1: \documentclass[10pt,twoside]{article}
2: \usepackage{nevin-pf}
3: \usepackage{url}
4:
5:
6: % $Id: FFK.tex 203 2003-06-12 18:49:33Z nevin $
7:
8: \newcommand{\FF}{{}_2F_1}
9: % \newcommand{\JF}[1]{\normalmarginpar%
10: % \marginpar{\raggedright\small\underline{JAF}: #1}}
11: \newcommand{\JF}[1]{}
12:
13: \hyphenation{Ha-da-mard}
14:
15:
16: \DeclareMathOperator{\Arg}{Arg}
17: %% For comments in the margin
18: \newcommand{\nkchanged}{\marginpar{\raggedright\small\sf 6/11}}
19: \newcommand{\malert}[1]{\marginpar{\raggedright\small\bf #1}}
20:
21:
22:
23: %% From Philippe
24: \def\frak{\mathfrak}
25: \def\bb{\mathbb}
26: \def\cal{\mathcal}
27: \def\E{{\bb{E}}
28: \def\Pr{\bb P}}
29: \def\ds{\displaystyle}
30: \def\C{\bb C}
31: \def\Z{\bb Z}
32: \pagestyle{myheadings}
33: \markboth{{\sc J.~A.~Fill, P.~Flajolet, and N.~Kapur}}{{\sc Singularity
34: Analysis and Tree Recurrences}}
35: \begin{document}
36: %%\definecolor{webgreen}{rgb}{0,0.5,0}
37:
38:
39: \title{\bf
40: Singularity Analysis, \\ Hadamard Products, \\ and Tree Recurrences}
41: \author{James Allen Fill$^{1}$ \and Philippe Flajolet${}^2$ \and Nevin
42: Kapur${}^1$\\
43: \small ${}^1$ The Johns Hopkins University,
44: Department of Mathematical Sciences,\\[-2truept]
45: \small 34th and Charles Streets,
46: Baltimore, MD 21218-2682 (USA)\\[-2truept]
47: \small \texttt{\{jimfill,nevin\}@jhu.edu},
48: \texttt{http://www.mts.jhu.edu/\~{}\{fill,kapur\}/}\\
49: \small ${}^2$ {\sc Algorithms} Project, INRIA, F-78153 Rocquencourt
50: (France) \\[-2truept]
51: \small \texttt{Philippe.Flajolet@inria.fr},
52: \texttt{http://algo.inria.fr/flajolet/}%
53: }
54:
55: \date{June~12, 2003}
56:
57: \maketitle
58:
59: \begin{abstract}
60: We present a toolbox for extracting asymptotic information on the
61: coefficients of combinatorial generating functions. This toolbox
62: notably includes a treatment of the effect of Hadamard products on
63: singularities in the context of the complex Tauberian technique
64: known as singularity analysis. As a consequence, it becomes possible
65: to unify the analysis of a number of divide-and-conquer algorithms,
66: or equivalently random tree models, including several classical
67: methods for sorting, searching, and dynamically managing equivalence
68: relations.
69: \end{abstract}
70: \emph{Keywords:} Singularity analysis, Hadamard products, tree
71: recurrences, asymptotic expansions, generating functions,
72: divide-and-conquer, shape functional, generalized polylogarithm,
73: contour integration, moment pumping \\
74: \emph{2000 Mathematics Subject Classification:} 05A16 (Primary);
75: 40E99, 68W40 (Secondary)%
76:
77:
78:
79: \bigskip
80:
81: %\marginal{What about $o$-transfers and $\sim$-transfers?}
82:
83: %% Section 1: Introduction
84:
85:
86: This study was
87: motivated by a desire to unify the analysis of a number of algorithms
88: and data structures of computer science. By analysis
89: we mean %% is meant
90: here (precise) average-case analysis of cost functions as introduced
91: by Knuth and illustrated in the collection~\cite{Knuth00} as well as
92: in his monumental series, \emph{The Art of Computer Programming} (see
93: especially~\cite{Knuth97,Knuth98a}). In the first part of this paper
94: (Section~\ref{DandC-sec} and~\ref{sec:some-special-tree}), we consider
95: a major paradigm of algorithmic design, the ``divide-and-conquer''
96: principle, which is closely related to families of random trees and
97: associated ``tree recurrences''. The basic framework is described in
98: Section~\ref{DandC-sec}, while lead examples are introduced
99: in Section~\ref{sec:some-special-tree} below. Our treatment rests on
100: combinatorial generating functions.
101:
102: % and we view a divide-and-conquer schema as effecting a well-defined
103: % transformation on singularities of associated generating functions.
104:
105: The central part of this paper (Sections~\ref{sec:sing-expans-diff}
106: and~\ref{sec:hadam-prod-transf}) is devoted to the process of
107: extracting coefficients, at least asymptotically, from generating
108: functions. Singularities have long been recognized to contain highly
109: useful information in this regard, and we start by recalling in
110: Section~\ref{sec:sing-expans-diff} the basic principles of the complex
111: Tauberian approach known as ``\emph{singularity analysis}''.
112: Applications to algorithms and trees require, in particular,
113: techniques for coping with generating functions that may be
114: constructed by a tower of several transformations. Here, we develop
115: the theory of composition of singularities under Hadamard products in
116: Section~\ref{sec:hadam-prod-transf}. (The reader only interested in
117: complex-analytic aspects can jump directly to
118: Sections~\ref{sec:sing-expans-diff} and~\ref{sec:hadam-prod-transf}.)
119:
120: The final part (Sections~\ref{sec:some-applications}
121: and~\ref{sec:appl-high-moments}) returns to the original problem of
122: analysing divide-and-conquer algorithms, taking full advantage of the
123: analytic results of previous sections. Tree recurrences and first
124: moments form the subject of Section~\ref{sec:some-applications}, where
125: full asymptotic expansions are derived for
126: expectations of costs. Section~\ref{sec:appl-high-moments} describes
127: possible extensions of the basic framework to the determination of variances
128: and higher moments as well as to some other random tree models.
129:
130:
131:
132:
133:
134:
135: \section{Introduction}\label{DandC-sec}
136:
137: ``Divide-and-Conquer'' is a major principle of algorithmic design in
138: computer science. An instance ($I$) of a problem to be solved is first
139: split into smaller subproblems ($I',I''$) that are solved recursively
140: by the same process; the partial solutions are then woven back to
141: yield a solution to the original problem. The abstract scheme is then
142: of the form:
143: \begin{equation}\label{eq1}
144: \hbox{\begin{tabular}{lll}
145: solve$(I)$ &:=& $(I',I'')$ := split$(I)$;\\
146: && $J'$ := solve$(I')$; $J''$ := solve$(I'')$;\\
147: && return weave$(J',J'')$.
148: \end{tabular}}
149: \end{equation}
150: (Problems of size smaller than a certain threshold are treated
151: directly without any recursive call.) Algorithms resorting to the
152: scheme~(\ref{eq1}) include classical sorting methods (mergesort,
153: quicksort, radix-exchange sort), data structures based on trees
154: (binary search trees, digital trees known as ``tries'', quadtrees for
155: multidimensional search, union--find trees) as well as various methods
156: used in computational geometry, distributed computation, and
157: communication theory. We refer the reader to classical books on data
158: structures, algorithms, and analysis of algorithms for details, for
159: instance,~\cite{CoLeRi90,GrKn82,Hofri95,Knuth98a,Mahmoud92,Mahmoud00,Sedgewick88,SeFl96,Szpankowski01,ViFl90}.
160:
161:
162:
163:
164:
165: In general, a class of probabilistic models $\frak{M}_n$ indexed by
166: the size~$n$ of the problem instance is assumed to reflect the nature
167: of data fed to the algorithm. A cost function---typically, the number
168: of certain operations performed by the algorithm---then becomes a
169: random variable~$X_n$ whose form is induced by~$\frak{M}_n$ and the
170: particular divide-and-conquer algorithm considered. The problem is
171: then to obtain characteristics of~$X_n$, for instance its mean, higher
172: moments, or even distributional information. The asymptotic
173: limit~$n\to\infty$ is usually considered, since an important
174: phenomenon of ``asymptotic simplification'' is to be expected in a
175: large number of situations.
176:
177:
178:
179:
180: Under natural conditions,
181: % (``randomness preservation''),
182: a recurrence that closely mimics the recursive structure
183: of~(\ref{eq1}) relates the random variables~$X_n$:
184: \begin{equation}\label{eq2}
185: X_n=t_n+X_{K_n}+\widetilde{X}_{n-a-K_n}.
186: \end{equation} % \marginal{Mention appropriate randomness.}
187: The interpretation is as follows: $t_n$ is a quantity\footnote{%
188: Some analyses require a randomly varying toll. For mean value
189: analysis, the distinction between deterministic and stochastic tolls
190: is, however, immaterial. }, called the ``toll'', that
191: represents the
192: cost incurred by splitting the initial instance and weaving back the
193: final solution; $K_n$ is the (random) size of the first subproblem, in
194: which case, the second subproblem has a size that is the complement of
195: $K_n$ to $n-a$, for some small constant $a$ (usually, $a=0$ or~$a=1$),
196: which is specific to the algorithm considered. The random variables
197: of type $X$ and $K$ are assumed to be independent, as are
198: the two $X$-sequences~$X$ and~$\widetilde{X}$ on the
199: right
200: in~(\ref{eq2}), and a subproblem of
201: size~$k$ is assumed to satisfy model~$\mathfrak{M}_k$---this property
202: is sometimes called ``randomness preservation'' and is satisfied by
203: many cases of algorithmic interest. A direct asymptotic treatment of
204: the recursive relation~(\ref{eq2}) binding random variables is
205: sometimes feasible; see the (metric) ``contraction method'' surveyed
206: by R\"osler and R\"uschendorf~\cite{RoRu01} and applied by
207: Neininger~\cite{Neininger02} to a subset of the problems discussed
208: here.
209:
210:
211: Turning to average-case analysis, the expected cost $f_n := \E(X_n)$
212: satisfies a \emph{recurrence} that is directly implied by~(\ref{eq2}):
213: \begin{equation}\label{eq3}
214: f_n=t_n+\sum_{k} p_{n,k} (f_k+f_{n-a-k}).
215: \end{equation}
216: with
217: % $t_n := \E(T_n)$. the expectation of the toll and the
218: the \emph{splitting probabilities} $p_{n,k}:=\Pr(K_n=k)$ being
219: determined by the model~$\frak{M}_n$ used. Trees are naturally
220: associated with recursive procedures, and, accordingly, the
221: recurrence~(\ref{eq3}) can be viewed as associated with a random tree
222: model of the following form: the root has size~$a$, the left subtree
223: has size~$k$ with probability $p_{n,k}$, and the right subtree has the
224: remaining quantity~$n-a-k$ as size. Then~(\ref{eq3})is
225: interpreted as giving the expectation of a cost function over the tree
226: structure that is induced by the family of tolls, $t_n$. For this
227: reason, a recurrence having the form~(\ref{eq3}) is called a
228: \emph{tree recurrence}. Tree recurrences are the main object of study
229: of this paper.
230:
231:
232:
233: One way to view the tree recurrence~(\ref{eq3}) is as a linear
234: transformation on sequences
235: \begin{equation}\label{eq4}
236: (f_n)= {\cal K}\left[(t_n)\right],
237: \end{equation}
238: that takes a toll sequence $(t_n)$ and returns the corresponding
239: average-cost sequence $(f_n)$. The functional $\cal K$ is fully
240: determined by the splitting probabilities $p_{n,k}$. A classical
241: approach to the derivation of explicit forms consists in introducing
242: \emph{generating functions} (GFs). Fix a sequence of
243: \emph{normalization constants} $\omega_n$ (that are problem-specific)
244: and define the generating functions
245: % \marginal{NOTATION:\ $f$ of $f(z)$?; $\omega_n$ or $w_n$?; $f$ of
246: % $\phi$ (to avoid confusion with ordinary GF)?}
247: \[
248: f(z):=\sum f_n \omega_n z^n, \qquad t(z):=\sum t_n \omega_n z^n.
249: \]
250: Then, the transformation~$\cal K$ induces another linear
251: transformation $\cal L$ on GFs:
252: \begin{equation}\label{eq5}
253: f(z)= {\cal L}\left[t(z)\right].
254: \end{equation}
255: With an adequate choice of the constants~$\omega_n$, explicit forms
256: of~$f_n$ can often be obtained, provided at least that toll sequences
257: are of a simple enough form.
258:
259: Our main objective is to develop generating-function methods
260: by which one can quantify the
261: way the \emph{asymptotic form} of (expected) costs relates to
262: properties of the \emph{toll sequence}.
263: %% , placing ourselves within the
264: %% framework of the generating function approach.
265: It is known that
266: asymptotic properties of number sequences (as the index~$n$ tends to
267: infinity) are closely related to the nature of
268: the~\emph{singularities} of the corresponding generating functions.
269: This suggests that we examine the way the operator~$\cal L$
270: operates on
271: scales of singular functions and view it as a ``\emph{singularity
272: transformer}''. Informally, there is a
273: transformation~$\widehat{\cal L}$, induced by~$\cal L$ and acting on
274: an asymptotic scale of functions singular at some fixed point~$z_0$.
275: Using $\operatorname{Sing}(f(z))$ to denote the expansion of~$f(z)$ at
276: the singularity~$z_0$, one has
277: \begin{equation}\label{eq6}
278: % \operatorname{Sing}(f(z))=\widehat{\cal
279: % L}\left[\operatorname{Sing}(t(z))\right]
280: \operatorname{Sing}(f(z))=\widehat{\cal L}\left[t(z)\right]
281: \end{equation}
282: Under fairly general conditions, there is a tight coupling between
283: singular expansions of a generating function and the asymptotic form
284: of its coefficients. The outcome of this process, justified by
285: \emph{singularity analysis}~\cite{FlOd90b,Odlyzko95}, is a direct
286: relation written figuratively as
287: \begin{equation}\label{eq7}
288: \operatorname{Asympt} ((f_n))=\widehat{\cal
289: K}\left[(t_n)\right],
290: \end{equation}
291: where~$\widehat{\cal K}$ depends on $(t_n)$ via the % singular
292: structure of its generating function~$t(z)$.
293:
294: The path we follow in this paper is the one given
295: by~(\ref{eq4})--(\ref{eq7}), which is then globally
296: summarized by the following diagram:
297: \begin{equation}\label{commdiag}
298: \begin{array}{ccccc}
299: \ds (t_n) \mathop{\longrightarrow}^{\ds\cal K} (f_n) &&&&
300: \ds (t_n) \mathop{\longrightarrow}^{\ds\widehat{\cal K}}
301: \operatorname{Asympt}((f_n)) \\
302: \\
303: \ds \Downarrow &&&&\ds \Uparrow
304: \\
305: \\
306: \ds t(z) \mathop{\longrightarrow}^{\ds\cal L} f(z) && \Longrightarrow &&
307: \ds t(z) \mathop{\longrightarrow}^{\ds\widehat{\cal L}}
308: \operatorname{Sing}(f(z))
309: \end{array}
310: \end{equation}
311: We propose to develop a collection of generic tools that
312: supplement the basic singularity analysis framework of Flajolet and
313: Odlyzko~\cite{FlOd90b}. In particular, we discuss in the next
314: sections the action on singularities of differential and integral
315: operators, as well as of Hadamard products. As a result, the
316: way~$\cal L$ operators associated with many recurrences transform
317: singularities can be analyzed precisely. This in turn yields a fairly
318: general classification of the asymptotic growth phenomena associated
319: to a variety of classical tree recurrences, including the ones of
320: binary search trees, binary trees, and union--find trees, which will
321: serve here as guiding examples.
322:
323: \medskip
324:
325: Most of the existing computer science literature is devoted to the
326: ``deterministic'' divide-and-conquer recurrences that correspond to a
327: splitting size $K_n$ that is deterministic, depending on~$n$
328: alone---typically, $K_n\equiv\lfloor n/2\rfloor$. In such a case, the
329: probability distribution $(p_{n,k})_{k=0}^n$ is supported at a single
330: point. The main asymptotic order of~$f_n$ is then given by what
331: Cormen, Rivest, and Leiserson have termed ``master theorems'':\
332: see~\cite{CoLeRi90,GrKn82,SeFl96}. (Usually, the finer
333: characteristics of the asymptotic regime involve fractal
334: fluctuations~\cite{FlGo94,SeFl96}.) What we consider here instead are
335: methods for dealing with ``stochastic'' divide-and-conquer
336: recurrences, where $K_n$ is a random variable (dependent on~$n$) with
337: support spread over a whole subinterval in~$(0,n)$. This stochastic
338: case is discussed by Roura in~\cite{Roura97}: Roura's arguments are
339: based on elementary real analysis, so that they are of quite a wide
340: scope, but his estimates are by nature mostly confined to first-order
341: asymptotics. In this article we show that, in the many cases of
342: practical interest where some strong complex-analytic structure is
343: present, full asymptotic expansions can be derived. Our treatment is
344: somewhat parallel in spirit to that of Knuth and Pittel whose
345: inspiring work~\cite{KnPi89} provided one of the initial
346: motivations\footnote{ See also Pittel's interesting recent
347: article~\cite{Pittel99} which appeared while our own work was still
348: in progress. } for the present study. An additional benefit of the
349: complex-analytic approach is that it often gives access to variances
350: and higher moments, in which case the limit distribution of
351: costs can be identified.
352:
353:
354:
355:
356: \section[Special tree recurrences]{Some ``special'' tree recurrences}\label{sec:some-special-tree}
357:
358: In this section, we briefly review some tree recurrences that
359: are of special interest in combinatorial mathematics and analysis of
360: algorithms.
361:
362: \subsection{The binary search tree recurrence}
363: One of the simplest model of random trees is defined as follows: To
364: determine a tree $T_n$ of size~$n\ge1$, take a root and append to it a
365: left subtree of size $k$ and a right subtree of size~$n-k$, where $k$
366: is uniformly distributed over the set of permissible
367: values~$\{0,1,\ldots n-1\}$; a tree of size~$0$ is the empty tree. In
368: earlier notations, this process corresponds to
369: \begin{equation}\label{bst1}
370: p_{n,k}\equiv \Pr(K_n = k) :=\frac{1}{n},\qquad\hbox{for}\quad
371: k=0,1,\ldots n-1.
372: \end{equation}
373: As is well known, the model defined by~\eqref{bst1} corresponds to
374: random trees defined by either the binary search tree data structure
375: or the quicksort
376: algorithm~\cite{Knuth98a,Mahmoud92,Mahmoud00,SeFl96,ViFl90}. The
377: corresponding tree recurrence~\eqref{eq3} is then
378: \begin{equation}\label{bst2}
379: f_n=t_n+\frac{2}{n}\sum_{k=0}^{n-1} f_k,
380: \end{equation}
381: with $f_0 := t_0$.
382:
383: Ordinary GFs determined by the choice of coefficients $\omega_n=1$ for
384: all~$n$ are then
385: \[
386: f(z):=\sum_{n\ge0} f_n z^n, \qquad t(z):=\sum_{k\ge0} t_n z^n,
387: \]
388: and standard rules for the manipulation of GFs translate~(\ref{bst2})
389: into a linear integral equation
390: \[
391: f(z)=t(z)+2\int_0^z f(w) \, \frac{dw}{1-w}.
392: \]
393: Differentiation yields the ordinary differential equation
394: \[
395: f'(z)=t'(z)+\frac{2}{1-z}f(z),
396: \]
397: which is then solved by the variation-of-constants method:
398: \begin{equation}\label{bst3}
399: f(z)=\cal L[t(z)], \qquad
400: \hbox{where}\quad \cal L[t(z)]:=(1-z)^{-2}\int_0^z
401: \left(\partial_w t(w)\right) (1-w)^2\, dw.
402: \end{equation}
403: In~(\ref{bst3}) we have assumed without loss of generality the initial
404: conditions $t_0=f_0=0$ (thanks to linearity and the fact that the
405: transform of~$t_n \equiv \delta_{n,0}$ is $n+1$). The
406: notation~$\partial_w$
407: borrowed from differential algebra is used to denote derivatives
408: whenever the operator nature of transformations is to be stressed.
409:
410:
411: It is instructive to follow what Greene and Knuth call the
412: ``repertoire'' approach~\cite{GrKn82}. This consists in building a
413: repertoire of the ($\cal K$ or $\cal L$) transforms of basic tolls,
414: then trying to determine the effect of a new toll by expressing it in
415: the basis of known tolls. What is convenient here is the class of
416: tolls
417: \[
418: t_n^\alpha:=\binom{n+\alpha}{\alpha} = \frac{(\alpha + 1)(\alpha + 2)
419: \cdots (\alpha + n)}{n!}, \qquad \hbox{i.e.,}\quad
420: t^\alpha(z)=(1-z)^{-\alpha-1}.
421: \]
422: Then, by~(\ref{bst3}) one finds, for~$\alpha\not=1$,
423: \[\begin{array}{lll}
424: f^{\alpha}(z)&=&\ds \frac{\alpha+1}{\alpha-1}\left[(1-z)^{-\alpha-1}-(1-z)^{-2}
425: \right]\\
426: f^\alpha_n&=&\ds
427: \frac{\alpha+1}{\alpha-1}\left[\binom{n+\alpha}{\alpha}-\binom{n+1}{1}\right],
428: \end{array}
429: \]
430: while $\alpha=1$ leads to
431: \[
432: f^{1}(z)=\frac{2}{(1-z)^2}\log\frac{1}{1-z}, \qquad
433: f^{1}_n=2(n+1)\left(H_{n+1}-1\right),
434: \]
435: with $H_n : =1+\tfrac12+\cdots+\tfrac1n$ the $n$th harmonic number.
436:
437: Stirling's formula implies asymptotically, for~$\alpha$ not a negative
438: integer,
439: % \marginal{$\alpha \not\in \{-1,-2,\ldots \}$.}
440: \[
441: \binom{n+\alpha}{\alpha} \sim \frac{n^{\alpha}}{\Gamma(\alpha+1)},
442: \]
443: with $\Gamma$ the Euler gamma function. Then what goes on
444: is summarized by the following table:
445: \begin{equation}\label{bst999}
446: \hbox{\begin{tabular}{lll}
447: \hline\hline
448: $\ds t_n=\binom{n+\alpha}{\alpha}$, $\alpha>1$
449: &&
450: % $f_n\sim \frac{\alpha+1}{\alpha-1}t_n$\\
451: $f_n\sim \frac{\alpha+1}{\alpha-1}\frac{n^{\alpha}}{\Gamma(\alpha+1)}$\\
452: $t_n=n+1$
453: &&
454: $f_n\sim 2 n\log n$\\
455: $\ds t_n=\binom{n+\alpha}{\alpha}$, $0<\alpha<1$
456: &&
457: $f_n\sim \frac{1+\alpha}{1-\alpha}n$.\\
458: \hline\hline
459: \end{tabular}}
460: \end{equation}
461: The discontinuity in the asymptotic regime of~$f$ at~$\alpha=1$, where
462: a logarithm appears, is noticeable.
463: Also, the tolls in the scale satisfying $t_n\ll n$ are seen to induce
464: costs that all collapse to linear functions.
465:
466: A full discussion of the binary search tree recurrence necessitates
467: determining the effect of toll functions like $\sqrt{n}$, $\log n$,
468: and
469: $1/n^2$, a task which is not entirely elementary. By the remarks of
470: the introduction, this involves determining the singularities of the
471: corresponding generating functions \emph{and}, in view
472: of~(\ref{bst3}), making explicit the way singular expansions get
473: composed under differentiation and integration
474: (Section~\ref{sec:sing-expans-diff}). This subject will then be taken
475: up again in Section~\ref{sec:binary-search-tree}; the particular case
476: of the toll~$t_n=\log n$ is of special importance and will be treated
477: in detail there.
