption{An (incomplete) binary tree having horizontal profile

304:   $[1,2,4,3,2]$ and   vertical profile $[2,2,4; 2,1,1]$.}

305: \label{fig-profils}

306: %\hrule

307: \end{figure}

308: 

309: We may also invoke an \emm a posteriori,\/ justification to the study

310: of these trees: the form of the \gfs\ we obtain is  remarkable,

311: whatever family of trees we consider, and suggests that there must be

312: some beautiful hidden combinatorics in these problems, which should be

313: explored further.

314: 

315: 

316: However, the main motivation for this work is the connection between

317: embedded trees and the \emm integrated superBrownian excursion,

318: (ISE). Choose one of the three families of trees, and consider the

319: following \emm random, probability distribution on $\rs$:

320: \beq

321: \label{random-measure}

322: \mu_n = \frac 1 {n+1} \sum_{j\in \zs} X_n(j) \delta_{cjn^{-1/4}},

323: \eeq

324: where $X_n(j)$ is the (random) number of nodes labelled $j$,

325: $\delta_x$ denotes the Dirac measure at $x$, and the constant $c$ 

326: equals $\sqrt 2$ for the first family, $\sqrt 3$ for the second one and

327: $1$ for the family of binary trees. Then $\mu_n$ is

328: known to converge weakly

329: to a limiting random probability distribution called the

330: ISE~\cite{aldous,marckert-mokka-snake,jf-homothetie,jf-janson}. See

331: Figure~\ref{fig-simulations} for simulations of $\mu_n$.

332: 

333: \begin{figure}[b]

334: \begin{center}

335: \includegraphics[height=3cm,width=3cm]{profil-v1.eps}

336: \hskip 2mm

337: \includegraphics[height=3cm,width=3cm]{profil-v2.eps}

338: \hskip 2mm

339: \includegraphics[height=3cm,width=3cm]{profil-v3.eps}

340: \hskip 2mm

341: \includegraphics[height=3cm,width=3cm]{profil-v4.eps}

342: 

343: \includegraphics[height=3cm,width=3cm]{profil-v5.eps}

344: \hskip 2mm

345: \includegraphics[height=3cm,width=3cm]{profil-v6.eps}

346: \hskip 2mm

347: \includegraphics[height=3cm,width=3cm]{profil-v7.eps}

348: \hskip 2mm

349: \includegraphics[height=3cm,width=3cm]{profil-v8.eps}

350: \end{center}

351: \caption{The plot of $X_n(j)$ vs. $j$ for random binary trees with

352:   $n=1000$ nodes.}

353: \label{fig-simulations}

354: \end{figure}

355: 

356: Our limit results provide some information  about the law of the

357: ISE. For instance, we prove  that $cM_nn^{-1/4}$,  the largest point

358: having a positive weight under $\mu_n$, converges in law to $N_{\hbox{\sc ise}}$, the

359: supremum of the support of the ISE. We denote this by

360: $$

361: cM_nn^{-1/4}\dconv N_{\hbox{\sc ise}}.

362: $$

363:  The results we obtain for the limit law of $M_nn^{-1/4}$ thus

364: translate into expressions of the moments, distribution function and

365: density of the supremum of the ISE. Note that the moments were already

366: obtained by Delmas~\cite{delmas}.

367: Our second limit result deals with the random variables

368: $X_n(\lfloor  \lambda n^{1/4} \rfloor)$. Observe that

369: \beq

370: \label{limit-density}

371: \mu_n(c\lambda-cn^{-1/4}, c\lambda]= \frac 1 {n+1} X_n(\lfloor  \lambda n^{1/4} \rfloor) .

372: \eeq

373:  This leads us to \emm conjecture,\/

374: that the random variable $Y(\lambda)$ involved in our local

375: limit law 

376: %(it describes the limit law of  $n^{-3/4} X_n(j)$) 

377: satisfies

378: \beq

379: \label{conv-densite}

380: Y(\lambda) \deq cf_{\hbox{\sc \small ise}}(c\lambda)

381:  \eeq

382: where $f_{\hbox{\sc \small ise}}$ is the (random) density of

383: the ISE. Similarly, 

384: $$

385: \mu_n[c\lambda, +\infty)= \frac 1 {n+1} X_n^+(\lceil  \lambda n^{1/4} \rceil), 

386: $$

387:  and we \emm prove,\/ that  the

388: random variable $Y^+(\lambda)$ involved in our global limit law

389: satisfies

390: $$

391: Y^+(\lambda) \deq g_{\hbox{\sc \small ise}}(c\lambda)

392: $$

393: where $g_{\hbox{\sc \small ise}}$ is the

394: (random) tail distribution function of the ISE. The results we obtain

395: about the laws of $Y(\lambda)$ and $Y^+(\lambda)$ thus translate into

396: formulas for the Laplace transforms of $f_{\hbox{\sc \small

397:     ise}}(\lambda)$ and  $g_{\hbox{\sc \small ise}}(\lambda)$ (the

398: formula for $f_{\hbox{\sc \small

399:     ise}}(\lambda)$ being conjectural).

400: 

401: Our conjecture on $f_{\hbox{\sc \small

402:     ise}}$  is naturally supported

403:   by the fact that the law of $Y(\lambda/c)/c$ is 

404:   independent of the tree family we start from. This is one of the

405:   reasons why we consider as many as three families of trees. The

406:   other reasons 

407:   involve the connections with planar maps, the remarkable form of the

408:   \gfs\ we obtain, and our unshakeable

409:   interest in binary trees.  The details of the calculations

410:   are only given for the first of the three families

411:   (Sections~\ref{section-snake-gf} 

412:   to~\ref{section-limit-global}), while  the results are merely stated

413:   for the other   two families   (Section~\ref{section-universality}).

414: 

415: Let us finally mention that the moments of the \emm center of mass,\/

416: of the ISE have recently been determined by two different

417: approaches~\cite{chassaing-janson,janson}. In our discrete setting,

418: this boils down to studying the convergence of the variable

419: $$

420: \frac 1{n^{5/4} }\sum _{j\in \zs} jX_n(j).

421: $$

422: 

423: \subsection{Overview of the paper}

424: The starting point of our approach is a series of exact enumerative

425: results dealing with our first class of trees: plane trees in which

426: the labels of  adjacent nodes differ by $\pm 1$. These

427: results are gathered in the next section. We obtain for 

428: instance an explicit expression for the bivariate \gf \ of labelled trees,

429: counted by the number of edges and the number of nodes labelled $j$

430: (for $j$ fixed). This section includes, and owes a lot to, some results

431: recently obtained by Bouttier, Di Francesco and

432: Guitter~\cite{bdg-geodesic,bdg-statistics} on the 

433: enumeration of  trees having no label greater than $j$. This

434: part of our work raises a number of challenging combinatorial

435: questions --- why are these expressions so simple? --- which are not

436: addressed in this paper. 

437: 

438: The limit behaviours of the random variables $M_n$, $X_n(\lfloor

439: \lambda n^{1/4}\rfloor )$ and

440: $X_n^+(\lambda n^{1/4})$ are respectively established in the next three

441: sections (Sections~\ref{section-max} 

442:   to~\ref{section-limit-global}). The main technique that we use is

443:   the ``analysis of singularities'' of Flajolet and

444: Odlyzko~\cite{flajolet-odlyzko}. It permits to extract 

445: the asymptotic behaviour of the coefficients of a \gf . This technique

446: has already proved useful in numerous occasions, in particular for

447: proving limit theorems that are similar in flavour to the ones

448: obtained in this paper: these theorems deal with the height of simply

449: generated trees and their profile, which are known to be related to the 

450: height of the Brownian excursion and its local

451: time~\cite{flajolet-binary-trees,drmota-profile}. This 

452: technique is carefully exemplified in Section~\ref{section-max} (which

453: is devoted to the maximal label) before the more difficult questions

454: of the local and global limit laws are attacked

455: (Sections~\ref{section-limit-local}

456: and~\ref{section-limit-global}). 

457: 

458: Finally, two other families of trees are briefly studied in

459: Section~\ref{section-universality}: trees with increments $0, \pm 1$

460: and naturally embedded binary trees.  The emphasis is put on their

461: enumerative properties, which turn out to be as remarkable and

462: surprising as those of our first family of trees. The limit laws we

463: obtain are (up to a scalar) the same as for the first family.

464: 

465: 

466: \bigskip

467: Let us conclude with some notation and a few definitions on \fps \ and

468: generating functions. 

469: Let $\GK$ be a field. We denote by $\GK[t]$ the ring of polynomials in

470: $t$ with coefficients in $\GK$, and by $\GK(t)$ the field of rational

471: functions in $t$ with coefficients in $\GK$. We denote by $\GK[[t]]$

472: the ring of \fps\ in $t$ 

473: with coefficients in $\GK$. If $A(t) \in \GK[[t]]$ and  $n\in\ns$,

474: the notation $[t^n]A(t)$ stands for the coefficient of $t^n$ in

475: $A(t)$.

476: The series $A(t)$ is said to be \emm algebraic over $\GK(t)$, if it satisfies a

477: non-trivial 

478: polynomial equation of the form $P(t,A(t))=0$, where $P$ is a

479: bivariate polynomial with coefficients in $\GK$. In this case, the

480: \emm degree,\/ of $A(t)$ 

481: is the smallest possible degree of $P$ (in its second variable).

482: 

483: Let $\A$ be a set of

484: discrete objects, equipped with a \emm size, that takes 

485: nonnegative integer values.  Assume

486: that for all $n \in \ns$, the number of objects of $\A$ of size $n$ is

487: finite, and denote this number by $a_n$. The {\em \gf\ of the objects of

488: $\A$, counted by their size,}\/ is the formal power series

489: $$

490: A(t)=\sum_{n\ge 0} a_n t^n.

491: $$

492: The above notions generalize in a straightforward way to multivariate

493: power series. Such series  arise naturally when enumerating objects

494: according to several parameters.

495: 

496: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

497: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

498: \section{Enumerative results}

499: \label{section-snake-gf}

500: We consider in this section (and  in the three following ones)

501: our first family of labelled plane trees: the root is labelled $0$,

502: and the labels of two adjacent nodes differ by $\pm 1$. 

503: 

504: \subsection{Trees with small labels}

505: The first enumerative problem we address has already been studied by

506: Bouttier, Di Francesco and

507: Guitter~\cite{bdg-geodesic,bdg-statistics}. It deals with the largest

508: label occurring in a tree.

509: For $j\in \ns$, let  $T_j\equiv T_j(t)$ be the \gf \ of labelled 

510: trees in which all labels are less than or equal to $j$. The

511: indeterminate $t$ keeps track of the number of edges. Let  $T\equiv

512: T(t)$ be the \gf \ of all labelled trees. Clearly, $T_j$ converges to

513: $T$ (in the space of \fps \ in $t$) as $j$ goes to infinity. It is

514: very easy to describe an infinite set of equations that completely

515: defines the collection of series $T_j$.

516: 

517: %

518: \begin{Lemma} The series $T$ satisfies

519: \label{lemma-Tj}

520: \beq

521: \label{eqT}

522: T=1+2tT^2.

523: \eeq

524: More generally, for $j\ge 0$,

525: %\beq\label{eqTj}

526: $$

527: T_j=1+t(T_{j-1}+T_{j+1})T_j

528: %\eeq

529: $$

530: while $T_j=0$ for $j<0$.

531: \end{Lemma}

532: \noindent

533: {\bf Proof.} The two ingredients of the proof will be useful for the 

534: other enumerative problems we  address below.

535: Firstly, replacing each label $k$ by $j-k$ shows that $T_j$ is also the

536: \gf\ of trees 

537: \emm rooted at $j$, and having only non-negative labels (we say that a tree

538: is rooted at $j$ if its root has label $j$).

539: Secondly, consider such a tree and assume it is not reduced to a

540: single node. The root has a leftmost child, which  is the root of a 

541: labelled subtree, rooted at $j\pm 1$ and having only non-negative labels.

542:    Deleting this subtree leaves a smaller tree rooted at

543: $j$, having only non-negative labels (see Figure~\ref{fig-decomp}). The

544:    result follows.

545: \cqfd

546: 

547: 

548: \begin{figure}[hbt]

549: \begin{center}

550: \input{tree-decomp.pstex_t}

551: \end{center}

552: \caption{The decomposition of plane labelled trees.}

553: \label{fig-decomp}

554: %\hrule

555: \end{figure}

556: 

557: 

558: The above lemma shows that the series $T$, counting labelled trees

559: by edges, is algebraic, and the short proof we have given provides a simple

560: combinatorial explanation for this property. What  is far less clear

561: --- but nevertheless true --- is that 

562: each of the series $T_j$ is algebraic too, as stated in the

563: proposition below, which we borrow

564: from~\cite{bdg-geodesic,bdg-statistics}. These series will be 

565: expressed in terms of the 

566: series $T\equiv T(t)$ and of the  unique formal power series

567: $Z\equiv Z(t)$, with constant term $0$, satisfying

568: \beq

569: \label{eqZt}

570: Z=t\, \frac{(1+Z)^4}{1+Z^2}.

571: \eeq

572: Observe that $T$ and $Z$ are related by:

573: \beq

574: \label{eqTZ}

575: T=\frac{(1+Z)^2}{1+Z^2}.

576: \eeq

577: 

578: 

579: \begin{Propo}[{\bf Trees with small labels \cite{bdg-geodesic,bdg-statistics}}]

580: \label{propo-snake-bounded}

581: Let $T_j\equiv T_j(t)$ be the \gf\ of trees having no label greater

582: than $j$. Then $T_j$ is algebraic of degree (at most) $2$. In particular,

583: $$

584: T_0=1-11\,t-{t}^{2}

585: +4\,t \left( 3+2\,t \right) {T_0}

586: -16\,{t}^{2}{{ T_0}}^{2}.

587: $$

588: Moreover,  for all $j \ge -1$,

589: \beq

590: \label{product-formTj-small}

591: T_j= T\,\frac{(1-Z^{j+1})(1- Z^{j+5})}{(1- Z^{j+2})(1- Z^{j+4})},

592: \eeq

593: where $Z\equiv Z(t)$ is given by~\eqref{eqZt}.

594: \end{Propo}

595: \noindent

596: {\bf Proof.} It is very easy to check, using~(\ref{eqZt}--\ref{eqTZ}),

597: that the above values of $T_j$ satisfy the recurrence relation of

598: Lemma~\ref{lemma-Tj} and the initial condition $T_{-1}=0$.

599: How to \emm discover, such a formula is another story, which is told

600: in~\cite{bdg-geodesic}. 

601: The remarkable product form of $T_j$ still awaits a 

602: combinatorial explanation. 

603: 

604: The equation satisfied by $T_0$ is obtained by eliminating $T$ and $Z$

605: from the case $j=0$ of~\eqref{product-formTj-small}. Then an induction

606: of $j$, based on Lemma~\ref{lemma-Tj}, implies that each $T_j$ is

607: quadratic (at most) over $\qs(t)$. \cqfd

608: 

609: 

610: \noindent

611: {\bf Remarks}\\

612: {\bf 1.} The product form~\eqref{product-formTj-small}, combined with

613: the facts that $T$ is quadratic over $\qs(t)$ and $Z$ is 

614: quadratic over $\qs(T)$, shows that $T_j$ belongs to an extension of

615: $\qs(t)$ of degree $4$. This is true,  but not optimal, since $T_j$ is

616: actually quadratic over $\qs(t)$. Hence this product form does not

617: give the best possible information on the degree of $T_j$.

618: \\

619: {\bf 2.} The trees counted by $T_0$ (equivalently, the trees having only

620: non-negative labels) are known to be in bijection with certain planar

621: maps called \emm Eulerian triangulations,~\cite{bdg-statistics}. Through

622: this bijection, the number of edges of the tree is sent to the number

623: of black faces of the triangulation. These triangulations are nothing

624: but the dual maps of the \emm bicubic, (that is, bipartite and

625: trivalent) maps, which were first enumerated by Tutte~\cite{census-maps}.

626: In particular, the

627: coefficients of $T_0(t)$ are remarkably simple:

628: $$

629: T_0(t)= 

630: \frac{(1-8t)^{3/2}-1+12t+8t^2}{32t^2}

631: =1+\sum_{n\ge 1} \frac{3. 2^{n-1}}{(n+1)(n+2)}{{2n}\choose n}t^n.

632: $$

633: 

634: 

635: \subsection{The number of nodes labelled $j$}

636: Let us now turn our attention to a bivariate counting problem.

637: For $j \in \zs$, let $S_j \equiv S_j(t,u)$ be the \gf\ of labelled trees,

638: counted by the number of edges (variable $t$) and the number of

639: nodes labelled $j$ (variable $u$).  Clearly, $S_j(t,1)=T(t)$ for

640: all $j$. Moreover, an obvious symmetry entails that $S_j=S_{-j}$. 

641: \begin{Lemma} For $j\not = 0$,

642: \label{lemma-Sj-rec}

643: \beq

644: \label{eqSj}

645: S_j=1+t(S_{j-1}+S_{j+1})S_j

646: \eeq

647: while for $j=0$,

648: \beq

649: \label{eqS0}

650: S_0= u + t(S_{-1}+S_{1})S_0 =  u +2 tS_{1}S_0.

651: \eeq

652: \end{Lemma}

653: %

654: \noindent

655: {\bf Proof.}

656: Observe that $S_j \equiv S_j(t,u)$ is also the \gf\ of labelled trees

657: rooted at $j$,

658: counted by the number of edges and the number of

659: nodes labelled $0$. The  decomposition of trees illustrated in Figure~\ref{fig-decomp} then provides the  lemma. The only difference between the cases

660:    $j=0$ and $j\not =0$ lies in the \gf\ of the tree reduced to a single node.

661: \cqfd

662: 

663: Again, the series $S_j(t,u)$ turn out to be algebraic, for reasons

664: that currently remain mysterious (from the combinatorics

665: viewpoint). They can be expressed in terms of the 

666: series $T$ and $Z$ given by~(\ref{eqZt}--\ref{eqTZ}).

667: %

668: \begin{Propo}[{\bf The number of nodes labelled $j$}]

669: \label{propo-snake-gf}

670: For any $j\in \zs$, the \gf \ $S_j\equiv S_j(t,u)$ that counts

671: labelled trees by the number of edges and the number of nodes labelled $j$ is

672: algebraic over $\qs(T,u)$ of degree at most $3$ (and hence has degree

673: at most $6$ over $\qs(t,u)$). More precisely,

674: \beq

675: \label{eqS0T}

676: %

677: \frac{(T-S_0)^2}{(u-1)^2}= 1- \frac{2(1-T^2)}{2+S_0-S_0T},

678: %

679: %{S_0}^{3} \left( T-1 \right) -2\,{S_0}^{2} \left( {T}^{2}-T+1 \right) 

680: %+ S_0 \left((u-1)^2(1-T)+T^3-T^2+4T\right)

681: %+2\,u{T}^{2} \left(u -2 \right)=0.

682: \eeq

683: and all the $S_j$ belong to $\qs(t,u,S_0)$.

684: Moreover,  for all $j \ge 0$,

685: \beq

686: \label{product-formSj}

687: S_j= T\,\frac{(1+\mu Z^j)(1+\mu Z^{j+4})}{(1+\mu Z^{j+1})(1+\mu Z^{j+3})},

688: \eeq

689: where $Z\equiv Z(t)$ is given by~\eqref{eqZt}

690: and $\mu\equiv \mu(t,u)$ is the unique \fps\ in $t$

691: satisfying

692: \beq

693: \label{eqMZ}

694: \mu= (u-1) \frac{(1+Z^2)(1+\mu Z)(1+\mu Z^2)(1+\mu Z^3)}

695: {(1+Z)(1+Z+Z^2)(1-Z)^3(1-\mu Z^2)}.

696: \eeq

697: The series $\mu(t,u)$  has polynomial coefficients in $u$,

698: and satisfies $\mu(t,1)=0$. It has degree $3$ over $\qs(Z,u)$ and

699: $12$ over $\qs(t,u)$. 

700: %The first terms in its expansion are:

701: %$$

702: %\mu(t,u)=(u-1)\Big(

703: %1+ut+u \left( u+6 \right) {t}^{2}+u \left( {u}^{2}+14\,u+34 \right)

704: %{t}^{3}+O(t^4)\Big) .

705: % u \left( {u}^{3}+24\,{u}^{2}+127\,u+200 \right) {t}^{4}\Big)+O(t^5).

706: %$$

707: \end{Propo}

708: At some point, we will need a closed form expression for $\mu$ in

709: terms of $Z$. Here is one.

710: \begin{Propo}

711: \label{propo-mu-closed-form}

712: Write

713: $$

714: v= \frac{(u-1) Z (1+Z^2)}{(1+Z)(1+Z+Z^2)(1-Z)^3}.

715: $$

716: Then the algebraic series $\mu$ involved in the

717: expression~\eqref{product-formSj} of $S_j$, and defined by~\eqref{eqMZ}, is

718: $$

719: \mu(t,u)= \frac 1 {Z^2} \left(    \frac 2

720: {1+v(1-Z)^2/3+2/3 \sqrt{3+v^2(1-Z)^4}\cos (\phi/3)}

721: -1 \right)

722: $$

723: where

724: $$

725: \phi= \arccos \left( \frac{-9v(1+4Z+Z^2)+v^3(1-Z)^6}{(3+v^2(1-Z)^4)^{3/2}}\right).

726: $$

727: \end{Propo}

728: \noindent

729:     {\bf Proof of Propositions~\ref{propo-snake-gf}

730:       and~\ref{propo-mu-closed-form}.}\footnote{All the 

731:       calculations in this paper have been 

732:       done using {\sc Maple}. We do not recommend  the reader to

733:       check them by hand.}

734: First, observe that the family of series $S_0, S_1, S_2, \ldots$ is

735: completely determined by~\eqref{eqSj} (taken for $j>0$) and the second

736: part of~\eqref{eqS0}. The fact that for any series $\mu \in \qs(u)[[t]]$, the

737: expression~\eqref{product-formSj} satisfies~\eqref{eqSj} for all $j>

738:       0$ is a straighforward 

739: verification, once $t$ and $T$ have been expressed in terms of $Z$

740: (see~\eqref{eqZt} and~\eqref{eqTZ}). The form

741: of~\eqref{product-formSj} is 

742: borrowed from~\cite{bdg-geodesic}. In order for~\eqref{product-formSj}

743: to be the correct expression of $S_j$, it remains to

744: satisfy the second part of~\eqref{eqS0}. This last condition provides a polynomial

745: equation relating $\mu$, $T$, $Z$, $t$  and $u$. In this equation,

746: replace $t$ and 

747: $T$ by their expressions  in terms of $Z$ (given

748: by~(\ref{eqZt}--\ref{eqTZ})). This gives exactly~\eqref{eqMZ}.

749: It

750: can be easily checked that $\mu$ has degree $6$ over $\qs(T,u)$ and

751: degree $12$ over $\qs(t,u)$.

752:  

753: \medskip

754: 

755: The  equation~\eqref{eqS0T} satisfied by $S_0$ is obtained  by

756: eliminating  $\mu$ and $Z$ (using~\eqref{eqMZ} and~\eqref{eqTZ}) from

757: the expression~\eqref{product-formSj} of $S_0$.  This equation  gives

758:  an equation of degree $6$ over $\qs(t,u)$ if one eliminates $T$

759: thanks to~\eqref{eqT}.

760: 

761: Now the equations~\eqref{eqS0}, \eqref{eqSj} and~\eqref{eqT}, combined with an

762: induction on $j$, imply that for $j\ge

763: 1$, the series $S_j$ belongs to the field  $\qs(T,u,S_0)$,

764: which has just been proved to be an extension of $\qs(T,u)$ of degree

765:   $3$. This concludes the proof of  Proposition~\ref{propo-snake-gf}.

766: 

767: \smallskip

768: Let us finally prove Proposition~\ref{propo-mu-closed-form}. The

769: equation~\eqref{eqMZ} that defines $\mu$ can be rewritten

770: $$

771: \mu= \frac v Z \frac{(1+\mu Z)(1+\mu Z^2)(1+\mu Z^3)}

772: {1-\mu Z^2}.

773: $$

774: Hence $\mu $ is the unique \fps \ in $v$ (with rational coefficients

775: in $Z$) that satisfies the above

776: equation and equals $0$ when $v$ is $0$. It is not hard to check that

777: the closed form expression we give satisfies these two conditions.

778: \cqfd

779: 

780: \noindent{\bf Remarks}\\

781: {\bf 1.} The product form~\eqref{product-formSj} of

782: Proposition~\ref{propo-snake-gf}  refines the

783: product form~\eqref{product-formTj-small} that deals with trees with small

784: labels. Indeed, when $u=0$, Eq.~\eqref{eqMZ} gives $\mu=-1$, and the

785: expression of $S_j(t,0)$ coincides, as it should, with the expression

786: of $T_{j-1}(t)$ given by Proposition~\ref{propo-snake-bounded}.\\

787: %

788: {\bf 2.} There exists an alternative way to

789: derive an equation for  $S_0$ from the system of Lemma~\ref{lemma-Sj-rec}.

790:  As was observed in~\cite[p.~645]{bdg-statistics} for the 

791: problem of counting trees with bounded labels, 

792: Eq.~\eqref{eqSj} implies that for $j\ge 1$,

793: $$

794: I(S_{j-1},S_j)=I(S_{j},S_{j+1})

795: $$

796: where the ``invariant'' function $I$ is given by

797: $$

798: I(x,y)= xy(1-tx)(1-ty)+txy-x-y.