478:
479:
480: \subsection{The uniform binary tree recurrence}
481: This recurrence is of the form ($n \geq 0$, with the convention~$f_0
482: := t_0$)
483: \begin{equation}\label{cat1}
484: f_n=t_n+\sum_{k=0}^{n-1} \frac{C_k C_{n-1-k}}{C_n}\left(f_k+f_{n-k}\right),
485: \qquad
486: \hbox{with}\quad
487: C_n :=\frac{1}{n+1}\binom{2n}{n},
488: \end{equation}
489: a Catalan number. It corresponds to the uniform model of binary
490: trees, where all the $C_n$ binary trees with~$n$ internal nodes are
491: taken with equal likelihood. Indeed, the number of trees of size~$n$
492: satisfies the recurrence
493: \begin{equation}\label{cat2}
494: C_n=\sum_{k=0}^{n-1} C_k C_{n-1-k}\quad (n \geq 1), \qquad
495: C_0=1,
496: \end{equation}
497: as seen from a root decomposition. The
498: quantity~$p_{n,k}=C_kC_{n-1-k}/{C_n}$ is then the probability that a
499: tree of size~$n$ has left and right subtrees of respective sizes $k$
500: and $n-1-k$.
501:
502: The GF of Catalan numbers satisfies a relation that is the image of
503: the recurrence~(\ref{cat2}), namely, $C(z)=1+zC(z)^2$, so that
504: \begin{equation}\label{cat3}
505: C(z)=\frac{1}{2z}\left(1-\sqrt{1-4z}\right).
506: \end{equation}
507: In order to solve~(\ref{cat1}) by generating functions, one should use
508: as
509: normalization constants the quantities~$\omega_n=C_n$, and introduce
510: \begin{equation}\label{catcost1}
511: t(z):=\sum_{n\ge0} t_n C_n z^n, \qquad
512: f(z):=\sum_{n\ge0} f_n C_n z^n.
513: \end{equation}
514: % \marginal{See note preceding~\eqref{eq5}.}
515: Then~(\ref{cat1}) translates into a linear algebraic equation,
516: \[
517: f(z)=t(z)+2zC(z)f(z),
518: \]
519: from which the form of the~$\cal L$ operator immediately results:
520: \begin{equation}\label{catcost2}
521: f(z)=\cal L[t(z)], \qquad
522: \hbox{where}\quad \cal L[t(z)]=\frac{1}{\sqrt{1-4z}} t(z).
523: \end{equation}
524: % As such, \marginal{Reword?}
525: This form makes it possible to analyze directly only a restricted
526: collection of tolls, for instance, ones of the form
527: $t_n^r:=(n+1)n\cdots(n-r+2)$ (by differentiation), or
528: $t_n^{-r}=1/((n+2)(n+3)\cdots(n+r+1))$ (by integration). However,
529: tolls of such simple forms as $\sqrt{n}$, $H_n$, and~$\log n$, are
530: left out of the scale of the $t_n^{\pm r}$.
531:
532:
533:
534:
535: Define the \emph{Hadamard product} of two entire series or two
536: functions analytic at the origin, $a$ and~$b$, as their termwise
537: product,
538: \begin{equation}\label{hadam-def}
539: a(z)\odot b(z)=\sum_{n\ge0} a_n b_n z^n,
540: \qquad
541: \hbox{if}\quad
542: a(z)=\sum_{n\ge0} a_n z^n,\quad
543: b(z)=\sum_{n\ge0} b_n z^n.
544: \end{equation}
545: Then, from~(\ref{catcost1}) and~(\ref{catcost2}), the cost functional
546: is expressed by the
547: modified transformation (of~$\cal L$ type)
548: \begin{equation}\label{cat99}
549: f(z)=\frac{\tau(z)\odot C(z)}{\sqrt{1-4z}}, \quad
550: \hbox{where}\quad
551: f(z)=\sum_n f_n C_nz^n, \quad\tau(z):=\sum_n t_n z^n\,.
552: \end{equation}
553: This now relates the \emph{ordinary} generating function $\tau(z)$ of
554: the tolls and the \emph{normalized} generating function $f(z)$ of the
555: costs (with the $\omega_n=C_n$ normalization) via a Hadamard product.
556:
557:
558:
559: Determining the way costs get transformed under this model then
560: necessitates a way to combine singular expansions under Hadamard
561: products. This is the central part of our article; see
562: Section~\ref{sec:hadam-prod-transf}, where a general theorem is
563: stated. The ``critical'' value for tolls at which a discontinuity in
564: the induced costs manifests itself is now at $t_n=\sqrt{n}$,
565: and\footnote{%
566: The notation $x=\Theta(y)$ expresses the inequalities $c_1y<x<c_2y$
567: for some constants $c_1,c_2$ satisfying $0<c_1<c_2<+\infty$. }
568: \begin{equation}\label{cat999}
569: \hbox{\begin{tabular}{lll}
570: \hline\hline
571: $t_n=n^{\alpha}$, $\alpha>1/2$ && $f_n=\Theta(n^{\alpha+\frac12})$\\
572: $t_n=n^{1/2}$ && $f_n=\Theta(n\log n)$\\
573: $t_n=n^{\alpha}$, $0 < \alpha < 1/2$ && $f_n=\Theta(n)$.\\
574: \hline\hline
575: \end{tabular}%
576: }
577: \end{equation}
578: This phenomenon observed in~\cite[Prop.~2]{FlSt87} [of
579: which~(\ref{cat999}) above corrects a few misprints] neatly
580: distinguishes the binary Catalan model from the binary search tree
581: model, as seen by comparing~(\ref{cat999}) to~(\ref{bst999}). A proof
582: accompanied by complete expansions will be given in the application
583: section: see Section~\ref{sec:binary-tree-recurr} below.
584:
585:
586:
587: \subsection{The union--find tree recurrence}
588: \label{sec:cayl-tree-recurr-1}
589: By a result attributed to Cayley, there are $U_n =n^{n-2}$ ``free''
590: unrooted trees (i.e., labeled connected acyclic graphs) on~$n$ nodes,
591: and, accordingly, $T_n=n^{n-1}$ rooted trees. Consider the model in
592: which initially each unrooted tree of size~$n$ is taken with equal
593: likelihood. Choose an edge at random amongst any of the possible
594: $n-1$ edges of the tree, orient it in a random way, then cut it. This
595: separates the tree into an ordered pair of smaller trees that are now
596: rooted. Continue the process with each of the resulting subtrees,
597: discarding the root. Assume that the cost incurred by selecting the
598: edge and splitting the tree is $t_n$. Then the total cost incurred
599: when starting from a random unrooted tree and recursively splitting it
600: till the completely disconnected graph is obtained satisfies the
601: recurrence $(n \geq 1)$
602: \begin{equation}\label{cay1}
603: f_n=t_n+\sum_{k=1}^{n-1}p_{n,k}(f_k+f_{n-k}),
604: \quad\hbox{where}\quad
605: p_{n,k}=\binom{n}{k}\frac{k^{k-1}(n-k)^{n-k-1}}{2(n-1)\, n^{n-2}}.
606: \end{equation}
607: (Proof: There are $n^{n-1}$ rooted trees on~$n$ nodes and the binomial
608: coefficient takes care of relabellings.) The recurrence~(\ref{cay1})
609: has been studied in great detail by Knuth and Pittel in~\cite{KnPi89},
610: an article that largely motivated our study. In fact, there are good
611: algorithmic reasons for considering the recurrence~(\ref{cay1}): if
612: time is reversed, then the recursion describes the evolution of a
613: random graph from totally disconnected to tree-like, when successive
614: edges are added at random. The latter is exactly the probabilistic
615: model involved in the ``union--find'' (or equivalence-finding)
616: algorithm~\cite{CoLeRi90,Sedgewick88,ViFl90}, for which detailed
617: analyses had been provided by Knuth and
618: Sch\"onhage~\cite{KnSc78} in 1978\footnote{%
619: Precisely, the model is known as the ``random spanning tree model''.
620: The derivation of our equation~(\ref{cay2}) closely mimics
621: Section~11 of~\cite{KnSc78}.}. (Note that this model is not the same
622: as the simply generated family of Cayley trees.)
623:
624: Let $T(z)$, $U(z)$ be the exponential generating functions of the
625: sequences~$(T_n)$, $(U_n)$, that is,
626: \[
627: T(z)=\sum_{n=1}^\infty n^{n-1}\frac{z^n}{n!}, \qquad
628: U(z)=\sum_{n=1}^\infty n^{n-2}\frac{z^n}{n!},
629: \]
630: It is a well-known fact of combinatorics that $T(z)$ satisfies the
631: functional relation $T(z)=ze^{T(z)}$, and one has $U=T-(T^2/2)$;
632: see~\cite{GoJa83,Knuth97,Stanley98}. Define now the generating
633: functions
634: \[
635: t(z)=\sum_{n\ge1}t_n(n^{n-1}-n^{n-2})\frac{z^n}{n!},\qquad
636: f(z)=\sum_{n\ge1}f_n n^{n-1}\frac{z^n}{n!},
637: \]
638: where the normalization constants for~$f(z)$ are $\omega_n=n^{n-1}/n!$
639: and, for convenience, a marginally different normalization,
640: $\omega'_n=n^{n-2}(n-1)/n!$, has been introduced in the case
641: of $t(z)$.
642: Then the recurrence~(\ref{cay1}) has the form of a binomial
643: convolution, so that the cost GF~$f(z)$ satisfies
644: \[
645: f(z)-\int_0^z f(w)\, \frac{dw}{w}=t(z)+f(z)T(z).
646: \]
647: By differentiation, this last relation transforms into a linear
648: differential equation of the first order, itself readily
649: solved by the
650: variation-of-constants method. Assuming (without loss of generality)
651: the initial condition $t_1=f_1=0$, the solution found is
652: \begin{equation}\label{cay2}
653: f(z)=\frac{T(z)}{1-T(z)}\int_0^z \partial_w t(w)\, \frac{dw}{T(w)}.
654: \end{equation}
655:
656: %% \malert{Reminder for later: If $t_1=f_1 \ne 0$ then $f$ is given
657: %% by~(6.32) in the original draft. It is no more difficult to handle
658: %% that case than it is to handle the simplified case presented here.}
659: In terms of the ordinary generating function of costs, namely,
660: % \marginal{Notation.}
661: \[
662: \tau(z):=\sum_{n\ge2}t_n z^n,
663: \]
664: equation~(\ref{cay2}) can be rephrased as an integral transform
665: involving a Hadamard product, namely,
666: \begin{equation}\label{cay3}
667: f(z)=\frac12\frac{T(z)}{1-T(z)}\int_0^z \partial_w \left(\tau(w)\odot
668: T(w)^2\right) \, \frac{dw}{T(w)}.
669: \end{equation}
670: The dominant singularity at~$z=e^{-1}$ of the Cayley tree function~$T$
671: is
672: well known to be of the square root type. Then the integral
673: transform~(\ref{cay3}) operates in a way that combines a Hadamard
674: product and ordinary products, as well as integration and
675: differentiation. This subject will be resumed in
676: Section~\ref{sec:cayl-tree-recurr}, after general theorems have been
677: established by which one can cope with such situations. The final
678: conclusions turn out to be qualitatively similar to what was observed
679: for the Catalan model in~(\ref{cat999}).
680:
681:
682:
683:
684:
685:
686: %% Section 3: Singular expansions and differentiation
687: \section{Singular expansions, differentiation, and integration}\label{sec:sing-expans-diff}
688:
689:
690: Singularities of generating functions encode very precise information
691: regarding the asymptotic behaviour of coefficients. In this section,
692: we first recall in Subsection~\ref{subsec:basicSA} the principles of a
693: process by which this information can be extracted: this is the
694: singularity analysis framework of~\cite{FlOd90b,Odlyzko95}. We then
695: prove that functions amenable to singularity analysis are closed under
696: integration and differentiation; see Subsection~\ref{subsec:diffint}.
697: These operations have already been seen to intervene in the
698: analysis of some of the major tree recurrences.
699: % belonging to the uniform tree and the Cayley tree types.
700:
701: \subsection{Basics of singularity analysis}\label{subsec:basicSA}
702:
703: Singularity analysis deals with functions that have isolated
704: singularities on the boundary of their disc of convergence and are
705: consequently continuable to wider areas of the complex plane. The
706: case of a unique dominant singularity suffices for the applications
707: treated here. (In addition, the case of finitely many dominant
708: singularities is easily reduced to this situation by using composite
709: contours and cumulating contributions arising from individual
710: singularities.) Given the obvious scaling rule,
711: \[
712: [z^n] f(z)=\rho^{-n}f(\rho z),
713: \]
714: one may restrict attention, whenever necessary, to the case where the
715: singularity is at~1. The scaling rule shows that the position of the
716: singularity [at~$\rho$ for $f(z)$] introduces an exponential scaling
717: factor ($\rho^{-n}$) multiplied by the coefficient of a function
718: singular at~1 [the function~$f(\rho z)$].
719:
720:
721: \begin{definition}\label{def:1}\em
722: A function defined by a Taylor series with radius of convergence
723: equal to~$1$ is \emph{$\Delta$-regular} if
724: % , with the sole exception of $z=1$,
725: it can be analytically continued in a domain
726: \begin{equation*}
727: \Delta(\phi,\eta) := \{z: |z| < 1 + \eta, |\Arg(z-1)| > \phi\},
728: \end{equation*}
729: for some $\eta > 0$ and $0 < \phi < \pi/2$. A function $f$ is said to
730: admit a \emph{singular expansion} at $z=1$ if it is $\Delta$-regular
731: and
732: \begin{equation}
733: \label{eq:27}
734: f(z) = \sum_{j=0}^J c_j(1-z)^{\alpha_j} + O(|1-z|^A)
735: \end{equation}
736: uniformly in $z \in \Delta(\phi,\eta)$,
737: for a sequence of complex numbers
738: $(c_j)_{0 \leq j \leq J}$ and an increasing sequence of real numbers
739: $(\alpha_j)_{0 \leq j \leq J}$ satisfying $\alpha_j < A$. It is said
740: to satisfy a singular expansion \emph{``with logarithmic terms''} if,
741: similarly,
742: \begin{equation}\label{eq:27ext}
743: f(z) = \sum_{j=0}^J c_j\left(L(z)\right)(1-z)^{\alpha_j} + O(|1-z|^A),
744: \qquad
745: L(z):=\log\frac{1}{1-z},
746: \end{equation}
747: where~each $c_j(\cdot)$ is a polynomial.
748: \end{definition}
749: Note that, by assumption, the $O(\cdot)$ error term in~\eqref{eq:27}
750: must hold uniformly in $z \in \Delta(\phi,\eta)$.
751: We also allow in the
752: usual way infinite asymptotic expansions
753: % \begin{equation*}
754: % f(z) \sim \sum_{k \geq 0} c_k(1-z)^{\alpha_k},
755: % \end{equation*}
756: representing an infinite collection of mutually compatible expansions
757: of type~\eqref{eq:27}.
758:
759: For the sake of notational simplicity, we shall mostly limit our
760: statements to the basic case~(\ref{eq:27}) and briefly comment on how
761: they extend to the logarithmic case~(\ref{eq:27ext}). The basic
762: theorem is the following:
763: \begin{theorem}[Basic singularity analysis~\cite{FlOd90b}]\label{thm:basic-sa}
764: If~$f(z)$ admits a singular expansion of the form~(\ref{eq:27})
765: valid in a $\Delta$-domain, then
766: \begin{equation}\label{xfer}
767: [z^n]f(z)=\sum_{j=0}^J
768: c_j\binom{n-\alpha_j-1}{-\alpha_j-1}+O(n^{-A-1}).
769: \end{equation}
770: \end{theorem}
771: (The proof of this and similar results is based on an extensive use of
772: Hankel contours; see the already cited references.) The last
773: expansion can be rephrased as a standard asymptotic expansion since,
774: for~$\alpha\not\in\{0,1,2\ldots\}$, one has
775: \[
776: \binom{n-\alpha-1}{-\alpha-1}\sim
777: \frac{n^{-\alpha-1}}{\Gamma(-\alpha)}
778: \left(1+\frac{\alpha(\alpha+1)}{2n}+
779: \frac{\alpha(\alpha+1)(\alpha+2)(3\alpha+1)}{24n^2}+\cdots\right),
780: \]
781: while all the terms corresponding to~$\alpha$ a nonnegative integer
782: have an asymptotically null contribution. When logarithmic terms are
783: present in the singular expansion, corresponding logarithmic terms
784: arise in the asymptotic expansion of coefficients. The calculations
785: are conveniently carried out by differentiation with respect to the
786: parameter~$\alpha$:
787: \[
788: [z^n](1-z)^{\alpha}L(z)^r=(-1)^r\frac{\partial^r}{\partial
789: \alpha^r}[z^n](1-z)^\alpha =(-1)^r\frac{\partial^r}{\partial
790: \alpha^r}\binom{n-\alpha-1}{-\alpha-1},
791: \]
792: which yields for instance ($\alpha\not\in\{0,1,2\ldots\}$):
793: \[
794: \begin{array}{lll}
795: \ds [z^n](1-z)^{\alpha}L(z)& =&\ds -\frac{\partial}{\partial \alpha}
796: \binom{n-\alpha-1}{-\alpha-1}
797: \\
798: &=&\ds \ds \binom{n-\alpha-1}{-\alpha-1}
799: \left(\frac{1}{-\alpha}+\frac{1}{1-\alpha}+\cdots+\frac{1}{n-1-\alpha}
800: \right)\\
801: &=& \ds\frac{n^{-\alpha-1}}{\Gamma(-\alpha)}\left(\log
802: n-\psi(-\alpha)+O\left(\frac{\log n}{n}\right)
803: \right).
804: % &\sim& \ds\frac{n^{-\alpha-1}}{\Gamma(-\alpha)}\left(\log
805: % n-\psi(-\alpha)-\frac{1}{2n}\left(1+2\alpha-\alpha(\alpha+1)(\log
806: % n-\Psi(-\alpha))\right)
807: % +\cdots\right).
808: %% \\[3truemm]
809: %% \ds [z^n](1-z)^{\alpha}L^2(z)& =&\ds \frac{\partial^2}{\partial \alpha^2}
810: %% \binom{n-\alpha-1}{-\alpha-1}
811: %% \\
812: %% &=& \ds\frac{n^{-\alpha-1}}{\Gamma(-\alpha)}\left(\left(\log
813: %% n-\psi(-\alpha)\right)^2-\psi'(-\alpha)+O\left(\frac{\log^2 n}{n}\right)
814: %% \right).
815: \end{array}\]
816: (Here $\psi$ is the logarithmic derivative of~$\Gamma$.)
817:
818:
819: The same proof techniques also make it possible to translate error
820: terms involving logarithmic terms; see~\cite{FlOd90b} for details. In
821: particular, the following transfer holds for~$A$ and~$B$ real numbers:
822: \begin{equation}\label{log-xfer}
823: O((1-z)^A L^B(z))\qquad\leadsto\qquad O(n^{-A-1}\log^B n).
824: \end{equation}
825:
826:
827: Finally, we shall make use of a result which renders amenable to
828: singularity analysis generating functions whose coefficients involve
829: powers of $n$ and its logarithms.
830:
831: \begin{definition}\em
832: The generalized polylogarithm~$\Li_{\alpha,r}$, where~$\alpha$ is an
833: arbitrary complex number and $r$ a nonnegative integer is defined
834: for~$|z|<1$ by
835: \[
836: \Li_{\alpha,r}(z) :=\sum_{n\ge1} (\log n)^r \frac{z^n}{n^\alpha},
837: \]
838: and the notation~$\Li_\alpha$ abbreviates~$\Li_{\alpha,0}$.
839: \end{definition}
840: In particular, one has $\Li_{1,0}(z)=\Li_1(z)=L(z)$, the usual
841: logarithm, cf.~(\ref{eq:27ext}). The singular expansion of the
842: polylogarithm, taken from~\cite{Flajolet99}, involves the Riemann
843: $\zeta$ function:
844: \begin{theorem}[Singularities of polylogarithms~\cite{Flajolet99}]\label{thm:li-sing}
845: The function~$\Li_{\alpha,r}(z)$ is $\Delta$--continuable and, for
846: $\alpha\not\in\{1,2,\ldots\}$, it satisfies the singular expansion
847: \begin{equation}\label{li-sing}
848: \Li_{\alpha,0}(z)\sim\Gamma(1-\alpha)w^{\alpha-1}+
849: \sum_{j\ge0}\frac{(-1)^j}{j!}\zeta(\alpha-j)w^j, \quad
850: w=-\log z=\sum_{\ell=1}^\infty \frac{(1-z)^\ell}{\ell}.
851: \end{equation}
852: For~$r>0$, the singular expansion of~$\Li_{\alpha,r}$ is obtained by
853: \[
854: \Li_{\alpha,r}(z)=(-1)^r\frac{\partial^r}{\partial
855: \alpha^r}\Li_{\alpha,0}(z),
856: \]
857: and corresponding termwise differentiation of~\eqref{li-sing} with
858: respect to~$\alpha$.
859: \end{theorem}
860: In particular, for~$\alpha<1$, the main asymptotic term
861: of~$\Li_{\alpha,r}$ is
862: \[
863: \Gamma(1-\alpha) (1-z)^{\alpha-1} L^r(z).
864: \]
865: Similar expansions hold when $\alpha$ is a positive integer;
866: see~\cite{Flajolet99} for details.
867:
868: \begin{example} \emph{Stirling's formul\ae.}
869: The factorial function, is attainable via the form
870: \[
871: \log n! =\log 1+\log 2+\cdots+\log n = [z^n]
872: \frac{1}{1-z}\Li_{0,1}(z),
873: \]
874: to which singularity analysis can be applied now that we have taken ordinary
875: generating functions. Theorem~\ref{thm:li-sing} yields
876: the singular expansion
877: \[
878: \frac{1}{1-z}\Li_{0,1}(z)\sim\frac{L(z)-\gamma}{(1-z)^2}
879: +\frac12\frac{-L(z)+\gamma-1+\log{2\pi}}{1-z}+\cdots,
880: \]
881: from which Stirling's formula can be read off, by
882: Theorem~\ref{thm:basic-sa}:
883: \[
884: \log n!\sim n\log n -n +\frac{1}{2}\log n+\log\sqrt{2\pi}+\cdots\,.
885: \]
886: [Stirling's constant~$\log\sqrt{2\pi}$ comes out as $-\zeta'(0)$.]
887: Similarly, the ``superfactorial function'',
888: \[
889: S(n) := 1^1\cdot 2^2\cdots
890: n^n\equiv\frac{(n!)^{n+1}}{1!\,2!\,\cdots\,n!},
891: \]
892: satisfies
893: \[
894: \log S(n) = [z^n]\frac{1}{1-z}\Li_{-1,1}(z),
895: \]
896: which gives rise to a second-order ``Stirling's formula'',
897: \[
898: S(n) \sim n^{\frac12n^2+\frac12n+\frac{1}{12}}e^{-\frac14 n^2} A,
899: \]
900: with
901: \[
902: A:= \exp \left(\frac1{12} - \zeta'(-1)\right) =
903: \exp\left(-\frac{\zeta'(2)}{2\pi^2}+\frac{\log(2\pi)+\gamma}{12}\right).
904: \]
905: (This last expansion, originally due to Glaisher, Jeffery, and Kinkelin,
906: goes back to the 1860s and it can be established by
907: Euler--Maclaurin summation; see Finch's book~\cite{Finch03} for
908: context and references.) The systematic character of the derivation
909: given here clearly applies to many similar functions.\hfill~$\Box$
910: \end{example}
911:
912:
913: Methods of the last example may be used more generally to determine
914: the Euler--Maclaurin constant relative to sums of the form $\sum (\log
915: n)^r/n^s$. The derivation by singularity analysis is quite systematic
916: and several formul{\ae} of Ramanujan can be obtained in this way, for
917: instance,
918: \[
919: \lim_{N\to\infty} \left(\sum_{n=1}^N \frac{\log^k
920: n}{n}-\frac{\log^{k+1}N}{k+1}\right)=A_k, ~\hbox{with}~
921: A_k:=\frac{(-1)^k}{k+1}\frac{d^{k+1}}{ds^{k+1}}\left((s-1)\zeta(s)\right)_{s=1}
922: ,\] involving the Stieltjes constants~$A_k$. See Berndt's account of
923: the problem in~\cite[p.~164]{Berndt85} and references therein.
924:
925: % \begin{example} {\em Differences.} Given a sequence $(f_n)$, its
926: % difference sequence is the sequence $(g_n)$ defined by
927: % \[
928: % g_n=\Delta^n f_0=\sum_{k=0}^n \binom{n}{k}(-1)^{n-k} f_k.
929: % \]
930: % It is often the case that massive cancellations take place,
931: % and these may be hard to capture. In terms of ordinary generating
932: % functions, this however involves a simple transformation,
933: % \[
934: % g(z)=\frac{1}{1=z}f\left(\frac{z]{1+z}\right),
935: % \qquad
936: % f(z):=\sum_{n\ge0}f_n z^n, \quad
937: % g(z):=\sum_{n\ge0} g_n z^n.
938: % \]
939: % Consider the differences of~$f_n=\sqrt{n+1}$. One has
940: % \[
941: % g(z)=\frac{1}{z}\Li_{1/2,0}\left(\frac{z}{1+z}\right).
942: % \]
943: % The singularity of~$g$ is now at~$z=-1$. Making use of the asymptotic
944: % expansion of $\Li(y)$ as $\Re(y)\to\-infty$ (this can be obtained,
945: % like in~\cite{Flajolet99} by means of Mellin-Lindel\"of
946: % representations), one finds
947: % the singular expansion of~$g$ at~$z=-1$.
948: % PF: I stop here. This would require technology MUCH beyond what we do
949: % here!
950: %
951: %
952:
953:
954: \subsection{Differentiation and integration}\label{subsec:diffint}
955:
956:
957: In preparation for our later treatment of Hadamard products, we need a
958: theorem that enables us to differentiate local expansions of analytic
959: functions around a singularity. Such a result cannot of course be
960: unconditionally true; see, for example,~\eqref{eq:56}. However, it
961: turns out that functions amenable to singularity analysis
962: satisfy this property. The statement that follows is an adaptation
963: suited to our needs of well-known differentiability properties of
964: complex asymptotic expansions (see especially Theorem~I.4.2 of Olver's
965: book \cite[p.~9]{Olver74}).
966: \begin{theorem}[Singular differentiation]
967: \label{thm:2A}
968: If $f(z)$ is $\Delta$-regular and admits a singular expansion near
969: its singularity in the sense of~\eqref{eq:27}, then for each
970: integer $r > 0$, $\tfrac{d^r}{dz^r}f(z)$ is also $\Delta$-regular
971: and admits an expansion obtained through term-by-term
972: differentiation:
973: \begin{equation*}
974: \frac{d^r}{dz^r} f(z) = (-1)^r\sum_{j=0}^J c_j
975: \frac{\Gamma(\alpha_j+1)}{\Gamma(\alpha_j+1-r)}
976: (1-z)^{\alpha_j-r} + O( |1-z|^{A-r}).
977: \end{equation*}
978: \end{theorem}
979: \begin{proof}
980: Clearly, all that is required is to establish the effect of
981: differentiation on error terms, which is expressed symbolically as
982: \begin{equation*}
983: \frac{d}{dz} O( |1-z|^A ) = O( |1-z|^{A-1} ).
984: \end{equation*}
985: By iteration, only the case of a single differentiation ($r=1$)
986: needs to be considered.
987:
988: \begin{figure}
989: \centering
990: % %%\includegraphics[width=0.8\linewidth]{pacman-diff}%
991: \includegraphics[width=5truecm]{geometry-diff.eps} \hspace*{1truecm}
992: \includegraphics[width=5truecm]{circles.eps}
993: \caption{The contour $\gamma(z)$ used in the proof of the
994: differentiation theorem: (left) the basic geometry;
995: (right) two sets of circles $\gamma(z)$.}%
996: \label{fig:diff}%
997: \end{figure}
998:
999: Let~$g(z)$ be a function that is regular in a domain
1000: $\Delta(\phi,\eta)$ where it is assumed to satisfy
1001: $g(z)=O(|1-z|^A)$ for $z \in \Delta$.
1002: Choose a subdomain
1003: $\Delta':=\Delta(\phi',\eta')$,
1004: where $\phi<\phi'<\frac\pi2$ and $0<\eta'<\eta$. By elementary
1005: geometry, for any sufficiently small~$\kappa>0$, the disc of radius
1006: $\kappa|z-1|$ centered at a value~$z\in\Delta'$
1007: lies entirely in~$\Delta$; see Figure~\ref{fig:diff}.
1008: We fix such a small
1009: value~$\kappa$ and let~$\gamma(z)$ represent the boundary
1010: of that disc oriented positively.
1011:
1012: The starting point is Cauchy's integral formula
1013: \begin{equation}
1014: \label{eq:35}
1015: g'(z) = \frac{1}{2\pi i} \int_C g(w) \frac{dw}{(w-z)^2},
1016: \end{equation}
1017: a direct consequence of the residue theorem. Here~$C$ should
1018: encircle~$z$ while lying inside the domain of regularity of~$g$,
1019: and we opt for the choice $C\equiv\gamma(z)$. Then trivial bounds
1020: applied to~(\ref{eq:35}) give:
1021: \[ \begin{array}{lll}
1022: |g'(z)|&=& O\left( |\!| \gamma(z)|\!|\cdot |1-z|^A
1023: |1-z|^{-2}\right)
1024: \\
1025: &=& \ds O\left(|1-z|^{A-1}\right).
1026: \end{array}
1027: \]
1028: The estimate involves the length of the contour, $|\!| \gamma(z)|\!|$,
1029: which is $O(|1-z|)$ by construction, as well as the bound on~$g$
1030: itself, which is $O(|1-z|^A)$ since all points of the contour are
1031: themselves at a distance exactly of the order of $|1-z|$ from~1.
1032: \end{proof}
1033:
1034:
1035: %
1036: % We select a small $\epsilon > 0$ and restrict $z$ to the domain
1037: % $\Delta(\phi+\epsilon,\eta)$.
1038: % The integral is estimated using the
1039: % contour $\gamma$
1040: % depicted in Figure~\ref{fig:diff}: $\gamma := \gamma_1 \cup \gamma_2
1041: % \cup \gamma_3 \cup
1042: % \gamma_4$, with
1043: % \begin{align*}
1044: % \gamma_1 &:= \{ w: |w-1| = \tfrac{1}{2}|z-1|, |\Arg(w-1)| \geq
1045: % \phi\},\\
1046: % \gamma_2 &:= \{ w: \tfrac{1}{2}|z-1| \leq |w-1|, |w| \leq 1+\eta,
1047: % \Arg(w-1) = \phi\},\\
1048: % \gamma_3 &:= \{ w: |w| = 1 + \eta, |\Arg(w-1)| \geq \phi\},\\
1049: % \gamma_4 &:= \{ w: \tfrac{1}{2}|z-1| \leq |w-1|, |w| \leq 1+\eta,
1050: % \Arg(w-1) = -\phi \}.
1051: % \end{align*}
1052: %
1053: % We then decompose the integral giving $g'(z)$ as $\int_\gamma =
1054: % \int_{\gamma_1} +
1055: % \int_{\gamma_2} + \int_{\gamma_3} + \int_{\gamma_4}$. Along
1056: % $\gamma_1$ we apply
1057: % trivial
1058: % % \JF{explanations: first because $g(w) = O(1)$, second because $|w -
1059: % % z| \geq |1 - z| -
1060: % % |1 - w| = \frac{1}{2} |1 - z|$, third is bound on $|\gamma_1|$}
1061: % upper bounds:
1062: % \begin{equation}
1063: % \label{eq:36}
1064: % \left|\int_{\gamma_1}\right| = O(1) \cdot O\left(\frac{1}{
1065: % |1-z|^2}\right) \cdot
1066: % O(|1-z| ) = O( |1-z|^{-1}).
1067: % \end{equation}
1068: % The integral along the external circle $\gamma_3$ satisfies trivially
1069: % \begin{equation}
1070: % \label{eq:37}
1071: % \left|\int_{\gamma_3}\right| = O(1) = O( | 1-z |^{-1}).
1072: % \end{equation}
1073: % If we set $|1-z|=\delta$, we find that the integrals along $\gamma_2$ and
1074: % $\gamma_4$ are of no bigger order than
1075: % \begin{equation*}
1076: % \int_{\delta / 2}^\infty = \frac{du}{\delta^2 + u^2} =
1077: % O\left(\frac{1}{\delta}\right) = O( |1-z|^{-1}).
1078: % \end{equation*}
1079: % Thus we have
1080: % \begin{equation}
1081: % \label{eq:38}
1082: % \left|\int_{\gamma_2}\right|, \; \left|\int_{\gamma_4}\right| = O\left(
1083: % |1 - z|^{-1} \right).
1084: % \end{equation}
1085: % Combining the estimates \eqref{eq:36}--\eqref{eq:38} yields the
1086: % asymptotic form of $g'(z)$ as $O( |1-z|^{-1})$, as required at~\eqref{eq:33}.
1087: % \end{proof}
1088:
1089:
1090: For instance, taking
1091: \[
1092: g(z) = \cos{\log{\left(\frac{1}{1-z}\right)}} \quad \text{and} \quad
1093: g'(z) = - \frac{1}{1-z}\sin{\log{\left(\frac{1}{1-z}\right)}},
1094: \]
1095: we correctly predict that $g(z) = O(1) \Rightarrow g'(z) = O(
1096: |1-z|^{-1})$. On the other hand, the apparent paradox given by the
1097: pair
1098: \begin{equation}
1099: \label{eq:56}
1100: g(z) = \cos{{\left(\frac{1}{1-z}\right)}} \qquad \text{and}\qquad
1101: g'(z) = - \frac{1}{(1-z)^2}
1102: \sin{{\left(\frac{1}{1-z}\right)}} ,
1103: \end{equation}
1104: is resolved by observing that in no nondegenerate sector around $z=1$
1105: do we have $g(z) = O(1)$.
1106:
1107: It is also well known that integration of asymptotic expansions is
1108: usually easier than differentiation. Here is a statement
1109: custom-tailored to our needs.
1110:
1111:
1112: \begin{theorem}[Singular integration] \label{thm:2B}
1113: Let $f(z)$ be $\Delta$-regular and admit a $\Delta$-expansion near
1114: its singularity in the sense of~\eqref{eq:27}. Then
1115: $\int_0^zf(t)\,dt$ is also $\Delta$-regular. Assume that none of
1116: the quantities $\alpha_j$ and $A$ equals~$-1$.
1117:
1118: $(i)$~If~$A<-1$, then the singular expansion of $\int f$ is
1119: \begin{equation}\label{eq:in1}
1120: \int_0^z f(t)\,dt=-\sum_{j=0}^J
1121: \frac{c_j}{\alpha_j+1}(1-z)^{\alpha_j+1}+O\left(|1-z|^{A+1}\right).
1122: \end{equation}
1123:
1124: $(ii)$~If~$A>-1$, then the singular expansion of $\int f$ is
1125: \[
1126: \int_0^z f(t)\, dt=-\sum_{j=0}^J
1127: \frac{c_j}{\alpha_j+1}(1-z)^{\alpha_j+1} +L_0
1128: +O\left(|1-z|^{A+1}\right),
1129: \]
1130: where the ``integration constant'' $L_0$ has the value
1131: \[
1132: L_0:=\sum_{\alpha_j<-1}\frac{c_j}{\alpha_j+1}+
1133: \int_0^1\Big[f(t)-\sum_{\alpha_j<-1}c_j(1-t)^{\alpha_j}\Big]\, dt.
1134: \]
1135: \end{theorem}
1136: \begin{remark*}
1137: The case where either some~$\alpha_j$ or~$A$ is $-1$ is easily treated
1138: by the additional rules
1139: \[
1140: \int_0^z (1-t)^{-1}\, dt=L(z), \qquad \int_0^z O(|1-t|^{-1})\, dt =
1141: O(L(z)).
1142: \]
1143: Similar rules consistent with elementary integration are applicable
1144: for powers of logarithms: they are derived from the easy identities
1145: (for $\alpha \ne -1$)
1146: \[
1147: \int_0^z (1-t)^\alpha L^r(t)\, dt=(-1)^r \frac{\partial^r}{\partial
1148: \alpha^r} \int_0^z (1-t)^{\alpha}\,dt = (-1)^{r+1}
1149: \frac{\partial^r}{\partial \alpha^r}
1150: \frac{(1-z)^{\alpha+1}}{\alpha+1},
1151: \]
1152: for~$r$ a positive integer. Furthermore, the corresponding
1153: $O$--transfers hold true. (The proofs are simple modifications of the
1154: one given below for the basic case.)
1155: \end{remark*}
1156: \begin{proof} The basic technique consists in integrating, term by
1157: term, the singular expansion of~$f$. We let~$r(z)$ be the remainder
1158: term in the expansion of~$f$, that is,
1159: \[
1160: r(z):=f(z)-\sum_{j=0}^J c_j(1-z)^{\alpha_j}.\] By assumption,
1161: throughout the $\Delta$-domain one has, for some positive
1162: constant~$K$,
1163: \[
1164: |r(z)|\le K|1-z|^A.
1165: \]
1166:
1167:
1168: \begin{figure}
1169: \centering
1170: %%\includegraphics[width=0.8\linewidth]{pacman-int}%
1171: \includegraphics[width=6truecm]{pacman.3}%
1172: \caption{The contour used in the proof of the integration theorem.}
1173: % Theorem \ref{thm:2B}.}%
1174: \label{fig:int}%
1175: \end{figure}
1176:
1177: $(i)$~\emph{Case $A<-1$.} By straight-line integration between~$0$
1178: and~$z$, one finds~(\ref{eq:in1}), as soon as it has been
1179: established that
1180: \[
1181: \int_0^z r(t)\, dt=O\left(|1-z|^{A+1}\right).
1182: \]
1183: By Cauchy's integral formula, we can choose any path of integration
1184: that stays within the region of analyticity of $r$. We choose the
1185: contour $\gamma := \gamma_1 \cup \gamma_2$, shown in
1186: Figure~\ref{fig:int}.
1187: Then\footnote{%
1188: The symbol $|dt|$ designates the differential line element~(often
1189: denoted by $ds$) in the corresponding curvilinear integral. }
1190: \begin{align*}
1191: \left|\int_{\gamma} r(t) \,dt \right|
1192: % &= \left| \int_{\gamma_1} r(t) \,dt +
1193: % \int_{\gamma_2} r(t) \,dt \right|\\
1194: &\leq \left| \int_{\gamma_1} r(t) \,dt \right| + \left|
1195: \int_{\gamma_2} r(t) \,dt \right|\\
1196: &\leq K \int_{\gamma_1} |1-t|^A \,|dt| + K
1197: \int_{\gamma_2} |1-t|^A|\,|dt|\\
1198: & = O(|1-z|^{A+1}).\\
1199: \end{align*}
1200: Both integrals are~$O(|1-z|^{A+1})$: for the integral
1201: along~$\gamma_1$, this results from explicitly carrying out the
1202: integration; for the integral along $\gamma_2$, this results from
1203: the trivial bound~$O(|\!|\gamma_2|\!| (1-z)^A)$.
1204:
1205: % \[
1206: % |1-z|^A |\log{(1-z)}|^B \int_{\gamma_2} |dt| =
1207: % O(\,|1-z|^{A+1}|\log{(1-z)}|^{B}\,).
1208: % \]
1209: % The last integral along $\gamma_1$ is a real integral. Define $x := 1 -
1210: % |1-z|$, so that the last integral along $\gamma_1$ equals
1211: % \begin{align*}
1212: % & \int_{0}^x (1-t)^{A} (\log{(1-t)^{-1}})^B \,dt\\
1213: % &= \frac{1}{|A|-1} \left[(1-x)^{A+1}(\log{(1-x)^{-1}})^B
1214: % - B\int_{0}^x (1-t)^A(\log{(1-t)^{-1}})^{B-1} \,dt\right]\\
1215: % &= \frac{1}{|A|-1} (1-x)^{A+1} (\log{(1-x)^{-1}})^B + o\left(\int_{0}^x (1-t)^{A}
1216: % (\log{(1-t)^{-1}})^B \,dt\right).
1217: % \end{align*}%
1218: % We have shown that
1219: % \begin{align*}
1220: % \int_{\gamma_1} |1-t|^A|\log{(1-t)}|^B\,dt &= (1 + o(1)) \frac{1}{|A|-1}
1221: % (1-x)^{A+1}(\log{(1-x)^{-1}})^B\\
1222: % &= (1 + o(1)) \frac{1}{|A|-1} |1-z|^{A+1} (\log{|1-z|^{-1}})^B\\
1223: % &= (1 + o(1)) \frac{1}{|A|-1} |1-z|^{A+1} |\log{(1-z)^{-1}}|^B\\
1224: % &= \Theta( |1-z|^{A+1} | \log{(1-z)^{-1}} |^B )\\
1225: % &= O( |1-z|^{A+1} | \log{(1-z)^{-1}} |^B ).
1226: % \end{align*}
1227: % %%% \NK{Not sure about the last couple of steps.}
1228: % Hence
1229: % \[ \int_{\gamma} f(t) \,dt = O(\,|1-z|^{A+1} |\log{(1-z)}|^B\,).\]
1230: %
1231: % The case $A=-1$ is similar. As above,
1232: % \[ \left|\int_{\gamma} f(t) \,dt\right| \leq K \int_{\gamma_1} |1-t|^{-1}
1233: % |\log{(1-t)}|^B \,dt + K \int_{\gamma_2} |1-t|^{-1} |\log{(1-t)}|^B \,|dt|.
1234: % \]
1235: % The integral here along $\gamma_2$ is
1236: % \[
1237: % |1-z|^{-1}|\log{(1-z)}|^B \int_{\gamma_2} \,|dt| = O(|\log{(1-z)}|^B) = o(
1238: % |\log{(1-z)}|^{B+1}).
1239: % \]
1240: % The integral along $\gamma_1$ is a real integral. If $B \neq -1$, then
1241: % setting $x := 1-|1-z|$, we integrate to get
1242: % \begin{multline*}
1243: % \int_{0}^x(1-t)^{-1}(\log{(1-t)^{-1}})^B\,dt \\
1244: % = \frac{1}{B+1}(\log{(1-x)^{-1}})^{B+1} = O(|\log{(1-z)}|^{B+1}).
1245: % \end{multline*}
1246: % If $B=-1$,
1247: % \begin{align*}
1248: % \left|\int_{\gamma_1} f(t)\,dt\right| &\leq O(1) + K \int_{1/2}^x
1249: % (1-t)^{-1}(\log{(1-t)^{-1}})^{-1}\,dt \\
1250: % &= O( |\log{\log{(1-z)^{-1}}}| ).
1251: % \end{align*}
1252: % Combining the two integrals gives us the result.
1253: $(ii)$~\emph{Case $A>-1$.} We let $f_-(z)$ represent the
1254: ``divergence part'' of~$f$ that gives rise to nonintegrability:
1255: \[
1256: f_-(z):=\sum_{\alpha_j<-1}c_j(1-z)^{\alpha_j}.
1257: \]
1258: Then with the decomposition $f=[f-f_-]+f_-$, integrations can be
1259: performed separately. First, one finds
1260: \[
1261: \int_0^z f_-(t)\,
1262: dt=-\sum_{\alpha_j<-1}\frac{c_j}{\alpha_j+1}(1-z)^{\alpha_j+1}
1263: +\sum_{\alpha_j<-1}\frac{c_j}{\alpha_j+1}.
1264: \]
1265: Next, observe that the asymptotic condition guarantees the existence
1266: of $\int_0^1$ applied to $[f-f_-]$, so that
1267: \[
1268: \int_0^z \left[f(t)-f_-(t)\right]\,dt =\int_0^1
1269: \left[f(t)-f_-(t)\right]\,dt +\int_1^z \left[f(t)-f_-(t)\right]\,dt.
1270: \]
1271: The first of these two integrals is a constant that
1272: contributes to~$L_0$.
1273: As to the second integral, term-by-term integration yields
1274: \[
1275: \int_1^z \left[f(t)-f_-(t)\right]\,dt=-\sum_{\alpha_j>-1}
1276: \frac{c_j}{\alpha_j+1} (1-z)^{\alpha_j + 1}+\int_1^z r(t)\, dt.
1277: \]
1278: The remainder integral is finite, given the growth condition on the
1279: remainder term, and, upon carrying out the integration along the
1280: rectilinear segment joining~$1$ to~$z$, trivial bounds show that it is
1281: indeed~$O(|1-z|^{A+1})$.
1282: \end{proof}
1283:
1284: %
1285: %
1286: % The case $A > -1$ is obtained by
1287: % observing that the asymptotic condition guarantees the existence of
1288: % $\int_0^1$ and by decomposing $\int_0^z = \int_0^1 + \int_1^z$. The contour
1289: % of integration is the line segment joining 1 to $z$. It is enough
1290: % to show that
1291: % \[
1292: % \left| \int_z^1 f(t)\,dt \right| = O(\,|1-z|^{A+1}|\log{(1-z)}|^{B}\,).
1293: % \]
1294: % Indeed, performing the change of variable $t = uz + (1-u)$, we can
1295: % bound the absolute value of this integral by $K$ times the
1296: % following:
1297: % \begin{align*}
1298: % & \int_0^1 |1-z|^A u^A |\log{(u(1-z))}|^B |1-z|\,du \\
1299: % &=
1300: % |1-z|^{A+1}|\log{(1-z)}|^{B} \int_0^1 u^A \left(
1301: % \frac{|\log{(u(1-z))}|}{|\log{(1-z)}|} \right)^B\, du \\
1302: % &=
1303: % O(|1-z|^{A+1}|\log{(1-z)}|^{B}),
1304: % \end{align*}
1305: % since the last integral is bounded.
1306: % \end{proof}
1307:
1308: %% Section 4: Hadamard products and transformation of singularities
1309: \section{Hadamard products and transformation of
1310: singularities}
1311: \label{sec:hadam-prod-transf}
1312: In this section we propose to examine the way singular expansions get
1313: composed under Hadamard products defined
1314: at~(\ref{hadam-def}). The Hadamard product is a bilinear form.
1315: So if we have a set of functions admitting known singular expansions,
1316: we need to establish their composition law, and this will give
1317: composition rules for finite terminating expansions
1318: (Subsection~\ref{subsec:had-pow}). In order to extend this to
1319: asymptotic expansions with error terms, we need to establish a theorem
1320: providing the shape of
1321: \[
1322: O( |1-z|^A ) \odot O( |1-z|^B ).
1323: \]
1324: This is the more demanding part of the analysis, which is the subject
1325: of Subsection~\ref{subsec:had-bigo}. Finally, in
1326: Subsection~\ref{subsec:had-all}, we provide a summary
1327: statement, Theorem~\ref{thm:hadmain}, to the effect that the class of
1328: functions amenable
1329: to singularity analysis is closed under Hadamard products and
1330: that the composition of singular expansions is effectively computable.
1331:
1332: \subsection{Composition of singular elements}\label{subsec:had-pow}
1333:
1334: The composition rule for polylogarithms is trivial, since
1335: \[
1336: \Li_{\alpha,r}(z) \odot \Li_{\beta,s}(z) = \Li_{\alpha+\beta,r+s}(z).
1337: \]
1338: However, polylogarithms do not have a simple composition rule with
1339: respect to ordinary products.
1340: %% \texttt{Though something can be done
1341: %% with Mellin \ldots} \JF{We should think about this comment of PhF.}
1342: We next turn to the composition rule for the basis formed by functions
1343: of the form $(1-z)^{a}$, where $a$ may be any
1344: real
1345: %%\JF{PhF said $\alpha$ is a real number, but can't it be complex?}%
1346: % PF: True: but in the paper we haven't dealt with complex exponent
1347: % as this would complicate notations with a lot of $\Re(\alpha)$'s
1348: number. From the expansion
1349: \begin{equation}
1350: \label{eq:39}
1351: (1-z)^{a} = 1 + \frac{-a}{1} z +
1352: \frac{(-a)(-a+1)}{2!} z^2 + \cdots
1353: \end{equation}
1354: around the origin,
1355: we get through term-by-term multiplication
1356: \begin{equation}
1357: \label{eq:40}
1358: (1-z)^{a} \odot (1-z)^{b} = \FF[-a,-b;1;z].