799: $$

800: But $S_j$ converges to $T$ as $j$ goes to infinity, in the set of

801: \fps\ in $t$. This implies

802: $$

803: I(S_{0},S_1)=I(T,T).

804: $$

805: Eliminating $S_1$ between the above equation and~\eqref{eqS0} gives an

806: equation between $S_0$, $T$ and $t$.

807: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

808: 

809: \subsection{The number of nodes labelled $j$ or more}

810: Let us finally study our third and last enumeration problem. For $j \in

811: \zs$, let $R_j \equiv R_j(t,u)$ be the \gf\ of labelled trees, 

812: counted by the number of edges (variable $t$) and the number of

813: nodes labelled $j$ at least (variable $u$). 

814: \begin{Lemma}

815: \label{lemma-Rj-rec}

816: The set of series $R_0, R_1, R_2, \ldots$ is completely determined by the

817: following equations: for $j\ge 1$,

818: \beq

819: R_j=1+tR_j(R_{j-1}+R_{j+1} ) 

820: \label{rec-Rj}

821: \eeq

822: and

823: \beq

824: \label{R0-R1}

825:  R_0(t,u)=uR_1(tu,1/u).

826: \eeq

827: More generally, for all $j \in\zs$, one has:

828: \beq

829: \label{eq-symmetry}

830:  R_{-j}(t,u)=uR_{j+1}(tu,1/u).

831: \eeq

832: \end{Lemma}

833: \noindent {\bf Proof.} For all $j\in \zs$, the series  $R_j\equiv

834: R_j(t,u)$ is also the \gf \ of 

835: trees rooted at $j$, counted by their number of edges

836: and the number of nodes having a non-positive label. The equation

837: satisfied by $j$, for $j \ge 1$, 

838: follows once again from the decomposition of trees illustrated in

839: Figure~\ref{fig-decomp}. It remains to prove the symmetry

840: relation~\eqref{eq-symmetry}. For any tree $\tau $, let $n_{\le 0}(\tau )$

841: denote the number of nodes of $\tau $ having a non-positive label. We 

842: use similar notations for the number of nodes having label at most

843: $j$, etc. Let $\T_{j,n}$ denote the set of trees rooted at

844: $j$ and having $n$ edges.

845: As observed above,

846: $$

847: R_{-j}(t,u)= \sum_{n\ge 0} t^n \sum_{\tau  \in \T_{-j,n}} u^{n_{\le 0}(\tau )}

848: = \sum_{n\ge 0} t^n \sum_{\tau  \in \T_{-j,n}} u^{n+1-n_{> 0}(\tau )},

849: $$

850: because a tree with $n$ edges has a total of $n+1$ nodes. A

851: translation of all labels by $-1$ gives

852: $$

853: R_{-j}(t,u)=u \sum_{n\ge 0} (tu)^n \sum_{\tau  \in \T_{-j-1,n}} u^{-n_{\ge

854:     0}(\tau )},

855: $$

856: while replacing each label $k$ by $-k$ finally gives

857: $$

858: R_{-j}(t,u)=u \sum_{n\ge 0} (tu)^n \sum_{\tau  \in \T_{j+1,n}} u^{-n_{\le

859:     0}(\tau )}=u R_{j+1}(tu,1/u).

860: $$

861: \cqfd

862: 

863: Again, the series $R_j$ are algebraic, and admit a closed form

864: expression in terms of $T$ and $Z$.

865: \begin{Propo}[{\bf  The number of nodes labelled $j$ or more}]

866: \label{propo-snake-gf-rep}

867: Let $j \in \zs$. The \gf\ $R_j(t,u)\equiv R_j$

868: that counts labelled trees by the number of edges and the number of

869: nodes labelled $j$ or more

870: is  algebraic of degree at most $2$ over $\qs(T(t),T(tu))$. Hence it

871: has degree at most $8$ over $\qs(t,u)$. More precisely, it belongs to

872: the extension of $\qs(T(t),T(tu))$ generated by 

873: $$

874: %\sqrt\Delta_1= 

875: \sqrt{(T+\tilde T)^2-4T\tilde T (T-1)(\tilde T -1)}

876: $$

877: where $T\equiv T(t)$ and $\tilde T \equiv T(tu)$. \\

878: Moreover,  for all $j \ge 0$,

879: \beq

880: \label{product-formRj}

881: R_j= T\,\frac{(1+\nu Z^j)(1+\nu Z^{j+4})}{(1+\nu Z^{j+1})(1+\nu Z^{j+3})},

882: \eeq

883: where $Z\equiv Z(t)$ is given by~\eqref{eqZt}

884: and $\nu\equiv \nu(t,u)$ is a \fps\ in $t$, with polynomial

885: coefficients in $u$, which is algebraic

886: %  of degree $2$ over $\qs(Z(t),Z(tu))$, 

887: of degree $4$ over $\qs(u,Z)$, and  of degree $16$

888: over $\qs(t,u)$. 

889: This series satisfies $\nu(t,1)=0$. 

890: The first terms in its expansion are:

891: $$

892: \nu(t,u)=(u-1)\Big(

893: 1+2\,ut+ \left(7\,u+ 6\,{u}^{2} \right) {t}^{2}

894: + \left( 32\,u+36\,{u}^{2}+23\,{u}^{3} \right) {t}^{3}+

895: O(t^4)\Big) .

896: $$

897: \end{Propo}

898: Before we prove this proposition, let us give something like a closed form for

899: $\nu$. Since $\nu$ has degree 4 over $\qs(u,Z)$, and $Z$ has degree 4

900: over $\qs(t)$, the series $\nu$ is in theory expressible in terms of

901: radicals... It turns that this expression is less terrible than one

902: could fear.

903: \begin{Propo}

904: \label{propo-nu}

905: Define the following four \fps\ in $t$ with polynomial coefficients

906: in $u$:

907: $$

908: \delta\equiv\delta(t,u)= 1-8(u-1) \frac{Z(1+Z^2)}{(1-Z)^4} 

909: = \frac{1-8tu}{1-8t},

910: $$

911: $$

912: V\equiv V(t,u) = \frac{1-\sqrt \delta} 4 =

913:  \frac{1-\sqrt {\frac{1-8tu}{1-8t}}} 4,

914: $$

915: $$

916: \Delta\equiv \Delta(t,u)= (1-V)^2 - \frac{4Z V^2}{(1+Z)^2},

917: $$

918: and 

919: $$

920: P=(1+Z) \frac{1-V-\sqrt \Delta}{2VZ}.

921: $$

922: Then $P$ has degree $16$ over $\qs(t,u)$, degree $2$ over $\qs(V,Z)$,

923: and satisfies the following ``Lagrangian'' equation:

924: $$

925: P=\frac V{1+Z} (1+P)(1+ZP).

926: $$

927: Moreover,  the algebraic series $\nu$ involved in the

928: expression~\eqref{product-formRj}  of $R_j$ is

929: $$

930: \nu= \frac P Z \frac {1-P(1+Z)-P^2(1+Z+Z^2)}{1+Z+Z^2+PZ(1+Z)-P^2Z^2}.

931: $$

932: \end{Propo}

933: \noindent

934: {\bf Proof of Proposition~\ref{propo-snake-gf-rep}.} We have already

935: checked, in the proof of 

936: Proposition~\ref{propo-snake-gf}, that for any \fps \ $\nu$ in $t$, the series

937: defined by~\eqref{product-formRj} for $j\ge 0$ satisfy  the recurrence

938: relation~\eqref{rec-Rj} for $j\ge 1$. It remains to prove that one can

939: choose $\nu$ so as to satisfy~\eqref{R0-R1}. For any \fps\

940: $A$ in $t$  having rational coefficients in $u$, we denote by

941: $\tilde A$ the series $\tilde A(t,u) =A(tu,1/u)$. Observe that $\tilde{\tilde

942:   A}=A$. With this notation, if $R_j$ is of the generic

943: form~\eqref{product-formRj}, 

944: the relation~\eqref{R0-R1} holds if and only if

945: \beq

946: \label{nu-nutilde}

947:  1+\nu =

948: u\, \frac{\tilde T} T\,\frac

949: {(1+\nu Z)(1+\nu Z^{3})(1+ \tilde \nu \tilde Z)(1+\tilde \nu \tilde Z^{5})}

950: {(1+\nu Z^{4})(1+\tilde \nu \tilde Z ^{2})(1+\tilde \nu \tilde Z^{4})}.

951: \eeq

952: Let $\rs_m[u]$ denote the space of polynomials in $u$, with real

953: coefficients, of degree at most $m$. 

954: Let $\rs_n[u][[t]]$ denote the set of \fps \ in $t$ with polynomial

955: coefficients in $u$ such that for all $m \le n$, the coefficient of

956: $t^m$ has degree at 

957: most $m$. Observe

958: that this set of series in stable under the usual operations on

959: series: sum, product, and quasi-inverse.

960: Write  $\nu=\sum_{n\ge 0} \nu_n(u) t^n$. We are going to prove, by

961: induction on $n$,  that~\eqref{nu-nutilde} determines uniquely each

962: coefficient $\nu_n(u)$, and that this coefficient belongs to $\rs_{n+1}[u]$.

963: 

964: First, observe that for any \fps\ $\nu$, the right-hand side of~\eqref{nu-nutilde}  is $u+

965: O(t)$. This implies $\nu_0(u)=u-1$. Now assume that our induction hypothesis

966: holds for all $m<n$. Recall that $Z$ is a multiple of $t$:

967: this implies that  $ \nu  Z$ belongs to  $\rs_n[u][[t]]$. The

968: induction hypothesis also implies that the coefficient 

969: of $t^m$ in $u\tilde \nu$ belongs to $\rs_{m+1}[u]$, for all $m<n$. 

970: Note that $\tilde Z = Z(tu)=tu+O(t^2)$ is a

971: multiple of $t$ and $u$ and also belongs to  $\rs_n[u][[t]]$. This

972: implies that  $\tilde \nu \tilde Z$ 

973: belongs to $\rs_n[u][[t]]$ too.

974:  The same is true for all the other series

975: occurring in the right-hand side of~\eqref{nu-nutilde}, namely $T,

976: \tilde T, Z, \tilde Z$. Given the closure properties of the set

977: $\rs_n[u][[t]]$, we conclude that the right-hand side

978: of~\eqref{nu-nutilde}, divided by $u$, belongs to this set. Moreover,

979: the fact that $Z$ and $\tilde Z$ are multiples of $t$ guarantees that

980: the coefficient of $t^n$ in this series only involves the $\nu_i(u)$

981: for $i<n$.

982:   By extracting the coefficient

983: of $t^n$ in~\eqref{nu-nutilde}, we conclude that $\nu_n(u)$ is

984: uniquely determined and belongs to  $u\rs_n[u]\subset\rs_{n+1}[u] $. 

985: 

986: This completes the proof of the existence and uniqueness of the series

987:  $\nu$  satisfying~\eqref{nu-nutilde}. 

988:  Also, setting $u=1$ (that is, $\tilde T=T$ and $\tilde Z=Z$) in this

989:  equation shows that $\nu(t,1)=0$.

990: 

991: Let us now replace $t$ by $tu$ and $u$ by $1/u$

992: in~\eqref{nu-nutilde}. This gives:

993: \beq

994: \label{nutilde-nu}

995:  1+\tilde \nu =\frac 1

996: u\, \frac{ T} {\tilde T} \,\frac

997: {(1+\tilde \nu \tilde Z)(1+\tilde \nu \tilde Z^{3})(1+ \nu  Z)(1+ \nu  Z^{5})}

998: {(1+\tilde \nu \tilde Z^{4})(1+ \nu  Z ^{2})(1+ \nu  Z^{4})}.

999: \eeq

1000: In the above two equations, replace $T$ by its expression~\eqref{eqTZ}

1001: in terms of $Z$. Similarly, replace $\tilde T$ by its expression

1002: in terms of $\tilde Z$. Finally, it follows from~\eqref{eqZt} and from

1003: the fact that $\tilde Z = Z(tu)$ that

1004: \beq

1005: \label{uZZtilde}

1006: u= \frac {\tilde Z} Z \, \frac{(1+Z)^4(1+\tilde Z^2)}

1007: {(1+\tilde Z)^4(1+Z^2)}.

1008: \eeq

1009: Replace $u$ by this expression in~\eqref{nu-nutilde} and~\eqref{nutilde-nu}.

1010: Eliminate $\tilde \nu$ between the resulting

1011: two equations: this gives a polynomial equation that relates $\nu, Z$

1012: and $\tilde Z$, of degree $2$ in $\nu$. The elimination of $\tilde Z$

1013: between this quadratic equation and~\eqref{uZZtilde} provides an

1014: equation of degree $4$ in $\nu$ that relates $\nu$ to $Z$ and $u$.

1015: Finally, the elimination of $ Z$ shows that $\nu$ is algebraic of

1016: degree $16$ over $\qs(t,u)$.

1017: 

1018: \medskip

1019: Let us now focus on the first part of the proposition.

1020: From the form~\eqref{product-formRj}, and the fact that $\nu$ has

1021: degree $4$ over $\qs(u,Z)$ and $Z$ has degree $4$ over $\qs(t)$, we

1022: conclude that the degree of $R_j$ over $\qs(t,u)$ is a divisor of

1023: $16$. Let us prove that is is actually a divisor of $8$. The proof

1024: goes as follows:

1025: \begin{enumerate}

1026: \item  Using the generic form~\eqref{product-formRj}, and the equations

1027: satisfied by $T, Z$ and $\nu$, we obtain a polynomial equation of

1028: degree $8$ over $\qs(t,u)$ for $R_0$.

1029: 

1030: 

1031: \item  Using~\eqref{eqT} to express $t$ in terms of $T$, and 

1032: $$

1033: u=\frac {T^2}{\tilde T ^2} \, \frac{1-\tilde T}{1-T},

1034: $$

1035: (which also follows from~\eqref{eqT}),

1036: we convert the equation satisfied by $R_0$ into a polynomial equation

1037: (still of degree $8$ in $R_0$) relating $R_0$  to $T$ and $\tilde

1038: T$. This equation factors into four quadratic polynomials in

1039: $R_0$. The factor that actually vanishes is identified by setting $u=1$

1040: (in which case $\tilde T=T=R_0$). 

1041: 

1042: 

1043: \item  From this equation, we conclude that  $R_0$ belongs

1044: to the extension of $\qs(T,\tilde T)$ generated by

1045: $$

1046: \sqrt\Delta_1= \sqrt{(T+\tilde T)^2-4T\tilde T (T-1)(\tilde T -1)}.

1047: $$  

1048: Observe that this extension of $\qs(t,u)$ is left invariant by the

1049: transformation $A\mapsto \tilde A$.

1050: 

1051: \item  From the fact that $R_1=u \tilde{R_0}$ (see~\eqref{eq-symmetry}),

1052:  we conclude that $R_1$ also belongs to $\qs(T,\tilde

1053:  T,\sqrt\Delta_1 )$.

1054: 

1055: \item  The recurrence relation~\eqref{rec-Rj} on the $R_j$ allows us to

1056: extends this to all $R_j$, for $j\ge 0$.

1057: 

1058: 

1059: \item  Finally,~\eqref{eq-symmetry} shows that our algebraicity result

1060: actually holds for all $R_j$, for $j \in \zs$.

1061: \end{enumerate}

1062: \cqfd

1063: 

1064: \noindent{\bf Proof of Proposition~\ref{propo-nu}.}

1065: In the course of the proof of Proposition~\ref{propo-snake-gf-rep}, we

1066: have obtained a polynomial equation $P(\nu,Z,u)=0$, of degree 4 in

1067: $\nu$, relating the series $\nu(t,u), Z(t)$, and the variable

1068: $u$. This equation is not written in the paper (it is a bit too big),

1069: but it follows from~\eqref{nu-nutilde} and~\eqref{nutilde-nu}. In this

1070: equation, replace $u$ by its expression in terms of $\delta$ and

1071: $Z$. Then replace $\delta$ by its expression in terms of $V$:  the

1072: resulting equation factors into two terms! Each of them is quadratic

1073: in $\nu$. In order to decide which of these factors cancels, one uses

1074: the fact that when $u=1$ (that is, $V=0$), the series $\nu$ must be

1075: $0$.

1076: It remains to solve a quadratic equation in $\nu$. Its discriminant is

1077: found to be $\Delta$, and one may find convenient to introduce the

1078: series $P$ which is Lagrangian in $V$.

1079: \cqfd

1080: 

1081: \noindent{\bf Remark.} Again, the product form~\eqref{product-formRj}

1082: of Proposition~\ref{propo-snake-gf-rep} includes as a special case

1083: the enumeration of trees with labels at most $j-1$, obtained when

1084: $u=0$. Indeed,~\eqref{nu-nutilde} shows that $\nu=-1$ when $u=0$,

1085: and~\eqref{product-formRj} 

1086:  then reduces to~\eqref{product-formTj-small}.

1087: 

1088: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1089: \section{The largest label, and the support of the ISE}

1090: \label{section-max}

1091: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1092: Let $\mathcal T _0$ denote the set of labelled trees (rooted at $0$),

1093: and let $\mathcal T _{0,n}$ denote the subset of $\mathcal T _0$ 

1094: formed by trees having $n$ edges.

1095: We  endow $\mathcal T _{0,n}$  with the

1096: uniform distribution. In other words, any of its elements occurs with

1097: probability

1098: $$

1099: \frac 1{2^n C_n}

1100: $$

1101: where $C_n= \frac{1}{n+1} {{2n} \choose n}$ is the $n$th Catalan

1102: number, and is well-known to be the number of (unlabelled) plane trees

1103: with $n$ edges.

1104: 

1105:  Let $M_n$ denote the random variable equal to the

1106: largest label occurring in a random tree of  $\mathcal T _{0,n}$. 

1107: The law of $M_n$ is related to the series $T_j$ studied in

1108: Proposition~\ref{propo-snake-bounded}:

1109: $$

1110: \PP\left(M_n\le j\right)=  \frac{[t^n] T_j}{2^n C_n}.

1111: $$

1112: Let us define a normalized version of $M_n$ by

1113: $$

1114: N_n= \frac{M_n}{n^{1/4}}.

1115: $$

1116: The aim of this section is to prove the convergence of $N_n$ in

1117: distribution\footnote{The above convention will be used throughout the

1118:   paper: if a random variable depending on $n$ is denoted by some

1119:   letter of the alphabet, then its suitably normalized version is

1120:   denoted by \emm the next letter, of the alphabet.}.

1121: %

1122: \begin{Theorem}

1123: \label{thm-max}

1124: As $n$ goes to infinity, the random variable $N_n$ converges

1125:  in distribution to a non-negative random

1126: variable $N$. The tail distribution function of $N$, defined by

1127: $ G(\lambda)=\PP(N > \lambda )$, satisfies

1128: $$

1129: G(\lambda)=

1130: \frac {12}{i\sqrt \pi}\int_ \Gamma\frac{v^5 e^{v^4}}{\sinh^2(\lambda v) }

1131: dv

1132: =\frac{6}{\sqrt \pi \lambda ^6}\int_0^\infty\frac{1-\cos u \cosh

1133:   u}{(\cosh u -\cos u)^2} u^5 e^{-u^4/(4\lambda^4)}du

1134: $$

1135: where the contour $\Gamma$ is formed of two half-lines:

1136: $$

1137: \Gamma= \{1-te^{-i\pi/4}, t\in (\infty,0]\} \cup \{1+te^{-i\pi/4},

1138: t\in [0,\infty)\}. 

1139: $$

1140: Equivalently, the variable $N$ has density

1141: $$

1142: f(\lambda)=\frac {24}{i\sqrt \pi}\int_ \Gamma\frac{\cosh(\lambda v)v^6

1143:   e^{v^4}}{\sinh^3(\lambda v) } dv

1144: =\frac{6}{\sqrt \pi \lambda ^{11}}\int_0^\infty\frac{1-\cos u \cosh

1145:   u}{(\cosh u -\cos u)^2} u^5(6\lambda^4-u^4) e^{-u^4/(4\lambda^4)}du

1146: $$

1147: with respect to the Lebesgue measure on $\rs_+$.

1148: The moments of $N$ are finite, and admit simple expressions:

1149: $$

1150: \EE(N)=\frac{3\sqrt \pi}{2\Gamma(3/4)},\quad

1151: \EE(N^2)=3\sqrt \pi,

1152: $$

1153:  and for  $k\ge 3$, 

1154: $$

1155: \EE(N^k)=\frac{24 \sqrt \pi k! \zeta(k-1)}{2^k \Gamma ((k-2)/4)}.

1156: $$

1157: Finally, the moments of $N_n=M_n/n^{1/4}$ converge to the moments of $N$.

1158: \end{Theorem}

1159: %

1160: \noindent The functions $G$ and $f$ are plotted in Figure~\ref{figure-densite}. 

1161: 

1162: \begin{figure}[phtb]

1163:   \begin{center}

1164:     \includegraphics[height=3cm,width=3cm]{Glimite.eps}

1165: \hskip 28mm

1166: \includegraphics[height=3cm,width=3cm]{densite.eps}

1167:     \caption{The tail distribution function $G$ and the

1168:     density $f$ of the    limit distribution $N$.} 

1169:     \label{figure-densite}

1170:   \end{center}

1171: \end{figure}

1172: 

1173: 

1174: 

1175: \noindent The proof of this theorem will be split into four

1176: subsections (Sections~\ref{section-distribution-function}

1177: to~\ref{section-moments-Nn}). In view of the following proposition,

1178: this theorem gives 

1179: the density, distribution function and moments of the supremum of  the

1180: support of the ISE. 

1181: 

1182: \begin{Propo}[{\bf The supremum of the support of the ISE}]

1183: \label{coro-support}

1184: Let $N_{{\hbox{\sc \small ise}}}$ denote the supremum of the support of

1185: the ISE

1186: $$

1187: N_{{\hbox{\sc \small ise}}}= \sup\{y: \mu_{\hbox{\sc \small ise}}(y,

1188: \infty)>0\}.

1189: $$

1190: Then 

1191: $ N_{{\hbox{\sc \small ise}}}$ has the same law as the random variable

1192: $\sqrt 2N$ described in  Theorem~{\em\ref{thm-max}.} 

1193: \end{Propo}

1194: 

1195: \noindent{\bf Remark.} The moments of $N_{\hbox{\sc \small ise}}$ 

1196: are thus

1197: $$

1198: \EE(N_{{\hbox{\sc \small ise}}})=\frac{3\sqrt \pi}{\sqrt 2\Gamma(3/4)},\quad

1199: \EE(N_{{\hbox{\sc \small ise}}}^2)=6\sqrt \pi,

1200: $$

1201: and for  $k\ge 3$, 

1202: $$

1203: \EE(N_{{\hbox{\sc \small ise}} }^k)=\frac{24 \sqrt \pi k! \zeta(k-1)}{\sqrt 2^k \Gamma ((k-2)/4)}.

1204: $$

1205: They were

1206: already obtained by Delmas~\cite{delmas} using a completely different

1207: (and continuous) approach.  The expressions he gives actually differ

1208: from ours by a factor $2^{k/4}$, due to a different choice of

1209: normalization. Note that the zeta function also appears in the moments

1210: of the maximum of the Brownian excursion, which follows a theta

1211: law~\cite{chung}. This law is known to describe the limiting

1212: normalized height of  simple trees~\cite{flajolet-binary-trees}.

1213: Finally, let us mention that another,  more complicated

1214:  expression of the density of the limiting

1215:  variable $N$ was obtained

1216:  in~\cite{bdg-geodesic} (maybe in a slightly less rigorous

1217:  way). Proposition~\ref{coro-support} is proved in

1218:  Section~\ref{section-supremum}. 

1219: 

1220: 

1221: \subsection{Convergence of the distribution function}

1222: \label{section-distribution-function}

1223: We prove in this section that the tail distribution function of $N_n$

1224: converges pointwise.

1225: Let $\lambda \ge 0$ and $j=\lfloor \lambda n^{1/4} \rfloor$. 

1226: The probability we are interested in is

1227: \beq

1228: \label{GU}

1229: \PP(N_n > \lambda )=\PP(M_n > \lambda n^{1/4})=\PP(M_n >j)=\frac{[t^n] U_j(t)}{2^n C_n},

1230: \eeq

1231: where

1232: \beq

1233: \label{Uj-expr}

1234: U_j(t)\equiv U_j = T- T_j=  

1235: \frac { \left( 1+Z \right) ^{2}{Z}^{j+1} \left( 1+Z+{Z}^{2} \right)  

1236: \left(1- Z \right) ^{2}}

1237: { \left( 1+{Z}^{2} \right) 

1238:  \left( 1-{Z}^{j+2} \right)  \left( 1-{Z}^{j+4} \right) }

1239:  \eeq

1240: is the \gf  \ of trees having at least one label greater than

1241: $j$. This algebraic series has a positive radius of

1242: convergence\footnote{So do all  algebraic power series}, 

1243: and by Cauchy's formula,

1244: \begin{eqnarray}

1245: [t^n] U_j& =&\frac 1 {2i\pi} \int_\C U_j(t)

1246: \frac{dt}{t^{n+1}} \nonumber\\

1247: &=&\frac {1} {2i\pi} \int_\C 

1248: \frac { \left( 1+Z \right) ^{2}{Z}^{j+1} \left( 1+Z+{Z}^{2} \right)  

1249: \left(1- Z \right) ^{2}}

1250: { \left( 1+{Z}^{2} \right) 

1251:  \left( 1-{Z}^{j+2} \right)  \left( 1-{Z}^{j+4} \right) }

1252: \frac{dt}{t^{n+1}},

1253: \label{cauchy}

1254: \end{eqnarray}

1255: for any contour $\mathcal C$ included in the analyticity domain of

1256: $U_j$ and enclosing positively the origin.

1257: 

1258: This leads us to study  the singularities of $U_j$, and therefore

1259:  those of  $Z$. 

1260:  We  gather in the following lemma a few  properties of this series.

1261: 

1262: \begin{Lemma}[{\bf Analytic properties of $Z$}]

1263: \label{lemma-Z}

1264: Let $Z\equiv Z(t)$ be the unique \fps \ in $t$ with constant term $0$

1265: satisfying~\eqref{eqZt}.

1266: This series has non-negative integer coefficients.

1267: It has radius of convergence $1/8$, and can be continued analytically on 

1268: the domain $\D=\cs\setminus[1/8, +\infty)$. In the neighborhood of

1269:   $t=1/8$, one has

1270: \beq

1271: \label{Z-asympt}

1272: Z(t)=1-2(1-8t)^{1/4} + O(\sqrt{1-8t}).

1273: \eeq

1274: Moreover,  $|Z(t)|<1$ on the domain $\D$.

1275: More precisely, the only roots of unity 

1276: that are  accumulation points of the set $Z(\D)$ are $1$ and

1277: $-1$, and they are only approached by $Z(t)$ when $t$ tends to $1/8$

1278: and when $|t|$ tends to $\infty$, respectively.

1279: \end{Lemma}

1280: \noindent

1281: {\bf Proof.}

1282: In order to establish the first statement, we observe that

1283: $$

1284: Z=W(1+Z)^2

1285: $$

1286: where $W\equiv W(t)$ is the only \fps\ in $t$ with constant term zero

1287: satisfying

1288: \beq

1289: \label{eqW}

1290: W=t+2W^2.

1291: \eeq

1292: These equations imply that both $W$ and $Z$ have non-negative integer

1293: coefficients.

1294: 

1295: \medskip

1296: The general approach for studying the singularities of

1297: algebraic series (see for instance~\cite{fla-poly-alg}) gives

1298: the second part of the lemma (up to~\eqref{Z-asympt}).  The polynomial

1299: equation defining $Z(t)$ has leading coefficient $t$ and

1300: discriminant  $4(1-8t)^3$, so that the

1301: only possible singularity of $Z$ is $1/8$. Alternatively,

1302: one can exploit the following closed form expression:

1303: \beq

1304: \label{Z-expr}

1305: Z(t)=\frac{\sqrt{1-4t+\sqrt{1-8t}}\left(\sqrt{1-4t+\sqrt{1-8t}} -\sqrt

1306:   2 (1-8t)^{1/4}\right) }{4t}. 

1307: \eeq

1308: 

1309: 

1310: Let us now come to the third part of the lemma, and prove that

1311:   $|Z(t)|$ never reaches $1$ on the

1312:   domain $\D$. Assume $Z(t)=e^{i\theta}$, with

1313:   $\theta\in[-\pi,\pi]$. From~\eqref{eqZt}, one has 

1314: $$

1315: t=t_\theta \quad \hbox{ where } \quad t_\theta=

1316:  \frac{\cos\theta}{8\cos^4(\theta/2)} \quad \hbox{ and } \quad

1317:  \theta\in(-\pi,\pi). 

1318: $$

1319: This shows that $t$ is real, and belongs to $(-\infty, 1/8)$. 

1320: But the expression~\eqref{Z-expr} of $Z(t)$ shows that $Z(t)$ is real,

1321: which contradicts the 

1322: hypothesis $Z(t)=e^{i\theta}$, unless $\theta=0$. But then $t=1/8$

1323: and does not belong to the domain $\D$. Hence the modulus of $Z$ never

1324: reaches $1$ on $\D$.

1325: One can actually prove that, for $\theta \in (-\pi, 0)$,

1326: $$

1327: Z(t_\theta)= \frac{1+ \sin \theta}{\cos \theta},

1328: $$

1329: but we do not need so much precision here.

1330: % Conversely, let us try to evaluate $Z(t_\theta)$ for

1331: % $t=\theta\in(-\pi,0)$ (by parity, this is sufficient for $t$ to

1332: % cover the whole interval $(-\infty, 1/8)$). Can it be equal to $

1333: % e^{i\theta}$? 

1334: %The four roots of the equation (in $Z$)

1335: %$$

1336: %Z=t-\theta \frac{(1+Z)^4}{1+Z^2}$$

1337: %are

1338: %$$

1339: %Z_{1,2}= e^{\pm i \theta} \quad\hbox{ and } \quad

1340: %Z_{3,4}= \frac{1\pm\sin \theta}{\cos \theta}.

1341: %$$

1342: %%

1343: %The expression~\eqref{Z-expr} shows that $Z(t)$ is real when $t\in

1344: %(-\infty, 1/8)$. Therefore the solutions $Z_{1,2}$ cannot coincide

1345: %with $Z(t)$. Moreover, for $|t|<1/8$, the value of $Z(t)$ is given by

1346: %a \fps\ with non-negative coefficients, so that $Z(t)$ is an \emm

1347: %increasing, function of $t$ on $[0,1/8)$. This rules out  the

1348: %possibility that $Z(t)=Z_4$ (the solution with a ``minus'' sign),

1349: %andwe conclude that $$ Z(t_\theta)=  \frac{1+ \sin \theta}{\cos

1350: %\theta}.$$This function of $\theta$ increases from $-1$ to $1$ as

1351: %$\theta$ goesfropm $-\pi$ to $0$, and thus has modulus less than $1$

1352: %for $t\in(-\infty, 0)$. Consequently, $Z(t\theta)\not = e^{i \theta}$

1353: %and more generally $|Z(t)|<= 1$ for $t \in \cs\setminus[1/8, +\infty)$.

1354: 

1355: Finally, if a sequence $t_n$ of $\D$ is such that $Z(t_n) \rightarrow

1356: e^{i\theta}$ as $n \rightarrow \infty$, with $\theta\in(-\pi, \pi]$,

1357: then either $\theta=\pi$ and, by~\eqref{eqZt}, the

1358: sequence $|t_n|$  tends to $\infty$, or $\theta\in (-\pi, \pi)$

1359:   and $t_n$ converges to $t_\theta$. But then by continuity, $Z(t_n)$

1360: actually converges to $Z(t_\theta)$, which, as argued above, only

1361: coincides with $e^{i\theta}$ when $\theta=0$, that is, $t_\theta=1/8$. In this

1362: case, $Z(t_n)\rightarrow 1$.

1363: \cqfd

1364: 

1365: 

1366: \begin{figure}[hbt]

1367: \begin{center}

1368: \input{contour.pstex_t}

1369: \end{center}

1370: \caption{The integration contour $\C_n$.}

1371: \label{fig-contour}

1372: %\hrule

1373: \end{figure}

1374: 

1375: Let us now go back to the evaluation of the tail distribution function

1376: of $N_n$ via the integral~\eqref{cauchy}.

1377: %The  argument is now routine for amateurs of singularity analysis.

1378:  We choose a contour $\C={\mathcal C}_n$ that depends on $n$ and consists of

1379:  two parts ${\mathcal C}_n^{(1)}$ and ${\mathcal C}_n^{(2)}$ (see

1380:  Figure~\ref{fig-contour}): 

1381: \begin{itemize}

1382: %

1383: \item[$\bullet$ ] ${\mathcal C}_n^{(1)}$ is an arc  of radius

1384: $r_n/8=(1+ \log^2 n /n)/8$, centered at the origin; note that its

1385:   radius tends to $1/8$ as $n$ goes to infinity,

1386: %

1387: \item[$\bullet$ ] ${\mathcal C}_n^{(2)}$ is a Hankel contour around $1/8$, at distance

1388: $1/(8n)$ of the real axis, 

1389: which meets  ${\mathcal C}_n^{(1)}$ at both ends; this contour shrinks

1390: around $1/8$  as $n$ goes to infinity; more precisely, as $t$ runs

1391: along $\C_n^{(2)}$, the variable $z$ defined by

1392: $$

1393: t=\frac 1 8 \left(1+ \frac z n\right)

1394: $$

1395: runs over the truncated Hankel contour $\Ha_n$ shown on the right

1396:  of Figure~\ref{fig-hankel}:

1397: $$

1398: {\Ha_n}= \left\{  {x-i}, \ x \in [0, x_n] \right\} \cup

1399: \left\{-{e^{i\theta}}, \  \theta \in [-\pi/2, \pi/2]\right\}

1400: \cup

1401: \left\{ {x+i}, \ x \in [0, x_n] \right\} 

1402: $$

1403: %

1404: where $(1+x_n/n)^2+1/n^2=r_n^2$, so that $x_n \le \log ^2 n$ 

1405: and $x_n=\log^2n+O(1/n)$.

1406: \end{itemize}

1407: We denote by $z_n=x_n+i$ the top right end of $\Ha_n$. This point

1408: tends to infinity as $n$ does.

1409: 

1410: 

1411: \begin{figure}[b]

1412: \begin{center}

1413: \input{hankel.pstex_t}

1414: \end{center}

1415: \caption{The Hankel contour $\Ha$ and its truncated version $\Ha_n$.}

1416: \label{fig-hankel}

1417: %\hrule

1418: \end{figure}

1419: 

1420:  The integral~\Ref{cauchy} on $\C=\mathcal C _n$ is the sum of the

1421: contributions of the 

1422: contours $\C_n^{(1)}$ and $\C_n^{(2)}$. We shall see that the dominant

1423: contribution is that of ${\mathcal C}_n^{(2)}$, because of the

1424: vicinity of the singularity at $t=1/8$. 

1425: %

1426: 

1427: Let us first bound carefully $Z(t)$ for $t \in \C_n$. Let $t_n \in

1428: \C_n$ be such that

1429: $$

1430: |Z(t_n)|=\max_{t\in \C_n} |Z(t)|.

1431: $$ 

1432: By Lemma~\ref{lemma-Z}, $|Z(t_n)|$ tends to $1$ as $n$

1433: grows. Moreover, every accumulation point $a$ of the sequence $t_n$

1434: satisfies $|a|\le 1/8$ and $|Z(a)|=1$. This forces $a=1/8$, and we

1435: conclude that $t_n\rightarrow 1/8$.  Write $t_n=(1-u_n)/8$. Then $u_n

1436: \rightarrow 0$, but $|u_n|\ge 1/n$. By~\eqref{Z-asympt},

1437: $$

1438: Z(t_n)=1-2u_n^{1/4} \left( 1 +o(1)\right).

1439: $$

1440: Let us write, for short, $v_n=1-Z(t_n)$. Then $v_n \rightarrow 0$ but

1441: \beq

1442: \label{1-Z-bound}

1443: |v_n| =2|u_n|^{1/4}\left( 1 +o(1)\right)\ge n^{-1/4}

1444: \eeq

1445: for $n$ large enough. Moreover,

1446: $$

1447: |\arg v_n| =\frac 1 4 |\arg(u_n)| +o(1) \le \frac \pi 4 + o(1),

1448: $$

1449: so that

1450: $$

1451: \cos (\arg v_n) \ge \frac 1 {\sqrt 2} +o(1).

1452: $$

1453: Finally,

1454: $$

1455: |Z(t_n)|^2= |1-v_n|^2 = 1-2|v_n| \cos (\arg v_n) + |v_n|^2

1456: \le 1 -\sqrt 2 |v_n|\left( 1 +o(1)\right),

1457: $$

1458: that is,

1459: $$

1460: |Z(t_n)|\le\displaystyle 1- \frac 1 {\sqrt 2} |v_n|\left( 1 +o(1)\right)

1461: \le \displaystyle 1- \frac 1 2 n^{-1/4} .

1462: $$

1463: The latter inequality follows from~\eqref{1-Z-bound}, and holds for

1464: $n$ {large enough}. Finally, for $t \in \C_n$,

1465: \beq

1466: \label{Z-bound}

1467: 1-|Z(t)| \ge \frac 1 2 n^{-1/4}.

1468: \eeq

1469: 

1470: 

1471: Let us now consider the integral on  the contour ${\mathcal C}_n^{(1)}$.

1472: By Lemma~\ref{lemma-Z}, the quantity

1473: $$

1474: \frac { \left( 1+Z \right) ^{2}{Z}^{j+1} \left( 1+Z+{Z}^{2} \right)  

1475: \left(1- Z \right) ^{2}}

1476: {  1+{Z}^{2}   }

1477: $$

1478:  is uniformly bounded on this contour by some constant

1479:  $c$, independant of $n$ and $t$. Moreover, 

1480: $$

1481: |1-Z^{j+2}|\ge 1 -|Z|^{j+2} \ge 1-|Z|\ge \frac 1 2 n^{-1/4}

1482: $$

1483: by~\eqref{Z-bound}.

1484: The same bound holds for the term $1-Z^{j+4}$.

1485: Therefore the modulus of the contribution of $\C_n^{(1)}$ in the

1486: integral~\eqref{cauchy}  is bounded by 

1487: \beq

1488: \label{C1-estimate}

1489:  4c\, 8^n n^{1/2} \,r_n^{-n}%=O(8^n\, e^{-\log^2n})

1490: =O(8^n n^{1/2-\log n})=o(8^n/n^m)

1491: \eeq

1492: for any $m>0$.

1493: 

1494: Let us now study the contribution of the contour ${\mathcal C}_n^{(2)}$.

1495: As $t$ varies along ${\mathcal C}_n^{(2)}$, the variable $z$ defined by

1496: $t=(1+z/n)/8$ varies along the contour 

1497: $\Ha_n$.

1498: As $n$ goes to infinity, this contour converges to the contour

1499: $\Ha$ shown on the left side of Figure~\ref{fig-hankel}. Let 

1500: $z\in\Ha$. Then $z\in\Ha_n$ for $n $ large enough, 

1501: $|z|\le |z_n|\sim \log^2n$, and,

1502: as $n$ goes to infinity,

1503: the following approximations hold 

1504: with error terms independent of $z$:

1505: \beq

1506: \label{estimates}\left\{

1507: \begin{array}{rcll}

1508: %

1509: \displaystyle 

1510: Z(t)&=&\displaystyle 1-2 (-z)^{1/4} n^{-1/4} + O\left( n^{-1/2} \log n \right)\\

1511: 1-Z(t)&=&\displaystyle  2 (-z)^{1/4} n^{-1/4}\left( 1 +  O( n^{-1/4} \sqrt{\log n})\right)\\

1512: %

1513: %

1514: \\

1515: Z(t)^j& =& \displaystyle \exp(-2\lambda (-z)^{1/4}) \left(1+O(n^{-1/4}{\log

1516:   n})\right)

1517: &\displaystyle (\hbox{recall } j=\lfloor \lambda n ^{1/4}\rfloor)\\

1518: %

1519: %

1520: \\

1521: \displaystyle t^{-n-1} &=&\displaystyle \displaystyle 8^{n+1} e^{-z} \left( 1+ O(\log^4 n /n)\right).

1522: \end{array}

1523: \right.

1524: \eeq

1525: Observe  that, for $z \in \Ha$, the real part of $(-z)^{1/4}$

1526: is bounded from below by a positive constant $\alpha$. 

1527: Hence

1528: $$

1529: |\exp(-2\lambda (-z)^{1/4})|=  \exp(-2\lambda \Re (-z)^{1/4}) \le 

1530: \exp(-2\lambda \alpha),

1531: $$

1532: so that $\exp(-2\lambda (-z)^{1/4}) $ does not approach $1$. 

1533: This allows us to write

1534: $$

1535: \frac 1 {1-Z^{j+2}} = \frac 1 { 1 -\exp(-2\lambda (-z)^{1/4})}  \left(1+O(n^{-1/4}{\log

1536:   n})\right).

1537: $$

1538: %so that $\exp(-2\lambda (-z)^{1/4}) $ does not approach $1$. 

1539: Hence, uniformly in $t \in \C_n^ {(2)}$, we have

1540: \begin{eqnarray*}

1541: U_j(t) t^{-n-1}&=&

1542: \frac { \left( 1+Z \right) ^{2}{Z}^{j+1} \left( 1+Z+{Z}^{2} \right)  

1543: \left(1- Z \right) ^{2}}

1544: { \left( 1+{Z}^{2} \right) 

1545:  \left( 1-{Z}^{j+2} \right)  \left( 1-{Z}^{j+4} \right) }

1546: t^{-n-1}\\

1547: &=& \frac{6. 8^{n+1}}{n^{1/2}} 

1548: \frac{\sqrt{-z}e^{-z}}{\sinh^2(\lambda (-z)^{1/4})} (1+ O( n^{-1/4} \log n))

1549: \end{eqnarray*}

1550: with $8t=1+z/n$. 

1551: Let us now integrate this over ${\mathcal C}_n^{(2)}$:

1552: \begin{eqnarray*}

1553: \displaystyle \int_{{\mathcal C}_n^{(2)}}

1554:  U_j(t)\frac{dt}{t^{n+1}}

1555: &=&\frac{6. 8^n}{n^{3/2}} \int_{\Ha_n}

1556: \frac{\sqrt{-z}e^{-z}(1+ O( n^{-1/4} \log n))}{\sinh^2(\lambda (-z)^{1/4})}

1557: dz\\

1558: &=& \frac{6. 8^{n}}{ n^{3/2}} \left(

1559: \int_{\Ha}

1560: \frac{\sqrt{-z}e^{-z}}{\sinh^2(\lambda (-z)^{1/4})} dz

1561: +o(1)\right).

1562: \end{eqnarray*}

1563: %

1564: We now put together our estimates of the integrals on $\C_n^{(1)}$

1565: (Eq.~\eqref{C1-estimate}) and $\C_n^{(2)}$ and obtain

1566: $$

1567: [t^n] U_j(t) =\frac{6. 8^{n} n^{-3/2}}{2i \pi} \left(

1568: \int_{\Ha  }

1569: \frac{\sqrt{-z}e^{-z}}{\sinh^2(\lambda (-z)^{1/4})}dz

1570: +o(1)\right).

1571: $$

1572: Using~\eqref{GU} and the estimation $C_n\sim 4^n n

1573: ^{-3/2}/\sqrt \pi$, this gives

1574: $$

1575: \PP(N_n > \lambda )\rightarrow

1576: \frac{3}{i \sqrt \pi} \int_{\Ha}

1577: \frac{\sqrt{-z} e^{-z}}{\sinh^2(\lambda (-z)^{1/4})} dz.

1578: $$

1579: 

1580: 

1581: 

1582: 

1583: The next step in our proof of Theorem~\ref{thm-max} is to set 

1584: $v=(-z)^{1/4}$ in the above integral. As $z$ runs on $\Ha$, the

1585: variable $v$ runs on the contour $\J$ of Figure~\ref{figure-contours}, and the corresponding 

1586: integral is easily seen to coincide with the integral on the contour

1587: $\Gamma$ defined in the statement of the theorem. This gives the first

1588: expression of $G(\lambda)$.

1589: 

1590: 

1591: 

1592: \begin{figure}[pbth]

1593:   \begin{center}

1594:     \includegraphics[scale=0.3]{contours.eps}

1595:     \caption{The contours $\Gamma$ (two half lines) and $\J$.}

1596:     \label{figure-contours}

1597:   \end{center}

1598: \end{figure}

1599: 

1600: 

1601: We now want to express $G(\lambda)$ as a real integral. We first observe

1602: that the integration contour $\Gamma$ can be replaced by its

1603: translated version

1604: $$

1605: \Gamma_0= \{-re^{-i\pi/4}, r\in (\infty,0]\} \cup \{re^{i\pi/4},

1606: r\in [0,\infty)\}. 

1607: $$

1608: This parametrization of $\Gamma_0$ by $r$ splits the integral into two

1609: real integrals, and one finds:

1610: \begin{eqnarray*}

1611: G(\lambda) & =&

1612: -  \frac {12}{\sqrt \pi}\int_ 0^\infty \left( \frac 1{\sinh^2(\lambda

1613:   r e^{i\pi/4})} + \frac 1{\sinh^2(\lambda

1614:   r e^{-i\pi/4})} \right) r^5 e^{-r^4} dr \\

1615: &=&

1616:  \frac {48}{\sqrt \pi}\int_ 0^\infty  

1617: \frac{1-\cos (\sqrt 2 \lambda r) \cosh

1618:   (\sqrt 2 \lambda r)}{(\cosh (\sqrt 2 \lambda r) -\cos (\sqrt 2 \lambda r))^2}

1619: r^5 e^{-r^4} dr .

1620: \end{eqnarray*}

1621: The expected expression of $G(\lambda)$ % the limit of the

1622: 				% distribution function 

1623: follows, upon setting $u=\sqrt 2 \lambda r$.

1624: 

1625: \subsection{The limit law and its density}

1626: We now want to prove that

1627: $G(\lambda)$ is the tail distribution function of a random variable.

1628: Since it is the limit of

1629: non-increasing functions, it is non-increasing. Its integral

1630: expressions show that it is a continuous, and even a differentiable

1631: function of $\lambda$ on $(0, +\infty)$. In order to conclude, we

1632: still need to prove that~\cite[Thm.~14.1]{billingsley}

1633: $$

1634: \lim_{\lambda \rightarrow \infty}G(\lambda)=0 \hbox{ and } 

1635: \lim_{\lambda \rightarrow 0}G(\lambda)=1.

1636: $$

1637: In order to prove the first statement, we use the second expression of

1638: $G(\lambda)$ given in the theorem. We note that the function

1639: $$

1640: u\mapsto \frac{1-\cos u \cosh   u}{(\cosh u -\cos u)^2}

1641: $$

1642: is well-defined, bounded  and continuous on $[0, +\infty)$. Moreover,

1643:   as $u $ goes to infinity,

1644: $$

1645: \left|\frac{1-\cos u \cosh   u}{(\cosh u -\cos u)^2}\right|

1646: =O(e^{-u}),

1647: $$

1648: so that the integral

1649: $$

1650: \int_0^\infty\left| \frac{1-\cos u \cosh

1651:   u}{(\cosh u -\cos u)^2}\right| u^5 du

1652: $$

1653: is convergent. The term $1/\lambda ^6$ in the expression of

1654: $G(\lambda)$ then implies the convergence of $G(\lambda)$ to $0$ as

1655: $\lambda \rightarrow \infty$.

1656: 

1657: In order to study the limit of  $G(\lambda)$ as $\lambda \rightarrow

1658: 0^+$, we consider instead the first expression of $G(\lambda)$. Since

1659: $x^2/\sinh^2(x)$ is analytic in the disk of radius $\pi$, with expansion

1660:  $1-x^2/3+ O(x^4)$, there exists a constant $c$ such that for $|v| \le

1661: \pi /(2\lambda)$, 

1662: \beq

1663: \label{int-expansion}

1664: \left| \frac 1 {\sinh ^2(\lambda v)}-\frac 1 {\lambda^2 v^2} +\frac 1

1665: 3\right|

1666: \le c \lambda^2 |v|^2.

1667: \eeq

1668: Let us write

1669: $$

1670: \int_ \Gamma\frac{v^5 e^{v^4}}{\sinh^2(\lambda v) } dv

1671: =\int_ \Gamma\left( \frac{1}{\sinh^2(\lambda v)}  -\frac 1 {\lambda^2

1672:     v^2} +\frac 1 3\right) v^5 e^{v^4}dv 

1673: + \int_ \Gamma \left( \frac {v^3}{\lambda ^2} - \frac {v^5} 3\right)

1674: e^{v^4}dv .

1675: $$

1676: Recall the Hankel expression of the reciprocal of the Gamma function,

1677: valid for any $s\in \cs$:

1678: \beq

1679: \label{gamma-inverse}

1680: \frac 1 {\Gamma (s)} =\frac 1{2i \pi}

1681: \int_{\Ha} (-z)^{-s} e^{-z}dz

1682: = \frac 2 {i \pi}  \int_\Gamma v^{3-4s} e^{v^4} dv.

1683: \eeq

1684: Consequently,

1685: $$

1686: \int_ \Gamma   {v^3} e^{v^4}dv = \frac{i\pi}{2\Gamma(0)}=0 ,

1687: \quad  \quad

1688: \int_ \Gamma   {v^5} e^{v^4}dv =

1689: \frac{i\pi}{2\Gamma(-1/2)}=-\frac{i\sqrt {\pi}} 4,

1690: $$

1691: and we can rewrite

1692: $$

1693: G(\lambda)= \frac{12}{i\sqrt {\pi}}

1694: \int_ \Gamma\frac{v^5 e^{v^4}}{\sinh^2(\lambda v) } dv

1695: =1 + \frac{12}{i\sqrt {\pi}}\int_ \Gamma\left( \frac{1}{\sinh^2(\lambda v)}  -\frac 1 {\lambda^2

1696:     v^2} +\frac 1 3\right) v^5 e^{v^4}dv .