1359: % 1 + \frac{\alpha \beta}{(1!)^2} z +
1360: % \frac{\alpha(\alpha+1)\beta(\beta+1)}{(2!)^2} z^2 + \cdots.
1361: \end{equation}
1362: Here $\FF$ represents
1363: the classical \emph{hypergeometric function} of Gauss defined
1364: % for $\gamma$ not a nonpositive integer
1365: by
1366: \begin{equation}
1367: \label{eq:43}
1368: \FF[\alpha,\beta;\gamma;z] = 1 + \frac{\alpha\beta}{\gamma} \frac{z}{1!} +
1369: \frac{\alpha(\alpha+1)\beta(\beta+1)}{\gamma(\gamma+1)}
1370: \frac{z^2}{2!} + \cdots.
1371: \end{equation}
1372: % so that~\eqref{eq:40} reduces to
1373: % \begin{equation}
1374: % \label{eq:41}
1375: % (1-z)^{-\alpha} \odot (1-z)^{-\beta} = \FF[\alpha,\beta;1;z].
1376: %\end{equation}
1377: From the transformation theory of hypergeometrics, see
1378: e.g.~\cite[p.~163]{MR56:12235}, we
1379: know that, in general, hypergeometric functions can
1380: be expanded in the vicinity of $z=1$ by means of the $z \mapsto 1 - z$
1381: transformation. Instantiation
1382: of this transformation with $\gamma=1$ yields
1383: \begin{multline}
1384: \label{eq:42}
1385: \FF[\alpha,\beta;1;z] =
1386: \frac{\Gamma(1-\alpha-\beta)}{\Gamma(1-\alpha)\Gamma(1-\beta)}
1387: \,\FF[\alpha,\beta;\alpha+\beta;1-z]\\
1388: + \frac{\Gamma(\alpha+\beta-1)}{\Gamma(\alpha)\Gamma(\beta)}
1389: (1-z)^{-\alpha-\beta+1}
1390: \,\FF[1-\alpha,1-\beta;2-\alpha-\beta;1-z].
1391: \end{multline}
1392: In other words, we can state the following proposition:
1393: \begin{proposition} \label{thm:3}
1394: When $a$, $b$, and $a + b$ are not integers,
1395: the Hadamard product
1396: \[ (1-z)^{a} \odot (1-z)^{b} \]
1397: has an infinite $\Delta$-expansion with exponent scale
1398: \[
1399: \{0,1,2,\ldots\} \cup \{a + b + 1, a + b + 2,\ldots\},
1400: \]
1401: namely,
1402: \[
1403: (1-z)^{a} \odot (1-z)^{b} \sim \sum_{k \geq 0} \lambda_k^{(a,b)}
1404: \frac{(1-z)^k}{k!} + \sum_{k \geq 0} \mu_k^{(a,b)}
1405: \frac{(1-z)^{a + b +1+k}}{k!},
1406: \]
1407: where the
1408: coefficients
1409: $\lambda$ and $\mu$ are given by
1410: \begin{align*}
1411: \lambda_k^{(a,b)} &= \frac{\Gamma(1 + a + b)}
1412: {\Gamma(1+a)\Gamma(1+b)}
1413: \frac{\rising{(-a)}{k}\rising{(-b)}{k}}{\rising{(-a-b)}{k}},\\
1414: \mu_k^{(a,b)} &=
1415: \frac{\Gamma(-a-b-1)}{\Gamma(-a)\Gamma(-b)}
1416: \frac{\rising{(1+a)}{k}
1417: \rising{(1+b)}{k}}{\rising{(2+a+b)}{k}}.
1418: \end{align*}
1419: Here $\rising{x}{k}$ is defined when $k$ is a nonnegative
1420: integer as $x(x+1)\cdots(x+k-1)$.
1421: \end{proposition}
1422: \begin{remark*}
1423: \label{rem:a-b-integer}
1424: The case where either $a$ or $b$ is an integer poses
1425: no difficulty: one has
1426: \begin{itemize}
1427: \item[---] $(1-z)^a \odot g(z)$ is a polynomial if $a=m$, where
1428: $m\in\Z_{\ge0}$;
1429: \item[---] $(1-z)^a \odot g(z)$ is a derivative if $a=-m$ where
1430: $m\in\Z_{>0}$, since
1431: \[
1432: (1-z)^{-m}\odot g(z)=\frac{1}{(m-1)!}\partial_z^{m-1}
1433: \left(z^{m-1}g(z)\right),
1434: \]
1435: and this case is covered by singular differentiation,
1436: Theorem~\ref{thm:2A}.
1437: \end{itemize}
1438: Notice that Proposition~\ref{thm:3} remains valid in these two cases
1439: with the natural convention that
1440: $1/\Gamma(-j) = 0$ when $j \in \Z_{\geq 0}$.
1441:
1442: The case where $a + b \in \Z$ needs transformation formul{\ae}
1443: that extend~(\ref{eq:42}) and are found explicitly in the books by
1444: Abramowitz and Stegun~\cite[pp.~559--560]{AbSt73} and by Whittaker and
1445: Watson~\cite[\S14.53]{WhWa27}.
1446: % and Prudnikov \emph{et al}~\cite{PrBrMa86}.
1447: \end{remark*}
1448: \begin{remark*}
1449: \label{rem:logarthmic-factors}
1450: The case of expansions with logarithmic terms
1451: is covered
1452: by ``differentiation under the integral sign'',
1453: as we now explain. Consider for instance
1454: the Hadamard product
1455: \[
1456: \left[(1-z)^{-\alpha}L(z)\right]\odot (1-z)^{-\beta}
1457: = \frac{\partial}{\partial_\alpha}\,
1458: \FF[\alpha,\beta;1;z],
1459: \]
1460: where we assume for convenience that none of
1461: $\alpha$, $\beta$, $\alpha + \beta$ is an integer. For any
1462: fixed~$\beta$ and any fixed~$z$,
1463: with, say, $z\in(0,1)$, both sides of~(\ref{eq:42})
1464: represent analytic functions
1465: of~$\alpha$. Thus, their derivatives
1466: with respect to~$\alpha$ are identical as functions
1467: of~$\alpha$. This induces a transformation formula, originally valid in the
1468: stated $z$-range, which involves modified
1469: hypergeometric functions (these have additional $\psi$-factors
1470: in their coefficients) obtained from the
1471: fundamental ${}_2F_1$ function by differentiation with respect to some
1472: of the parameters. The modified functions then do exist in extended
1473: regions of the complex $z$--plane as shown by taking the
1474: classical Barnes
1475: representations in terms of contour integrals
1476: (see, e.g.,~\cite[\S14.5]{WhWa27}) and then differentiating under the
1477: integral sign. The net effect of this discussion is
1478: that the fundamental transformation~(\ref{eq:42}) supports
1479: differentiation with respect to~$\alpha,\beta$ and that the
1480: formally derived transformations provide analytically valid
1481: composition formul{\ae} for Hadamard products
1482: \begin{equation}\label{hadlog}
1483: [(1-z)^{-\alpha}L^k(z)] \odot [(1-z)^{-\beta} L^\ell(z)]
1484: \end{equation}
1485: of the base functions.
1486: \end{remark*}
1487:
1488: In practice, for all the cases described above, one may often proceed
1489: as follows: \textit{(i)}~take advantage of the a priori \emph{existence} of a
1490: singular expansion of~$f\odot g$, with $f(z)=(1-z)^{a}$,
1491: $g(z)=(1-z)^{b}$ or some of their derivatives, that is valid
1492: for~$z$ in a $\Delta$-region (here the slit complex plane);
1493: \textit{(ii)}~compute an asymptotic expansion of the coefficients of $f\odot
1494: g$ by multiplication of the asymptotic expansions of~$f_n$ and~$g_n$
1495: as obtained via singularity analysis; \textit{(iii)}~reconstruct a singular
1496: function that matches asymptotically $f_ng_n$ by using singularity
1497: analysis in the reverse direction. In
1498: Subsection~\ref{subsec:had-all}, this process is formalized by the
1499: ``Zigzag Algorithm'' and illustrated by the return of P\'olya's
1500: drunkard.
1501:
1502: \smallskip
1503:
1504: Globally, we are facing a situation where polylogarithms are simple for
1505: Hadamard products and relatively complicated for ordinary products,
1506: with the dual situation occurring in the case of power functions.
1507: Each particular situation is likely to dictate whether calculations
1508: are best expressed in a basis of standard singular functions like
1509: $\{(1-z)^{a}L(z)^k\}$ or
1510: with polylogarithms, $\{\Li_{\alpha,k}(z)\}$.
1511:
1512: \subsection{Composition of error terms}\label{subsec:had-bigo}
1513: We now examine how $O(\cdot)$ terms get composed
1514: under Hadamard products. The task is easier when
1515: the resulting function gets large at its singularity
1516: as shown by Proposition~\ref{lem:3}. Fortunately,
1517: thanks to the results of
1518: Section~\ref{sec:sing-expans-diff}
1519: regarding differentiation and integration, all cases can be reduced to this
1520: one: see Proposition~\ref{thm:4} below.
1521:
1522: {\sloppy
1523: The starting point is a general integral formula
1524: due to Hadamard for \mbox{$(f \odot g)(z)$}, where
1525: \[
1526: f(z) = \sum_{n \geq 0} f_nz^n \qquad \text{and} \qquad g(z) = \sum_{n \geq
1527: 0} g_nz^n.
1528: \]}%
1529: Assume that $f$ and $g$ are analytic in the unit disc
1530: and let $z$ be a complex number satisfying $|z| < 1$.
1531: Consider the integral
1532: \begin{equation}
1533: \label{eq:45}
1534: I = \frac{1}{2\pi i} \int_{\gamma_0} f(w)g\left(\frac{z}{w}\right)
1535: \frac{dw}{w},
1536: \end{equation}
1537: taken (counterclockwise) along a contour~$\gamma_0$
1538: % (see Figure~\ref{fig:5-1a})
1539: which
1540: is simply a circle of radius $\rho$ centered at the origin such that
1541: $|z| < \rho < 1$. In this way, both factors in the integrand are analytic
1542: functions of~$w$ along the contour.
1543: Evaluating the integral~(\ref{eq:45})
1544: by expanding the functions, we find
1545: \[
1546: I = \sum_{n \geq 0} f_ng_n z^n.
1547: \]
1548: This is the classical formula of Hadamard for Hadamard products,
1549: \begin{equation}
1550: \label{eq:45bis}
1551: (f \odot g)(z) = \frac{1}{2\pi i} \int_{\cal C}
1552: f(w)g\left(\frac{z}{w}\right)
1553: \frac{dw}{w},
1554: \end{equation}
1555: valid, by analyticity, for any simple contour~$\cal C$ such that
1556: each~$w\in\cal C$ satisfies $|z|< |w| <1$.
1557:
1558:
1559: \begin{proposition} \label{lem:3}
1560: Assume that $f(z)$ and $g(z)$ are $\Delta$-regular in $\Delta(\psi_0, \eta)$
1561: and that
1562: \[
1563: f(z) = O(|1-z|^a)\mbox{\rm \ \ and\ \ }g(z) = O(|1-z|^b), \quad
1564: z\in\Delta(\psi_0,\eta),
1565: \]
1566: where $a$ and $b$
1567: satisfy $a+b+1 < 0$. Then the Hadamard product $(f \odot g)(z)$ is
1568: regular in a (possibly smaller)
1569: $\Delta$-domain, call it~$\Delta'$, where it admits the expansion
1570: \begin{equation}
1571: \label{eq:44}
1572: (f \odot g)(z) = O( |1-z|^{a+b+1} ).
1573: \end{equation}
1574: \end{proposition}
1575:
1576: % \begin{figure}%[htbp]%
1577: % \centering
1578: % \includegraphics{hadamard.1}%
1579: % \caption{The contour $\gamma_0$.}%
1580: % \label{fig:5-1a}%
1581: % \end{figure}
1582:
1583: % We first give the proof in the simpler case when%
1584: % \JF{Is this stronger assumption used? If so, where?, and what
1585: % about the general
1586: % $\Delta$-regular case?}
1587: % $f(z)$ and $g(z)$
1588: % are analytic in the whole of the complex plane slit along $z \geq 1$.
1589:
1590: \begin{figure}
1591: % \hrule
1592: % \vspace*{-1.25truecm}
1593: \begin{center}
1594: \includegraphics[width=6truecm]{hcont-dir.ps}\hspace*{-1.5truecm}%
1595: \raisebox{16pt}{\includegraphics[width=4.0truecm]{hcont-inv.ps}}\hspace*{-1.15truecm}%
1596: \includegraphics[width=6truecm]{hcont-both.ps}
1597: \end{center}
1598: \caption{\label{hadgeom-fig} The geometry of Hadamard domains:
1599: (left) boundary of a $\Delta$-domain ($1+\eta=1.5$);
1600: (middle) boundary of $\Delta^{-1}$;
1601: (right) an allowable domain in \mbox{$\Delta\cap z\Delta^{-1}$} for
1602: application of Hadamard's formula is the unshaded subset of~$\Delta$
1603: ($|z|=1.25$).}%
1604: \end{figure}
1605:
1606: % \def\nonD{\overline{(z\Delta^{-1})}
1607:
1608: \begin{proof} We first observe\footnote{%
1609: This part of the argument is an adaptation to our needs of a
1610: famous result first due to Hadamard regarding the continuation of
1611: Hadamard products; see for instance the description
1612: in~\cite[Vol.~II, p.~300]{Bieberbach31}
1613: or~\cite[Sec.~88]{Dienes58}. Accordingly, we limit ourselves to a
1614: succinct discussion only meant to set the stage for the precise
1615: estimates starting at~(\ref{condh2}).} that $f \odot g$
1616: is continuable to \emph{certain} points $z$ such that $|z| > 1$.
1617: (Precisely, as shown below, it admits a continuation in a
1618: $\Delta$-domain.) Indeed, because of the analytic continuation
1619: properties of $f$ and $g$, both~$f(w)$ and~$g(z/w)$ are analytic
1620: functions of~$w$ in the domain $\Delta\cap(z\Delta^{-1})$,
1621: where $\Delta^{-1}$ denotes $\{w^{-1}: w\in\Delta\}$;
1622: see Figure~\ref{hadgeom-fig} for a rendering. In other words, the
1623: allowed domain of values of~$w$ is $\Delta$ stripped of the internal
1624: domain $(z\Delta^{-1})^c$, where
1625: $(\cdot)^c$ represents
1626: complementation. Fix then some~$z_1$ outside the unit disc but
1627: within~$\Delta$, and choose a simple contour~$\gamma_1$ inside both
1628: $\Delta$ and $z_1\Delta^{-1}$. Let $I(z)$ be the integral
1629: of~(\ref{eq:45}) and~(\ref{eq:45bis}) taken along this fixed
1630: contour~$\gamma_1$. (The feasibility of finding a
1631: suitable~$\gamma_1$ is suggested by Figure~\ref{hadgeom-fig}, at
1632: least when~$|z_1|$ remains close enough to~1 and
1633: $z_1$ is to the left of~1; % far enough from~1 and to its left;
1634: % since the disc inside
1635: % $\overline{(z_1\Delta^{-1})}$ is of radius $z_1/(1+\eta)$;
1636: a particular contour adapted to the case where~$z_1$ is close to~1 and
1637: possibly to its right will be constructed explicitly in the proof
1638: below.) Now, when $z$ moves radially along the segment $(0,z_1)$, the
1639: quantity~$I(z)$ defines an analytic function of~$z$ that does coincide
1640: with~$(f\odot g)(z)$ as soon as $|z|\le 1$ [this results from the
1641: ``standard'' formula~(\ref{eq:45bis})]. Thus analytic continuation of
1642: $f\odot g$, from within the unit disc to some~$z_1$ lying outside of
1643: the unit disc
1644: % , hence globally to the~$\Delta$-domain
1645: is granted. The argument shows at the same time that Hadamard's
1646: formula~(\ref{eq:45bis}) remains a valid representation of~$f\odot g$
1647: along such a contour~$\gamma_1$ or any of its deformations legally
1648: granted by analyticity.
1649:
1650:
1651: \smallskip
1652:
1653: % In order to obtain an asymptotic estimate
1654: % of $f \odot g$, we must fix the geometry of the integration
1655: % contour, call it $\gamma$, to be adopted in Hadamard's
1656: % formula~(\ref{eq:45bis}). Without
1657: % loss of generality, we assume that $z$ lies in the upper half-plane
1658: % and stays close-enough to~1. Set~$\delta:=|1-z|$, which
1659: % is the main quantity that governs the
1660: % scaling of the contour.
1661: % The main idea consists in adopting a contour that avoids~1 by staying
1662: % inside~$\Delta$ at a distance about $\delta$ from~1,
1663: % then directs itself towards~$z$, which it then environs
1664: % from the outside at a distance
1665: % of about~$\delta$, then rejoins an arc of a circle of some suitable
1666: % radius~$R<1$, and finally closes back on the rest of the contour
1667: % near~1; see Figure~\ref{fig:5-1c}.
1668: % The two parts of the contour that lie near~1 and~$z$ are expected to
1669: % capture most of the contribution to the main integral
1670: % of~(\ref{eq:45bis}), since they
1671: % correspond to the regions where~$f(w)$ and~$g(z/w)$ become singular
1672: % functions of~$w$. Care is needed in fixing details of
1673: % the geometry of the contour since one needs to ensure simultaneously
1674: % $w\in\Delta$ and $w\in z\Delta^{-1}$.
1675: %
1676:
1677: % \marginpar{PF: have we specified~$\delta$?}
1678:
1679:
1680: % We assume that~$f$ and~$g$ are both analytic in a common
1681: % domain~$\Delta(\psi_0,\eta)$ and set
1682: We next turn to estimating the growth at its singularity of~$h:=f
1683: \odot g$. It suffices to prove the estimate~(\ref{eq:44}) on $h$
1684: for~$z$ belonging to a restricted domain
1685: $\Delta':=\Delta(\psi_1,\eta_1)$, where we shall take
1686: \begin{equation}\label{condh2}
1687: \eta_1=c_1\eta,\qquad \left(\frac\pi2-\psi_1\right)
1688: =c_1\left(\frac\pi2-\psi_0\right),
1689: \end{equation}
1690: for some small positive constant~$c_1$. Notice also that it suffices
1691: to establish the estimate of~(\ref{eq:44}) for
1692: \begin{equation}\label{condh22}
1693: |z-1|<\eta_1=c_1\eta
1694: \end{equation}
1695: with~$z\in\Delta'$, since~$h$, being analytic in the rest
1696: of~$\Delta'$, is certainly bounded there.
1697:
1698:
1699: \begin{figure}%[htbp]%
1700: \centering
1701: \includegraphics[width=7.5truecm]{hadc0.8pl0.1xI.ps}\hspace*{-1.75truecm}
1702: \includegraphics[width=7.5truecm]{hadc1.2pl0.6xI.ps}
1703:
1704: % \vspace*{0.5truecm}
1705: \caption{The geometry of the contour $\gamma$.}%
1706: \label{fig:5-1c}%
1707: \end{figure}
1708:
1709: The main geometric objects from which the contour is built are as
1710: follows. First consider the circle centered at the origin
1711: \begin{equation}\label{condh4}
1712: C_0:=\left\{w: |w|=R\right\},\qquad R:=1-c_2\eta
1713: \end{equation}
1714: for some small constant~$c_2$ (independent of~$z$).
1715: % An arc of the circle $|z|=R$ is part of~$\gamma$ and
1716: % all points~$w$ of~$\gamma$ will be taken to satisfy $|w|\ge R$.
1717: Set~$\delta=|z-1|$, which is the main parameter
1718: governing the scaling of the contour~$\gamma$. We also consider the
1719: circle
1720: \begin{equation}\label{condh5}
1721: C_z:=\left\{ w: |w-z| = c_3\delta \right\},
1722: \end{equation}
1723: for some small positive~$c_3$.
1724: % An arc of the circle $C_z$ is also part of~$\gamma$.
1725: Finally, the contour~$\gamma$ includes parts of the two tangents
1726: $T,T'$ to the circle~$C_0$ issuing from~$z$; see
1727: Figure~\ref{fig:5-1c}. The contour is then precisely specified as
1728: \[
1729: \gamma=\gamma_0 \cup \gamma_T \cup \gamma_z \cup \gamma_{T'},
1730: \]
1731: where $\gamma_T$ is the segment of~$T$ formed of points in between
1732: $C_0$ and~$z$ that are exterior to~$C_0$ and~$C_z$, and
1733: similarly for $\gamma_{T'}$. The component~$\gamma_0$ is the part of
1734: the circle~$C_0$ that lies on the ``southwest'' of~0 and joins
1735: with~$T,T'$; the component~$\gamma_z$ is the part of the circle~$C_z$
1736: that lies on the ``northeast'' of~$z$ and joins with~$T$, $T'$. The
1737: constants~$c_1,c_2,c_3$ are to be specified later and
1738: they can be taken as small as needed.
1739:
1740:
1741:
1742: The fundamental constraint to be satisfied is that~$\gamma$ should lie
1743: entirely within~$\Delta\cap(z\Delta^{-1})$ when~$z$ stays
1744: within~$\Delta'$: for~$w\in\gamma$, this ensures simultaneously
1745: $w\in\Delta$ and $z/w\in\Delta$, hence the validity of the Hadamard
1746: integral~(\ref{eq:45bis}). By a priori choosing~$c_1$ (which
1747: limits~$z$) and~$c_3$ (which controls the radius of~$C_z$) both small
1748: enough, the condition $\gamma\subseteq\Delta$ is granted
1749: by elementary geometry. (E.g., the circle $C_z$ will not extend too much
1750: to the right of $\Re(w)=1$ and will therefore be ``compatible'' with
1751: the indentation of~$\Delta$ at~1.) Next, one should have
1752: $\gamma\cap{(z\Delta^{-1})^c}=\emptyset$. This requires in
1753: particular choosing the radius~$R$ in~(\ref{condh4}) larger than $z(1
1754: + \eta)^{-1}$, which
1755: is at most $(1+c_1\eta)/(1+\eta)$ since $z$ has been restricted to
1756: $|z-1|<c_1\eta$ by~(\ref{condh22}). This geometric condition
1757: expressed as
1758: \begin{equation}\label{condh6}
1759: \frac{1+c_1\eta}{1+\eta}<1-c_2\eta
1760: \end{equation}
1761: is granted as soon as $c_1,c_2$ are both taken small enough. [E.g.,
1762: it suffices that both $c_1$, $c_2$ be less than
1763: $\frac12(1+\eta)^{-1}$.] We
1764: henceforth assume these smallness conditions on~$c_1,c_2,c_3$ to be
1765: satisfied.
1766: Finally, the contour should avoid the apex\footnote{%
1767: By the ``apex'' of~$(z\Delta^{-1})^c$, we mean the
1768: complement in~$(z\Delta^{-1})^c$ of the largest circular
1769: disc centered at the origin which is contained in
1770: $(z\Delta^{-1})^c$. } of the domain $(z\Delta^{-1})^c$.
1771: Define the ``viewing angle'' of a point~$P$ exterior to a circle~$C$
1772: as the angle betwen the two tangents to~$C$ issuing from~$P$. For a
1773: circle of radius~$r$ and a point at distance~$d$ from the center, this
1774: angle is $2\arcsin(r/d)$. In particular the point~$z$ itself views the
1775: circle~$C_0$ of radius~$R$ under the angle $2\arcsin({R}/{|z|})$, and this
1776: viewing angle is bounded from below by
1777: \[
1778: 2\arcsin \left(\frac{R}{1+c_1\eta}\right)=2\arcsin\left(
1779: \frac{1-c_2\eta}{1+c_1\eta}\right),
1780: \]
1781: since the farthest~$z$ can get from the origin is by
1782: assumption~$1+c_1\eta$. It then suffices to choose~$c_1,c_2$ so that
1783: \begin{equation}\label{view}
1784: 2\arcsin\left(
1785: \frac{1-c_2\eta}{1+c_1\eta}\right)>2\psi_0
1786: \end{equation}
1787: (e.g., decide $c_2=c_1$, then decrease~$c_1=c_2$ until the inequality
1788: in~(\ref{view}) is satisfied) in order to ensure that the angle under
1789: which~$z$ views the circle~$C_0$ exceeds~$2\psi_0$. Since~$C_0$
1790: encloses the inner disc of~$(z\Delta^{-1})^c$ with
1791: which it is concentric, and since the angle at~$z$ of the apex
1792: of~$(z\Delta^{-1})^c$ is $2\psi_0$, there results that the
1793: angle at~$z$ between~$\gamma_{T}$ and~$\gamma_{T'}$ encompasses the
1794: apex of~$(z\Delta^{-1})^c$; see Figure~\ref{geomangle-fig}.