1697: $$

1698: Let us  cut the above integral into two parts,  $|v|\le

1699: \pi/(2\lambda)$ and  $|v|>\pi/(2\lambda)$. The first part is easily

1700: seen to tend to $0$ as $\lambda$ does, thanks

1701: to~\eqref{int-expansion}. For the second part, we observe that for

1702: $\lambda |v| >\pi/2$,

1703: $$

1704: \left| \frac 1 {\sinh ^2(\lambda v)}-\frac 1 {\lambda^2 v^2} +\frac 1

1705: 3\right|

1706: $$

1707: is bounded (by a constant independent of $\lambda$ and $v$), that the

1708: integral of $v^5e^{v^4}$ on $\Gamma$ is absolutely convergent, and

1709: that the contour $\{v\in \Gamma : |v|> \pi/(2\lambda)\}$ ``shrinks to

1710: $\infty$'' as $\lambda \rightarrow 0$. We finally conclude that $

1711: G(\lambda)$ tends to $1$ as $\lambda \rightarrow 0$.

1712: 

1713: Consequently, there exists a random variable $N$ having

1714: distribution function $1-G(\lambda)$, and $N_n$ converges in law to

1715: $N$. Since $G$ is differentiable, 

1716: $N$ has a density with respect to the Lebesgue measure on $\rs_+$,

1717: which is $f(\lambda)=-G'(\lambda)$. The two expressions of $G$ given

1718: in the theorem provide the two expressions of $f$.

1719: 

1720: 

1721: \subsection{The moments of $N$}

1722: Let us first prove that for all $k\ge 0$, the tail distribution function

1723: of $N$ satisfies

1724: \beq

1725: \label{G-infty}

1726: G(\lambda)= o(\lambda^{-k}) \hbox{ as } \lambda \rightarrow \infty.

1727: \eeq

1728: This is easily seen to imply the existence of moments of $N$ of all

1729: orders. In order to prove the above bound, we write

1730: $$

1731: G(\lambda)= \frac{24}{i\lambda \sqrt \pi} \int_\Gamma \frac

1732: {v^4(5+4v^4)e^{v^4}}

1733: {e^{2\lambda v}-1} dv.

1734: $$

1735: This is obtained from the first expression of $G( \lambda)$ using an

1736: integration by parts. Now, for $\lambda>0$ and $v\in \Gamma$,

1737: $$

1738: |e^{2\lambda v}-1|\ge |e^{2\lambda v}|-1= e^{2\lambda \Re(v)}-1

1739: \ge  e^{2\lambda }-1.

1740: $$

1741: From this, and from the term $e^{v^4}$ in the integral, we conclude

1742: that there exists a constant $c$ such that

1743: $$

1744: G(\lambda)\le \frac{c}{ e^{2\lambda }-1}.

1745: $$

1746: The bound~\eqref{G-infty}

1747: follows. This bounds also guarantees that for $k\ge 1$,

1748: \beq

1749: \label{moments-tail}

1750: \EE(N^k)= k\int_0^{\infty} \lambda^{k-1} G(\lambda) d\lambda.

1751: \eeq

1752: 

1753: \noindent

1754: {\bf The generic case: $k\ge 3$}.

1755: Recall the following integral representations of the Riemann zeta function: for

1756: $\Re(s)>1$,

1757: $$

1758: \zeta(s) =\frac 1 {\Gamma(s)} \int_0^\infty \frac{w^{s-1}}{e^w-1}dw

1759: =

1760: \frac 1 {4\Gamma(s+1)} \int_0^\infty \frac{w^s}{\sinh^2(w/2)} dw

1761: =

1762: \frac {2^{s-1}} {\Gamma(s+1)} \int_0^\infty \frac{y^s}{\sinh^2(y)} dy.

1763: $$

1764: The second expression follows from the first one after an integration

1765: by parts.

1766: 

1767: Let us now combine~\eqref{moments-tail} with the first expression of

1768: $G( \lambda)$:

1769: \beq

1770: \label{ENk-first}

1771: \EE(N^k)= \frac {12k}{i\sqrt\pi} \int_0^\infty \lambda^{k-1} d\lambda

1772: \int_\Gamma \frac{v^5e^{v^4 }}{\sinh^2(\lambda v)} dv.

1773: \eeq

1774: Assume for the moment that we can exchange the order of integration

1775: (this will be justified later). Exchange the integrals, and replace the

1776: variable $\lambda$ by $y/v$, where $y$ is a new variable:

1777: $$

1778: \EE(N^k)= \frac {12k}{i\sqrt\pi}

1779: \int_\Gamma v^{5-k}e^{v^4 }dv

1780:  \int_{v\rs_+}\frac{y^{k-1}}{\sinh^2(y)}dy.

1781: $$

1782: For $k\ge 3$, the function $y\mapsto {y^{k-1}}/{\sinh^2(y)}$ is

1783: meromorphic on $\cs$, with poles at $ik\pi$ for $k \in \zs$ and

1784: $k\not = 0$. From this, and from the strong decay of this function as

1785: $\Re(y) \rightarrow  \infty$, it follows that the integral on $y$ is

1786: actually independent of the choice of $v\in \Gamma$. In particular, it

1787: is equal to its value at $v=1$, which is

1788: $$

1789:  \int_0^\infty\frac{y^{k-1}}{\sinh^2(y)}dy = \frac{4 \Gamma(k)

1790:  \zeta(k-1)}{2^k},

1791: $$

1792: as recalled above. The integral on $v$ is then evaluated in terms of

1793: the Gamma function using~\eqref{gamma-inverse}, and the expected

1794: expression of $\EE(N^k)$ follows.

1795: 

1796: It remains to justify the exchange of integrals in~

1797: \eqref{ENk-first}. Observe that 

1798: $$

1799: |\sinh(y)|= |e^y -e^{-y}|/2 \ge \left( |e^y| -|e^{-y}|\right)/2

1800: =\sinh(\Re(y)),

1801: $$

1802: so that for $v \in \Gamma$,

1803: $$

1804: \frac 1{|\sinh^2 (\lambda v)|} \le \frac 1{\sinh^2 (\lambda )}.

1805: $$

1806: Moreover, the integral of $v^5e^{v^4 }$ along $\Gamma$ is absolutely

1807: convergent, and so is the integral of $\lambda^{k-1}/\sinh^2 (\lambda

1808: )$

1809: over $R_+$. It follows that the integral~\eqref{ENk-first}, once

1810: converted into two real integrals,  is

1811: absolutely convergent, so that the  integrals can be exchanged.

1812: %in~\eqref{ENk-first}. 

1813: 

1814: \medskip

1815: \noindent

1816: {\bf The case $k=1$.} We cannot apply exactly the same procedure as

1817: above, because the integral of  $1/\sinh^2 (\lambda)$ over $\rs_+$ is

1818: divergent.

1819: However, in view of~\eqref{gamma-inverse}, we can write

1820: $$

1821: G(\lambda) = \frac {12}{i\sqrt \pi} \int_\Gamma v^5e^{v^4} 

1822: \left( \frac 1 {\sinh^2(\lambda v)} - \frac 1 {\lambda^2 v^2}\right).

1823: $$

1824: Also, replacing $\Gamma$ by $\Gamma_0$ in the latter integral does not

1825: change its value.

1826: The technique is then the same as above:

1827: \begin{eqnarray}

1828: \EE(N)&= &\frac {12}{i\sqrt\pi} \int_0^\infty d\lambda

1829: \int_{\Gamma_0} {v^5e^{v^4 }}

1830: \left( \frac 1 {\sinh^2(\lambda v)} - \frac 1 {\lambda^2 v^2}\right)

1831: dv

1832: \label{EN-first}\\

1833: &=& \frac {12}{i\sqrt\pi}

1834: \int_{\Gamma_0} v^{4}e^{v^4 }dv

1835:  \int_{v\rs_+}\left( \frac 1 {\sinh^2(y)} - \frac 1

1836:      {y^2}\right)dy \nonumber

1837: \end{eqnarray}

1838: (assuming we can change the order of integration). Again, the integral

1839: on $y$ is independent of $v$, and equal to

1840: $$

1841: \int_0^\infty \left( \frac 1 {\sinh^2(y)} - \frac 1

1842:      {y^2}\right)dy = \left[ \frac 1 y - \frac 2

1843:        {e^{2y}-1}\right]_0^\infty

1844: =-1.

1845: $$

1846: Using again~\eqref{gamma-inverse} to evaluate the integral on $v$, one

1847: finds

1848: $$

1849: \EE(N)= -\frac {6\sqrt \pi}{\Gamma(-1/4)}=

1850: \frac{3\sqrt\pi}{2\Gamma(3/4)}.

1851: $$

1852: In order to justify the exchange of integrals in~\eqref{EN-first}, we

1853: wish to prove that~\eqref{EN-first}

1854: is absolutely convergent. In order to do so, we split the integral

1855: over $\Gamma_0$ into two real integrals, corresponding respectively to

1856: $v=re^{i\pi/4}$ and $v=re^{-i\pi/4}$. We are thus led to prove that

1857: $$

1858: \int_0^\infty d\lambda

1859: \int_0^\infty {r^5e^{-r^4 }}

1860:  \left|\frac 1 {\sinh^2(\lambda re^{i\pi/4})} - \frac 1 {i\lambda^2 r^2}\right|

1861: dr

1862: $$

1863: is absolutely convergent (and a similar result when $i$ is

1864: replaced by $-i$). But we can exchange the order of

1865: integration in this integral of \emm positive, functions. Doing so,

1866: and setting $\lambda= y/r$ as above, proves that this integral is

1867: finite.

1868: 

1869: \smallskip

1870: \noindent{\bf The case $k=2$.}

1871: Let us start from another expression of $G(\lambda)$, obtained by

1872: writing $v= w/\lambda$:

1873: $$

1874: G(\lambda)= \frac{12}{i\sqrt \pi \lambda ^6 }\int_{\lambda \Gamma}

1875: \frac{w^5}{\sinh^2(w)} e^{w^4/\lambda^4} dw

1876: = \frac{12}{i\sqrt \pi \lambda ^6 }\int_{ \Gamma}

1877: \frac{w^5}{\sinh^2(w)} e^{w^4/\lambda^4} dw.

1878: $$

1879: The second expression follows from the analyticity properties of the

1880: integrand. Now, take $\epsilon >0$, and let us evaluate

1881: \begin{eqnarray}

1882: \int_\epsilon ^\infty \lambda G(\lambda)d\lambda

1883: &= &\frac{12}{i\sqrt \pi}\int_\epsilon ^\infty \frac 1 {\lambda^5} d\lambda

1884: \int_{ \Gamma}

1885: \frac{w^5}{\sinh^2(w)} e^{w^4/\lambda^4} dw

1886: \label{N2-first}\\

1887: &= &\frac{12}{i\sqrt \pi}\int_{ \Gamma}

1888: \frac{w^5}{\sinh^2(w)}dw

1889: \int_\epsilon ^\infty \frac {e^{w^4/\lambda^4}} {\lambda^5}d\lambda

1890: \nonumber\\

1891: &= &\frac{12}{i\sqrt \pi}\int_{ \Gamma}

1892: \frac{w^5}{\sinh^2(w)}

1893: \left[ -\frac{e^{w^4/\lambda^4}}{4w^4} \right]_\epsilon ^\infty dw\nonumber\\

1894: &= &\frac{3}{i\sqrt \pi}\int_{ \Gamma} \frac{w}{\sinh^2(w)}

1895: \left( e^{w^4/\epsilon^4}-1\right) dw.\nonumber

1896: \end{eqnarray}

1897: The absolute convergence of integrals that legitimates  the exchange

1898: of integrals in~\eqref{N2-first} is, this time, obvious (thanks to the

1899: fact that $\lambda >\epsilon$). 

1900: Now, the analyticity of the function $w\mapsto {w}/{\sinh^2(w)}$ for

1901: $\Re(w)>0$, and its strong decay as $\Re(w)\rightarrow \infty$, imply

1902: that

1903: $$

1904: \int_{ \Gamma} \frac{w}{\sinh^2(w)}dw=0.

1905: $$

1906: Hence

1907: $$

1908: \begin{array}{llllllll}

1909: \displaystyle \int_\epsilon ^\infty \lambda G(\lambda)d\lambda

1910: &=&

1911: \displaystyle \frac{3}{i\sqrt \pi}\int_{ \Gamma} \frac{w

1912:   e^{w^4/\epsilon^4}}{\sinh^2(w)}dw

1913: &=&\displaystyle \frac{3\epsilon^2}{i\sqrt \pi}\int_{ \Gamma} \frac{v

1914:   e^{v^4}}{\sinh^2(\epsilon v)}dv\\

1915: &=&\displaystyle  \frac{3}{i\sqrt \pi}\int_{ \Gamma} \frac{

1916:   e^{v^4}}{v}dv + o(1)

1917: &=&\displaystyle \frac{3\sqrt\pi} 2 & \hbox{ (by \eqref{gamma-inverse})} .

1918: \end{array}

1919: $$

1920: Now, observe that

1921: $$

1922: 2\int_\epsilon ^\infty \lambda G(\lambda)d\lambda = \EE(N^2 \1_{N>

1923:   \epsilon}) - \epsilon^2 G(\epsilon).

1924: $$

1925:  The announced expression of the second moment of $N$ follows.

1926: 

1927: 

1928: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1929: \subsection{Convergence of the moments of $N_n$}

1930: \label{section-moments-Nn}

1931: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1932: In this section, we prove that the moments of $N_n=M_n/n^{1/4}$

1933: converge to the corresponding moments of $N$. In order to do so, we first

1934: express $\EE(M_n^k)$ as the coefficient of $t^n$ in a certain

1935: series. Then, we apply the general consequences of the analysis of

1936: singularities: if this series is regular enough (with a precise

1937: meaning of regular), one can derive the asymptotic behaviour of its

1938: coefficients from the singular behaviour of the series near its

1939: dominant singularities~\cite{flajolet-odlyzko}.

1940: 

1941: 

1942: Recall that the series $U_j$, given by~\eqref{Uj-expr}, counts the trees that

1943: contain at least one label larger than $j$. Hence $U_{j-1}-U_j$ counts

1944: the trees having maximal label $j$. Also, note that

1945: \beq

1946:  \label{UV}

1947: U_j= V(Z^j)-V(Z^{j+2}),

1948: \eeq

1949: where

1950: $$

1951: V(x)= \frac{xZ(1+Z)(1-Z^3)}{(1+Z^2)(1-xZ^2)}.

1952: $$

1953: Consequently, for $k\ge 1$,

1954: \beq

1955: \label{EMnk}

1956: \EE(M_n^k)= \frac 1 {2^n C_n} \sum_{j\ge 1} j^k [t^n] (U_{j-1}-U_j)

1957: = \frac 1 {2^n C_n}   [t^n]\sum_{j\ge 0} \left((j+1)^k-j^k\right)U_j.

1958:  \eeq

1959: For $k=1$, this gives

1960: $$

1961:  {2^n C_n}\EE(M_n)= [t^n]  \sum_{j\ge 0} \left(V(Z^j)-V(Z^{j+2})\right)= 

1962:  [t^n]\left( V(1)+V(Z)\right)

1963: = [t^n] \frac{Z(1+2Z+2Z^2)}{1+Z^2}.

1964: $$

1965: By Lemma~\ref{lemma-Z}, the latter series is analytic in

1966: $\cs\setminus[1/8, \infty)$.  

1967: %

1968: The generic consequences of the

1969:   analysis of singularities apply: one can derive the asymptotic

1970:   behaviour of the coefficients from the singular behaviour of the

1971:   series~\cite{flajolet-odlyzko}. Given that, when $t\rightarrow 1/8$,

1972: $$

1973: \frac{Z(1+2Z+2Z^2)}{1+Z^2}

1974: = \frac 5 6 -6 (1-8t)^{1/4} + O(\sqrt{1-8t}),

1975: $$

1976: the behaviour of the $n$th coefficient of this series is

1977: $$

1978:  [t^n] \frac{Z(1+2Z+2Z^2)}{1+Z^2} = -6 \frac {8^n

1979:  n^{-5/4}}{\Gamma(-1/4)} (1+o(1))

1980: =

1981: \frac 3 2  \frac {8^n

1982:  n^{-5/4}}{\Gamma(3/4)}(1+o(1)).

1983: $$

1984: It remains to divide by $2^n C_n \sim 8^n n^{-3/2}/ \sqrt \pi$ to

1985: conclude that

1986: $$

1987: \EE(M_n n^{-1/4}) \rightarrow \frac{3 \sqrt \pi}{2\Gamma (3/4)},

1988: $$

1989: which is also the first moment of $N$.

1990: 

1991: Now, by combining the expression~\eqref{UV} of $U_j$ and~\eqref{EMnk},

1992: one obtains, for $k\ge 2$,

1993: \beq

1994: \label{EMk}

1995: 2^nC_n \EE(M_n^k)=[t^n]\left( V(1)+(2^k-1)V(Z) +\sum_{j\ge 2}

1996: \left((j+1)^k-j^k-(j-1)^k+(j-2)^k\right) V(Z^j)\right).

1997: \eeq

1998: Observe that $(j+1)^k-j^k-(j-1)^k+(j-2)^k$ is a polynomial in $j$ of

1999: degree $k-2$ and leading coefficient $2k(k-1)$. Let

2000: $$

2001: A_\ell(t) = \sum_{j\ge -1}(j+2)^\ell V(Z^j).

2002: $$

2003: We are going to prove

2004: that, for $\ell\in\ns$,

2005: \beq

2006: \label{a-asympt}

2007: \displaystyle a_n(\ell):= [t^n] A_\ell(t)=\left\{

2008: \begin{array}{lll}

2009: \displaystyle \frac 3 4 \frac {8^n} n  &

2010: \hbox{ if }  \ell=0,\\

2011: \displaystyle \frac{3. 8^n \ell ! \zeta(\ell+1) n^{\ell/4 -1}}

2012: {2^\ell \Gamma(\ell/4)} & \hbox{ if } \ell \ge 1.

2013: \end{array}\right.

2014: \eeq

2015: Assume for the moment this is proved, and let us conclude about the

2016: limiting moments of $N_n=M_n n^{-1/4}$. First, we observe that for $j \ge 0$,

2017: $V(Z^j)$ has (only) a fourth root singularity, so

2018: that the coefficient of $t^n$ in $V(Z^j)$ grows like $8^n n^{-5/4}$,

2019: up to a multiplicative constant. This observation, combined

2020: with~\eqref{EMk} and the

2021: above asymptotics of $ a_n(\ell)$, implies that the dominant term in the

2022: asymptotic  behaviour  of $2^nC_n\EE(M_n^k)$ is that of $2k(k-1)

2023: a_n(k-2)$. After normalizing by $2^n C_n n^{k/4}$, this gives

2024: $$

2025: \EE(M_n^kn^{-k/4}) \rightarrow

2026: \left\{

2027: \begin{array}{lll}

2028: \displaystyle 3\sqrt \pi&

2029: \hbox{ if }  k=2,\\

2030: \displaystyle \frac{24 \sqrt \pi k!\zeta(k-1)}

2031: {2^k \Gamma((k-2)/4)} & \hbox{ if }k \ge 3.

2032: \end{array}\right.

2033: $$

2034: These limiting moments are exactly those of $N$.

2035: 

2036: \medskip

2037: It remains to study the asymptotic behaviour of the numbers

2038: $a_n(\ell)$ (for $\ell$ fixed, and $n$ going to infinity). We have:

2039: $$

2040: \begin{array}{llll}

2041:  \displaystyle A_\ell(t)&=&\displaystyle \frac{(1+Z)(1-Z^3)}{Z(1+Z^2)} 

2042: \sum_{j\ge 1}j^\ell   \frac {Z^{j}}{1-Z^{j}}\\

2043: &=&\displaystyle \frac{(1+Z)(1-Z^3)}{Z(1+Z^2)} \sum_{j\ge 1, m\ge 1}j^\ell Z^{jm}\\

2044: &=&\displaystyle \frac{(1+Z)(1-Z^3)}{Z(1+Z^2)} \sum_{N\ge 1} Z^N \sigma_\ell(N)

2045: \end{array}

2046: $$

2047: where

2048: $$ 

2049: \sigma_\ell(N)=\sum_{j|N} j^\ell.

2050: $$

2051: The function

2052: $$

2053: D_ \ell(z)=  \sum_{N\ge 1} z^N \sigma_\ell(N)

2054: $$

2055: is easily seen to have radius of convergence $1$. Moreover, as $z$

2056: tends to $1$ in such a way $|\arg(1-z) |<\phi < \pi/2$,

2057: $$

2058: D_ \ell(z)\sim \left\{

2059: \begin{array}{lll}

2060: \displaystyle \frac 1{1-z} \log \left(\frac 1 {1-z}\right) 

2061: & \hbox{ if } \ell=0,\\

2062: \displaystyle \frac{\ell !\zeta(\ell+1)}{(1-z)^{\ell+1}} &

2063:     \hbox { if } \ell \ge 1

2064: \end{array}

2065: \right.

2066: $$

2067: (this can be obtained using a Mellin

2068: transform~\cite{flajolet-binary-trees,flajolet-gourdon-dumas}).  

2069: The above expression of $A_\ell(t)$, combined with Lemma~\ref{lemma-Z}

2070: and these properties of $D_\ell(z)$, shows that $A_\ell(t)$ is analytic

2071: in the domain $\D=\cs\setminus[1/8, \infty)$.

2072: %, and has no other singularity that  $1/8$. 

2073: Moreover, since $|\arg(1-Z)| \le \pi /4 +o(1)$ as $t

2074:   \rightarrow 1/8$ in $\D$, we can use the above estimates of

2075:   $D_\ell(z)$. This gives

2076: $$

2077: A_\ell(t) \sim

2078: \left\{

2079: \begin{array}{lll}

2080: \displaystyle -3 \log 2 + \frac 3 4  \log \left(\frac 1 {1-8t}\right) 

2081: & \hbox{ if } \ell=0,\\

2082: \displaystyle \frac{3 \ell !\zeta(\ell+1)}{2^\ell (1-8t)^{\ell/4}} &

2083:     \hbox { if } \ell \ge 1.

2084: \end{array}

2085: \right.

2086: $$

2087:  The generic results derived from the

2088:   analysis of singularities apply, and give the asymptotic

2089:  behaviour~\eqref{a-asympt} of the numbers $a_n(\ell)$. 

2090: This concludes the proof of Theorem~\ref{thm-max}.

2091: 

2092: 

2093:  \subsection{The supremum of the support of the  ise}

2094: \label{section-supremum}

2095: Let us finally prove Proposition~\ref{coro-support}. 

2096: The following  argument 

2097: %is due to Jean-Fran\c cois Marckert. It 

2098: requires a

2099:  detour via discrete snakes and Brownian snakes. We refer

2100:  to~\cite{legall,marckert-mokka-snake,jf-janson} for definitions and

2101:  notation\footnote{We warn the reader that normalizations change from

2102:  one paper to another.}. In

2103:  particular, we  use the following integral representation of the

2104:  random measure $\mu_{\hbox{\sc ise}}$: %~\cite{legall,marckert-mokka-snake}: 

2105:  for any  continuous  bounded function $g$ on $\rs$,

2106: \beq

2107: \label{ise-rep}

2108: \int_\rs g(y) d\mu_{\hbox{\sc \small ise}}(y)= \int_0^1 g( r(t))dt

2109: \eeq

2110:  where $r(.)$ is a random process, continuous on $ [0,1]$, called the

2111:  head of the Brownian  snake. In other words, $\mu_{\hbox{\sc \small

2112:      ise}}$ is the \emm occupation measure, of the process $r$. (Again, the

2113:  definition of $r$ varies from 

2114:  one paper to the other. The above formula fixes our normalization of $r$.) 

2115: 

2116: 

2117:  The random variable $N_n= M_n n^{-1/4}$ coincides with

2118:  $\max(r_n)$, where $r_n$ is the (normalized) head of the discrete

2119:  snake associated with our tree family. The random process $\sqrt 2 r_n$

2120:  converges weakly to $r$, the head of the Brownian

2121:  snake~\cite{marckert-mokka-snake}.  Since $\max$ 

2122:  is a continuous functional on $\C[0,1]$, this implies that $

2123: \sqrt 2  N_n= \sqrt 2 \max (r_n)$  converges in distribution to $\max

2124:  (r)$. Thus $\max( r)$ has  density $f(\lambda/\sqrt 2)/\sqrt

2125:  2$, where $f$ is defined in Theorem~\ref{thm-max}. 

2126: % In particular,  $\max(r)$ does not charge points.

2127: 

2128: \begin{figure}[htb]

2129: \begin{center}

2130: \input{fepsilon.pstex_t}

2131: \end{center}

2132: \caption{The functions $f_{\lambda,\epsilon}$, $g_{\lambda,\epsilon}$ and  $h_{\lambda,\epsilon}$.}

2133: \label{fig-epsilon}

2134: \end{figure}

2135: 

2136: It remains to prove that $\max (r)$ is equal (in distribution) to 

2137: $N_{\hbox{\sc \small ise}}$, the supremum of the support of the  ISE.

2138: Let $\lambda\in \rs$ and $\epsilon >0$. Let

2139: $f_{\lambda, \epsilon} $ % and $g_{\lambda,\epsilon}$ 

2140: be the function plotted on the left-hand side of

2141: Figure~\ref{fig-epsilon}. We have  

2142: $$

2143: N_{\hbox{\sc \small ise}} \le \lambda \Longleftrightarrow 

2144: \mu_{\hbox{\sc \small ise}}(-\infty, \lambda]=1 \Longleftrightarrow\int_\rs

2145: f_{\lambda,1} (y) d\mu_{\hbox{\sc \small ise}}(y)=1.