1795: In this way, the apex of~$(z\Delta^{-1})^c$ is avoided.
1796:
1797: \begin{figure}
1798:
1799: \begin{center}
1800: \includegraphics[width=4truecm]{geomangle.eps}
1801: \end{center}
1802:
1803: \caption{\label{geomangle-fig}
1804: Apex avoidance condition: The angle at~$z$ of contour~$\gamma$ is
1805: constructed to
1806: be larger than the angle at~$z$ of the apex of
1807: $(z\Delta^{-1})^c$.}
1808: \end{figure}
1809:
1810:
1811: Last, for~$\lambda$ any of the four contours of which~$\gamma$ is
1812: comprised, let $I(\lambda)$ be the integral of~(\ref{eq:45}) taken
1813: along contour~$\lambda$. The circular arc $\gamma_{z}$
1814: has all its points at a distance $c_3\delta$ from~$z$,
1815: so that there
1816: \[
1817: |1-w|=\Theta(\delta),\quad |z-w|=\Theta(\delta),\quad
1818: f(w)g\left(\frac{z}{w}\right)=O\left(\delta^{a+b}\right).
1819: \]
1820: Therefore, by trivial bounds,
1821: \begin{equation}\label{hfin1}
1822: I({\gamma_{z}})=O\left(\delta^{a+b+1}\right).
1823: \end{equation}
1824: On the other hand, along $\gamma_{0}$ the functions
1825: $f(w)$ and $g(z/w)$ stay away from their singularities, so that
1826: \begin{equation}\label{hfin2}
1827: I({\gamma_{0}})=O(1).
1828: \end{equation}
1829: There remains only to estimate the contribution along
1830: the two connecting segments~$\gamma_{T}$ and $\gamma_{T'}$. The two
1831: situations are similar (upon interchanging the roles of~$a$ and~$b$).
1832: It is then easily seen that the contribution along the
1833: ray stemming from $z$ is bounded from above by a multiple of an
1834: integral of the form
1835: \begin{equation}\label{anint}
1836: \int_{c_3\delta}^{+\infty} t^a |t-z_0|^b\, dt
1837: \end{equation}
1838: where~$z_0$ is a complex number at a distance $\Theta(\delta)$ from
1839: the real line. (The quantity~$t$ parameterizes the tangent line~$T$
1840: or~$T'$.) The last integral is $O(\delta^{a+b+1})$ as
1841: results from the change of variables~$t=\delta\tau$. Consequently,
1842: one finds
1843: \begin{equation}\label{hfin3}
1844: I({\gamma_{T}})+I({\gamma_{T'}})=O\left(\delta^{a+b+1}\right).
1845: \end{equation}
1846: Putting together all the estimates of~(\ref{hfin1}), (\ref{hfin2}),
1847: (\ref{hfin3}) yields the desired result.
1848: \end{proof}
1849:
1850:
1851:
1852:
1853:
1854:
1855: %The main idea is that the turgescent
1856: %\JF{(laughing) We will need to clean this up! Also, as NK and I
1857: %discussed, this isn't
1858: %true for \emph{all} of~$\gamma_2$ and~$\gamma_4$, but rather just the
1859: %part of each close
1860: %to~$z$.}
1861: %part is at a distance $\Theta(\delta)$ from both $z$ and 1, where we
1862: %have set $\delta :=
1863: %|1 - z|$.
1864:
1865: % Define
1866: % \begin{align*}
1867: % \theta &:= \Arg{z}, & \phi &:= \arcsin{\tfrac{c\delta}{R}},\\
1868: % \alpha_1 & := Re^{i(\theta - \phi)}, & \beta_1 &:= z - c \delta e^{i(\theta +
1869: % \frac{\pi}{2})},\\
1870: % \alpha_2 & := Re^{i(\theta + \phi)}, &\beta_2 &:= z + c \delta e^{i(\theta +
1871: % \frac{\pi}{2})},
1872: % \end{align*}
1873: % where we can (it is not hard to see) and do choose $0 <c \leq
1874: % \tfrac14\sin\psi_0$ small enough that $\gamma$ (and its interior) stays
1875: % entirely within the interior of $\Delta(\psi_0,\eta)$. (The reason for
1876: % the restriction $c \leq \tfrac14\sin\psi_0$ will become clear later.)
1877: %
1878: %
1879: %
1880: % % \begin{figure}%[ht]%
1881: % % \centering
1882: % % \includegraphics{hadamard.2}%
1883: % % \caption{The contour $\gamma$.}%
1884: % % \label{fig:5-1b}%
1885: % % \end{figure}
1886: % % \begin{figure}%[htbp]%
1887: % % \centering
1888: % % \includegraphics{hadamard.3}%
1889: % % \caption{The geometry of the contour $\gamma$.}%
1890: % % \label{fig:5-1c}%
1891: % % \end{figure}
1892: %
1893: % %%\input{choice-of-c}
1894: % %%{\ttfamily%
1895: % %% Here's my current idea on \underline{The choice of $c$} (see
1896: % %% Figure~\ref{fig:choice-of-c}) Define $r := \min\{ |z-v|: v \in
1897: % %% \partial\Delta(\phi,\eta)\}$, so that any circle of radius less than $r$ centered
1898: % %% at $z$ will stay entirely within $\Delta(\phi,\eta)$. Also, $\exists d > 0$,
1899: % %% such that the line segment joining 0 to $z$ can be displaced
1900: % %% parallel to itself by a distance $d$ and still stay within
1901: % %% $\Delta(\phi,\eta)$. Choose $c\delta \leq \min(r,d)$, or $c \leq \min(r,d)/\delta$ and
1902: % %% this will ensure that the contour $\gamma$ stays entirely within
1903: % %% $\Delta(\phi,\eta)$.}
1904: % %% To find $d$, have a look at Figure~\ref{fig:choice-of-d}.
1905: % %%\input{choice-of-c2}
1906: %
1907: %
1908: % Then the contour $\gamma$ is $\gamma_1 \cup \gamma_2 \cup \gamma_3
1909: % \cup \gamma_4$ where
1910: % \begin{align*}
1911: % \gamma_1 &= \{ w: |w| = R,\; \Arg{w} \in [0,\theta - \phi] \cup [\theta
1912: % + \phi, 2\pi]\},\\
1913: % \gamma_2 &= \{ w: w = (1-t)\alpha_1 + t\beta_1,\; 0 \leq t \leq 1\},\\
1914: % \gamma_3 &= \{ w: |w-z| = \delta/2,\; \Arg(w - z) \in [0,\pi]\},\\
1915: % \gamma_4 &= \{ w: w = t\alpha_2 + (1-t)\beta_2,\; 0 \leq t \leq 1 \}.
1916: % \end{align*}
1917: % By the asymptotic conditions on $f$ and $g$, we know that there are
1918: % positive constants $K_1$ and $K_2$ such that $f(w) \leq K_1 |1-w|^a$
1919: % and $g(w) \leq K_2 |1-w|^b$ for all $w$ on $\gamma$. We also note that $R
1920: % \leq |w| \leq 1+\eta$, so we can find a constant $K$ such that
1921: % \begin{align}
1922: % \label{eq:49}
1923: % |(f \odot g)(z)| &= \left|\frac{1}{2\pi i} \int_{\gamma}\,f(w)
1924: % g\left(\frac{z}{w}\right) \frac{dw}{w}\right| \notag\\
1925: % & \leq \frac{K}{2\pi} \int_{\gamma}\,|1 - w|^a |w - z|^b\,|dw| \notag\\
1926: % & \leq \frac{K}{2\pi} \left( \int_{\gamma_1} + \int_{\gamma_2} +
1927: % \int_{\gamma_3} + \int_{\gamma_4} \right) |1 - w|^a |w - z|^b\,|dw|.
1928: % \end{align}
1929: % We now show that each of the four integrals in~\eqref{eq:49} is
1930: % $O(\delta^{a + b + 1} = |1-z|^{a+b+1})$, completing the proof
1931: % of~\eqref{eq:44}.
1932: % \begin{enumerate}
1933: % \item Uniformly along $\gamma_1$, the integrand is $\Theta(1)$.
1934: % Hence the integral is
1935: % $\Theta(1)$, which is $O(|1-z|^{a+b+1})$ since $a+b+1 < 0$.
1936: % \item Along $\gamma_2$, we can express the variable of integration $w$ as%
1937: % \[
1938: % w \equiv w(u) = (\alpha_1 - \beta_1)u + \beta_1, \quad 0 \leq u \leq 1,
1939: % \]
1940: % so that
1941: % \[ dw = (\alpha_1 - \beta_1)du.\]
1942: % Note that $|\beta_1 - \alpha_1| = (1 + o(1))(1-R)$.
1943: % The integral of interest is then bounded by a constant times
1944: % \[
1945: % \int_{0}^\infty |1 - w(u)|^a|w(u) - z|^b\,du .
1946: % \]
1947: %
1948: % We split this integral into two pieces, one where $u \leq M\delta$, say
1949: % $\gamma_{21}$, and the other where $u > M\delta$, say $\gamma_{22}$, where $M$ is a
1950: % suitably chosen large constant. We will show that each of these
1951: % integrals is $O(\delta^{a+b+1})$. To this end, observe that
1952: % \begin{align}
1953: % 1 - w &= 1 - \beta_1 + (\beta_1 - \alpha_1)u \notag\\
1954: % &= 1 - z + c\delta e^{i(\theta + \tfrac\pi2)} + u |\beta_1 - \alpha_1|e^{i\theta} \notag\\
1955: % &= e^{i\theta}( e^{-i\theta} - |z| + i c\delta + u|\beta_1 -
1956: % \alpha_1| ), \label{eq:57}
1957: % %%\intertext{whence}
1958: % %% \label{eq:48}
1959: % %%|1-w| &= |e^{-i\theta} - |z| + ic\delta + u|\beta_1-\alpha_1| |.
1960: % \end{align}
1961: % We defer to Lemma~\ref{lemma:3.5} a calculation that establishes
1962: % $|1-w| = \Omega(\delta)$ uniformly along all of $\gamma_2$. To get an
1963: % upper bound, using the triangle inequality,
1964: % \begin{align*}
1965: % |1-w| & \leq |1-\beta_1| + |\beta_1-w|\\
1966: % & \leq \delta + c\delta + |\beta_1-w| = (c+1)\delta + |\beta_1-w|.
1967: % \end{align*}
1968: % Restricting attention to $\gamma_{21}$ gives $|\beta_1-w| \leq
1969: % |\beta_1-\alpha_1|M\delta$. Hence, uniformly
1970: % along $\gamma_{21}$, $|1-w| = \Theta(\delta)$.
1971: %
1972: % For $|w-z|$, a trivial lower bound is $|w-z| \geq |\beta_1-z| =
1973: % c\delta$. Along $\gamma_{21}$, an upper bound is established by the triangle
1974: % inequality as
1975: % \[
1976: % |w-z| \leq |\beta_1-z| + |w-\beta_1| \leq c\delta + |\beta_1-\alpha_1|M\delta.
1977: % \]
1978: % Hence $|w-z| = \Theta(\delta)$ uniformly along $\gamma_{21}$, so that the
1979: % integral along $\gamma_{21}$ is $\Theta(\delta^{a+b+1})$ since the
1980: % integrand is uniformly $\Theta(\delta^{a+b})$ and the path of integration is of
1981: % length $\Theta(\delta)$.
1982: %
1983: % For the integral along $\gamma_{22}$, we will establish that $|1-w| =
1984: % \Theta(u)$ and $|w-z| = \Theta(u)$ for suitably large $M$. The
1985: % desired bound on the integral will then follow, since $a+b+1 < 0$.
1986: %
1987: % We use~\eqref{eq:57} and the fact that the real part of a complex
1988: % number is a lower bound on its absolute value to get
1989: % %%We can lower-bound $|1-w|$ by the real part of
1990: % %%the right-hand side of~\eqref{eq:57} as
1991: % \[
1992: % |1-w| \geq \cos{\theta} - |z| + |\beta_1-\alpha_1|u.\
1993: % \]
1994: % Writing $z=1+ \delta e^{i\psi}$, we note that
1995: % \[
1996: % \cos\theta = \frac{1+\delta\cos\psi}{|z|} \geq
1997: % \frac{1+\delta\cos\psi}{1+\delta} \geq
1998: % \frac{1-\delta}{1+\delta} \geq 1 - 2\delta.
1999: % \]
2000: % Hence
2001: % \begin{align*}
2002: % |1-w| &\geq 1 - 2\delta - (1+\delta) + (1 + o(1))(1-R)u \\
2003: % &= (1 + o(1))(1-R)u -3\delta.
2004: % \end{align*}
2005: % Since $u \geq M\delta$, if we choose $M$ large enough, we get $|1-w| =
2006: % \Omega(u)$.
2007: %
2008: % For the upper bound,
2009: % \begin{align*}
2010: % |1-w| & \leq |w-\beta_1| + |1-\beta_1|\\
2011: % & = |\alpha_1 - \beta_1|u + |1-\beta_1|\\
2012: % & \leq (1 + o(1))(1-R)u + (c+1)\delta\\
2013: % & = O(u).
2014: % \end{align*}
2015: % Hence $|1-w| = \Theta(u).$
2016: %
2017: % The estimate for $|w-z|$ is similar. For the upper bound, observe that
2018: % \begin{align*}
2019: % w-z &= \beta_1 - z + (\alpha_1 - \beta_1)u,
2020: % \intertext{whence}
2021: % |w-z| &\leq |\beta_1-z| + |\alpha_1-\beta_1|u\\
2022: % &= c\delta + (1 + o(1))(1-R)u = O(u).
2023: % \end{align*}
2024: % For the lower bound, we express
2025: % \[ w-z = -c\delta e^{i(\theta+\tfrac{\pi}{2})} + u|\alpha_1-\beta_1|e^{i\theta} =
2026: % e^{i\theta}(u|\alpha_1-\beta_1| - ic\delta), \]
2027: % so that
2028: % \begin{align*}
2029: % |w-z| &= |u|\alpha_1-\beta_1| - ic\delta|\\
2030: % &\geq u|\alpha_1 - \beta_1|\\
2031: % &= (1 + o(1))(1-R)u.
2032: % \end{align*}
2033: % This establishes $|w-z| = \Omega(u)$ and hence $|w-z| = \Theta(u)$.
2034: %
2035: % \item Along $\gamma_3$, $|w-z| = c\delta$, and $(1-c)\delta = \delta -
2036: % c\delta \leq |1-w| \leq \delta + c\delta= (1+c)\delta$ using the triangle
2037: % inequality. The
2038: % contour of integration has length $\pi c\delta$. Hence the integral is
2039: % $\Theta(\delta^{a+b+1}) = O( |1-z|^{a+b+1} )$.
2040: % \item Along $\gamma_4$, the argument is similar to (and somewhat simpler
2041: % than) that for the integral
2042: % along $\gamma_2$.
2043: % \end{enumerate}
2044: % \end{proof}
2045: % We now present the proof of the lower-bound on $|1-w|$ along
2046: % $\gamma_2$ used in
2047: % the proof of Proposition~\ref{lem:3}.
2048: % \begin{lemma}
2049: % \label{lemma:3.5}
2050: % In the notation of the proof of Proposition~\ref{lem:3}, $|1-w| =
2051: % \Omega(\delta)$ uniformly along all of $\gamma_2$.
2052: % \end{lemma}
2053: % \begin{proof}
2054: % If $\sin\theta > 2c\delta$, we lower-bound $|1-w|$
2055: % using~\eqref{eq:57} and the fact that the imaginary part of a
2056: % complex number is a lower bound on the its absolute value,
2057: % \[
2058: % |1-w| \geq -\Im(e^{-i\theta} - |z| + ic\delta + u|\beta_1-\alpha_1|)
2059: % = \sin\theta - c\delta
2060: % > c\delta.
2061: % \]
2062: % Otherwise, we use the real part as a lower bound:
2063: % \[
2064: % |1-w| \geq \Re(e^{-i\theta} - |z| + ic\delta + u|\beta_1-\alpha_1|)
2065: % \geq \cos\theta - |z| + u|\beta_1-\alpha_1|.
2066: % \]
2067: % We choose $c$ small enough so that if $\Re(z) = 1 - |z|\cos\theta < 0$, then
2068: % $\sin\theta > 2c\delta$. Indeed, if $\Re(z) > 1$, then $z = 1 +
2069: % \delta e^{i\psi}$ where $\psi_0 < \psi < \pi/2$, and so
2070: % \[
2071: % \sin\theta = \frac{\delta\sin\psi}{|z|} > \frac{\delta\sin\psi_0}{2}
2072: % \]
2073: % and any $c \leq \frac{1}{4}\sin\psi_0$ will suffice. So we may
2074: % assume $$\sin\theta \leq 2c\delta \text{ and } 1 - |z|\cos\theta
2075: % \geq 0.$$
2076: %
2077: % Now $\cos^2\theta \geq 1 - 4c^2\delta^2$, whence $\cos\theta = 1 -
2078: % o(\delta)$, and we need only prove that $1 - |z| \geq 0$ and $1 -
2079: % |z| \geq \Omega(\delta)$. To this end,
2080: % observe that $\delta^2 = |1-z|^2 = (1-|z|\cos\theta)^2 +
2081: % (|z|\sin\theta)^2$, so that
2082: % \begin{align*}
2083: % 1 &= |z|\cos\theta + (1 - |z|\cos\theta)\\
2084: % &\geq |z|\sqrt{1- 4c^2\delta^2} + \delta\sqrt{1-4c^2|z|^2}\\
2085: % &\geq |z|(1 - 4c^2\delta^2) + \delta(1 - 4c^2|z|^2),
2086: % \intertext{and hence, for all sufficiently small $\delta$,}
2087: % 1 - |z| &\geq \delta(1 - 4c^2|z|^2) - 4|z|c^2\delta^2\\
2088: % &\geq \delta(1-16c^2) - 8c^2\delta^2 \quad \text{(since
2089: % $|z| \leq 2$ for $\delta \leq 1$)}\\
2090: % &\geq p\delta,
2091: % \end{align*}
2092: % where $p$ is a positive constant depending only on $\psi_0$, since
2093: % $c \leq \tfrac14\sin\psi_0 < \tfrac14$.
2094: % \end{proof}
2095: \begin{remark*}
2096: \label{rem:intlog}
2097: The proof technique of Proposition~\ref{lem:3}
2098: tolerates the presence of logarithmic factors,
2099: in which case it suffices to
2100: develop the corresponding estimates for the basic
2101: integral~(\ref{anint}). We find in this way,
2102: when $a+b+1<0$, the estimate
2103: \[
2104: O\left(|1-z|^a |L^k(z)| \right)\odot O\left(|1-z|^b |L^\ell(z)| \right)
2105: =O\left(|1-z|^{a+b+1} |L^{k+\ell}(z)| \right).
2106: \]
2107: The contour $\gamma$ used in the proof
2108: is also susceptible to many variations. For instance, one may deform it
2109: slightly to include a ``hook'' near $w=1$,
2110: in which case the modified contour may be used to estimate
2111: more finely the singular behaviour of Hadamard products.
2112: \end{remark*}
2113: We can then extend the asymptotic range
2114: covered by Proposition~\ref{lem:3} as follows.
2115: \begin{proposition} \label{thm:4}
2116: Assume that $f(z)$ and $g(z)$ are $\Delta$-regular and that
2117: for~$z\in\Delta$,
2118: \[
2119: f(z) = O(|1-z|^a)\mbox{\rm \ \ and\ \ }g(z) = O(|1-z|^b).
2120: \]
2121: \textit{(i)} If $k < a + b +1 < k+1$ for some integer $-1 \leq k <
2122: \infty$, then for~$z\in\Delta'$:
2123: \begin{equation*}
2124: \label{eq:46}
2125: (f \odot g)(z) = \sum_{j=0}^k \frac{(-1)^j}{j!}
2126: (f\odot{}g)^{(j)}(1)(1-z)^j + O(|1-z|^{a+b+1}).
2127: \end{equation*}
2128: \par\noindent
2129: $(ii)$ If $a+b+1$ is a nonnegative integer then for~$z\in\Delta'$:
2130: \begin{equation*}
2131: \label{eq:47}
2132: (f \odot g)(z) = \sum_{j=0}^{a+b} \frac{(-1)^j}{j!}
2133: (f\odot{}g)^{(j)}(1)(1-z)^j
2134: + O( |1-z|^{a+b+1}|L(z)|).
2135: \end{equation*}
2136: \end{proposition}
2137: \begin{proof}
2138: Let $\partial=\partial_z$ denote the operator $\tfrac{d}{dz}$ and let $\vartheta$ denote
2139: the Euler operator $z\partial$. Observe that
2140: \[
2141: \vartheta(f\odot{}g) = (\vartheta{}f) \odot{}g =
2142: f\odot (\vartheta g),
2143: \]
2144: which yields
2145: \[
2146: \vartheta^{k+1}(f\odot{}g)=(\vartheta^{k+1} f)\odot{}g.
2147: \]
2148: The differentiation properties of Theorem~\ref{thm:2A} imply [with $k
2149: := a + b + 1$ in Case~(ii)] that
2150: $\vartheta^{k+1}f(z)$ is $O(|1-z|^{a-k-1})$. Thus, Proposition~\ref{lem:3}
2151: applies, to the effect that
2152: \[
2153: \left( \vartheta^{k+1}(f\odot{}g) \right)(z) = O\left(|1-z|^{a+b-k}\right).
2154: \]
2155: \smallskip
2156: On the other hand, the operator~$\vartheta^{-1}$ is (for $h$ in the
2157: image of $\vartheta$)
2158: \[
2159: \left( \vartheta^{-1} h \right)(z) :=P_0+\int_0^z h(t)\, \frac{dt}{t},
2160: \]
2161: for some integration constant~$P_0$.
2162: It is then possible to recover~$h=f\odot g$ through successive
2163: integrations,
2164: by making use of Theorem~\ref{thm:2B}.
2165:
2166: \smallskip
2167:
2168: Case~\textit{(i)}. By definition of $k$, one has
2169: $-1<a+b-k<0$. Repeated integrations then show that
2170: \begin{equation}\label{hback1}
2171: (f\odot{}g)(z)=P(z)+O\left(|1-z|^{a+b+1}\right),
2172: \end{equation}
2173: for some polynomial~$P(z)$ of degree~$k$ that encapsulates
2174: the sequence of integration constants.
2175: Equation~(\ref{hback1})
2176: yields qualitatively the form of the statement.
2177: The polynomial $P(z)$ is then
2178: automatically determined as the first $(k+1)$ terms of
2179: the Taylor expansion of~$f\odot g$ at~$1$, which is precisely what
2180: our assertion expresses.
2181:
2182: \smallskip
2183:
2184: Case~\textit{(ii)}. In this case, the first integration step requires
2185: integrating a term~$O(|1-z|^{-1})$, which leads to the
2186: logarithmic form of the statement. (See also the comments following
2187: Theorem~\ref{thm:2B}.)
2188: \end{proof}
2189:
2190:
2191: %
2192: % Case~$(i)$. We first show that
2193: % \begin{equation}
2194: % \label{eq:62}
2195: % (f\odot{}g)(z) = \sum_{j=0}^k \frac{(-1)^j}{j!}
2196: % [\theta^j(f\odot{}g)](1)(-\ln{}z)^j + O(|1-z|^{a+b+1}).
2197: % \end{equation}
2198: % The proof is by induction on $k$. The basis case of
2199: % $k=-1$ is a special case of Proposition~\ref{lem:3}.
2200: % and by Theorem~\ref{thm:2A} that
2201: % \[
2202: % (\theta{}f)(z) = zf'(z) = O(|1-z|^{a-1}).
2203: % \]
2204: % If $k \geq 0$, then by the induction hypothesis
2205: % \[
2206: % [\theta(f\odot{}g)](\zeta) = (\theta{}f\odot{}g)(\zeta)
2207: % = \sum_{j=0}^{k-1} \frac{(-1)^j}{j!}
2208: % [\theta^j(\theta{}f\odot{}g)](1) (-\ln{}\zeta)^j + O( |1-\zeta|^{a+b}).
2209: % \]
2210: % We divide by $\zeta$ and then integrate both sides
2211: % with respect to $\zeta$ from $z$ to $1$ to get
2212: % \begin{align*}
2213: % (f \odot g)(z) &= (f \odot g)(1) + \sum_{j=0}^{k-1}
2214: % \frac{(-1)^{j+1}}{(j+1)!} [\theta^{j+1}(f\odot{}g)](1)
2215: % (-\ln{}z)^{j+1} \notag\\
2216: % & \qquad {}+ O(|1-z|^{a+b+1})\notag\\
2217: % &= \sum_{j=0}^k \frac{(-1)^j}{j!}
2218: % [\theta^j(f\odot{}g)](1)(-\ln{}z)^j +
2219: % O(|1-z|^{a+b+1}),\label{eq:61}
2220: % \end{align*}
2221: % as required.
2222: %
2223: % To obtain the expansion in the asserted form, we use the
2224: % following two well-known facts:
2225: % \begin{enumerate}
2226: % \item For fixed nonnegative integers $j$ and $k$, as $z \to 1$ we
2227: % have
2228: % \[
2229: % (-\ln{}z)^j = j! \sum_{r=0}^k\stirlingone{r}{j}
2230: % \frac{(1-z)^r}{r!} + O(|1-z|^{k+1}),
2231: % \]
2232: % where $\stirlingone{r}{j}$ denotes a Stirling number\footnote{
2233: % Better ref.: GKP~\cite{GrKnPa89}} of the
2234: % first kind (see equation~1.2.9-(26) in~\cite{Knuth97}).
2235: % \item The operators $D$ and $\theta$ satisfy
2236: % \NK{reference for the second?}
2237: % \[
2238: % z^rD^r = \sum_{j=1}^r(-1)^{r-j}
2239: % \stirlingone{r}{j}\theta^j, \quad r \geq 1.
2240: % \]
2241: % \end{enumerate}
2242: % Using these facts in~\eqref{eq:62} and recalling that $a+b+1 <
2243: % k+1$, we get
2244: % \begin{align*}
2245: % (f\odot{}g)(z) &= \sum_{j=0}^k (-1)^j[\theta^j(f\odot{}g)](1)
2246: % \sum_{r=0}^k \stirlingone{r}{j} \frac{(1-z)^r}{r!} +
2247: % O(|1-z|^{a+b+1})\\
2248: % &= (f\odot{}g)(1) + \sum_{r=1}^k (-1)^r \frac{(1-z)^r}{r!}
2249: % \sum_{j=1}^k (-1)^{r-j}
2250: % \stirlingone{r}{j}[\theta^j(f\odot{}g)](1)\\
2251: % &\qquad + O(|1-z|^{a+b+1})\\
2252: % &= \sum_{r=0}^k (-1)^r
2253: % \frac{(1-z)^r}{r!}\left.[w^r(f\odot{}g)^{(r)}(w)]\right\rvert_{w=1}(1)
2254: % + O(|1-z|^{a+b+1})\\
2255: % &= \sum_{j=0}^k \frac{(-1)^j}{j!}
2256: % (f\odot{}g)^{(j)}(1)(1-z)^j + O(|1-z|^{a+b+1}),
2257: % \end{align*}
2258: % as desired.
2259: % \par\noindent
2260: % (b) The structure of the proof is the same as that of
2261: % part~(a). One first proves a result similar to~\eqref{eq:62}.
2262: % Again, the proof is by induction on $k := a+b+1$. When $k=0$,
2263: % after applying Theorem~\ref{thm:2A} and Proposition~\ref{lem:3}
2264: % to obtain $\theta(f \odot g)(z) = O(|1-z|^{-1})$ we can apply
2265: % Theorem~\ref{thm:2B} to get $(f \odot g)(z) = O(|\log(1 - z)|)$.
2266: % The induction step is similar to that in (a) above. Finally,
2267: % using the same argument as for part~(a), we obtain the expansion
2268: % in the asserted form.
2269: % \end{proof}
2270:
2271:
2272: \subsection{Composition rules}\label{subsec:had-all}
2273: At this stage, we can summarize the state of affairs regarding
2274: Hadamard products by the following general statement.
2275:
2276:
2277: \begin{theorem}[Hadamard composition of singularities]\label{thm:hadmain}
2278: Let $f(z)$ and $g(z)$ be two functions that are $\Delta$--regular
2279: with expansions of the type~\eqref{eq:27}:
2280: \[
2281: f(z)=\sum_{m=0}^M c_m(1-z)^{\alpha_m}+O(|1-z|^A),
2282: \qquad
2283: g(z)=\sum_{n=0}^N d_n(1-z)^{\beta_n}+O(|1-z|^B).
2284: \]
2285: Then, the Hadamard product $(f\odot g)(z)$ is
2286: also $\Delta$--regular. Its singular expansion is computable
2287: by bilinearity, using the composition rules of Proposition~\ref{thm:3}
2288: and the remarks thereafter,
2289: with error terms provided by Propositions~\ref{lem:3} and~\ref{thm:4}:
2290: \[
2291: (f\odot g)(z)=\sum_{m,n} c_m d_n [(1-z)^{\alpha_m}\odot(1-z)^{\beta_n}]
2292: +P(1-z)+
2293: O\left(|1-z|^{C}\right),
2294: \]
2295: where~$C := 1+\min(\alpha_0+B,A + \beta_0)$
2296: and~$P$ is a polynomial of degree less than~$C$.
2297: \end{theorem}
2298:
2299: The polynomial~$P$ is accessible via the Taylor expansion of
2300: $h-h_{\operatorname{sing}}$, where $h_{\operatorname{sing}}$
2301: represents the sum of all the elements in the asymptotic expansion
2302: of~$h := f \odot g$ at~$z=1$ that are singular.
2303: This theorem then validates the following algorithm,
2304: which is often helpful in
2305: computations done by hand when composing functions % singular expansions
2306: under Hadamard products.
2307: \begin{itemize}\def\Asympt{\operatorname{Asympt}}
2308: \item[]
2309: \emph{\bf ``Zigzag'' Algorithm}. [Computes the singular expansion of
2310: $f\odot g$ up to $O(|1-z|^C)$. ]\\[1truemm]
2311: {\bf 1.} Use singularity analysis to determine
2312: separately the asymptotic
2313: expansions $\Asympt(f_n)$, $\Asympt(g_n)$ of~$f_n=[z^n]f(z)$
2314: and~$g_n=[z^n]g(z)$ into descending powers of~$n$.
2315: \\
2316: {\bf 2.} Perform the resulting product and reorganize
2317: as~$\Asympt(f_ng_n)$.
2318: \\
2319: {\bf 3.} Choose a basis $\mathcal{B}$ of singular functions,
2320: for instance, the standard basis
2321: $\mathcal{B}=\left\{(1-z)^\alpha L(z)^k\right\}$,
2322: or the polylogarithm basis,
2323: $\mathcal{B}=\left\{ \Li_{\beta,k}(z)\right\}$.
2324: Construct a function~$H(z)$ expressed
2325: in terms of~$\mathcal{B}$ whose singular behaviour is such that
2326: the asymptotic form of its coefficients, $\Asympt(H_n)$, is compatible
2327: with~$\Asympt(f_ng_n)$ up to the needed error terms.
2328: \\
2329: {\bf 4.} Output the singular expansion of
2330: $f\odot g$ as the quantity $H(z)+P(z)+O(|1-z|^C)$, where $P$ is a
2331: polynomial in~$(1-z)$ of degree less than~$C$.
2332: \end{itemize}
2333: The reason for the addition of a polynomial in Step~{\bf 4},
2334: is that integral powers of~$(1-z)$ do not leave a trace
2335: in coefficient asymptotics since their contribution is
2336: asymptotically null.
2337: (An example of such ``hidden''
2338: analytic terms already appears
2339: in the composition rule for powers given in Proposition~\ref{thm:3}.)
2340: The Zigzag Algorithm is then principally useful for
2341: determining the divergent part of expansions.
2342: If needed, the coefficients
2343: in the polynomial~$P$
2344: can be expressed as
2345: values of the function $f\odot g$ and
2346: its derivatives at~1 once it has been
2347: stripped of its nondifferentiable terms.
2348: (This is analogous to the situation prevailing in Proposition~\ref{thm:4}.)
2349:
2350: \begin{example} {\em The return of P\'olya's drunkard.}
2351: In the $d$--dimensional lattice~$\mathbb{Z}^d$ of points with integer
2352: coordinates, the drunkard performs a random walk starting from the
2353: origin with steps in~$\{-1,+1\}^d$, each taken with equal
2354: likelihood. The probability that the drunkard is back at the
2355: origin after $2n$ steps is
2356: \begin{equation}\label{qdrunk}
2357: q_n^{(d)}=\left(\frac{1}{2^{2n}}\binom{2n}{n}\right)^d,
2358: \end{equation}
2359: since the walk is a product~$d$ independent
2360: 1--dimensional walks. The probability that $2n$ is the epoch of the
2361: \emph{first} return to the origin is the quantity~$p_{n}^{(d)}$,
2362: which is determined implicitly by
2363: \begin{equation}\label{pdrunk}
2364: \left(1-\sum_{n=1}^\infty p_{n}^{(d)}z^n\right)^{-1}
2365: =\sum_{n=0}^\infty q_{n}^{(d)}z^n,
2366: \end{equation}
2367: as results from the convolution equations
2368: expressing the decomposition of loops into primitive loops.
2369: In terms of the associated ordinary generating functions $P$ and~$Q$, this
2370: relation thus reads as $(1-P(z))^{-1}=Q(z)$.
2371:
2372: The asymptotic analysis of the $q_n$'s is straightforward;
2373: the one of
2374: the $p_n$'s is more involved and is of interest in connection with
2375: recurrence and transience of the random walk;
2376: see, e.g.,~\cite{DoSn84,Lawler91}.
2377: The Hadamard closure theorem provides a
2378: direct access to this problem. Define
2379: \[
2380: \lambda(z):=\sum_{n\ge0}\frac{1}{2^{2n}}\binom{2n}{n}z^n
2381: \equiv \frac{1}{\sqrt{1-z}}.
2382: \]
2383: Then, Equations~(\ref{qdrunk}) and~(\ref{pdrunk}) imply:
2384: \[
2385: P(z)=1-\frac{1}{\lambda(z)^{\odot d}}, \qquad
2386: \hbox{where}\quad
2387: \lambda(z)^{\odot d}:=\lambda(z)\odot\cdots \odot\lambda(z)
2388: \ \hbox{($d$ times)}.
2389: \]
2390: The singularities of $P(z)$ are found to be
2391: as follows.
2392:
2393: \smallskip
2394: % \begin{itemize}
2395: %\item[]
2396: \underline{$d=1$}: No Hadamard product is involved and
2397: \[
2398: P(z)=1-\sqrt{1-z},\qquad\hbox{implying}\quad
2399: p_n^{(1)}=\frac{1}{n2^{2n-1}}\binom{2n-2}{n-1}\sim\frac{1}{2\sqrt{\pi
2400: n^3}}.
2401: \]
2402: (This agrees with the classical
2403: combinatorial solution expressed in terms of Catalan numbers.)
2404:
2405:
2406: \smallskip
2407: % \item[]
2408: \underline{$d=2$}: By the Hadamard closure theorem, the function
2409: $Q(z)=\lambda(z)\odot\lambda(z)$ admits a priori
2410: a singular expansion at~$z=1$ that is
2411: composed solely of elements of the form $(1-z)^\alpha$ possibly
2412: multiplied by integral
2413: powers of the logarithmic function~$L(z)$. From a computational
2414: standpoint (cf.\ the Zigzag Algorithm),
2415: it is then best to start from the coefficients themselves,
2416: \[
2417: q_n^{(2)}\sim \left(\frac{1}{\sqrt{\pi n}}-\frac{1}{8\sqrt{\pi
2418: n^3}}+\cdots\right)^2
2419: \sim
2420: \frac{1}{\pi}\left(\frac{1}{n}-\frac{1}{4n^2}+\cdots\right),
2421: \]
2422: and reconstruct the only singular expansion that is compatible,
2423: namely
2424: \[
2425: Q(z)=\frac{1}{\pi}L(z)+K+O((1-z)^{1-\epsilon}),
2426: \]
2427: where~$\epsilon > 0$ is an arbitrarily small constant
2428: and $K$ is fully determined as the limit as $z\to1$ of $Q(z)-\pi^{-1}
2429: L(z)$. Then it can be seen that the function~$P$ is
2430: $\Delta$--continuable. (Proof: Otherwise, there would
2431: be complex poles arising from zeros of the function~$Q$ on the unit
2432: disc, and this would entail
2433: in~$p_n^{(2)}$ the presence of terms
2434: oscillating around~0, a fact that contradicts the necessary positivity
2435: of probabilities.) The singular expansion of~$P(z)$ at~$z=1$ results
2436: immediately from that of~$Q(z)$:
2437: \[
2438: P(z)\sim 1-\frac{\pi}{L(z)}+\frac{\pi^2 K}{L^2(z)}+\cdots.
2439: \]
2440: so that, by the extension of Theorem~\ref{thm:basic-sa} to arbitrary
2441: powers of logarithms as given in~\cite{FlOd90b,Odlyzko95}, one has
2442: \[\begin{array}{lll}
2443: p_n^{(2)} &=&\ds \frac{\pi}{n\log^2
2444: n}
2445: -2\pi\frac{\gamma+\pi K}{n\log^3 n}
2446: +O\left(\frac{1}{n\log^4n}\right)
2447: \\
2448: K&=&\ds 1+\sum_{n=1}^\infty
2449: \left(16^{-n} \binom{2n}{n}^2-\frac{1}{\pi n}\right)\\
2450: &\doteq& 0.8825424006106063735858257\,.
2451: \end{array}
2452: \]
2453: %% This can be gotten in Mathematica by using
2454: %% \[
2455: %% K = \lim_{z \to 1} [{}_2F_1( 1/2, 1/2; 1; z ) - \pi^{-1} L(z) ]
2456: %% \]
2457: (See the study by Louchard \emph{et al.}~\cite[Sec.~4]{LoScToZi94} for
2458: somewhat similar calculations.)
2459:
2460: \smallskip
2461: % \item[]
2462: $d=3$: This case is easy since $Q(z)$ remains finite at its
2463: singularity $z=1$ where it admits an expansion in powers of
2464: $(1-z)^{1/2}$, to the effect that
2465: \[
2466: q_n^{(3)}\sim \left(\frac{1}{\sqrt{\pi n}}-\frac{1}{8\sqrt{\pi
2467: n^3}}+\cdots\right)^3
2468: \sim \frac{1}{\pi^{3/2}}
2469: \left(\frac{1}{n^{3/2}}-\frac{3}{8n^{5/2}}+\cdots
2470: \right).
2471: \]
2472: The function~$Q(z)$ is a priori $\Delta$-continuable
2473: and its singular expansion can be reconstructed
2474: from the form of coefficients:
2475: \[
2476: Q(z)\mathop{\sim}_{z\to1}
2477: Q(1)-\frac{2}{\pi}\sqrt{1-z}+O(|1-z|),\]
2478: leading to
2479: \[
2480: P(z)=\left(1-\frac{1}{Q(1)}\right)-\frac{2}{\pi Q^2(1)}\sqrt{1-z}
2481: +O(|1-z|).
2482: \]
2483: By singularity analysis, the last expansion gives
2484: \[
2485: \begin{array}{lll}
2486: p_n^{(3)}&=&\ds
2487: \frac{1}{\pi^{3/2}Q^2(1)}\,\frac{1}{n^{3/2}}+O\left(\frac{1}{n^2}\right)
2488: \\
2489: Q(1) &=& \ds\frac{\pi}{\Gamma\left(\frac{3}{4}\right)^4}
2490: ~ \doteq~1.3932039296856768591842463.
2491: \end{array}\]
2492: A complete asymptotic expansion in powers $n^{-3/2},n^{-5/2},\ldots$
2493: can be obtained by the same devices.
2494: In particular this improves the error term above to~$O(n^{-5/2})$.
2495: The explicit form of~$Q(1)$
2496: results from its expression as the generalized hypergeometric
2497: ${}_3F_2[\frac12,\frac12,\frac12;1,1;1]$,
2498: which evaluates by Clausen's theorem and
2499: Kummer's identity to the square of a
2500: complete elliptic integral.
2501: (See the papers by Larry Glasser for context, for
2502: instance~\cite{Glasser72}; nowadays,
2503: {\sf Maple} and {\sf Mathematica} even provide this value
2504: automatically).
2505:
2506: Higher dimensions are treated similarly, with logarithmic terms
2507: surfacing in asymptotic expansions
2508: for all even dimensions.~\hfill~$\Box$
2509: \end{example}
2510:
2511: We observe that, without the developments of the present paper,
2512: the precise asymptotic structure of such sequences is not obvious.
2513: Methods of the last example may be used to provide
2514: a rigorous setting to certain asymptotic enumeration
2515: results stated by physicists, where back-and-forth equivalences
2516: between singular expansions of functions
2517: and asymptotic expansions of coefficients are often used without
2518: much justification.
2519: See for instance the works of Guttmann and
2520: collaborators~\cite{BoGu97,GuPr93} and Chyzak's
2521: numeric-symbolic study~\cite{Chyzak97}
2522: relative to special self-avoiding polygons.
2523:
2524: %% Section 5: Some applications
2525: \section{Applications: first moments}\label{sec:some-applications}
2526:
2527: Thanks to the extended singularity
2528: analysis toolkit,
2529: we are now in a position to analyze the tree recurrences that were introduced
2530: in Section~\ref{sec:some-special-tree}.
2531: For each of the three models, two types of tolls are
2532: to be considered:
2533: \[
2534: t_n^{\alpha}:=n^\alpha \text{ (with $\alpha > 0$)},\qquad
2535: t_n^{\log}=\log n,
2536: \]
2537: and we assume in both cases that $t_0=0$.
2538: The corresponding ordinary generating functions are the
2539: polylogarithm~$\Li_\alpha\equiv \Li_{\alpha,0}$ and the specific
2540: $\Li_{0,1}$, whose singular expansions have been
2541: already recalled as Theorem~\ref{thm:li-sing}.
2542: In each case, a linear transform~$\cal L$ relates
2543: the generating function of costs, $f(z)$, to a generating function of
2544: tolls, either $t(z)$ (normalized) or $\tau(z)$ (``raw'').
2545: Theorems on composition of singularities make it possible to
2546: follow step
2547: by step the elementary operations of which~$\cal L$ is composed
2548: and determine the effect of the $\cal L$ transform on singularities in
2549: a systematic manner. Given that computations are ``automatic'', we
2550: will mostly focus our discussion on main terms and on the global shape
2551: of singular expansions, leaving some of the details as exercises to
2552: the reader---or better, to a computer algebra engine.
2553:
2554: The net outcome in each of the three tree models under consideration
2555: is the following: for large tolls, the cost is driven by the toll itself;
2556: for small tolls, the cost is of linear growth and, in a sense,
2557: ``freely'' caused by
2558: the recursion itself, that is, driven by the cumulation of costs
2559: due to small subtrees; in between,
2560: there is a threshold value of the toll where a ``resonance''
2561: takes place between the toll and the recursion, leading to the emergence of a
2562: logarithmic factor. Such facts parallel what is familiar in the context
2563: of inhomogeneous linear differential equations, where either the free
2564: regime or the forced regime dominates, with logarithmic terms
2565: being created precisely by resonances.
2566:
2567:
2568: % \emph{Notation}. Throughout our calculations, $c$ will denote an
2569: % unspecified constant, possibly zero and possibly different at each
2570: % appearance.
2571:
2572:
2573: \subsection{The binary search tree recurrence}\label{sec:binary-search-tree}
2574:
2575: For the binary search tree model,
2576: there is an integral transform $\mathcal L$ that relates the
2577: ordinary generating function of tolls, $t(z)$, and the ordinary generating
2578: function of the induced costs, $f(z)$: it is given by~(\ref{bst3})
2579: according to which $f(z)=\mathcal L[t(z)]$, where (with~$t_0=f_0=0$)
2580: \begin{equation}\label{Lbst}
2581: \mathcal L[t(z)]=\frac{1}{(1-z)^2}\int_0^z t'(w)(1-w)^2\, dw.
2582: \end{equation}
2583: Consequently, the computation is entirely \emph{mechanical}\footnote{%
2584: In the {\sc Maple} system for symbolic computations
2585: about two dozen instructions suffice to implement calculations,
2586: once use is made of Bruno Salvy's package {\sf equivalent}
2587: dedicated to the asymptotic analysis of coefficients of generating
2588: functions~\cite{Salvy91}.
2589: It suffices to program the polylogarithm expansions
2590: (Theorem~\ref{thm:li-sing}),
2591: use the system capabilities for series expansions, differentiation, and
2592: integration (Theorems~\ref{thm:2A} and~\ref{thm:2B}), and conclude by
2593: an appeal to Salvy's program that implements the basic transfers of
2594: Theorem~\ref{thm:basic-sa}.
2595: } and
2596: it only needs the theorems relating to integration, differentiation,
2597: and polylogarithms (Theorems~\ref{thm:2A} and~\ref{thm:2B}) in
2598: conjunction with basic
2599: singularity analysis (Theorem~\ref{thm:basic-sa}).
2600: Our derivation below constitutes an alternative to
2601: parallel results by Neininger~\cite{Neininger02},
2602: Chern~\emph{et al.}~\cite{ChHw01,ChHwTs02}, and Fill and
2603: Kapur~\cite{FK-transfer}, who employ elementary but
2604: perhaps less transparent
2605: methods (typically, the approximation of discrete sums by
2606: integrals).
2607: % Furthermore, the easy access to higher order
2608: % asymptotic expansions proves rewarding in the study of limit
2609: % distributions discussed in the next section.
2610:
2611:
2612:
2613: \begin{theorem} Under the binary search tree model,
2614: the expected values of the costs
2615: induced by tolls of type $t_n^\alpha$
2616: $(\alpha > 0$) and $t_n^{\log}$ admit full asymptotic expansions
2617: in descending powers of~$n$ and integral powers of~$\log n$.
2618: The main terms are summarized by the following table:
2619: \[
2620: \begin{array}{cl|cl}
2621: \hline
2622: \hbox{\em Toll $(t_n)$} & & \hbox{\em Cost $(f_n)$} & \\
2623: \hline\hline
2624: n^\alpha & (2<\alpha) & \ds \frac{\alpha+1}{\alpha-1}n^\alpha &
2625: {}+O(n^{\alpha-1}) \\[.15in]
2626: n^2 & & \ds 3n^2 &{}-6n\log n+(10-6\gamma )n+O(\log n) \\[.1in]
2627: n^\alpha & (1<\alpha<2) & \ds \frac{\alpha+1}{\alpha-1}n^\alpha &
2628: {}+K_\alpha n+O(n^{\alpha-1}) \\
2629: n & & 2n\log n &\ds {}+2(\gamma-1)n+2\log n+2\gamma+1+O\left(\frac1n
2630: \right) \\
2631: n^\alpha & (0<\alpha<1) & K_\alpha n &
2632: \ds {}+\frac{\alpha+1}{\alpha-1}n^\alpha+K_\alpha+o(1)\\
2633: \log n & & K'_0 n & \ds {}-\log
2634: n+(K'_0-2)-\frac{1}{2n}+\frac{1}{9n^2} +O\left(\frac{1}{n^3}\right).\\
2635: % -\frac{1}{72n^3}+O(\frac{1}{n^4}) \\
2636: % last term was first inferred numerically as 1/(72*n^3)
2637: % then checked by ``equivalent''.