2146: $$

2147: Thanks to~\eqref{ise-rep}, this gives

2148: $$

2149: N_{\hbox{\sc \small ise}} \le \lambda \Longleftrightarrow 

2150:  \int_0^1 f_{\lambda,1} ( r(t)) dt =1

2151: $$

2152: Taking probabilities yields to

2153: $$

2154: \PP( N_{\hbox{\sc \small ise}} \le \lambda) = 

2155:  \PP\left( 

2156: \int_0^1 f_{\lambda,1} (r(t)) dt =1

2157: \right)

2158: =\PP(\max(r) \le \lambda),

2159: $$

2160: since $r$ is almost surely continuous.

2161: 

2162: 

2163: 

2164: 

2165: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2166: \section{A local limit law}

2167: \label{section-limit-local}

2168: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2169: 

2170: 

2171: For $j \in \zs$, let $X_n(j)$ denote the random variable equal to the

2172: number of nodes having label $j$ in a random tree of  $\mathcal T

2173: _{0,n}$. This quantity is related to the series $S_j(t,u)$ studied in

2174: Proposition~\ref{propo-snake-gf}. In particular,

2175: $$

2176: %P\left(X_n(j)=m\right)=  \frac{[u^mt^n] S_j}{2^n C_n}.

2177: \EE\left(e^{aX_n(j)}\right)=  \frac{[t^n] S_j(t,e^a)}{2^n C_n}.

2178: $$

2179: Also, observe that

2180: $$

2181: X_n(j)=0 \Longleftrightarrow M_n <j,

2182: $$

2183: where $M_n$ is the largest label, studied in the previous section.

2184: Let us define a normalized version of $X_n(j)$ by

2185: $$

2186: Y_n(j)= \frac{X_n(j)}{n^{3/4}}.

2187: $$

2188: Let $\lambda \in \rs$. The aim of this section is to prove that

2189: $Y_n(\lfloor \lambda n^ {1/4}\rfloor)$ converges in distribution, as

2190: $n$ goes to infinity,  to a random variable $Y(\lambda)$ that we

2191: describe by its Laplace transform.  This is achieved in

2192: Theorem~\ref{thm-conv} below, but we first want to present two

2193: consequences of this theorem, which have a simpler formulation. The

2194: first consequence deals with the case $\lambda=0$. Recall

2195: that, up to a normalization by $n^{3/4}$, the random variable

2196: $Y_n(0)$ gives the number of nodes labelled $0$ in a tree rooted at $0$.

2197: %

2198: \begin{Propo}[{\bf The number of nodes labelled $0$}]

2199: \label{limit-0}

2200: As $n$ goes to infinity, the random variable $3Y_n(0)/\sqrt 2$

2201: converges in distribution to $T^{-1/2}$, where 

2202:  $T$ follows a unilateral stable law of parameter $2/3$.

2203: The convergence of the moments holds as well: for $k\ge 0$,

2204: $$

2205: \EE\left( Y_n(0)^k\right) \rightarrow

2206: \left(\frac{\sqrt 2} 3\right)^k \frac{\Gamma(1+3k/4)}{\Gamma(1+k/2)}=

2207: \left(\frac{\sqrt 2} 3\right)^k \EE(T^{-k/2}). 

2208: $$

2209: \end{Propo}

2210: \noindent  

2211: This proposition will be proved in Section~\ref{moments-proof}.

2212: I am indebted to Alain Rouault, who recognized that the above moments

2213: were related to $T$.

2214:  Recall that $T$ is given by its Laplace transform:

2215: $$

2216: \EE(e^{-aT})= e^{-a^{2/3}} \hbox{ for } a\ge 0.

2217: $$

2218: 

2219: 

2220: \smallskip

2221: 

2222: The second consequence of Theorem~\ref{thm-conv} is an explicit

2223: expansion in $\lambda$ of the limiting first moment of $Y_n(j)$.

2224: \begin{Propo}[{\bf The first moment}]

2225: \label{propo-first}

2226:  Let $\lambda \in \rs$.  Denote $j

2227: =\lfloor \lambda n^{1/4}\rfloor$. Then, as $n$ goes to infinity,

2228: $$

2229: \EE\left( Y_n( j) \right)

2230: \rightarrow

2231: \frac 1 {\sqrt\pi}\sum_{m\ge 0} \frac{(-2|\lambda|)^m}{m!} \cos %\left(

2232: \frac{(m+1)\pi}4%\right)

2233: \Gamma\left(\frac{m+3}4\right).

2234: $$

2235: This function of $\lambda$ is plotted on Figure~{\em\ref{fig-profil}}.

2236: \end{Propo}

2237: \noindent

2238: Similar, but more and more complicated expressions may be written for

2239: the next moments of $Y(\lambda)$. This proposition will be proved in

2240: Section~\ref{first-proof}.

2241: Let us, finally, state our main theorem, from which the two above

2242: propositions derive.

2243: 

2244: \begin{figure}[pbth]

2245:   \begin{center}

2246:     \includegraphics[height=5cm,width=5cm]{profil-vertical-moyen.eps}

2247:     \caption{The average number of nodes labelled $\lfloor

2248:     \lambda\sqrt n\rfloor $ in

2249:     a tree of size $n$, when $n\rightarrow \infty$.}

2250:     \label{fig-profil}

2251:   \end{center}

2252: \end{figure}

2253: 

2254: 

2255: \begin{Theorem} [{\bf A local limit law}]

2256: \label{thm-conv}

2257: Let $\lambda\ge 0$.

2258: The sequence $Y_n(\lfloor \lambda n^{1/4}\rfloor)$ converges in

2259: distribution to a non-negative 

2260: random variable $Y(\lambda)$ whose Laplace transform is given, for

2261: $|a| < 4/\sqrt 3$,  by

2262: $$

2263: \EE\left( e^{aY(\lambda)}\right)=L(\lambda,a)

2264: $$

2265: where

2266: $$

2267: L(\lambda, a)=

2268: 1+ \frac {48}{i\sqrt \pi }\int_{\Gamma}

2269: \frac{A(a/v^3)e^{-2\lambda v}}

2270: {(1+A(a/v^3)e^{-2\lambda v})^2}v^{5} e^{v^4} dv,

2271: $$

2272: $A(x)\equiv A$ is the unique solution of

2273: \beq

2274: \label{A-eq}

2275: A=\frac x{24} \frac{(1+A)^3}{1-A}

2276: \eeq

2277: satisfying $A(0)=0$,

2278: and the integral is taken over

2279: $$

2280: \Gamma= \{1-te^{-i\pi/4}, t\in (\infty,0]\} \cup \{1+te^{-i\pi/4},

2281: t\in [0,\infty)\}. 

2282: $$

2283: More precisely, the Laplace transform of $Y_n(\lfloor \lambda n^{1/4}\rfloor)$

2284: converges pointwise to $L(\lambda, \cdot)$ on the interval  $(-

2285: 4/\sqrt 3, 4/\sqrt 3)$. The convergence of moments holds as well.

2286: \end{Theorem}

2287: It is believed (or known?) that the random measure $\mu_{{\hbox{\sc

2288:       \small ise}}}$ is almost surely absolutely continuous with

2289:       respect to the Lebesgue measure on

2290:       $\rs$. Eq.~\eqref{limit-density} leads us to the following conjecture.

2291: \begin{Conj}[{\bf The density of the ISE}]

2292: \label{conj-density}

2293:  There exists a random continuous process

2294: $f_{{\hbox{\sc \small ise}}}(\lambda)$, defined for $\lambda\in \rs$,

2295:  such that 

2296: $\mu_{{\hbox{\sc \small ise}}}=f_{{\hbox{\sc \small ise}}}\hbox{Leb}$, where $\hbox{Leb}$ denotes the Lebesgue measure

2297:  on  $\rs$. Moreover,  $f_{{\hbox{\sc \small ise}}}(\lambda)$ satisfies

2298: $$

2299: f_{{\hbox{\sc \small ise}}}(\lambda)\deq \frac 1 {\sqrt 2} \,

2300: Y\left(\frac{|\lambda|}{\sqrt 2}\right), 

2301: $$

2302: where the law of  $Y(\lambda)$ is given in

2303: Theorem~{\em\ref{thm-conv}.}

2304: \end{Conj}

2305: \noindent{\bf Comments}\\

2306: {\bf 1.} The limit random variable $Y(\lambda)$ equals $0$

2307: with a \emm positive probability,  as soon as $\lambda>0$. Indeed, by

2308: the  portmanteau Theorem~\cite[Thm.~11.1.1]{dudley}, 

2309: $$

2310: \PP(Y(\lambda)=0) \ge \limsup \PP(Y_n(\lfloor \lambda n^{1/4}\rfloor

2311: )=0) 

2312: = \limsup \PP(M_n< \lfloor \lambda n^{1/4}\rfloor).

2313: $$

2314: But, by Theorem~\ref{thm-max}, 

2315: $$

2316: \PP(M_n< \lfloor \lambda n^{1/4}\rfloor) \rightarrow 1-G(\lambda) >0.

2317: $$

2318: {\bf 2.} Let us add a few words on the series $A(x)$ defined

2319: by~\eqref{A-eq}, in order to convince ourselves that the integral

2320: giving $L(\lambda, a)$ is well-defined.

2321:  Clearly, the expansion of $A(x)$ at $x=0$ has

2322: non-negative coefficients. Looking at the discriminant of the equation that

2323: defines $A$ shows that $A$ has radius of convergence at least $4/\sqrt

2324: 3$. Moreover,  it is

2325: easy to prove that $A(4/\sqrt 3)=2-\sqrt 3=0.26\ldots$ Consequently, $|A(x)|$ is

2326: bounded by $2-\sqrt 3$ for $|x|\le 4/\sqrt 3$.\\

2327:  Since $|v|\ge 1$ for $v \in

2328: \Gamma$, the modulus of $A(a/v^3)$ is bounded from above by $2-\sqrt

2329: 3$. Moreover, $\Re(v)\ge 1$, so that $|e^{-2\lambda v}|\le

2330: e^{-2\lambda}<1$. Hence 

2331: $$

2332: \frac{A(a/v^3)e^{-2\lambda v}}

2333: {(1+A(a/v^3)e^{-2\lambda v})^2}

2334: $$

2335: is uniformly bounded on $\Gamma$, and $L(\lambda, a)$ is well-defined.

2336: 

2337: Note that the series $A(x)$ admits the following closed form expression:

2338: \beq

2339: \label{A-explicit}

2340: A(x)= \frac 2 {1+\frac 2 {\sqrt 3} \cos(\frac{\arccos(-x\sqrt 3 /4)} 3)}-1.

2341: \eeq

2342: This can be checked by proving that this expression

2343: satisfies~\eqref{A-eq} and the initial condition $A(0)=0$. 

2344: 

2345: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2346: \subsection{Proof of Theorem~\ref{thm-conv}}

2347: \label{local-proof}

2348: Let $\lambda \ge 0$ and $j=\lfloor \lambda n^{1/4}\rfloor$. Let us

2349: first express the Laplace transform of $Y_n(j)$  in terms of the 

2350: \gfs\ $S_j(t,u)$ of Proposition~\ref{propo-snake-gf}:

2351: \beq

2352: \label{moment-series}

2353: \EE\left( e^{aY_n(j)}\right)= \EE\left( e^{an^{-3/4}X_n(j)}\right)

2354: =\frac{[t^n] S_j(t, e^{an^{-3/4}})}{2^nC_n}.

2355: \eeq

2356: 

2357: Again, we will evaluate this Laplace transform thanks to the analysis

2358: of singularities~\cite{flajolet-odlyzko}.

2359: % The expression of $S_j$ is given in Proposition~\ref{propo-snake-gf}. 

2360: We wish to use again the integration contour $\C_n$ of

2361: Figure~\ref{fig-contour}. This requires to prove that $S_j(t,u)$ is 

2362: analytic in a neigborhood of  this contour (for $n$ large and $u=

2363: e^{an^{-3/4}}$). This is guaranteed by the following lemma. This

2364: lemma naturally includes some properties of the series $\mu$ involved

2365: in the product form~\eqref{product-formSj} of $S_j$.

2366: We denote by  $\I_n$ the part of the complex plane

2367: enclosed by $\C_n$ (including $\C_n$ itself).

2368: %

2369: \begin{Lemma}[{\bf Analytic properties of $\mu$ and $S_j$}]

2370: \label{lemma-Sj}

2371: Let $a$ be a real number such that $|a|<4/\sqrt 3$.

2372: Then there exists $\epsilon >0$ such that for $n$

2373: large enough, the series $\mu(t,u_n)$, with $u_n=e^{an^{-3/4}}$, is

2374: analytic in the domain

2375: $$

2376: \E_n =\{t : \left|t- 1/ 8 \right| >  1/ ((8+\epsilon)n)\}\setminus[ 1/ 8  , +\infty) .

2377: $$

2378: In particular, $\mu(t,u_n)$ is analytic in a neighborhood of  $\I_n$. 

2379: Its modulus in  $\I_n$ is smaller than $\alpha$, for some $\alpha<1$

2380: independent of $a$ and $n$.

2381: %

2382: The series $S_j(t,u_n)$ is also analytic in a neighborhood of  $\I_n$. 

2383: \end{Lemma}

2384: \noindent

2385: {\bf Proof.}

2386: The lemma is clear if $a=0$: in this case, $u_n=1$, the series

2387: $\mu(t,u_n)$ vanishes, and the series $S_j$ reduces to the size \gf\

2388: of labelled trees, namely $T$, which is analytic in

2389: $\cs\setminus[1/8,\infty)$.

2390: %

2391: We now assume that $a\not = 0$ and $|a|< 4/\sqrt 3$. This guarantees

2392: that $A(a)$ is well-defined, where the series $A$ is defined in

2393: Theorem~\ref{thm-conv}. 

2394: 

2395: 

2396: Let us first study the singularities of the series $\bar \mu\equiv

2397: \bar \mu(z,u)$ defined as the unique \fps \ in $z$ satisfying

2398: $$

2399: \bar \mu= (u-1) \frac{(1+z^2)(1+\bar \mu z)(1+\bar \mu z^2)(1+\bar \mu z^3)}

2400: {(1+z)(1+z+z^2)(1-z)^3(1-\bar \mu z^2)}.

2401: $$

2402: Note that $\bar \mu$ has polynomial coefficients in $u$, and vanishes

2403: when $u=1$.   

2404: Assume that $u$ is a fixed real

2405: number close to, but different from, $1$. Recall that, as all

2406: algebraic \fps , $\bar\mu(t,u)$ has a positive radius of convergence.

2407: Let us perform a classical analysis to detect its possible

2408: singularities. These singularities are found in the

2409: union of two sets $\S_1$ and $\S_2$:

2410: 

2411: $\bullet$   $\S_1$ is the set of  non-zero roots of the dominant

2412: coefficient of the 

2413: equation defining $\bar \mu$. That is,  $\S_1=\{\pm i\}$ ,

2414: 

2415: 

2416: $\bullet$   $\S_2$ is the set of the roots of the discriminant  of the

2417: equation defining $\bar \mu$. For $u=1+x$ and $x$ small, these roots

2418: are found to be

2419: $$

2420: z=\pm 1,\quad  z=-1+O(x),\quad z=e^{\pm 2i\pi/3} +O(x),\quad

2421: z=1+ \omega 12 ^{1/6}x^{1/3} + O(|x|^{2/3}),

2422: $$

2423: where $\omega$ satisfies $\omega ^6=1$. (The term $\omega$ allows us

2424: to write loosely 

2425: $x^{1/3}$ without saying which determination of the cubic root we take.)\\

2426: %

2427: Observe that the moduli of all these ``candidates for singularities''

2428: go to $1$ as $x$ goes to $0$. 

2429: 

2430: Now the series $\mu=\mu(t,u)$ involved

2431: in the expression~\eqref{product-formSj} of $S_j(t,u)$ satisfies

2432: $$

2433: \mu(t,u)= \bar \mu(Z(t), u)

2434: $$

2435: where $Z(t)$ is defined by~\eqref{eqZt}. In other words, we could have

2436: defined the series $\bar \mu$ by

2437: $$

2438: \bar \mu(z,u) = \mu\left(\frac{z(1+z^2)}{(1+z)^4}, u\right).

2439: $$

2440: Recall that $Z$ is analytic

2441: in the domain

2442: $\D=\cs\setminus [1/8, \infty)$.  Take $u=u_n=

2443: e^{an^{-3/4}}=1+x$, with $x=an^{-3/4}(1+o(1))$. By Lemma~\ref{lemma-Z}, for $n$

2444: large, the only values of $\S_1 \cup \S_2$ that may be reached by

2445: $Z(t)$, for $t \in\D $, are of the  form

2446: $$

2447: z=1+ \omega 12^{1/6}a^{1/3}n^{-1/4} + O(n^{-1/2}).

2448: $$

2449: In view of~\eqref{eqZt}, these values of $Z(t)$ are reached for

2450: $$

2451: t= \frac 1 8 - \frac{\omega ^4 (12)^{2/3}}{128} \frac{a^{4/3} } n + O(n^{-5/4}).

2452: $$

2453: Since $|a|< 4/\sqrt 3$, these values of $t$ are at distance less than

2454: $1/((8+\epsilon)n)$ of $1/8$, for some $\epsilon >0$, and hence

2455: outside the domain $\E_n$. Consequently, 

2456: $\mu(t,u_n)$ is analytic inside $\E_n$.

2457: 

2458: \medskip

2459: We now want to bound $\mu(t,u_n)$ inside $\I_n$. Let $t_n\in \I_n$ be such

2460: that

2461: $$

2462: |\mu(t_n,u_n)|= \max _{t\in \I_n} |\mu(t,u_n)|.

2463: $$

2464: In particular, $|\mu(t_n,u_n)| \ge |\mu((1-1/n)/8,u_n)|$. In order to

2465: evaluate the latter quantity, note that $Z((1-1/n)/8)=1-2n^{-1/4}+O(n

2466: ^{-1/2})$. Thanks to the closed form expression of $\mu$ given in

2467: Proposition~\ref{propo-mu-closed-form}, and to the

2468: expression~\eqref{A-explicit} of the series $A$, we see that

2469: $\mu((1-1/n)/8,u_n)\rightarrow A(a)$. Since $a\not = 0$, $A(a) \not

2470: =0$, and for $n$ large enough,

2471: \beq

2472: \label{mu-lower}

2473: |\mu(t_n,u_n)| \ge |\mu((1-1/n)/8,u_n)|=|A(a)|+o(1)>0.

2474: \eeq

2475: Recall that all the sets $\I_n$ are included in a ball of finite

2476: radius centered at the origin. Let $\alpha$ be an accumulation point

2477: of the sequence $t_n$. Then  $|\alpha|\le 1/8$. 

2478: 

2479: Assume first that  $\alpha \not = 1/8$. Then there exists $N$ such

2480: that $\alpha$ is in $\E_n$  for all $n\ge N$, that is, in the

2481: analyticity domain of $\mu(\cdot,u_n)$. Let $t_{n_1}, t_{n_2}, \ldots$

2482: converge to $\alpha$.  By continuity of $\mu$ in

2483: $t$ and $u$, we have

2484: $$

2485: \mu(t_{n_i},u_{n_i}) \rightarrow  \mu(\alpha,1)=0.

2486: $$

2487: This contradicts~\eqref{mu-lower}. Hence the only accumulation point

2488: of $t_n$ is $1/8$, and $t_n$ converges to $1/8$.

2489: Let us thus write

2490: $$

2491: t_n= \frac 1 8 \left( 1 -\frac {x_n} n \right).

2492: $$

2493: We have $x_n=o(n)$, but also $|x_n|>1$ since $t_n$ belongs to $\I_n$.

2494: We wish to estimate $\mu(t_n, u_n)$.

2495: %, using~\eqref{eqMZ}. 

2496: From the singular behaviour of $Z$ (Lemma~\ref{lemma-Z}), we derive

2497: $$

2498: Z(t_n)= 1-2  \left( \frac {x_n} n \right)^{1/4} + O\left(  \left(

2499: \frac {x_n} n \right)^{1/2}\right).

2500: $$

2501: Moreover,

2502: $$

2503: u_n-1=an^{-3/4} \left( 1 + O(n^{-3/4})\right).

2504: $$

2505: This gives

2506: $$

2507: \frac{u_n-1}{(1-Z)^3}= \frac{a}{8x_n^{3/4}}\left(1+ O\left(  \left(

2508: \frac {x_n} n \right)^{1/4}\right)\right).

2509: $$

2510:  If the sequence $x_n$ was

2511: unbounded, then there would exist a subsequence $x_{n_i}$ converging to

2512: infinity. Then $({u_{n_i}-1})/{(1-Z)^3}$ would tend to $0$.   The closed

2513: form expression of $\mu$ given in 

2514: Proposition~\ref{propo-mu-closed-form} implies that $\mu(t_{n_i},u_{n_i})$

2515: would tend to $0$, contradicting~\eqref{mu-lower}. Hence the sequence

2516: $x_n$ is bounded, and one derives from the explicit expressions of

2517: $\mu$ and $A$ that

2518: $$

2519: \mu(t_n,u_n)= A(ax_n^{-3/4}) + o(1).

2520: $$

2521: Since $A$ is bounded by  $2-\sqrt 3$ inside its disk of convergence,

2522: $|\mu(t_n,u_n)|$ is certainly smaller than some $\alpha$ for

2523: $\alpha<1$ and $n$ large enough.

2524: This concludes the proof of the second statement of Lemma~\ref{lemma-Sj}.

2525: 

2526: \medskip

2527: By continuity of $\mu(t,u_n)$, this function of $t$ is still bounded by $1$

2528: (in modulus) is a neighborhood of $\I_n$. Recall also that the modulus

2529: of $Z(t)$ never reaches $1$ for $t \in \cs \setminus [1/8,

2530:   \infty)$. The form~\eqref{product-formSj} then implies that

2531:   $S_j(t,u_n)$ is an analytic function of $t$ in a neigbourhood of $\I_n$.

2532: \cqfd

2533: 

2534: Let us now go back to the expression~\eqref{moment-series} of the

2535: Laplace transform of $Y_n(j)$. Thanks to the lemma we have just

2536: proved, we can use the Cauchy formula to extract the coefficient of

2537: $t^n$ in $S_j(t,u_n)$. We use the following expression of $S_j$:

2538: $$

2539: S_j=T+ T\frac{(1-Z)^2(1+Z+Z^2)\mu Z^j}{(1+\mu Z^{j+1})(1+\mu Z^{j+3})},

2540: $$

2541: which is easily derived from~\eqref{product-formSj}. Thus

2542: $$

2543: [t^n] S_j(t,u_n)= 2^nC_n + \frac 1{2i\pi} \int_{\C_n}

2544:  T\frac{(1-Z)^2(1+Z+Z^2)\mu Z^j}{(1+\mu Z^{j+1})(1+\mu Z^{j+3})}

2545: \frac{dt}{t^{n+1}}.

2546: $$

2547: Again, we split the contour $\C_n$ into two parts $\C_n^{(1)}$ and

2548: $\C_n^{(2)}$, shown in Figure~\ref{fig-contour}. As in the proof of

2549: Theorem~\ref{thm-max}, the contribution of

2550: $\C_n^{(1)}$ is easily seen to be $o(8^n/n^m)$ for all $m>0$, thanks

2551: to the results of Lemmas~\ref{lemma-Z} and~\ref{lemma-Sj}.  On

2552: $\C_n^{(2)}$, one has 

2553: $$

2554: t= \frac 1 8 \left( 1 + \frac z n\right)

2555: $$

2556: where $z$ lies in the truncated Hankel contour $\Ha_n$. Conversely,

2557: let $z \in\Ha$. Then $z \in\Ha_n$ for $n$ large enough, and, in

2558: addition to the estimations~\eqref{estimates} already used in the proof of

2559: Theorem~\ref{thm-max}, one finds

2560: \beq

2561: \label{muA}

2562: \mu(t,u_n)= A(a(-z)^{-3/4})(1+o(1)),

2563: \eeq

2564: where $A(x)$ is the series defined by~\eqref{A-eq}. After a few

2565: reductions, one finally obtains

2566: $$

2567: [t^n] S_j(t,u_n)= 2^nC_n +

2568: \frac {12. 8^n n^{-3/2} }{i  \pi}\int_{\Ha }

2569: \frac{A(a(-z)^{-3/4})\exp(-2\lambda (-z)^{1/4})\sqrt{-z}e^{-z}}

2570: {(1+A(a(-z)^{-3/4})\exp(-2\lambda (-z)^{1/4}))^2}dz + o(8^n n^{-3/2}).

2571: $$

2572: It remains to normalize by $2^nC_n=8^n n^{-3/2}/\sqrt \pi$, and then to

2573: set $v=(-z)^{1/4}$ to obtain the expected

2574: expression for the limit of the Laplace transform of $Y_n(j)$, with

2575: $j=\lfloor \lambda n ^{1/4}\rfloor$.

2576: 

2577: \medskip

2578: The limit Laplace transform $L(\lambda,a)$ is clearly continuous at

2579: $a=0$, and equals $1$ at this point. 

2580: A version of L\'evy's continuity theorem~\cite[Thm.~9.8.2]{dudley}

2581: adapted to Laplace transforms  implies that the sequence 

2582: $Y_n(j)$ converges in distribution to a limit random variable

2583: $Y(\lambda)$ having Laplace transform $L(\lambda, \cdot)$.

2584: 

2585: \medskip

2586: From the convergence of the Laplace transform in a neighbourhood of

2587: the origin, it is easy to derive that for every $k$,  the sequence of

2588: random variables $Y_n(j)^k$ is \emm uniformly integrable,. 

2589: But then the convergence in distribution implies the convergence of the

2590: moments~\cite[Thm.~5.4]{billingsley2}. This concludes the proof of

2591: Theorem~\ref{thm-conv}. 

2592: 

2593: 

2594: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2595: \subsection{Proof of Proposition~\ref{limit-0}}

2596: \label{moments-proof}

2597: When $\lambda=0$, the limiting Laplace transform reduces to

2598: $$

2599: L(0, a)=

2600: 1+ \frac {48}{i\sqrt \pi }\int_{\Gamma}

2601: \frac{A(a/v^3)}

2602: {(1+A(a/v^3))^2}v^{5} e^{v^4} dv

2603: =

2604: 1+ \frac {12}{i\sqrt \pi }\int_{\Gamma}

2605: \frac{\chi(a/v^3)}

2606: {1+\chi(a/v^3)}v^{5} e^{v^4} dv

2607: $$

2608: where $\chi(x)$ is the unique series in $x$ satisfying

2609: $$

2610: \chi=\frac x 6(1+\chi)^{3/2}.

2611: $$

2612: The Lagrange inversion formula~\cite[p.~38]{stanley-vol2} gives, for $k \ge 1$,

2613: $$

2614: [x^k]  \frac{\chi(x)} {1+\chi(x)}

2615: = \frac 1 {6^k} \frac{\Gamma(3k/2-1)}{k!\Gamma(k/2)}.

2616: $$

2617: Consequently,

2618: $$

2619: L(0, a)=1+ \frac {12}{i\sqrt \pi }

2620: \int_\Gamma

2621: \sum_{k\ge 1} \frac 1 {6^k} \frac{\Gamma(3k/2-1)}{k!\Gamma(k/2)} a^k

2622: v^{5-3k} e^{v^4} dv.

2623: $$

2624: The convergence is absolute, so that we can exchange the sum and the

2625: integral:

2626: $$

2627: L(0, a)=1+ \frac {12}{i\sqrt \pi }

2628: \sum_{k\ge 1} \frac 1 {6^k}

2629: \frac{\Gamma(3k/2-1)}{k!\Gamma(k/2)} a^k

2630: \int_\Gamma

2631: v^{5-3k} e^{v^4} dv.

2632: $$

2633: Using~\eqref{gamma-inverse}, and picking the coefficient of $a^k$, we

2634: find that the 

2635: $k$th moment of the random variable $Y(0)$ is

2636: $$

2637: \EE(Y(0)^k)=\frac {\sqrt \pi}{ 6^{k-1}}

2638:  \frac

2639: {\Gamma(3k/2-1)}{\,\Gamma(k/2)\Gamma((3k-2)/4)}.

2640: $$

2641: The duplication formula,

2642: \[2^{2s-1} \Gamma(s) \Gamma(s+1/2) = \sqrt \pi \, \Gamma(2s),

2643: \]

2644: applied to  $s=(3k-2)/4$, finally gives

2645: $$

2646:  \EE(Y(0)^k)=

2647:  \left( \frac{\sqrt 2} 3\right)^k

2648: \frac {\Gamma(1+3k/4)}{\,\Gamma(1+k/2)} =\lim_{n\rightarrow \infty} 

2649: \EE\left( Y_n( 0) ^k\right).

2650: $$

2651: Since $Y(0)$ has a Laplace transform, it is uniquely determined by its

2652: moments~\cite[Thm.~30.1]{billingsley}. But

2653: $$m_k= \frac

2654: {\Gamma(1+3k/4)}{\,\Gamma(1+k/2)}

2655: $$

2656: is known to be the  $k$th

2657: moment  of $T^{-1/2}$, where $T$ follows a unilateral stable law

2658: of parameter $2/3$

2659: (see~\cite[p.~111]{chaumont-yor}). Proposition~\ref{limit-0} follows.

2660: 

2661: 

2662: 

2663: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2664: \subsection{Proof of Proposition~\ref{propo-first}}

2665: \label{first-proof}

2666: We have derived above the moments of $Y(0)$ from the expression of its

2667: Laplace transform.  This extends to the moments of $Y(\lambda)$,

2668: for $\lambda >0$: for $k\ge 1$,

2669: $$

2670: \EE(Y(\lambda)^k)= \frac {48. k!}{i\sqrt \pi }\int_{\Gamma}

2671: [a^k]\frac{A(a/v^3)e^{-2\lambda v}}

2672: {(1+A(a/v^3)e^{-2\lambda v})^2}v^{5} e^{v^4} dv.

2673: $$

2674: Since $A(x)=x/24 + O(x^2)$, the case $k=1$ of the above identity reads

2675: $$

2676: \EE(Y(\lambda))= \frac {2}{i\sqrt \pi }\int_{\Gamma}

2677: {e^{-2\lambda v}} v^{2} e^{v^4} dv.

2678: $$

2679: In the above expression, expand the exponential as a series. The

2680: convergence of the sum and integral is absolute, so that one can

2681: exchange them. This gives:

2682: $$

2683: \EE(Y(\lambda))= \frac {2}{i\sqrt \pi }\sum_{m\ge 0} 

2684: \frac{(-2\lambda)^m}{m!}\int_{\Gamma}

2685:  v^{m+2} e^{v^4} dv.

2686: $$

2687: Using~\eqref{gamma-inverse} (which is valid for any $s$ with the

2688: convention $1/\Gamma(-n)=0$ for $n\in \ns$), this can be rewritten as

2689: $$

2690: \EE(Y(\lambda))

2691: ={\sqrt \pi }

2692: \sum_{m\ge 0} \frac{(-2\lambda)^m}{m!\Gamma((1-m)/4)} 

2693: =\frac 1 {\sqrt \pi }

2694: \sum_{m\ge 0} \frac{(-2\lambda)^m}{m!}\Gamma\left(\frac{m+3} 4\right) \cos\left(

2695: \frac {(m+1)\pi}4\right).

2696: $$

2697: The last equality follows from the complement formula,

2698: \beq

2699: \label{complement}

2700: \Gamma(s) \Gamma(1-s)= \frac \pi{\sin (\pi s)}.

2701: \eeq

2702: 

2703: 

2704: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2705: \section{A global limit law, and the distribution function of the ISE}

2706: \label{section-limit-global}

2707: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

2708: In Section~\ref{section-limit-local}, we have derived from

2709: Proposition~\ref{propo-snake-gf} some \emm 

2710: local, limit results; for instance, a limit law for $X_n(0)/n^{3/4}$,

2711: the (normalized) number of nodes labelled $0$. In this section, we

2712: proceed with a similar study, which aims at deriving from

2713: Proposition~\ref{propo-snake-gf-rep} a \emm global, limit result

2714: --- in particular, 

2715: the limit law of $X_n^+(0)/n$, the normalized number of nodes having a

2716: non-negative label. The technique is copied on

2717: Section~\ref{section-limit-local}, and we do not give all the details.

2718: 

2719: For $j \in \zs$, let $X^+_n(j)$ denote the random variable equal to the

2720: number of nodes having label at least $j$ in a random tree of

2721: $\mathcal T _{0,n}$. 

2722: Let us define a normalized version of $X^+_n(j)$ by

2723: $$

2724: Y^+_n(j)= \frac{X^+_n(j)}{n}.

2725: $$

2726: These quantities are related to the series $R_j(t,u)$ studied in

2727: Proposition~\ref{propo-snake-gf-rep}. In particular,

2728: $$

2729: \EE\left(e^{aY_n^+(j)}\right)= \EE\left(e^{an^{-1}X_n^+(j)}\right)=

2730: \frac{[t^n] R_j(t,e^{a/n})}{2^n C_n}. 

2731: $$

2732: We extend the definition of $X^+_n$ and $Y^+_n$ to real values

2733: in a natural way by setting $X^+_n(x)=X^+_n(\lceil x\rceil)$ and

2734: $Y^+_n(x)=Y^+_n(\lceil x\rceil)$. 

2735: Let $\lambda \ge 0$.  The aim of this section is to prove that

2736: $Y_n^+( \lambda n^ {1/4})$ converges in distribution, as

2737: $n$ goes to infinity,  to a random variable $Y^+(\lambda)$ that we

2738: describe by its Laplace transform.  This is achieved in

2739: Theorem~\ref{thm-conv-rep} below,  but we first want to present two

2740: consequences of this theorem, which have a simpler formulation. The

2741: first consequence is a striking limit law for $Y^+_n(0)$. Recall

2742: that, up to a normalization by $n$, this random variable gives

2743: the number of nodes having a non-negative label in a tree rooted at $0$.

2744: %

2745: \begin{Propo}[{\bf The number of non-negative nodes}]

2746: \label{limit-0-G}

2747:  As $n$ goes to infinity, the random variable $Y^+_n(0)$ converges in law to

2748: the uniform distribution on $[0,1]$.

2749: \end{Propo}

2750: \noindent  

2751: This proposition will be proved in Section~\ref{moments-proof-G}.

2752: The second consequence of Theorem~\ref{thm-conv-rep} is an explicit

2753: expansion in $\lambda$ of the limiting first moment of $Y_n^+( \lambda n^{1/4})$.

2754: \begin{Propo}[{\bf The first moment}]

2755: \label{propo-first-G}

2756:  Let $\lambda \ge 0$. Then, as $n$ goes to infinity,

2757: $$

2758: \EE\left( Y_n^+( \lambda n^{1/4}) \right)

2759: \rightarrow

2760: \frac 1 {2 \sqrt\pi}\sum_{m\ge 0} \frac{(-2\lambda)^m}{m!} \cos \left(

2761: \frac{m\pi}4\right)

2762: \Gamma\left(\frac{m+2}4\right).

2763: $$

2764: \end{Propo}

2765: \noindent

2766: This proposition will be proved in Section~\ref{first-proof-G}.

2767: 

2768: 

2769: Let us, finally, state our main theorem, from which the two above

2770: propositions derive. 

2771: \begin{Theorem} [{\bf A global limit law}]

2772: \label{thm-conv-rep}

2773: Let $\lambda\ge 0$.% and $j=\lfloor \lambda n^{1/4}\rfloor$. 

2774: The sequence $Y_n^+(\lambda n^{1/4})$ converges in distribution to a

2775: random variable $Y^+(\lambda)$ whose Laplace transform is given, for

2776: $|a|<1$,  by

2777: $$

2778: \EE\left( e^{aY^+(\lambda)}\right)=G(\lambda,a),

2779: $$

2780: where

2781: $$

2782: G(\lambda, a)=

2783: 1+ \frac {48}{i\sqrt \pi }\int_{\Gamma}

2784: \frac{B(a/v^4)e^{-2\lambda v}}

2785: {(1+B(a/v^4)e^{-2\lambda v})^2}v^{5} e^{v^4} dv,

2786: $$

2787: \beq

2788: \label{B-eq}

2789: B(x)=-\frac{(1-D)(1-2D)}{(1+D)(1+2D)}, \quad  \quad 

2790: D=\sqrt{\frac{1+\sqrt{1-x}}2},

2791: \eeq

2792: and the integral is taken over

2793: $$

2794: \Gamma= \{1-te^{-i\pi/4}, t\in (\infty,0]\} \cup \{1+te^{-i\pi/4},

2795: t\in [0,\infty)\}. 

2796: $$

2797: Moreover, the Laplace transform of  $Y_n^+(\lambda n^{1/4})$ converges

2798: pointwise to $G(\lambda,\cdot)$ on the interval $(-1,1)$. The

2799: convergence of moments holds as well.

2800: 

2801: The sequence $Y_n^+(-\lambda n^{1/4})$ converges in distribution to

2802: the random variable $1- Y^+(\lambda)$.

2803: \end{Theorem}

2804: %

2805: This theorem will be proved in the next subsection. In view of the

2806: following proposition, it tells us about the law of the distribution

2807: function of the ISE.

2808: \begin{Propo}[{\bf The tail distribution function of the ISE}]

2809: \label{coro-tail}

2810: Let $g_{{\hbox{\sc \small ise}}}(\lambda)= \mu_{{\hbox{\sc \small

2811:       ise}}}(\lambda, +\infty)$ denote the tail

2812: distribution function  of the 

2813: ISE. Then for $\lambda \ge 0$,

2814: $$

2815: g_{{\hbox{\sc \small ise}}}(\lambda)

2816: % = \mu_{{\hbox{\sc \small      ise}}}(\lambda, +\infty)

2817: \deq  Y^+(\lambda /\sqrt 2), 

2818: $$

2819: where the law of the variable $Y^+(\lambda )$ is given in

2820: Theorem~{\em\ref{thm-conv-rep}}. In particular,  $g_{{\hbox{\sc \small

2821:       ise}}}(0)$ is uniformly distributed on $[0,1]$. The random

2822: variable $g_{{\hbox{\sc \small ise}}}(-\lambda)$ 

2823: has the same distribution as $1-Y^+(\lambda /\sqrt 2)$.

2824:  \end{Propo}

2825: \noindent{\bf Comments}\\

2826: {\bf 1.} The law of $g_{{\hbox{\sc \small ise}}}(0)$ was already given by

2827: Aldous~\cite[Eq.~(12)]{aldous}.

2828:  \\

2829: {\bf 2.} 

2830: Let us add a few words on the series $B$ and $D$ to convince

2831: ourselves that the integral giving $G(\lambda, a)$ is well-defined as

2832: long as $|a|<1$. Let $E(x)=1-D(x)$. Then $E$ admits the following expansion:

2833: $$

2834: E(x)= 1-\sqrt{1-\frac{1-\sqrt{1-x}} 2} =

2835: 2 \sum_{n\ge 1} C_{n-1} \left( \frac

2836: {1-\sqrt{1-x}} 8\right)^n

2837: $$

2838: where $C_n$ is the $n$th Catalan number. Similarly,

2839: $$

2840: {1-\sqrt{1-x}}= 2 \sum_{n\ge 1} C_{n-1} 4^{-n} x^n,

2841: $$

2842: and these two identities imply that $E(x)$ has non-negative

2843: coefficients. Moreover, its radius of convergence is easily seen to

2844: be $1$, so that $|E(x)|\le E(1)=1-1/\sqrt 2 $ for $|x|\le 1$.

2845: Moreover, expressing $B$ in terms of  $E$ gives:

2846: $$

2847: B= \frac{E(1-2E)}{(2-E)(3-2E)},

2848: $$

2849: which shows that $B(x)$ is also analytic for $|x|<1$ and satisfies

2850: in this domain

2851: $$

2852: |B(x)|\le \frac{E(1)(1+2E(1))}{(2-E(1))(3-2E(1))}=22\sqrt 2-31

2853: =0.11...

2854: $$

2855: For $v \in \Gamma$, $|v|\ge 1$ and $\Re(v)\ge 1$. This implies that 

2856: $$

2857: \frac{B(a/v^4)e^{-2\lambda v}}

2858: {(1+B(a/v^4)e^{-2\lambda v})^2}

2859: $$

2860: is uniformly bounded on $\Gamma$, and $G(\lambda, a)$ is well-defined.

2861: 

2862: \subsection{Proof of Theorem~\ref{thm-conv-rep}}

2863: Let $j= \lceil \lambda n^{1/4}\rceil$. 

2864: Given that the product forms for the series $S_j$ and $R_j$ are very

2865: similar, it is not surprising that we use an approach copied on

2866: that of the previous section. We start from

2867: $$

2868: \EE(e^{aY_n^+(j)})=\EE(u_n^{X_n^+(j)})= \frac{[t^n] R_j(t,u_n)}{2^n C_n},

2869: $$

2870: with $u_n= e^{a/n}$. For technical reasons, we choose to modify

2871: slightly the integration contour of

2872: Figure~\ref{fig-contour}. The Hankel part of this contour, which was

2873: lying at distance $1/8$ of the real axis, is now moved a bit further,

2874: at distance $1/6$ of the real axis. More precisely, the new contour

2875: $\overline \C_n$ consists of

2876: two parts $\overline \C_n^{(1)}$ and  $\overline \C_n^{(2)}$ such that

2877: \begin{itemize}

2878: %

2879: \item[$\bullet$ ] $\overline \C_n^{(1)}$ is an arc  of radius

2880: $(1+ \log^2 n /n)/8$, centered at the origin;

2881: %

2882: \item[$\bullet$ ] $\overline \C_n^{(2)}$ is a Hankel contour around $1/8$, at distance

2883: $1/(6n)$ of the real axis, 

2884: which meets  $\overline \C_n^{(1)}$ at both ends.

2885: \end{itemize}

2886: %

2887: We first need to prove that the series

2888: $R_j(t,u_n)$ is analytic in a neighborhood of   $\overline{\I}_n$, the region

2889: lying inside the integration contour $\overline\C_n$. The following lemma

2890: is the counterpart of Lemma~\ref{lemma-Sj}.

2891: %

2892: \begin{Lemma}[{\bf Analytic properties of $\nu$ and $R_j$}]

2893: \label{lemma-Rj}

2894: Let $a$ be a real number such that $|a|<1$.

2895: Then 

2896: $\nu(t,u_n)$ is analytic in a neighborhood of $\overline{\I}_n$. 

2897: Its modulus in  $\overline{\I}_n$ is smaller than $\alpha$, for some $\alpha<1$

2898: independent of $a$ and $n$.

2899: %

2900: The series $R_j(t,u_n)$ is also analytic in a neighborhood of

2901: $\overline{\I}_n$. 

2902: \end{Lemma}

2903: \noindent

2904: {\bf Proof.} Again, the lemma is obvious if $a=0$. We thus assume

2905: $a\not =0$ and $|a|<1$.

2906: 

2907: Let us first study the singularities of the series  $\bar \nu\equiv

2908: \bar \nu(z,u)$ defined by

2909: $$

2910: \bar \nu(z,u)= \nu\left( \frac{z(1+z^2)}{(1+z)^4},u\right).

2911: $$

2912: According to Proposition~\ref{propo-snake-gf-rep}, $\bar \nu$  is a \fps \ in

2913: $z$ with polynomial coefficients in $u$, and by~\eqref{eqZt}, one has:

2914: $$

2915: \nu(t,u)= \bar \nu(Z(t), u).

2916: $$

2917: In the course of the proof of Proposition~\ref{propo-snake-gf-rep}, we

2918: have obtained a polynomial equation $P(\nu,Z,u)=0$, of degree 4 in

2919: $\nu$, relating  $\nu(t,u), Z(t)$ and the variable

2920: $u$. This equation is not written in the paper (it is a bit too big),

2921: but it can be easily obtained using the expression of $\nu$ given in

2922: Proposition~\ref{propo-nu}. By definition of $\bar \nu$, we have

2923: $P(\bar \nu, z, u)=0$.

2924: 

2925:  Assume that $u$ is a fixed real

2926: number close to $1$. That is, $u=1+x$, with $x$ small. In order to

2927: study the singularities of $\bar \nu$, we look again at the zeroes of

2928: the leading coefficient of $P$, and at the zeroes of its

2929: discriminant. This gives several candidates for the

2930: singularities of $\bar \nu(z,u)$, which we classify in three series

2931: according to their behaviour when $x$ is small. First, some

2932: candidates tend to a limit that is different from 1,

2933: $$

2934: z=- 1, \quad z= \pm i,  \quad 

2935: z= e^{\pm 2i\pi/3},  \quad 

2936: z= e^{\pm 2i\pi/3} + O(x), \quad 

2937: z=-1+O(x).

2938: $$

2939: Then, some candidates tend to 1 and lie at distance at most

2940: $|x|^{1/4}$ of 1 (up to a multiplicative constant):

2941: $$

2942: z= 1+ \omega (cx)^{1/4} + O(\sqrt{|x|}),

2943: $$

2944: %z=1,

2945: %z=1+2 \omega |x|^{1/4} + O(\sqrt{|x|}),

2946: %z=1+ \omega\left( \frac{64 |x|}3\right)^{1/4} + O(\sqrt{|x|}),

2947: %z=1+ e^{i\pi/4} \omega\left( \frac{16 |x|}3\right)^{1/4} + O(\sqrt{|x|}),

2948: %$$

2949: where $\omega$ is a fourth root of unity and  $c$ is in the set $\{ 0,

2950: 16,  {64 }/3, -16/3\}$. 

2951: Finally, some candidates tend to 1 but lie further away from 1 (more

2952: precisely, at distance $|x|^{1/6}$):

2953: $$

2954: z=1+ 2 e^{i\pi/6} \omega' x^{1/6} + O({|x|^{1/3}}),

2955: $$

2956: where $\omega'$ is a sixth root of unity.

2957: 

2958: Let us now consider $\nu(t,u)=\bar \nu(Z(t),u)$ with $u=u_n=

2959: e^{a/n}=1+x$, where $x=a/n(1+o(1))$. Recall that $Z$ is

2960: analytic in $\cs \setminus [1/8, \infty)$. By Lemma~\ref{lemma-Z}, the 

2961:   series $Z(t)$ never approaches any root of unity different from

2962:   1. Hence for $n$ large enough, $Z(t)$ never reaches any of

2963:   the candidates $z$ of the first series.

2964: 

2965: The candidates of the second series are of the   form

2966: $$

2967: z= 1 + \omega  (ac/n)^{1/4} + O( n^{-1/2})

2968: $$

2969: for some constant $c$, with $|c|\le 64/3$, depending on the candidate.

2970: By~\eqref{eqZt}, $Z(t)$ may only reach these values for

2971: $$

2972: t=\frac 1 8 -  \frac{ac}{128n} + O(n^{-5/4}).

2973: $$

2974: Since $|a|<1$, there exists $\epsilon>0$ such that these  values lie

2975: at distance less that $1/((6+\epsilon)n)$ of $1/8$, that is, outside a

2976: neighborhood of the domain $\overline{\I}_n$.

2977: 

2978: The candidates of the third series are more worrying: $Z(t)$ may

2979: reach them for

2980: \beq

2981: \label{t-worrying}

2982: t=\frac 1 8 - \frac{\omega ''}{8}  \left(a/n\right)^{2/3} + O(n^{-5/6}), 

2983: \eeq

2984: where $\omega''$ is a cubic root of unity,

2985: and these values may lie inside $\overline{\I}_n$.

2986: If $a>0$ and $\omega''=e^{\pm 2i\pi/3}$, or if $a<0$ and $\omega''=e^{

2987:   2i\pi/3}$, the modulus of the above value of $t$ is 

2988: found to be $1/8(1+ cn^{-2/3} + o(n^{-2/3}))$, for some \emm positive,

2989: constant $c$: this is larger than the radius of the contour $\overline

2990: \C_n$, which implies that $t$ lies outside a neighborhood of

2991: $\overline{\I}_n$. However, if 

2992:  $a>0$ and $\omega''=1$, or  if $a<0$ and  $\omega''=1$ or $e^{-

2993:   2i\pi/3}$,  the above value of $t$  lies definitely  inside

2994: $\overline{\I}_n$.  Its modulus is $1/8(1- cn^{-2/3} +

2995: o(n^{-2/3}))$, for some  positive constant $c$. 

2996: 

2997: In order to rule out the possibility that $\nu(t,u_n)$ has such a

2998: singularity, we are going to prove, by having a close look at the 

2999: expression of $\nu$ given in Proposition~\ref{propo-nu}, that the

3000: radius of convergence of $\nu(t,u_n)$ is at least $1/8-O(1/n)$.

3001: We use below the notation of Proposition~\ref{propo-nu}.

3002: 

3003: Clearly, the series $V(t,u_n)$ has radius of convergence $\min

3004: (1/8,1/(8u_n))$. In particular, this radius is at least

3005: $\rho_n:=1/(8(1+|x|))$ (with $u_n=1+x$).

3006:  Moreover,  the series $V$ admits the following

3007: expansion 

3008: $$

3009: V(t,1+x)= \frac 1 4 \left( 1-\sqrt{1-\frac{8tx}{1-8t}}\right)

3010: = \frac 1 2 \sum_{n \ge 1} C_{n-1} \left( \frac{2tx}{1-8t}\right)^n,

3011: $$

3012: where $C_n$ is the $n$th Catalan number. This shows that $V(t,1+ |x|)$

3013: is a series in $t$ with \emm positive, coefficients and that for all

3014: $t$ such that $|t|\le \rho_n$,

3015: $$

3016: \left|V(t,1+x)\right| \le V(|t|, 1+ |x|) \le V\left( \frac 1{

3017:   8(1+|x|)},   1+ |x|\right) = \frac 1 4 .

3018: $$

3019: The next step is to prove that  $\Delta(t,u_n)$ never vanishes for $|t|\le

3020: \rho_n$. Indeed,

3021: $$

3022: \Delta=

3023: %(1-V)^2-\frac{4 ZV^2}{(1+Z)^2} = 

3024: (1-V)^2 - 4WV^2,

3025: $$

3026: where $W\equiv W(t)$ is the formal power series in $t$ defined by~

3027: \eqref{eqW}. This series has radius $1/8$, and non-negative

3028: coefficients.