2638: \hline
2639: \end{array}
2640: \]
2641: \end{theorem}
2642: \begin{proof} For the case $\alpha$ a nonnegative integer,
2643: the integration can be carried out in finite terms since the
2644: generating function of tolls is rational. For instance, the case
2645: $\alpha=1$ corresponds to the well-known analysis of
2646: Quicksort and binary search tree
2647: algorithms~\cite{Knuth98a,Mahmoud92,SeFl96,ViFl90}.
2648:
2649: \def\leadsto{\ \longrightarrow\ } For $t_n^\alpha$, it suffices to
2650: examine the effect of the $\mathcal L$ transform on singular
2651: elements of the form $c(1-z)^\beta$; e.g., for the main term
2652: corresponding to $t_n=n^\alpha$, we should take $\beta=-\alpha-1$.
2653: The $\mathcal L$ transformation reads as a succession of operations,
2654: ``differentiate, multiply by~$(1-z)^2$, integrate, multiply
2655: by~$(1-z)^{-2}$''---all are covered by our previous theorems. The
2656: chain on any particular singular element starts as
2657: \[
2658: c(1-z)^\beta \mathop{\leadsto}^{\partial} c\beta(1-z)^{\beta-1}
2659: \mathop{\leadsto}^{\times(1-z)^2} c\beta(1-z)^{\beta+1}.
2660: \]
2661: At this stage, integration intervenes. Assume that $\beta+1\not=-1$.
2662: (Otherwise, a logarithm appears.) According to Theorem~\ref{thm:2B},
2663: and ignoring integration constants for the moment, integration gives
2664: \[
2665: c\beta(1-z)^{\beta+1}\mathop{\leadsto}^{\int}
2666: -c\frac{\beta}{\beta+2}(1-z)^{\beta+2}
2667: \mathop{\leadsto}^{\times(1-z)^{-2}}
2668: -c\frac{\beta}{\beta+2}(1-z)^{\beta} .
2669: \]
2670: Then this singular element corresponds to a contribution
2671: \[
2672: -c\frac{\beta}{\beta+2}\binom{n-\beta-1}{-\beta-1},
2673: \]
2674: which is of order $O(n^{-\beta-1})$. (The treatment of logarithmic
2675: terms is entirely similar.)
2676:
2677: The derivation above
2678: % suffices to justify all the cases listed in the
2679: % statement, but it
2680: has left aside the determination of the integration constants. These
2681: are given by the second case of Theorem~\ref{thm:2B}, which provides
2682: in particular access to the constants~$K_\alpha$ and $K'_0$. The
2683: constant term in the asymptotic expansion of the integral is of the
2684: form
2685: \[
2686: {\bf K}[t]:=\int_0^1
2687: \left[t'(w)(1-w)^2-\left(t'(w)(1-w)^2\right)_-\right]\,dw,
2688: \]
2689: where $f_-$ represents the sum of the singular terms in~$f$ having
2690: exponent $<-1$, as in the proof of Theorem~\ref{thm:2B}. In the
2691: singular expansion of~$f(z)$, this integration constant gets further
2692: multiplied by~$(1-z)^{-2}$; the resulting linear term in the
2693: asymptotic expansion of $f_n$ is then plainly
2694: \[
2695: {\bf K}[t]\cdot (n+1).\] In particular, if the growth of~$t_n$ is
2696: smaller than $n$, then, the divergence part is absent and ${\bf K}[t]$
2697: reduces to
2698: \[
2699: {\bf K}[t]=\int_0^1 t'(w)(1-w)^2\,dw=2\sum_{n=1}^\infty
2700: \frac{t_n}{(n+1)(n+2)},
2701: \]
2702: as follows from expanding the integrand around~0 and integrating the
2703: resulting series. This yields the following values for $\alpha<1$:
2704: \begin{equation}\label{eq:theKs}
2705: K_\alpha=2\sum_{n=1}^\infty \frac{n^\alpha}{(n+1)(n+2)},
2706: \qquad
2707: K'_0=2\sum_{n=1}^\infty \frac{\log n}{(n+1)(n+2)},
2708: \end{equation}
2709: while for~$1<\alpha<2$,
2710: \[
2711: K_\alpha=2\sum_{n=1}^\infty
2712: \frac{n^\alpha-\Gamma(\alpha+1)\binom{n+\alpha}{\alpha}}{(n+1)(n+2)}.
2713: \]
2714: The theorem is finally established.
2715: \end{proof}
2716: \begin{remark*}
2717: The slowly convergent series expressions
2718: of $K_\alpha,K'_0$ can be rephrased as definite
2719: integrals, thanks to Mellin transform techniques. The starting point
2720: is the easy formal identity,
2721: \begin{equation}\label{mellin0}
2722: \sum_{n\ge1} c_n n^{-s} = \frac{1}{\Gamma(s)}
2723: \int_0^\infty \left(\sum_{n\ge1} c_n e^{-nx}\right) x^{s-1}\, dx.
2724: \end{equation}
2725: The constant~$K_\alpha$ with~$\alpha<1$ corresponds to~$s=1-\alpha$
2726: and
2727: $c_n=n/[(n+1)(n+2)]$, for which
2728: the integrand admits of closed form since
2729: \[
2730: \sum_{n=1}^\infty
2731: \frac{n z^n}{(n+1)(n+2)}= \frac{1}{z^2}\left[(2-z)L(z)
2732: -2z \right].
2733: \]
2734: From there, the constant $K'_0$ is attained as
2735: $\left.\frac{d}{d\alpha}K_\alpha\right|_{\alpha=0}$.
2736: A final change of variables $x=-\log t$ then yields
2737: an integral representation for
2738: ``Fill's first logarithmic constant'' ($\gamma$ is Euler's constant):
2739: \begin{eqnarray}\label{fillcbst}
2740: K'_0&=&-\gamma-2\int_0^1\left[(t-2)\log(1-t)-2t\right]
2741: \left(\log\log\frac1t\right) \, \frac{dt}{t^3} \\
2742: &=&
2743: 1.20356491674961033428628333814873131775552838577096
2744: . \nonumber
2745: \end{eqnarray}
2746: The last estimate to 50D improves on the earlier 3D evaluation
2747: of~Fill \cite{Fill96}. The cost induced by~$t^{\log}$ is of
2748: particular interest as
2749: it is precisely the entropy of the distribution of binary search
2750: trees; see the account and first estimates in the book by Cover and
2751: Thomas~\cite[p.~72--74]{CoTh91}, as well as pointers to
2752: self-organizing search in Fill's article~\cite{Fill96}.
2753: In his doctoral dissertation~\cite[Section~5.1]{kapur03:_addit},
2754: Kapur has extended the methods and estimates
2755: to~$m$-ary search trees.
2756: \end{remark*}
2757:
2758:
2759: % the shape functional for $m$-ary
2760: % search trees~\cite[Section~5.1]{kapur03:_addit}.
2761:
2762:
2763: % \underline{The case $t_n = n^\alpha$.}
2764: % Consider first the case $t_n = n^\alpha$ with $\alpha > 0$.
2765: % Recalling~\eqref{bst3}, we first determine the singular expansion of
2766: % $t'$ using Theorem~\ref{thm:2A} and~\eqref{eq:1}. This leads to
2767: % \begin{equation}
2768: % \label{eq:2}
2769: % t'(z) (1-z)^2 = \Gamma(1+\alpha) (1+\alpha) (1-z)^{-\alpha} +
2770: % O(|1-z|^{-\alpha + 1}).
2771: % \end{equation}
2772: %
2773: % If $\alpha < 1$ then $t'(z) (1-z)^2$ is integrable, so
2774: % by~\eqref{bst3} and Theorem~\ref{thm:2B} we have
2775: % \begin{equation*}
2776: % f(z) = K (1-z)^{-2} + O(|1-z|^{-\alpha-1}),
2777: % \end{equation*}
2778: % where
2779: % \begin{equation}\label{eq:4}
2780: % K_\alpha := \int_{0}^1 \left( \partial_w\Li_{-\alpha,0}(w) \right)
2781: % (1-w)^2 \, dw = \int_0^1 \Li_{-(\alpha+1),0}(w) (1-w)^2 \, \frac{dw}w.
2782: % \end{equation}
2783: % By singularity analysis
2784: % \begin{equation*}
2785: % f_n = K_\alpha n + O(n^\alpha).
2786: % \end{equation*}
2787: %
2788: % If $\alpha=1$ then $t(z) = z(1-z)^{-2}$. Then routine calculations yield
2789: % \begin{equation*}
2790: % f(z) = 2(1-z)^{-2} L(z) - z (1-z)^{-2},
2791: % \end{equation*}
2792: % which leads to
2793: % \begin{equation*}
2794: % f_n = 2n \log{n} + O(n).
2795: % \end{equation*}
2796: %
2797: % Finally, if $\alpha > 1$ then applying Theorem~\ref{thm:2B}
2798: % to~\eqref{eq:2} gives
2799: % \begin{equation*}
2800: % f(z) = \frac{\alpha + 1}{\alpha - 1} \Gamma(\alpha + 1)
2801: % (1-z)^{-\alpha - 1} + R(z),
2802: % \end{equation*}
2803: % where
2804: % \begin{equation*}
2805: % R(z) =
2806: % \begin{cases}
2807: % O(|1-z|^{-2}) & 1 < \alpha < 2 \\
2808: % O(|1-z|^{-2}|L(z)|) & \alpha = 2 \\
2809: % O(|1-z|^{-\alpha}) & \alpha > 2.
2810: % \end{cases}
2811: % \end{equation*}
2812: % By singularity analysis then
2813: % \begin{equation*}
2814: % f_n = \frac{\alpha + 1}{\alpha - 1} n^\alpha + r(n),
2815: % \end{equation*}
2816: % where
2817: % \begin{equation*}
2818: % r(n) =
2819: % \begin{cases}
2820: % O(n) & 1 < \alpha < 2 \\
2821: % O(n \log{n} ) & \alpha = 2 \\
2822: % O(n^{\alpha-1}) & \alpha > 2.
2823: % \end{cases}
2824: % \end{equation*}
2825: %
2826: % \medskip\noindent
2827: % \underline{The case $t_n=\log{n}$}
2828: % \medskip
2829: %
2830: % Next consider the case $t_n = \log{n}$. Using~\eqref{bst3},
2831: % \eqref{eq:3}, and Theorem~\ref{thm:2A},
2832: % \begin{equation*}
2833: % t'(z) (1-z)^2 = L(z) + O(1).
2834: % \end{equation*}
2835: % Now by Theorem~\ref{thm:2B},
2836: % \begin{equation*}
2837: % f(z) = K (1-z)^{-2} + O(|1-z|^{-1} | L(z) |),
2838: % \end{equation*}
2839: % where
2840: % \begin{equation}\label{eq:5}
2841: % K := \int_0^1 \left( \partial_w \Li_{0,1}(w) \right ) (1-w)^2 \,
2842: % dw = \int_0^1 \Li_{-1,1}(w) (1-w)^2 \, \frac{dw}{w}.
2843: % \end{equation}
2844: % By singularity analysis,
2845: % \begin{equation*}
2846: % f_n = K n + O(\log{n}).
2847: % \end{equation*}
2848: %
2849: % The constants $K_\alpha$ and $K$ defined at~\eqref{eq:4}
2850: % and~\eqref{eq:5}, respectively, can be expressed as infinite series:
2851: % \begin{equation*}
2852: % \int_0^1 t'(w) (1-w)^2 \, dw = \int_0^1 \left( \sum_{n=1}^\infty n t_n
2853: % w^{n-1} \right) (1-w)^2 \, dw = 2 \sum_{n=1}^\infty \frac{t_n}{(n+1)(n+2)}.
2854: % \end{equation*}
2855: %
2856: % We note that by using a complete asymptotic expansion of the
2857: % generalized polylogarithm, rather than just the lead order and
2858: % remainder as in~\eqref{eq:1} and~\eqref{eq:3}, it is possible to
2859: % obtain a complete asymptotic expansion for the corresponding expected
2860: % cost. See Chapter~5 of~\cite{kapur03:_addit} for details.
2861: % % \marginal{\raggedright Reminder: We need to add literature references
2862: % % (e.g., \cite{MR1871558}, \cite{Fill96}, \cite{Neininger02})}
2863: %
2864:
2865:
2866:
2867: \subsection{The uniform binary tree recurrence}\label{sec:binary-tree-recurr}
2868:
2869: This section examines the uniform binary tree model
2870: that surfaces recurrently in combinatorics.
2871: Here, we put on a firm basis a classification of the expected costs
2872: corresponding the tolls $t_n^\alpha$ and $t_n^{\log}$
2873: which was outlined (with several typographical errors) in
2874: an article by Flajolet and Steyaert~\cite{FlSt87}.
2875: The particular case of the toll~$t_n=n$ has, like for binary search
2876: trees, a dignified history as it corresponds to path length
2877: in Catalan trees and to area under Dyck paths, whose first distributional
2878: analyses go back to Louchard and Tak\'acs~\cite{Louchard84,Takacs91}.
2879:
2880: Our starting point is~(\ref{cat99}) according to which the
2881: generating function of costs~$f(z)=\sum C_nf_nz^n$ normalized by the
2882: Catalan numbers~$C_n$ and the ordinary generating function of costs
2883: $\tau(z)=\sum t_n z^n$ are related by $f(z)=
2884: \overline{\cal L}[\tau (z)]$, where
2885: \begin{equation}\label{Lcat}
2886: \overline{\cal L}[\tau(z)]=\frac{1}{\sqrt{1-4z}}\left(\tau(z)\odot C(z)\right),
2887: \end{equation}
2888: with
2889: \[
2890: C(z)=\sum_{n\ge0} C_n z^n =\frac{1}{2z}\left(1-\sqrt{1-4z}\right),
2891: \qquad
2892: C_n=\frac{1}{n+1}\binom{2n}{n}.
2893: \]
2894: We state:
2895: \def\K{\overline{K}}
2896: \begin{theorem} Under the uniform binary tree model,
2897: the expected values of the costs
2898: induced by tolls of type $t_n^\alpha$
2899: $(\alpha > 0$) and $t_n^{\log}$ admit full asymptotic expansions
2900: in descending powers of~$n$ and integral powers of~$\log n$.
2901: The main terms are summarized by the following table:
2902: \[
2903: \begin{array}{cl|cl}
2904: \hline
2905: \hbox{\em Toll $(t_n)$} & & \hbox{\em Cost $(f_n)$} & \\
2906: \hline\hline
2907: n^\alpha & (\frac32 < \alpha) & \ds
2908: \frac{\Gamma(\alpha-\frac12)}{\Gamma(\alpha)}n^{\alpha+\frac12}&
2909: {}+O(n^{\alpha-\frac12}) \\[.15in]
2910: n^{3/2} & &
2911: \ds \frac{1}{\Gamma(3/2)}n^2 &{}+O(n\log n)
2912: \\[.15in]
2913: n^\alpha & (\frac12<\alpha<\frac32) & \ds
2914: \frac{\Gamma(\alpha-\frac12)}{\Gamma(\alpha)}n^{\alpha+ \frac12}&
2915: \ds {}+O(n) \\[.15in]
2916: n^{1/2} & & \ds \frac{1}{\sqrt{\pi}} n\log n &\ds
2917: {}+O(n) \\[.15in]
2918: n^\alpha & (0<\alpha<\frac12) & \K_\alpha n &
2919: \ds {}+O(1)
2920: \\[.15in]
2921: \log n & & \K'_0 n & \ds {}-2\sqrt{\pi}n^{1/2}+O(1).
2922: \\
2923: \hline
2924: \end{array}
2925: \]
2926: \end{theorem}
2927: \begin{proof}
2928: For the tolls $t_n^\alpha$, all that is required is to determine the
2929: singular expansion of
2930: \[
2931: \tau(z)\odot C(\frac{z}{4}) = \sum_{n=1}^\infty \frac{n^\alpha}{(n+1)}
2932: \binom{2n}{n}\left(\frac{z}{4}\right)^n.
2933: \]
2934: (For convenience, the singularity has been scaled to~1.)
2935: We use the Zigzag Algorithm
2936: presented in Subsection~\ref{subsec:had-all}.
2937: The known asymptotic expansion of the Catalan numbers is
2938: \[
2939: 4^{-n}C_n\sim\frac{1}{\sqrt{\pi}} {n^{-3/2}}
2940: \left(1-\frac{9}{8n}+\frac{145}{128n^2}-\cdots\right).
2941: \]
2942: Multiply this by $n^\alpha$ to get the expansion of $n^\alpha C_n4^{-n}$.
2943: The terms now involve the scale
2944: $\{n^{\alpha-\frac32}, n^{\alpha-\frac52},\ldots\}$.
2945: Assume that $\alpha$ is not a half-integer [i.e., \mbox{$\alpha \not\in
2946: (\frac12\Z) \setminus \Z$}]; see below
2947: for the contrary case. Then
2948: the basis of functions $\mathcal{B}=\{(1-z)^{-\alpha+k+ \frac12}\}$,
2949: where~$k$ ranges over the integers, has the property that
2950: the coefficients of its generic element
2951: are~$O(n^{\alpha-k-\frac32})$; in particular,
2952: \[
2953: [z^n](1-z)^{-\alpha+
2954: \frac12}\sim\frac{n^{\alpha-\frac32}}{\Gamma(\alpha-\frac12)}
2955: \left(1+\frac{(2\alpha-1)(2\alpha-3)}{8n}+\cdots\right).
2956: \]
2957: We can thus find a singular function $H(z)$ whose coefficients
2958: match asymptotically those of $\tau(z) \odot C(z/4)$, which is of the
2959: form
2960: \[
2961: H(z)=\frac{\Gamma(\alpha- \frac12)}{\sqrt{\pi}}
2962: (1-z)^{-\alpha+ \frac12}\left(1+c_1(1-z)+c_2(1-z)^2+\cdots\right),
2963: \]
2964: for some effectively computable sequence~$(c_j)$.
2965: The singular expansion of $\tau(z)\odot C(z/4)$ is
2966: then the sum of the expansion of~$H$ above and of a
2967: power series in~$(1-z)$, call it $P(z)$, that can be determined
2968: according to the principles of Section~\ref{sec:hadam-prod-transf}.
2969:
2970: % for instance,
2971: % \[
2972: % c_1=-(\alpha^2-2\alpha+3)/(2\alpha-1)(2\alpha-3).
2973: % \]
2974: % This is basically OKay
2975: % but it leaves aside the ``analytic part'' of the singular expansion
2976: % that doesn't come out of the process. One has to be careful..
2977:
2978: The singular expansion of $f(z/4)$ is that of
2979: $H(z)+P(z)$ divided by~$\sqrt{1-z}$, so that, by transfer, we get
2980: \[
2981: [z^n]f\left(\frac{z}{4}\right)\sim
2982: \frac{\Gamma(\alpha-\frac12)}{\sqrt{\pi}\Gamma(\alpha)}
2983: n^{\alpha-1}\left(1+\frac{c'_1}{n}+\frac{c'_2}{n^2}+\cdots\right)
2984: +[z^n]\frac{P(z)}{\sqrt{1-z}},
2985: \]
2986: for some sequence~$(c'_j)$, where the ``hidden'' analytic part
2987: $P(z)$ arises from the ``hidden'' analytic
2988: component in $\tau(z)\odot C(z/4)$.
2989: After dividing by~$C_n4^{-n}$, one finds finally:
2990: \begin{equation}\label{zoba}
2991: f_n \sim \frac{\Gamma(\alpha-\frac12)}{\Gamma(\alpha)}
2992: n^{\alpha+\frac12}\left(1+\frac{c''_1}{n}+\cdots\right)+R_n
2993: ,
2994: % \qquad
2995: % c''_1={\frac {4\,{\alpha}^{2}-10\,\alpha-3}{2\,\alpha-3}},
2996: \end{equation}
2997: where the ``hidden'' remainder term $R_n$ is of the form
2998: \[
2999: R_n\sim d_{-1}n+d_0+\frac{d_1}{n}+\cdots\,.
3000: \]
3001: This last estimate provides all the entries in the table above,
3002: whenever~$\alpha$ is not a half-integer, as it
3003: suffices to merge the two expansions of~(\ref{zoba}). In addition, when
3004: $0<\alpha<\frac12$, the series defining $t(z/4)$ converges at the
3005: singularity~1. Thus, the dominant asymptotic term of $f(z/4)$ is
3006: $t(1/4)/\sqrt{1-z}$, that is,
3007: \[
3008: f\left(\frac z4\right)\sim \frac{\K_\alpha}{\sqrt{1-z}},\qquad
3009: \K_\alpha:=\sum_{n=1}^\infty \frac{n^\alpha}{n+1}
3010: \frac{1}{4^n}\binom{2n}{n}.
3011: \]
3012:
3013: When $\alpha$ is a half-integer, logarithmic terms appear due to the
3014: presence of inverse integral powers of~$n$ in the coefficients
3015: of $t(z/4)$, but the derivation is otherwise similar.
3016: For instance at $\alpha=\frac12$, one has
3017: \[
3018: 4^{-n}\sqrt{n}C_n\sim \frac{1}{\sqrt{\pi}}\,\frac{1}{n}
3019: +O\left(\frac{1}{n^2}\right),
3020: \]
3021: which shows that
3022: \[
3023: t(z/4)=H(z)+P_0+O((1-z)^{1-\epsilon}),
3024: \qquad
3025: H(z)=\frac{1}{\sqrt{\pi}}L(z),\]
3026: resulting in the stated estimate.
3027:
3028:
3029: %
3030: % \underline{The case $t_n = n^\alpha$}
3031: % \medskip
3032: %
3033: % We first consider the toll $t_n = n^\alpha$ with $\alpha > 0$. Our
3034: % starting point is~\eqref{cat99} and the singular expansion for
3035: % $\tau(z)$ given at~\eqref{eq:1}. Recall that
3036: % \begin{equation}
3037: % \label{eq:12}
3038: % C(z) = \frac{1}{2z}(1- \sqrt{1-4z}) = 2 - 2(1-4z)^{1/2} + O(|1-4z|)
3039: % \end{equation}
3040: % as $z \to 1/4$ in $\mathbb{C} \setminus [1/4,+\infty)$. Also,
3041: % \[
3042: % \tau(z) \odot C(z/4) = \sum_n t_n \frac{C_n}{4^n} z^n = \sum_n t_n C_n
3043: % \left(\frac{z}{4}\right)^n = (\tau \odot C)(z/4)
3044: % \]
3045: % and $\tau(z)\odot C(z/4) = \tau(z)\odot(C(z/4)-2)$.
3046: %
3047: % When $\alpha<1/2$, by Proposition~\ref{thm:4}(a) we have
3048: % \[
3049: % (\tau \odot C)(z/4) = K + O(|1-z|^{\tfrac12 - \alpha}),
3050: % \]
3051: % so that
3052: % \[
3053: % f(z) = K(1-4z)^{-1/2} + O(|1-4z|^{-\alpha}),
3054: % \]
3055: % where
3056: % \[
3057: % K := \sum_{n=1}^\infty \frac{t_nC_n}{4^n}.
3058: % \]
3059: % It follows from singularity analysis and the estimate
3060: % \[
3061: % C_n = \frac{4^n}{\sqrt{\pi}n^{3/2}}\left(1 + O(\tfrac1n)\right)
3062: % \]
3063: % that
3064: % \[
3065: % f_n = K(n+1) + O(n^{\alpha+\tfrac12}).
3066: % \]
3067: %
3068: % If $\alpha \geq 1/2$ then we make use of the estimate
3069: % \begin{equation}
3070: % \label{eq:6}
3071: % (1-z)^{1/2} = \frac1{\Gamma(-1/2)}[ \Li_{3/2,0}(z) - \zeta(3/2) ]
3072: % + O(|1-z|),
3073: % \end{equation}
3074: % a consequence of Theorem~1 of~\cite{Flajolet99}, so that
3075: % \[
3076: % \tau(z) \odot (1-z)^{1/2} = \Li_{-\alpha,0}(z) \odot (1-z)^{1/2} =
3077: % \frac{1}{\Gamma(-1/2)} \Li_{\tfrac32-\alpha,0}(z) + R_1(z),
3078: % \]
3079: % where
3080: % \begin{equation}\label{eq:11}
3081: % R_1(z) =
3082: % \begin{cases}
3083: % c + O(|1-z|^{1-\alpha}) & 1/2 \leq \alpha < 1 \\
3084: % O(| L(z) | ) & \alpha = 1 \\
3085: % O(|1-z|^{1-\alpha}) & \alpha > 1.
3086: % \end{cases}
3087: % \end{equation}
3088: % (Recall our notational convention discussed just prior to Section~\ref{sec:binary-search-tree}.) Hence
3089: % \begin{equation}
3090: % \label{eq:9}
3091: % (\tau \odot C)(z/4) =
3092: % -\frac2{\Gamma(-1/2)}\Li_{\tfrac32-\alpha,0}(z) +
3093: % R_2(z),
3094: % \end{equation}
3095: % where $R_2$, like $R_1$, satisfies the
3096: % right side of~\eqref{eq:11} (with a possibly different $c$). When
3097: % $\alpha=1/2$, this gives us
3098: % \[
3099: % (\tau \odot C)(z/4) = -\frac2{\Gamma(-1/2)}L(z) + O(1),
3100: % \]
3101: % and again, using singularity analysis and our estimate for $C_n$,
3102: % we conclude that
3103: % \[
3104: % f_n = \frac{1}{\sqrt\pi} n\log{n} + O(n).
3105: % \]
3106: % When $\alpha>1/2$, we can use the singular expansion for
3107: % $\Li_{3/2-\alpha,0}(z)$ to conclude
3108: % \begin{equation}
3109: % \label{eq:13}
3110: % (\tau \odot C)(z/4) = \frac1{\sqrt\pi}
3111: % \Gamma(\alpha-\tfrac12)(1-z)^{-\alpha+\tfrac12} + R_3(z),
3112: % \end{equation}
3113: % where $R_3$ also satisfies the estimate~\eqref{eq:11} (with a possibly
3114: % different $c$). Thus
3115: % \[
3116: % f_n = \frac{\Gamma(\alpha-\tfrac12)}{\Gamma(\alpha)}
3117: % n^{\alpha+\tfrac12} + r(n),
3118: % \]
3119: % where
3120: % \begin{equation*}
3121: % r(n) =
3122: % \begin{cases}
3123: % O(n) & 1/2 < \alpha < 1 \\
3124: % O(n \log{n} ) & \alpha = 1 \\
3125: % O(n^\alpha) & \alpha > 1.
3126: % \end{cases}
3127: % \end{equation*}
3128: %
3129: %
3130: % \medskip\noindent
3131: % \underline{The case $t_n=\log{n}$}
3132: % \medskip
3133: %
3134:
3135: Finally, when $t_n = \log{n}$, we have $\tau(z) = \Li_{0,1}(z) =
3136: O(|1-z|^{-1-\epsilon})$ for any $\epsilon > 0$. % by~\eqref{eq:3}.
3137: Thus, by Proposition~\ref{thm:4}(i),
3138: \begin{equation*}
3139: (\tau \odot C)(z/4) = \K'_0+ O\Bigl(|1-z|^{\tfrac12 - \epsilon}\Bigr),
3140: % \end{equation*}
3141: % where
3142: % \begin{equation*}
3143: \qquad
3144: \K'_0 := \sum_{n=1}^\infty (\log n) \frac{C_n}{4^n}.
3145: \end{equation*}
3146: Singularity analysis and the estimate for $C_n$
3147: yield $f_n = \K'_0 n + O\Bigl(n^{\tfrac12 + \epsilon}\Bigr)$.
3148: Carrying higher-order terms, we get the
3149: mean of the shape functional,
3150: \begin{equation}
3151: \label{eq:24}
3152: \mu_n = \K'_0 n - 2\sqrt\pi n^{1/2} + O(1),
3153: \end{equation}
3154: which agrees with the estimate in Theorem~3.1 of~\cite{Fill96}
3155: and improves the remainder estimate.
3156: \end{proof}
3157:
3158: The Mellin technique of~(\ref{mellin0}) is once more applicable
3159: to the determination of ``Fill's
3160: second logarithmic constant''~$\K'_0$. It provides
3161: the value:
3162: \[\begin{array}{ccc}
3163: \K'_0&:=&\ds \sum_{k\ge1} \frac{\log k}{(k+1){4^k}}\binom{2k}{k}.
3164: \\
3165: &=& \ds -\gamma-\int_0^1 \frac{1}{\sqrt{1-t}(1+\sqrt{1-t})^2}\,
3166: \left(\log\log\frac1t\right)\, dt,\\
3167: \\
3168: &=& 2.0254384677765738877135187391417652470652930617658.
3169: \end{array}
3170: \]
3171: The subject of costs on binary trees is considered in greater depth
3172: in~\cite{fill03:_limit_catal} by
3173: applying the techniques developed in this paper. There, some higher-order
3174: estimates,
3175: asymptotics for moments of each order, and limiting distributions are
3176: derived when the toll sequence is either $n^\alpha$ or
3177: $\log{n}$.
3178:
3179: Our methods can also be used to treat more generally the case of
3180: all simple families of trees in the sense of Meir and
3181: Moon~\cite{MeMo78}, of which Catalan trees are a
3182: special case. This generalization is the subject of ongoing
3183: work.
3184:
3185: \subsection{The union--find tree recurrence}
3186: \label{sec:cayl-tree-recurr}
3187:
3188: \def\K{\widehat{K}}
3189:
3190: In this subsection, we revisit the Knuth--Pittel--Sch{\"o}nhage
3191: recurrence corresponding to the destruction of free labelled trees and
3192: dually to the management of equivalence
3193: relations~\cite{KnPi89,KnSc78}. The main result of this section is
3194: essentially a rephrasing of the main results of Knuth and Pittel
3195: in~\cite{KnPi89}, to which we add the possibility of determining
3196: complete asymptotic expansions. Like before, the starting point is
3197: the integral transform~(\ref{cay3}) (adjusted for the fact that $t_1^\alpha =
3198: f_1^\alpha = 1 \ne 0$), which relates the ordinary generating function of
3199: tolls~$\tau(z)$ to the normalized generating function of costs~$f(z)$
3200: via $f(z)=\cal L[\tau(z)]$, where
3201: \begin{equation}\label{Lcay}
3202: \cal L [\tau(z)] = t_1 z T'(z) +
3203: \frac12\frac{T(z)}{1-T(z)}\int_0^z \partial_w
3204: \left(\tau(w)\odot
3205: T^2(w)\right) \, \frac{dw}{T(w)}.
3206: \end{equation}
3207: There~$T(z)$ is the Cayley tree function whose
3208: singular expansion at the (unique) dominant
3209: singularity $z=e^{-1}$ is well known: one has the shape
3210: \begin{equation}\label{eq:10}
3211: T(z) \sim 1 - \sqrt{2}(1-ez)^{1/2} + c_1(1-ez) + \cdots
3212: \end{equation}
3213: as $z \to e^{-1}$ in any sector of angle $<2\pi$; see
3214: also~\cite[Eq.~(3.16)]{KnPi89}. (The paper by Corless \emph{et
3215: al.}~\cite{CoGoHaJeKn96} is a definitive reference regarding the
3216: tree function.) As noted earlier, the case of union--find tree
3217: recurrences
3218: combines all the composition results developed in this
3219: paper.
3220: % in the sector $-\pi + \epsilon < \arg{(1-ez)} < \pi -
3221: % \epsilon$ (for any $\epsilon > 0$).
3222: \begin{theorem} Under the union--find tree recurrence model,
3223: the expected values of the costs
3224: induced by tolls of type $t_n^\alpha$
3225: $(\alpha > 0$) and $t_n^{\log}$ admit full asymptotic expansions
3226: in descending powers of~$n$ and integral powers of~$\log n$.
3227: The main terms are summarized by the following table:
3228: \[
3229: \begin{array}{cl|cl}
3230: \hline
3231: \hbox{\em Toll $(t_n)$} & & \hbox{\em Cost $(f_n)$} & \\
3232: \hline\hline
3233: n^\alpha & (\frac32 < \alpha) & \ds
3234: \frac{\Gamma(\alpha-\frac12)}{\sqrt2\Gamma(\alpha)}n^{\alpha+\frac12}&
3235: {}+O(n^{\alpha-\frac12}) \\[.15in]
3236: n^{3/2} & &
3237: \ds \frac{1}{\sqrt2 \Gamma(3/2)}n^2 &{}+O(n\log n)
3238: \\[.15in]
3239: n^\alpha & (\frac12<\alpha<\frac32) & \ds
3240: \frac{\Gamma(\alpha-\frac12)}{\sqrt 2\Gamma(\alpha)}n^{\alpha+\frac12}&
3241: \ds {}+O(n) \\[.15in]
3242: n^{1/2} & & \ds \frac{1}{\sqrt{2\pi}} n\log n &\ds
3243: {}+O(n) \\[.15in]
3244: n^\alpha & (0<\alpha<\frac12) & (1 + \frac12\K_\alpha) n &
3245: \ds {}+O(n^{\alpha + \frac12})
3246: \\[.15in]
3247: \log n & & \frac12\K'_0 n & \ds {}+O(\sqrt{n}).
3248: \\
3249: \hline
3250: \end{array}
3251: \]
3252: \end{theorem}
3253:
3254: \begin{proof}
3255: \def\A{{\mathcal{A}}}\def\B{{\mathcal{N}}} %we've used \cal B earlier
3256: % with other meanings, so I changed from \def\B{{\mathcal{B}}}
3257: We shall content ourselves with indicating the way
3258: full asymptotic expansions can be determined within the generating
3259: function framework. (Detailed computations are left as an exercise for
3260: the reader.)
3261: In what follows, we set $Z=(1-z)$ and let~$\A$
3262: denote an unspecified entire series in powers of~$Z$,
3263: not necessarily the same at each
3264: occurrence.
3265: For instance, one may summarize
3266: diversely the expansion~(\ref{eq:10}) of~$T(z/e)$
3267: as
3268: \[
3269: T(z/e)\sim 1-\sqrt{2}Z^{1/2}+Z\A+Z^{3/2}A\sim \A+\A Z^{1/2},
3270: \]
3271: and so on. We shall also let $\B$ denote
3272: generically a series in descending powers
3273: of~$1/n$.
3274:
3275: We consider first the case of the toll~$t_n^\alpha$
3276: and assume for simplicity that $\alpha$ is not a half-integer:
3277: $\alpha\not\in(\frac12\Z) \setminus \Z$.
3278: The polylogarithm expansions grant us
3279: {a~priori} that the generating
3280: function~$\tau(z)$ lies in the class of functions amenable to
3281: singularity analysis, with
3282: \[
3283: \tau(z)\sim Z^{-\alpha-1}\A+ \A.
3284: \]
3285: Therefore, the Hadamard product $(\tau(z)\odot
3286: T^2(z/e))$ is also amenable. The coefficients of the
3287: latter function are of the form
3288: $n^{\alpha-\frac32}\B$, as follows from the fact that
3289: $[z^n]\tau(z)=n^\alpha$ and $[z^n]T^2(z/e)\sim n^{-3/2}\B$
3290: (by the singular expansion of~$T^2$). Thus, converting back this
3291: information to the function, we find
3292: \[
3293: \tau(z)\odot T^2(z/e) \sim Z^{-\alpha+\frac12}\A+\A,
3294: \qquad
3295: \partial_z(\tau(z)\odot T^2(z/e))\sim Z^{-\alpha-\frac12}\A+\A.
3296: \]
3297: What we have done here is to apply the Zigzag Algorithm of
3298: Section~\ref{sec:hadam-prod-transf} and the differentiation theorem.
3299: Then multiplication by~$1/T(z/e) \sim \A+Z^{1/2}\A$ shows
3300: that
3301: \[
3302: \frac{1}{T(z/e)}\partial_z[\tau(z)\odot T^2(z/e)]
3303: \sim Z^{-\alpha-\frac12}\A+Z^{-\alpha}\A+\A+Z^{\frac12}\A.
3304: \]
3305: Integration of this last expansion corresponds to
3306: increasing all exponents by~1. Finally one should multiply by
3307: $T(z/e)(1-T(z/e))^{-1}$ which is of type
3308: \mbox{$Z^{-1/2}\A+\A$}. This completes our handling of the second term
3309: on the right in~(\ref{Lcay}). Also,
3310: \[
3311: \frac{z}e T'(z/e) \sim Z^{-\frac12}\A + \A.
3312: \]
3313: The end result is then
3314: \[
3315: f(z/e)\sim Z^{-\alpha}\A+Z^{-\alpha+\frac12}\A+Z^{-\frac12}\A+\A.
3316: \]
3317: The dominant term is $Z^{-\alpha}$ when~$\alpha>\frac12$ whereas it is
3318: $Z^{-1/2}$ when $\alpha<\frac12$.
3319:
3320: At the same time, it is a simple task to trace the
3321: coefficients of main terms. For $\alpha>\frac12$, the
3322: main term of~$f(z/e)$ is $Z^{-\alpha}$,
3323: and one finds successively
3324: \begin{equation*}\label{eq:cayley18}
3325: \begin{array}{lll}\ds \tau(z) \odot T^2(z/e)&\sim&
3326: \ds \sqrt\frac2\pi \Gamma(\alpha -
3327: \tfrac12) (1-z)^{-\alpha + \frac12},\\[.15in]
3328: f(z/e)&\sim& \ds\frac{\Gamma(\alpha - \frac12)}{2\sqrt\pi}
3329: (1-z)^{-\alpha},
3330: \end{array}
3331: \end{equation*}
3332: where the last equation implies, via singularity analysis,
3333: an estimate of expected costs:
3334: \begin{equation*}
3335: f_n^\alpha \sim \frac{\Gamma(\alpha - \frac12)}{\sqrt2\Gamma(\alpha)}
3336: n^{\alpha + \frac12}.
3337: \end{equation*}
3338:
3339: For $\alpha<\frac12$, the main term is $Z^{-1/2}$
3340: and its coefficient is seen to arise from both terms on
3341: the right in~(\ref{Lcay}):\ we have
3342: \begin{equation*}
3343: f^\alpha(z/e) \sim \frac{1}{\sqrt{2}}\left(1 +
3344: \frac{\K_\alpha}{2} \right) (1-z)^{-1/2} ,
3345: \end{equation*}
3346: where $\K_\alpha=\K[n^\alpha]$ and the functional~$\K$ is
3347: \begin{equation}
3348: \label{eq:16}
3349: \K[t] := \int_0^{1/e}
3350: \partial_w ( \tau(w) \odot T^2(w) )\, \frac{dw}{T(w)}.
3351: \end{equation}
3352: Error terms can be similarly traced:
3353: in the case of~$f^\alpha(z/e)$, it is
3354: of type~$Z^{-\alpha}$ if~$0<\alpha<\frac12$, of type~$Z^{-1/2}$
3355: if $\frac12<\alpha<1$, and so on.
3356: The end results are summarized in the statement of the theorem.
3357:
3358:
3359: For half-integer $\alpha$, a logarithmic term appears.
3360: For instance, in the case $\alpha=\frac12$, this fact is
3361: associated to the shape of the coefficients
3362: \[
3363: [z^n](\tau(z)\odot T^2(z/e))\sim
3364: \frac{-2\sqrt{2}}{\Gamma(-1/2)}\,\frac{1}{n}
3365: +\frac{1}{n^2}\B,
3366: \]
3367: resulting in a singular expansion with a logarithmic term:
3368: \[
3369: \tau(z)\odot T^2(z/e)=-\sqrt{\frac{2}{\pi}}\log Z+c+Z\A+(Z\log Z)\A,
3370: \]
3371: for some~$c$.
3372:
3373:
3374: For the logarithmic toll [note that now $t_1=0$, so that
3375: the first term in~(\ref{Lcay}) does not contribute] we
3376: have~$\tau(z)=\Li_{0,1}(z)$, the integral in~(\ref{Lcay}) is
3377: convergent and, in the same way as for the case $\alpha < 1/2$, we get
3378: \begin{equation*}
3379: f(z) = \frac{\K'_0}{2\sqrt2}
3380: (1-ez)^{-1/2} + O(|1-ez|^{-\epsilon}), \end{equation*}
3381: which implies
3382: \begin{equation*}
3383: f_n = \frac12\K'_0 n + O(n^{\frac12 + \epsilon}),
3384: \end{equation*}
3385: with $\K'_0=\K[\log n]$
3386: and $\K$ defined at~\eqref{eq:16}.
3387: \end{proof}
3388:
3389:
3390: % \underline{The case $t_n = n^\alpha$}
3391: % \medskip
3392: %
3393: % As our final application we will use~\eqref{cay3} and the extended
3394: % singularity analysis toolkit to derive the asymptotic behavior of
3395: % $f_n$ defined by the recurrence~\eqref{cay1} when $t_1=0$ and $t_n =
3396: % n^\alpha$ for $n \geq 2$ with $\alpha > 0$. The singular expansion of
3397: % the ordinary generating function of $(t_n)$ is given by~\eqref{eq:1}.
3398: % From the implicit function theorem and Lagrangian inversion we know
3399: % that $T(z)$ is singular at $z = e^{-1}$ and admits a singular
3400: % expansion
3401: % \marginal{\raggedright We probably
3402: % should rewrite so that all expansions are as argument $\to 1$. But
3403: % wait for revision of \S\S~1 and~2.} (In
3404: % this section all our estimates will be valid as $z \to
3405: % e^{-1}$ in such a sector.) Then
3406: % \begin{equation*}
3407: % T^2(z) = 1 - 2\sqrt2(1-ez)^{1/2} + O(|1-ez|).
3408: % \end{equation*}
3409: %
3410: % Consider first the case $\alpha < 1/2$. By Proposition~\ref{thm:4},
3411: % \begin{equation*}
3412: % \tau(z) \odot T^2(z) = \tau(z) \odot (T^2(z) - 1) = c +
3413: % O(|1-ez|^{\tfrac12 - \alpha}),
3414: % \end{equation*}
3415: % where $c$ is a constant.
3416: % Hence, by Theorem~\ref{thm:2A},
3417: % \begin{equation*}
3418: % \partial_z ( \tau(z) \odot T^2(z) ) \frac1{T(z)} =
3419: % O(|1-ez|^{-\tfrac12-\alpha}).
3420: % \end{equation*}
3421: % Now, by Theorem~\ref{thm:2B},
3422: % \begin{equation}\label{eq:7}
3423: % \int_0^z \partial_w ( \tau(w) \odot T^2(w) ) \frac1{T(w)} \, dw =
3424: % \kappa + O(|1-ez|^{\frac12 - \alpha}),
3425: % \end{equation}
3426: % where
3427: % \begin{equation}
3428: % \label{eq:16}
3429: % \kappa := \int_0^{1/e} \partial_w ( \tau(w) \odot T^2(w) ) \frac1{T(w)}
3430: % \, dw.
3431: % \end{equation}
3432: % Furthermore
3433: % \begin{equation}\label{eq:8}
3434: % \frac{T(z)}{1-T(z)} = \frac1{\sqrt2} (1-ez)^{-1/2} + O(1),
3435: % \end{equation}
3436: % so using~\eqref{eq:7} and~\eqref{eq:8} in~\eqref{cay3}, we get
3437: % \begin{equation*}
3438: % f(z) = \frac{\kappa}{2\sqrt2} (1-ez)^{-1/2} + O(|1-ez|^{-\alpha}).
3439: % \end{equation*}
3440: % Using singularity analysis and Stirling's approximation we get
3441: % \begin{equation*}
3442: % f_n = e^n
3443: % \left[
3444: % \frac{\kappa}{2\sqrt2} \frac{n^{-1/2}}{\Gamma(1/2)} + O(n^{\alpha-1})
3445: % \right] \frac{n!}{n^{n-1}} = \frac{\kappa}{2} n + O(n^{\alpha +
3446: % \frac12}).
3447: % \end{equation*}
3448: %
3449: % Next, consider the case $\alpha > 1/2$. Compare~\eqref{eq:12}
3450: % and~\eqref{eq:10}. Just as we got~\eqref{eq:9}, we get
3451: % \begin{equation*}
3452: % \tau(z) \odot T^2(z) = -\frac{2\sqrt2}{\Gamma(-1/2)} \Li_{\tfrac32 -
3453: % \alpha,0}(ez) + R_1(z),
3454: % \end{equation*}
3455: % where
3456: % \begin{equation}
3457: % \label{eq:14}
3458: % R_1(z) =
3459: % \begin{cases}
3460: % c + O(|1-ez|^{1-\alpha}) & 1/2 < \alpha < 1 \\
3461: % O(|L(ez)|) & \alpha = 1 \\
3462: % O(|1-ez|^{1-\alpha}) & \alpha > 1.
3463: % \end{cases}
3464: % \end{equation}
3465: % Another singular expansion then yields [compare~\eqref{eq:13}]
3466: % \begin{equation*}
3467: % \tau(z) \odot T^2(z) = \sqrt\frac2\pi \Gamma(\alpha -
3468: % \tfrac12) (1-ez)^{-\alpha + \tfrac12} + R_2(z),
3469: % \end{equation*}
3470: % where $R_2$, like $R_1$, satisfies~\eqref{eq:14} (with a possibly
3471: % different $c$). By Theorem~\ref{thm:2A},
3472: % \begin{equation*}
3473: % \partial_z (\tau(z) \odot T^2(z)) = \partial_z
3474: % \left\{
3475: % \sqrt\frac2\pi \Gamma(\alpha -
3476: % \tfrac12) (1-ez)^{-\alpha + \tfrac12}
3477: % \right\} + R_3(z),
3478: % \end{equation*}
3479: % where (for any $\epsilon > 0$)
3480: % \begin{equation}
3481: % \label{eq:15}
3482: % R_3(z) =
3483: % \begin{cases}
3484: % O(|1-ez|^{-\alpha}) & \alpha \ne 1 \\
3485: % O(|1-ez|^{-\epsilon-1}) & \alpha = 1.
3486: % \end{cases}
3487: % \end{equation}
3488: % Recalling~\eqref{eq:10}, the integrand in~\eqref{cat3} has asymptotic
3489: % expansion
3490: % \begin{equation*}
3491: % \partial_z (\tau(z) \odot T^2(z)) \frac{1}{T(z)} = \partial_z
3492: % \left\{
3493: % \sqrt\frac2\pi \Gamma(\alpha -
3494: % \tfrac12) (1-ez)^{-\alpha + \tfrac12}
3495: % \right\} + R_4(z),
3496: % \end{equation*}
3497: % where $R_4$, like $R_3$, satisfies~\eqref{eq:15}. Integrating and applying
3498: % Theorem~\ref{thm:2B},
3499: % \begin{equation*}
3500: % \int_0^z \partial_w ( \tau(w) \odot T^2(w) ) \, \frac{dw}{T(w)} =
3501: % \sqrt\frac2\pi \Gamma(\alpha-\tfrac12)
3502: % (1-ez)^{-\alpha + \tfrac12} + R_5(z),
3503: % \end{equation*}
3504: % where
3505: % \begin{equation*}
3506: % R_5(z) =
3507: % \begin{cases}
3508: % O(1) & 1/2 < \alpha < 1 \\
3509: % O(|1-ez|^{-\epsilon}) & \alpha=1 \\
3510: % O(|1-ez|^{-\alpha+1}) & \alpha > 1.
3511: % \end{cases}
3512: % \end{equation*}
3513: % Using this and~\eqref{eq:8} in~\eqref{cat3},
3514: % \begin{equation}
3515: % \label{eq:cayley18}
3516: % f(z) = \frac{\Gamma(\alpha - \tfrac12)}{2\sqrt\pi} (1-ez)^{-\alpha}
3517: % + R_6(z),
3518: % \end{equation}
3519: % where
3520: % \begin{equation*}
3521: % R_6(z) =
3522: % \begin{cases}
3523: % O(|1-ez|^{-1/2}) & 1/2 < \alpha < 1 \\
3524: % O(|1-ez|^{-1/2-\epsilon}) & \alpha = 1 \\
3525: % O(|1-ez|^{-\alpha + \tfrac12}) & \alpha > 1.
3526: % \end{cases}
3527: % \end{equation*}
3528: % Applying the singularity analysis transfer theorem and using
3529: % Stirling's approximation, we get
3530: % \begin{equation*}
3531: % f_n = \frac{\Gamma(\alpha - \tfrac12)}