3029: Hence for all $t$ such that $|t|\le 1/8$, one has $|W(t)|\le W(1/8)=

3030: 1/4$. Consequently, for  $|t|\le \rho_n$,

3031: $$

3032: |\Delta(t,1+x)|\ge (1-|V(t,1+x)|)^2 - 4|W(t)||V(t,1+x)|^2 \ge \left( 1 - \frac 1

3033: 4\right)^2 - \frac 1 {16} = \frac 1 2.

3034: $$

3035: Hence $\Delta(t,u_n)$ does not vanish in the centered disk of

3036: radius $\rho_n$. It follows that the series $P(t,u_n)$ is analytic

3037: inside this disk. 

3038: 

3039: According to the expression of $\nu$  given

3040: in Proposition~\ref{propo-nu}, the series $\nu(t,u_n)$ is meromorphic

3041: for   $|t|\le \rho_n$. The final question we need to answer is whether

3042: $\nu$ has poles in this disk, and where. Returning to the polynomial

3043: $P$ such that $P(\nu, Z, u)=0$ shows that this can only happen if the

3044: coefficient of $\nu^4$ in this polynomial vanishes. But this can only

3045: occur if $z=Z(t)$ has one of the following forms:

3046: $$

3047: z=\pm i, \quad z= e^{\pm 2i\pi/3} +O(x), \quad z=-1 + O(x), \quad z= 1+ \omega

3048: (64x/3)^{1/4} + O(x^{1/2}).

3049: $$

3050:  As argued above, only the last value of $z$ is likely to be reached

3051: by $Z(t)$, and this may only occur if

3052: $$

3053: t=\frac 1 8 - \frac{1}{6} (a/n) + O(n^{-5/4}).

3054: $$

3055: Consequently, the radius of $\nu(t,u_n)$ is at least $1/8-O(1/n)$,

3056: and this proves that the values~\eqref{t-worrying} that have been

3057: shown to lie in the centered disk of radius $1/8$, 

3058: are not, after all, singularities of $\nu(t,u_n)$. This completes our

3059: proof that $\nu(t,u_n)$ is analytic in a neighborhood of  $\overline{\I}_n$.

3060: 

3061: \medskip

3062: We now want to bound $\nu(t,u_n)$ inside $\overline\I_n$. From now on,

3063: we can walk safely in the steps of the proof of Lemma~\ref{lemma-Sj}.

3064: Let $t_n\in \overline\I_n$ be such that

3065: $$

3066: |\nu(t_n,u_n)|= \max _{t\in \overline\I_n} |\nu(t,u_n)|.

3067: $$

3068: We first give a lower bound for this quantity, by estimating

3069: $\nu(t,u_n)$ for $t=1/8-1/(6n)$. This is easily done by combining the

3070: closed form expressions of $\nu$ (Proposition~\ref{propo-nu}) and $B$

3071: (Theorem~\ref{thm-conv-rep}). One obtains:

3072: $$

3073: |\nu(t_n,u_n)| \ge |\mu(1/8-1/(6n), u_n)|=|B(3a/4)|+o(1)>0.

3074: $$

3075: This lower bound is then used to rule out the possibility that the

3076: sequence $t_n$ has an accumulation point different from $1/8$. Thus

3077: $t_n$ converges to $1/8$, and one can write

3078: $$

3079: t_n= \frac 1 8 \left( 1 -\frac {x_n} n \right).

3080: $$

3081: We have $x_n=o(n)$, but also $|x_n|>4/3$ since $t_n$ belongs to

3082: $\overline \I_n$. We want to estimate $\nu(t_n, u_n)$.

3083: Since

3084: $$

3085: Z(t_n)= 1-2  \left( \frac {x_n} n \right)^{1/4} + O\left(  \left(

3086: \frac {x_n} n \right)^{1/2}\right)

3087: $$

3088: and

3089: $$

3090: u_n-1=a/n  \left( 1 + O(1/n)\right),

3091: $$

3092: one has

3093: $$

3094: \frac{u-1}{(1-Z)^4}= \frac{a}{16x_n}\left(1+ O\left(  \left(

3095: \frac {x_n} n \right)^{1/4}\right)\right).

3096: $$

3097: The closed form expressions of $\nu$ and $B$ imply that the sequence

3098: $x_n$ is bounded and

3099: $$

3100: \nu(t_n,u_n)= B(a/x_n) + o(1).

3101: $$

3102: Since $B$ is bounded by  $0.12$ inside its disk of convergence,

3103: $|\nu(t_n,u_n)|$ is certainly smaller than some $\alpha$ for

3104: $\alpha<1$ and $n$ large enough.

3105: This concludes the proof of the second statement of Lemma~\ref{lemma-Rj}.

3106: 

3107: \medskip

3108: By continuity of $\nu(t,u_n)$, this function of $t$ is still bounded by $1$

3109: (in modulus) is a neighborhood of $\overline \I_n$. 

3110: Recall also that the modulus

3111: of $Z(t)$ never reaches $1$ for $t \in \cs \setminus [1/8,

3112:   \infty)$. The form~\eqref{product-formRj} then implies that

3113:   $R_j(t,u_n)$ is an analytic function of $t$ in a neigborhood of

3114:   $\overline \I_n$. 

3115: \cqfd

3116: 

3117: 

3118: Once this rather painful lemma is at last established, the rest of the

3119: proof of Theorem~\ref{thm-conv-rep} copies  the end of the

3120: proof of Theorem~\ref{thm-conv}, with $S_j, \mu$ and $A$ respectively

3121: replaced by $R_j$,  $\nu$ and $B$. The counterpart of~\eqref{muA} is

3122: $$

3123:  \nu(t,u_n)= B(-a/z) (1+o(1)).

3124: $$

3125: Recall that the Hankel part of the contour $\overline \C_n$ is now at

3126: distance $1/(6n)$ of the real axis. Hence, when $n$ goes to infinity,

3127: one finds

3128:  $$

3129: [t^n] R_j(t,u_n)= 2^nC_n +

3130: \frac {12. 8^n n^{-3/2} }{i  \pi}\int_{4/3 \Ha }

3131: \frac{B(-a/z)\exp(-2\lambda (-z)^{1/4})\sqrt{-z}e^{-z}}

3132: {(1+B(-a/z)\exp(-2\lambda (-z)^{1/4}))^2}dz + o(8^n n^{-3/2}).

3133: $$

3134: After normalizing by $2^n C_n$ and setting $v=(-z)^{1/4}$, this gives

3135: $$

3136: \EE(e^{aY_n^+(j)}) \rightarrow

3137: 1+ \frac {48}{i\sqrt \pi }\int_{(4/3)^{1/4}\Gamma}

3138: \frac{B(a/v^4)e^{-2\lambda v}}

3139: {(1+B(a/v^4)e^{-2\lambda v})^2}v^{5} e^{v^4} dv,

3140: $$

3141: but the analyticity properties of the integrand allow us to replace

3142: the integration contour by $\Gamma$.

3143: 

3144: 

3145: 

3146: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3147: \subsection{Proof of Proposition~\ref{limit-0-G}}

3148: \label{moments-proof-G}

3149: When $\lambda=0$, the limiting Laplace transform reduces to

3150: $$

3151: G(0, a)=

3152: 1+ \frac {48}{i\sqrt \pi }\int_{\Gamma}

3153: \frac{B(a/v^4)}

3154: {(1+B(a/v^4))^2}v^{5} e^{v^4} dv

3155: =

3156: 1+ \frac {4}{3 i\sqrt \pi }\int_{\Gamma}

3157: \frac{\chi(a/v^4)(3-\chi(a/v^4))}

3158: {1+\chi(a/v^4)}v^{5} e^{v^4} dv

3159: $$

3160: where $\chi(x)$ is the unique \fps\ in $x$  satisfying

3161: $$

3162: \chi=\frac x 4(1+\chi)^{2}.

3163: $$

3164: The Lagrange inversion formula gives, for $k \ge 1$,

3165: $$

3166: [x^k]  \frac{\chi(x)(3-\chi(x))} {1+\chi(x)}

3167: = \frac {6} {4^k} \frac{(2k-2)!}{(k-1)!(k+1)!}.

3168: $$

3169: Consequently,

3170: $$

3171: G(0, a)=1+ \frac {8}{i\sqrt \pi }\int_{\Gamma}

3172: \sum_{k\ge 1} \frac 1 {4^k} \frac{(2k-2)!}{(k-1)!(k+1)!} a^k

3173: v^{5-4k} e^{v^4} dv.

3174: $$

3175: The convergence is absolute, so that we can exchange the sum and the

3176: integral:

3177: $$

3178: G(0, a)=1+\frac {8}{i\sqrt \pi } \sum_{k\ge 1} \frac 1 {4^k}

3179:  \frac{(2k-2)!}{(k-1)!(k+1)!} a^k

3180: \int_{\Gamma}v^{5-4k} e^{v^4} dv.

3181: $$

3182: Using~\eqref{gamma-inverse}, and picking the coefficient of $a^k$, we find that the

3183: $k$th moment of the random variable $Y^+(0)$ is

3184: $$

3185: \EE(Y(0)^k)=  \frac {8}{i\sqrt \pi } 

3186:  \frac 1 {4^k}

3187:  \frac{(2k-2)!}{(k-1)!(k+1)}\frac {i \pi}{2\Gamma(k-1/2)} = \frac 1 {k+1}.

3188: $$

3189: The unique distribution having  its $k$th

3190: moment equal to $1/(k+1)$  is the uniform distribution on

3191: $[0,1]$. Proposition~\ref{limit-0-G} follows. 

3192: 

3193: 

3194: 

3195: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3196: \subsection{Proof of Proposition~\ref{propo-first-G}}

3197: \label{first-proof-G}

3198: We have derived above the moments of $Y^+(0)$ from the expression of its

3199: Laplace transform.  This extends to the moments of $Y^+(\lambda)$,

3200: for $\lambda >0$: for $k\ge 1$,

3201: $$

3202: \EE(Y^+(\lambda)^k)= \frac {48. k!}{i\sqrt \pi }\int_{\Gamma}

3203: [a^k]\frac{B(a/v^4)e^{-2\lambda v}}

3204: {(1+B(a/v^4)e^{-2\lambda v})^2}v^{5} e^{v^4} dv.

3205: $$

3206: Since $B(x)=x/48 + O(x^2)$, the case $k=1$ of the above identity gives

3207: $$

3208: \EE(Y^+(\lambda))= \frac {1}{i\sqrt \pi }\int_{\Gamma}

3209: {e^{-2\lambda v}} v e^{v^4} dv.

3210: $$

3211: In the above expression, expand the exponential as a series. The

3212: convergence of the sum and integral is absolute, so that one can

3213: exchange them. This gives:

3214: $$

3215: \EE(Y^+(\lambda))= \frac {1}{i\sqrt \pi }\sum_{m\ge 0} 

3216: \frac{(-2\lambda)^m}{m!}\int_{\Gamma}

3217:  v^{m+1} e^{v^4} dv.

3218: $$

3219: Using~\eqref{gamma-inverse}, this can be rewritten as

3220: $$

3221: \EE(Y^+(\lambda))=\frac {\sqrt \pi } 2

3222: \sum_{m\ge 0} \frac{(-2\lambda)^m}{m!\Gamma((2-m)/4)} 

3223: =\frac 1 {2 \sqrt \pi }

3224: \sum_{m\ge 0} \frac{(-2\lambda)^m}{m!}\Gamma\left(\frac{m+2} 4\right) \cos\left(

3225: \frac {m\pi}4\right).

3226: $$

3227: The last equality follows from the complement formula~\eqref{complement}.

3228: 

3229: 

3230: \subsection{The distribution function of the {\sc ISE}}

3231: \label{section-coro-tail}

3232: \noindent Let us finally prove Proposition~\ref{coro-tail}.

3233: 

3234: Let $\mu_n$ be a sequence of random probability measures on $\rs$ converging

3235: weakly to a probability measure $\mu$. Let $F_n$ denote the (random)

3236: distribution function of $\mu_n$: for $ \lambda \in \rs$, 

3237: $$

3238: F_n(\lambda) = \mu_n(-\infty, \lambda].

3239: $$

3240: Similarly, let $F$ denote the distribution function of $\mu$. It is

3241: not very hard 

3242: to prove that, for all $\lambda\in\rs$ such that $\mu\{\lambda\}=0$,

3243: $F_n(\lambda)$ converges in distribution to $F(\lambda)$.

3244: (We prove this in the appendix of the paper, but it is certainly

3245: written somewhere in the literature.)

3246: 

3247: Let us now apply this general result to our context. The probability

3248: measure $\mu_n$ is given by~\eqref{random-measure}, with $c=\sqrt

3249: 2$. It is known to 

3250: converge to the random measure $\mu_{\hbox{\sc \small ise}}$. 

3251: Assume for the moment  that this measure does not assign a positive weight

3252: to any point. Then, with the above notation, $F_n(\lambda)$ converges in

3253: distribution to $F(\lambda)$, for all $\lambda\in \rs$.  But, given

3254: the definition~\eqref{random-measure} of $\mu_n$,

3255: $$

3256: F_n(\lambda)

3257: %=\mu_n(-\infty, \lambda]

3258: = 1-\mu_n(\lambda, \infty)= 1

3259: -\frac 1{n+1} X_n^+( \lambda n^{1/4}/\sqrt 2)  

3260: + \frac 1{n+1} X_n( \lambda n^{1/4}/\sqrt 2),

3261: $$

3262: where the definition of $X_n$ is extended to all reals  by

3263: $X_n(x)=0$ if $x\in \rs\setminus \zs$. By Theorems~\ref{thm-conv} and~\ref{thm-conv-rep}, 

3264: the right-hand side converges in distribution to $1-Y^+(\lambda/\sqrt

3265: 2)$.  Consequently, the tail distribution function of the ISE (that

3266: is, $\mu_{\hbox{\sc \small ise}}(\lambda, \infty)$) has the

3267: same law as $Y^+(\lambda/\sqrt2)$.

3268: 

3269: It remains to prove  that $\mu_{\hbox{\sc \small ise}}$ does not weight

3270:   points positively (almost surely). Let $\lambda\in \rs$. Then

3271: \beq

3272: \label{E-P}

3273: \PP(\mu_{\hbox{\sc \small ise}}\{\lambda\}>0)=0 \Longleftrightarrow \EE(\mu_{\hbox{\sc \small ise}}\{\lambda\})=0.

3274: \eeq

3275: Let $\epsilon >0$, and let $h_{\lambda, \epsilon}$ be the function

3276: plotted on the right-hand side of Figure~\ref{fig-epsilon}.

3277: Then

3278: $$

3279: \begin{array}{lllll}

3280: \mu_{\hbox{\sc \small ise}}\{\lambda\}&=&

3281: \displaystyle \lim_{\epsilon \rightarrow 0^+} \int_\rs h_{\lambda, \epsilon}(y) d\mu_{\hbox{\sc \small ise}}(y)\\

3282: &=&\displaystyle\lim_{\epsilon \rightarrow 0^+}\int_0^1 h_{\lambda, \epsilon}(

3283: r(t)) dt & (\hbox{by } \eqref{ise-rep})\\

3284: &=& \displaystyle\int_0^1 \1_{\lambda=  r(t)} dt.

3285: \end{array}

3286: $$

3287: Taking expectations, we obtain

3288: $$

3289: \EE(\mu_{\hbox{\sc \small ise}}\{\lambda\})

3290: = \EE\left(\int_0^1 \1_{\lambda=  r(t)}

3291: dt\right) = \int_0^1 \PP(\lambda=  r(t)) dt.

3292: $$

3293: But $\PP(\lambda=  r(t))=0$ for all $t\in (0,1)$ and $\lambda$, since

3294: $r(t)$ has a density with respect to the Lebesgue measure for all

3295: $t$. By~\eqref{E-P}, we conclude that $\mu_{\hbox{\sc \small

3296:     ise}}$ does not weight points positively.

3297: 

3298: 

3299: 

3300: The last statement of Proposition~\ref{coro-tail} is then easily proven,

3301: using a the symmetry of $\mu_{\hbox{\sc \small ise}}$ and the fact

3302: that it does not assign a positive probability to any point. 

3303: 

3304: 

3305: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3306: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3307: \section{Other tree models and universality}

3308: \label{section-universality}

3309: \subsection{Trees with increments $0, \pm 1$}

3310: We consider in this section a slight variation on the previous family

3311: of trees: the increments of the labels along edges may now be $0, \pm

3312: 1$. This family of trees has attracted a lot of interest in relation to

3313: planar

3314: maps~\cite{bdg-statistics,chassaing-schaeffer,cori-vauquelin,jf-mokka-brownianmap}. 

3315: \subsubsection{Enumerative results}

3316: \label{section-enumerative-0pm1}

3317: As above, let $T_j \equiv T_j(t)$ be the \gf\ of labelled trees

3318: in which all labels are at most $j$, counted by their number of

3319: edges. Let $S_j \equiv S_j(t,u)$ be the 

3320: \gf\ of labelled trees, counted by the number of edges (variable $t$)

3321: and the number of nodes labelled $j$ (variable $u$). Finally, let $R_j

3322: \equiv R_j(t,u)$ be the 

3323: \gf\ of labelled trees, counted by the number of edges

3324: and the number of nodes having label $j$ at least. 

3325: As above, it is easy to write an infinite system of equations defining

3326: any of the families $T_j$, $S_j$ or $R_j$. The only difference with

3327: our first family of trees  is that

3328: a third case arises in the decomposition of trees illustrated

3329: by Figure~\ref{fig-decomp}: the leftmost child of the root

3330: may have label $j$. In particular,  the \gf \  $T\equiv T(t)$ counting

3331: plane labelled trees now satisfies

3332: $$

3333: T=1+3tT^2,

3334: $$

3335: while  for $j\ge 0$,

3336: \beq

3337: \label{rec-TjP}

3338: T_j=1+t(T_{j-1}+T_j+T_{j+1})T_j.

3339: \eeq

3340: The equations of Lemmas~\ref{lemma-Sj-rec} and~\ref{lemma-Rj-rec} are

3341: modified in a similar way. 

3342: %

3343: The three infinite systems of equations thus obtained can be solved

3344: using the same 

3345: techniques as in Section~\ref{section-snake-gf}. The solutions are

3346: expressed in terms of the above

3347: series $T\equiv T(t)$ and of the  unique formal power series

3348: $Z\equiv Z(t)$, with constant term $0$, satisfying

3349: \beq

3350: \label{eqZtP}

3351: Z=t\, \frac{(1+4Z+Z^2)^2}{1+Z+Z^2}.

3352: \eeq

3353: Observe that $T$ and $Z$ are related by:

3354: %\beq

3355: %\label{eqTZP}

3356: $$

3357: T=\frac{1+4Z+Z^2}{1+Z+Z^2}.

3358: $$ %\eeq

3359: We state without proof the counterparts of

3360: Propositions~\ref{propo-snake-bounded},~\ref{propo-snake-gf} 

3361: and~\ref{propo-snake-gf-rep}. 

3362: %

3363: \begin{Propo}[{\bf Trees with small labels \cite{bdg-geodesic,bdg-statistics}}]

3364: \label{propo-snake-boundedP}

3365: Let $T_j\equiv T_j(t)$ be the \gf\ of trees having no label greater

3366: than $j$. Then $T_j$ is algebraic of degree (at most) $2$. In particular,

3367: $$

3368: T_0=1-16\,t

3369: +18\,t  {T_0}

3370: -27\,{t}^{2}{{ T_0}}^{2}.

3371: $$

3372: Moreover,  for all $j \ge -1$,

3373: %\beq\label{product-formTjP}

3374: $$

3375: T_j= T\,\frac{(1- Z^{j+1})(1- Z^{j+4})}{(1- Z^{j+2})(1-Z^{j+3})},

3376: %\eeq

3377: $$

3378: where $Z\equiv Z(t)$ is given by~\eqref{eqZtP}.

3379: \end{Propo}

3380: \noindent

3381: {\bf Remarks}\\

3382: {\bf 1.} As observed in~\cite[p.~645]{bdg-statistics}, there is an ``invariant'' function

3383: attached to equations of the form~\eqref{rec-TjP}:

3384:  for $j\ge 0$,

3385: $$

3386: I(T_{j-1},T_j)=I(T_{j},T_{j+1})

3387: $$

3388: where  $I$ is now given by

3389: $$

3390: I(x,y)= xy(1-t(x+y))-x-y.

3391: $$

3392: As explained in the remark that follows

3393: Propositions~\ref{propo-snake-gf} and~\ref{propo-mu-closed-form},

3394: this can be used to derive rapidly from~\eqref{rec-TjP} the value of 

3395: $T_0$.\\

3396: %

3397: {\bf 2.} As was the case for trees with increments $\pm1$,

3398:   the trees counted by $T_0$ (equivalently, the trees having only

3399: non-negative labels) are closely related to planar

3400: maps. More precisely, there is a one-to-one correspondence between

3401: non-negative trees having $n$ edges and planar maps having $n$

3402: edges~\cite{chassaing-schaeffer,cori-vauquelin}. 

3403: The coefficients of $T_0(t)$ are also remarkably simple:

3404: $$

3405: T_0(t)= 

3406: \frac{(1-12t)^{3/2}-1+18t}{54t^2}

3407: =\sum_{n\ge 0} \frac{2. 3^{n}}{(n+1)(n+2)}{{2n}\choose n}t^n.

3408: $$

3409: A combinatorial explanation for the algebraicity of $T_0$ is given

3410: in~\cite{cori-vauquelin}.

3411: 

3412: %

3413: \begin{Propo}[{\bf The number of nodes labelled $j$}]

3414: \label{propo-snake-gfP}

3415: For any $j\in \zs$, the \gf \ $S_j\equiv S_j(t,u)$ that counts

3416: labelled trees by the number of edges and the number of nodes labelled $j$ is

3417: algebraic of degree at most $4$ over $\qs(T,u)$ (and hence has degree

3418: at most $8$ over $\qs(t,u)$). 

3419: More precisely,

3420: $$

3421: %

3422: \frac{9T^4(u-1)^2}{(T-S_0)^2}= 9T^2-2T(T-1)(2T+1)S_0+(T-1)^2S_0^2

3423: %\eeq

3424: $$

3425: and all the $S_j$ belong to $\qs(t,u,S_0)$.

3426: Moreover,  for all $j \ge 0$,

3427: %\beq\label{product-formSjP}

3428: $$

3429: S_j= T\,\frac{(1+\mu Z^j)(1+\mu Z^{j+3})}{(1+\mu Z^{j+1})(1+\mu Z^{j+2})},

3430: %\eeq

3431: $$

3432: where $Z\equiv Z(t)$ is given by~\eqref{eqZtP}

3433: and $\mu\equiv \mu(t,u)$ is the unique \fps\ in $t$

3434: satisfying

3435: %\beq\label{eqMZP}

3436: $$

3437: \mu= (u-1) \frac{(1+Z+Z^2)(1+\mu Z)^2(1+\mu Z^2)^2}

3438: {(1+Z)^2(1-Z)^3(1-\mu^2 Z^3)}.

3439: %\eeq

3440: $$

3441: The series $\mu(t,u)$  has polynomial coefficients in $u$,

3442: and satisfies $\mu(t,1)=0$. It has degree $4$ over $\qs(Z,u)$ and $16$

3443: over $\qs(t,u)$. 

3444: \end{Propo}

3445: 

3446: \begin{Propo}[{\bf  The number of nodes labelled $j$ or more}]

3447: \label{propo-snake-gf-repP}

3448: Let $j \in \zs$. The \gf\ $R_j(t,u)\equiv R_j$

3449: that counts labelled trees by the number of edges and the number of

3450: nodes labelled $j$ or more

3451: is  algebraic over $\qs(t,u)$, of degree at most $8$. 

3452:  It has degree

3453: at most $2$ over $\qs(T,\tilde T)$, where $T\equiv T(t)$ and $\tilde

3454: T\equiv T(tu)$. More precisely, it belongs to the

3455: extension of $\qs(T,\tilde T)$ generated by 

3456: $$

3457: \sqrt{4(T+\tilde T)^2-T\tilde T (4+3T\tilde T)}.

3458: $$

3459: Moreover,  for all $j \ge 0$,

3460: %\beq\label{product-formRjP}

3461: $$

3462: R_j= T\,\frac{(1+\nu Z^j)(1+\nu Z^{j+3})}{(1+\nu Z^{j+1})(1+\nu Z^{j+2})},

3463: %\eeq

3464: $$

3465: where $Z\equiv Z(t)$ is given by~\eqref{eqZtP}

3466: and $\nu\equiv \nu(t,u)$ is a \fps\ in $t$, with polynomial

3467: coefficients in $u$, which is algebraic

3468: %  of degree $2$ over $\qs(Z(t),Z(tu))$, 

3469: of degree $4$ over $\qs(u,Z)$, and  of degree $16$

3470: over $\qs(t,u)$. 

3471: This series satisfies $\nu(t,1)=0$.

3472: The first terms in its expansion are:

3473: $$

3474: \nu(t,u)=(u-1)\Big(

3475: 1+3\,ut+ \left(15\,u+ 14\,{u}^{2} \right) {t}^{2}

3476: + \left( 104\,u+117\,{u}^{2}+83\,{u}^{3} \right) {t}^{3}+

3477: O(t^4)\Big) .

3478: $$

3479: \end{Propo}

3480: 

3481: \subsubsection{Limit laws}

3482: We now endow the set of labelled trees having $n$ edges with the

3483: uniform distribution, and consider the same random variables as for

3484: our first family of trees: $M_n$,  the largest label,  $X_n(j)$, the

3485: number of nodes having label $j$, % in a random tree of  rooted at $0$,

3486: and finally  $X_n^+(j)$, the

3487: number of nodes having label $j$ at least.

3488: 

3489: Again, we can prove that 

3490:  $M_nn^{-1/4}$ converges in law to  $N_{{\hbox{\sc \small

3491:    ise}}}/\sqrt 3 $,  where $N_{{\hbox{\sc \small

3492:    ise}}}$ is the supremum of the support of the ISE, and that

3493:  for all $\lambda\in \rs$, the sequence $X_n^+( \lambda n^{1/4})/ n$

3494:  converges in law to  $g_{{\hbox{\sc \small ise}}}(\sqrt{3 }\lambda 

3495: )$ where $g_{{\hbox{\sc \small ise}}}$ is the tail

3496: distribution function of the ISE. The arguments are the same as for

3497:  our first class of trees (Sections~\ref{section-supremum}

3498:  and~\ref{section-coro-tail}). 

3499: 

3500: Hence we could  just as well have started from the enumerative results of

3501: Section~\ref{section-enumerative-0pm1}, rather than from those of

3502: Section~\ref{section-snake-gf}, to characterize the laws of

3503: $N_{{\hbox{\sc \small  ise}}} $ and $g_{{\hbox{\sc \small

3504:       ise}}}(\lambda)$ (Propositions~\ref{coro-support}

3505: and~\ref{coro-tail}). More remarkably, we have performed on

3506: $X_n(j)$ an analysis similar to that of

3507: Section~\ref{section-limit-local}, and obtained \emm the same, local

3508: limit law. In other words, for all $\lambda \ge 0$,

3509: the sequence  $X_n(\lfloor \lambda n^{1/4}\rfloor) n^{-3/4}$ converges

3510: in law to  $

3511: \sqrt 3  f_{{\hbox{\sc \small ise}}}(\sqrt 3 \lambda )

3512: $

3513: where $f_{{\hbox{\sc \small ise}}}$ is the conjectured

3514: density of the ISE, given in Conjecture~\ref{conj-density}.

3515: 

3516: 

3517: In all three cases, the convergence of the moments holds as well.

3518: 

3519: 

3520: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3521: \subsection{Naturally embedded binary trees}

3522:  We study in the section the  incomplete binary

3523: trees\footnote{The author has obtained similar, but slightly heavier

3524:   results for embedded  \emm complete,\/  binary trees.}

3525: carrying their natural labelling, as shown on the right of

3526: Figure~\ref{fig-three-families}. Such trees are either

3527: empty, or have a root, to 

3528: which a left and right subtree (both possibly empty) are

3529: attached. A (minor) difference with the two previous  families of trees

3530: is that the main enumeration parameter is now the number of nodes

3531: rather than the number of edges.

3532: 

3533: \subsubsection{Enumerative results}

3534: \label{section-enumerative-binary}

3535:  Let $T_j \equiv T_j(t)$ be the \gf\ of (naturally labelled) binary trees

3536: in which all labels are at most $j$, counted by their number of

3537:  nodes. Let $S_j \equiv S_j(t,u)$ be the 

3538: \gf\ of binary trees, counted by the number of  nodes (variable $t$)

3539: and the number of nodes labelled $j$ (variable $u$). Finally, let $R_j

3540: \equiv R_j(t,u)$ be the 

3541: \gf\ of binary  trees, counted by the number of nodes

3542: and the number of nodes having label $j$ at least. 

3543: It is easy to write an infinite system of equations defining

3544: any of the families $T_j$, $S_j$ or $R_j$. The decomposition of trees

3545: that was crucial in Section~\ref{section-snake-gf} is now replaced by the

3546:  decomposition of Figure~\ref{fig-decomp-binary}. 

3547: % Observe that there is now  no choice for   the labels of the

3548: % children of the root. In particular,  

3549: The \gf \ $T\equiv T(t)$ counting 

3550: naturally labelled binary trees satisfies

3551: $$

3552: T=1+tT^2,

3553: $$

3554: (as it should!) while  for $j\ge 0$,

3555: \beq

3556: \label{rec-TjB}

3557: T_j=1+tT_{j-1}T_{j+1}.

3558: \eeq

3559: Note that the initial condition is now $T_{-1}=1$ (accounting for

3560: the empty tree). 

3561: 

3562: \begin{figure}[hbt]

3563: \begin{center}

3564: \input{tree-decomp-binary.pstex_t}

3565: \end{center}

3566: \caption{The decomposition of naturally labelled binary trees rooted

3567:   at $j$.}

3568: \label{fig-decomp-binary}

3569: %\hrule

3570: \end{figure}

3571: The equations of Lemmas~\ref{lemma-Sj-rec} and~\ref{lemma-Rj-rec}

3572: respectively become:

3573: \beq

3574: \label{rec-Sj-binary}

3575: S_j=\left\{

3576: \begin{array}{llll}

3577: 1+tS_{j-1}S_{j+1} & \hbox{if } j\not = 0, \\

3578: 1+tuS_1^2  & \hbox{if } j = 0,

3579: \end{array}

3580: \right.

3581: \eeq

3582: while

3583: \beq

3584: \label{rec-RjB}

3585: R_j= 1+tR_{j-1}R_{j+1}  \quad \hbox{for } j\ge1,

3586: \eeq

3587: and

3588: \beq

3589: \label{eq-symmetryB}

3590: R_{-j}(t,u)=R_{j+1}(tu,1/u)\quad \hbox{for all } j \in \zs.

3591: \eeq

3592: 

3593: %

3594: 

3595: 

3596: The three infinite systems of equations thus obtained can be solved

3597: using the same 

3598: techniques as in Section~\ref{section-snake-gf}. The solutions are

3599: expressed in terms of the above

3600: series $T\equiv T(t)$ and of the  unique formal power series

3601: $Z\equiv Z(t)$, with constant term $0$, satisfying

3602: \beq

3603: \label{eqZtB}

3604: Z=t\, {\frac{ \left( 1+{Z}^{2} \right) ^{2

3605: }} {  1-Z+{Z}^{2}  }}.

3606: \eeq

3607: Observe that $T$ and $Z$ are related by:

3608: $$

3609: T=\frac{1+Z^2}{1-Z+Z^2}.

3610: $$ %\eeq

3611: We state without proof the counterparts of

3612: Propositions~\ref{propo-snake-bounded},~\ref{propo-snake-gf} 

3613: and~\ref{propo-snake-gf-rep}. Once again, the results below are dying

3614: for combinatorial explanations!

3615: \begin{Propo}

3616: Let $T_j\equiv T_j(t)$ be the \gf\ of binary trees having no label greater

3617: than $j$. Then $T_j$ is algebraic of degree (at most) $2$. In particular,

3618: $$

3619: T_0= \frac {(1-4t)^{3/2}-1+8t-2t^2}{2t(1+t)}.

3620: $$

3621: Moreover,  for all $j \ge -1$,

3622: %\beq\label{product-formTjB}

3623: $$

3624: T_j= T\,\frac{(1- Z^{j+2})(1-Z^{j+7})}{(1- Z^{j+4})(1- Z^{j+5})},

3625: %\eeq

3626: $$

3627: where $Z\equiv Z(t)$ is given by~\eqref{eqZtB}.

3628: \end{Propo}

3629: \noindent

3630: It is easy to check that the above series $T_j$ satisfy the

3631: equations~\eqref{rec-TjB} and the initial condition $T_{-1}=1$. The

3632: method we used to 

3633: \emm discover,  this product form  is again borrowed

3634: from~\cite{bdg-geodesic}.

3635: \smallskip

3636: 

3637: \noindent{\bf Remark.}

3638: For this family of trees as well, we have found  an ``invariant'' function

3639: attached to equations of the form~\eqref{rec-TjB}:

3640:  for $j\ge 0$,

3641: $$

3642: I(T_{j-1},T_j)=I(T_{j},T_{j+1})

3643: $$

3644: where 

3645: $$

3646: I(x,y)=\left( x+y \right) {t}^{2}+{\frac { \left( {x}^{2}-x-y+{y}^{2} \right) t}{xy}}+{\frac {-1+x+y}{xy}}.

3647: $$

3648: This can be used to derive rapidly from~\eqref{rec-TjB} the value of 

3649: $T_0$.

3650: 

3651: \begin{Propo}[{\bf The number of nodes labelled $j$}]

3652: \label{propo-snake-gfB}

3653: For any $j\in \zs$, the \gf \ $S_j\equiv S_j(t,u)$ that counts

3654: binary trees by the number of nodes and the number of nodes labelled $j$ is

3655: algebraic of degree at most $4$ over $\qs(T,u)$ (and thus has degree

3656: at most $8$ over $\qs(t,u)$). More precisely,

3657: $$

3658: %

3659: \frac{T^2(u-1)^2}{u(T-S_0)^2}

3660: =

3661: \frac{(T-1)^4S_0^2-2TS_0 (T-1)^2(3-9T+7T^2)+T^2(T^2+T-1)^2}

3662: {(T-1)(S_0-1)(T^2+TS_0-S_0)^2}

3663: %\eeq

3664: $$

3665: and all the series $S_j$ belong to $\qs(t,u,S_0)$.

3666: Moreover,  for all $j \ge 0$,

3667: %\beq\label{product-formSjB}

3668: $$

3669: S_j= T\,\frac{(1+\mu Z^j)(1+\mu Z^{j+5})}{(1+\mu Z^{j+2})(1+\mu Z^{j+3})},

3670: %\eeq

3671: $$

3672: where $Z\equiv Z(t)$ is given by~\eqref{eqZtB}

3673: and $\mu\equiv \mu(t,u)$ is the unique \fps\ in $t$

3674: satisfying

3675: %\beq\label{eqMZB}

3676: $$

3677: \mu= (u-1) \frac{Z(1+\mu Z)^2(1+\mu Z^2)(1+\mu Z^6)}

3678: {(1+Z)^2(1+Z+Z^2)(1-Z)^3(1-\mu^2 Z^5)}.

3679: %\eeq

3680: $$

3681: The series $\mu(t,u)$  has polynomial coefficients in $u$,

3682: and satisfies $\mu(t,1)=0$. It has degree $4$ over $\qs(Z,u)$ and $16$

3683: over $\qs(t,u)$. 

3684: \end{Propo}

3685: \noindent{\bf Comment on the proof.} The proof is similar to the proof of

3686: Proposition~\ref{propo-snake-gf-rep} until one computes the equation

3687: satisfied by $S_0$. But then, the relation $S_0=1+tS_1^2$ does not

3688: allow us to conclude that $S_1$ belongs to $\qs(t,u,S_0)$. Instead, we

3689: compute the algebraic equation satisfied by $S_1$. It is found to have

3690: degree $4$ over $\qs(u,T)$. The above relation between $S_0$ and $S_1$

3691: shows that $S_0$ belongs to the extension of $\qs(t,u)$ generated by

3692: $S_1$. Comparing the degrees implies finally that

3693: $\qs(t,u,S_0)=\qs(t,u,S_1)$. Then~\eqref{rec-Sj-binary} shows, by

3694: induction on $j$, that all the series $S_j$ belong to this field.

3695: 

3696: 

3697: 

3698: \begin{Propo}[{\bf  The number of nodes labelled $j$ or more}]

3699: \label{propo-snake-gf-repB}

3700: Let $j \in \zs$. The \gf\ $R_j(t,u)\equiv R_j$

3701: that counts binary trees by the number of nodes and the number of

3702: nodes labelled $j$ or more

3703: is  algebraic over $\qs(t,u)$.

3704: % , of degree at most $16$.  It has degree at most $4$ over

3705: % $\qs(T,\tilde T)$, where $T\equiv T(t)$ and $\tilde T\equiv T(tu)$. 

3706: More precisely, $R_0$ has degree $16$ over

3707: $\qs(t,u)$ and degree $4$ over  $\qs(T,\tilde T)$, with $\tilde T=

3708: T(tu)$, and all the series

3709: $R_j$ belong to  $\qs(T,\tilde T, R_0)=\qs(t,u,R_0)$.

3710: Moreover,  for all $j \ge 0$,

3711: %\beq\label{product-formRjB}

3712: $$

3713: R_j= T\,\frac{(1+\nu Z^j)(1+\nu Z^{j+5})}{(1+\nu Z^{j+2})(1+\nu Z^{j+3})},

3714: %\eeq

3715: $$

3716: where $Z\equiv Z(t)$ is given by~\eqref{eqZtP}

3717: and $\nu\equiv \nu(t,u)$ is a \fps\ in $t$, with polynomial

3718: coefficients in $u$, which is algebraic

3719: %  of degree $2$ over $\qs(Z(t),Z(tu))$, 

3720: of degree $8$ over $\qs(u,Z)$ and  $32$ over $\qs(t,u)$. 

3721: This series satisfies $\nu(t,1)=0$. 

3722: The first terms in its expansion are:

3723: $$

3724: \nu(t,u)=(u-1)\Big(

3725: t+(u+1)t^2+(2u^2+3u+3)t^3+O(t^4)\Big) .

3726: $$

3727: \end{Propo}

3728: \noindent{\bf Comment on the proof.} The proof is similar to the proof of

3729: Proposition~\ref{propo-snake-gf-rep} until one computes the equation

3730: satisfied by $R_0$. One finds that $R_0$ has degree $4$ over

3731: $\qs(T,\tilde T)$, and degree 16 over

3732: $\qs(t,u)$. Using~\eqref{eq-symmetryB}, one then derives an equation

3733: satisfied by $R_1$. Strangely enough, it turns out that the minimal

3734: polynomials of  $R_0$ and $R_1$ over $\qs(T,\tilde T)$ (or over

3735: $\qs(t,u)$) \emm are the same,. The two series are 

3736: of course different:

3737: $$

3738: \begin{array}{lll}

3739: R_0(t,u)&=&1+tu+u(1+u)t^2+u(2u^2+2u+1)t^3+u(1+u)(4u^2+u+2)t^4+O(t^5),\\

3740: R_1(t,u)&=&1+t+(1+u)t^2+(u^2+2u+2)t^3+(1+u)(2u^2+u+4)t^4+O(t^5).

3741: \end{array}

3742: $$

3743: 

3744: Let  $P(x)$ be the minimal polynomial of $R_0$ and $R_1$ over

3745: $\GK\equiv\qs(T,\tilde T)$.

3746: We want to prove that $R_1$ belongs to the extension of

3747: $\GK$ generated by $R_0$. Note that this property does not simply

3748: follow from the fact that $R_0$ and $R_1$ are conjugate roots of

3749: $P$. For instance, for a \emm generic, polynomial $P$ of degree 4 over $\GK$,

3750: with Galois group $\Sn_4$, the four extensions of $\GK$ generated by the

3751: roots of $P$ are  different. We are going to determine the Galois

3752: group of our polynomial $P$, using the general strategy described

3753: in~\cite[p.~141--142]{bastida}. The 

3754: resolvent cubic  of $P$, which we denote $R$ below, is 

3755: found to factor into a linear term and a quadratic one. Hence the

3756: Galois group of $R$ over $\GK$ has order 2. This implies

3757:  that the Galois group $G$ of $P$ over $\GK$ is either the cyclic

3758:  group of order 4 or 

3759:  the dihedral group of order 8. In the former case, the four

3760:  extensions of $\GK$ generated by the roots of $P$ coincide (and are

3761:  equal to the splitting field of $P$) and we are done. In the latter case,

3762:  there exists a labelling of the four roots of $P$, say $X_0,X_1,

3763: X_2, X_3$, such that the group $G$, seen as a subgroup of the

3764: permutations of $\{0,1,2,3\}$, is

3765: $$G=\{ \hbox{id}, (0,1), (2,3), (0,1) (2,3), (0,2) (1,3), (0,3) (1,2),

3766: (0, 2,1,3), (0,3,1,2)\}.$$

3767: Then

3768: 

3769: $\bullet$ the simple extensions of $\GK$ generated by the $X_i$ satisfy

3770:     $\GK(X_0)=\GK(X_1)$ and     $\GK(X_2)=\GK(X_3)$,

3771: 

3772: $\bullet$ the root of the resolvent that belongs to $\GK$

3773: is $Y=X_0X_1+X_2X_3$.

3774: 

3775: 

3776: 

3777: The root of $R$ that belongs to $\GK=\qs(T,\tilde T)$ is found to be 

3778: $Y=t^{-3} /(u(1+u)) + O(t^{-2})$. We already know two roots of $P$,

3779: namely $R_0$ and $R_1$, which are equal to $1+O(t)$. The other two

3780: roots are respectively of the form $Q_2=-t^{-2}/u+ O(t^{-1})$ and

3781: $Q_3=-t^{-1}/(1+u)+ O(1)$.  

3782: From the value of $Y$, we conclude that the above properties hold with

3783: $X_0=R_0$, 

3784: $X_1=R_1$, $X_2=Q_2$ and  $X_3=Q_3$. In particular, $R_0$ and $R_1$

3785: belong to the same extension of degree 4 of $\qs(T,\tilde T)$. 

3786: 

3787: Then, an induction on $j$, based on~\eqref{rec-RjB}, implies that for

3788: all $j\ge 2$, the series $R_j$ belongs to the extension of

3789: $\GK=\qs(T,\tilde T)$ generated by $R_0$. Since $R_1(t,u)=R_0(tu,1/u)$ and

3790: $\GK(R_0)=\GK(R_1)$, the field $\GK(R_0)$ is invariant under the

3791: transformation $A(t,u) \mapsto  A(tu,1/u)$. This property,

3792: combined with~\eqref{eq-symmetryB}, implies that for $j \ge 1$, the

3793: series $R_{-j}$ belongs to $\GK(R_0)$.

3794: 

3795: 

3796: 

3797: \subsubsection{Limit laws}

3798: We now endow the set of binary trees having $n$ nodes with the

3799: uniform distribution, and consider the same random variables as for

3800: above: $M_n$,  the largest label,  $X_n(j)$, the

3801: number of nodes having label $j$, 

3802: and  $X_n^+(j)$, the

3803: number of nodes having label $j$ at least.

3804: 

3805: Again, we can prove that 

3806:  $M_nn^{-1/4}$ converges in law to  $N_{{\hbox{\sc \small

3807:    ise}}}$,  where $N_{{\hbox{\sc \small

3808:    ise}}}$ is the supremum of the support of the ISE, and that

3809:  for all $\lambda\in \rs$, the sequence $X_n^+( \lambda n^{1/4})/ n$

3810:  converges in law to  $g_{{\hbox{\sc \small ise}}}(\lambda 

3811: )$ where $g_{{\hbox{\sc \small ise}}}$ is the tail

3812: distribution function of the ISE. The arguments are the same as for

3813:  our first class of trees (Sections~\ref{section-supremum}

3814:  and~\ref{section-coro-tail}). 

3815: Hence we could  just as well have started from the enumerative results of

3816: Section~\ref{section-enumerative-binary}, rather than from those of

3817: Section~\ref{section-snake-gf}, to characterize the laws of

3818: $N_{{\hbox{\sc \small  ise}}} $ and $g_{{\hbox{\sc \small

3819:       ise}}}(\lambda)$ (Propositions~\ref{coro-support}

3820: and~\ref{coro-tail}). 

3821: 

3822: More remarkably, we have performed on

3823: $X_n(j)$ an analysis similar to that of

3824: Section~\ref{section-limit-local}, and obtained \emm the same, local

3825: limit law. In other words, for all $\lambda \ge 0$,

3826: the sequence  $X_n(\lfloor \lambda n^{1/4}\rfloor) n^{-3/4}$ converges

3827: in law to  $

3828:  f_{{\hbox{\sc \small ise}}}(\lambda )

3829: $

3830: where $f_{{\hbox{\sc \small ise}}}$ is the conjectured

3831: density of the ISE, given in Conjecture~\ref{conj-density}.

3832: %

3833: 

3834: 

3835: In all three cases, the convergence of the moments holds as well.

3836: 

3837: 

3838: \bigskip

3839: 

3840: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

3841: 

3842: \noindent{\bf Acknowledgements.} I am extremely grateful to

3843: Jean-Fran\c cois Marckert, who not only suggested the topic of this

3844: paper, but also spent a lot of time explaining me the probabilistic

3845: implications of the limit results I had obtained. My thanks also go to

3846: Alain Rouault, who identified the moments occurring in

3847: Proposition~\ref{limit-0}, and to Philippe Flajolet, Jean-François Le

3848: Gall and Guy Louchard 

3849: for providing useful  references.

3850: 

3851: \bibliographystyle{plain}

3852: \bibliography{ise}

3853: 

3854: 

3855: \bigskip

3856: \noindent{\bf Appendix: convergence of random distribution functions}\\

3857: We want to prove the result stated without proof at the beginning of

3858: the proof of Section~\ref{section-coro-tail}. 

3859: 

3860: Recall that a sequence of real random variables $Z_n$ converges in law to

3861: another random variable $Z$ if and only if for all $x \in \rs$ such

3862: that $\PP(Z=x)=0$,

3863: $$

3864: \lim_n \PP(Z_n\le x) =\PP(Z\le x).

3865: $$

3866: This implies   the so-called \emm

3867: portmanteau, inequality: for \emm all,\/ $x \in \rs$,

3868: \beq

3869: \label{conv-law}

3870: \PP(Z<x)\le \liminf \PP(Z_n\le x)\le \limsup \PP(Z_n\le x) \le\PP(Z\le x).

3871: \eeq 

3872: 

3873: 

3874: Let us now use the notation of Section~\ref{section-coro-tail}. The

3875:  convergence of $\mu_n$ to $\mu$ implies  that for any 

3876:  bounded  Lipschitz function $f$ on $\rs$~\cite[p.~71--74]{VV}:

3877: $$

3878: \int_\rs f(x) d\mu_n(x) \dconv \int_\rs f(x) d\mu(x).

3879: $$

3880: Let $\lambda\in \rs$ and let  $f_{\lambda,\epsilon}$ and $g_{\lambda,\epsilon}$

3881: be the functions plotted in Figure~\ref{fig-epsilon}. Then

3882: $$

3883: \int_\rs g_{\lambda,\epsilon}(y) d\mu_n(y) \le 

3884: F_n(\lambda) = \mu_n(-\infty, \lambda] 

3885: \le \int_\rs f_{\lambda,\epsilon}(y) d\mu_n(y).

3886: $$

3887:  Hence, for all $x\in \rs$,

3888: $$

3889: \PP\left(\int_\rs f_{\lambda,\epsilon}(y) d\mu_n(y) \le x\right)

3890: \le \PP(F_n(\lambda)\le x)\le \PP\left(\int_\rs

3891: g_{\lambda,\epsilon}(y) d\mu_n(y) \le x\right) .

3892: $$

3893: Since $\mu_n$ converges to $\mu$, and $g_{\lambda,\epsilon}$ is a

3894: bounded Lipschitz function,

3895: $$%Z_n=

3896:  \int_\rs g_{\lambda,\epsilon}(y) d\mu_n(y) \dconv

3897: % converges in law to 

3898:  \int_\rs g_{\lambda,\epsilon}(y) d\mu(y) .

3899: $$

3900: A similar result holds for the integral involving $f_{\lambda,\epsilon}$.

3901: Thus~\eqref{conv-law} implies

3902: $$

3903: \PP\left(\int_\rs f_{\lambda,\epsilon}(y) d\mu(y) < x\right) \le 

3904: \liminf \PP(F_n(\lambda)\le x)\le \limsup \PP(F_n(\lambda)\le x)\le

3905:  \PP\left(\int_\rs g_{\lambda,\epsilon}(y) d\mu(y) \le x\right) .

3906: $$

3907: The integral occurring in the

3908: rightmost expression of this inequality is bounded from below by $\mu(-\infty,

3909: \lambda-\epsilon]$, while the integral involving

3910: $f_{\lambda,\epsilon}$ is bounded from above by $\mu(-\infty,

3911: \lambda+\epsilon]$.  Hence

3912: $$

3913: \PP(\mu(-\infty, \lambda+\epsilon]<x)  \le 

3914: \liminf \PP(F_n(\lambda)\le x)\le \limsup \PP(F_n(\lambda)\le x)\le

3915: \PP(\mu(-\infty, \lambda-\epsilon]\le x).

3916: $$

3917: Taking the limit $\epsilon \rightarrow 0^+$ gives:

3918: $$

3919: \PP(\mu(-\infty, \lambda]<x)  \le 

3920: \liminf \PP(F_n(\lambda)\le x)\le \limsup \PP(F_n(\lambda)\le x)\le

3921: \PP(\mu(-\infty, \lambda)\le x).

3922: $$

3923: If, in addition, the measure $\mu$ does not assign a positive

3924: probability to $\lambda$, the

3925: rightmost expression in the above inequality equals 

3926: $\PP(\mu(-\infty, \lambda] \le x)$. The inequality becomes

3927: $$

3928: \PP(F(\lambda)< x) \le \liminf \PP(F_n(\lambda) \le x)

3929: \le \limsup \PP(F_n(\lambda) \le x)\le \PP(F(\lambda)\le x),

3930: $$

3931: so that for all $x$ such that $\PP(F(\lambda)= x)=0$,

3932: $$

3933: \lim\PP(F_n(\lambda) \le x) = \PP(F(\lambda)\le x).

3934: $$

3935: That is,  $F_n(\lambda)$ converges in law to $F(\lambda)$.

3936: 

3937: %\tableofcontents

3938: \end{document}

3939: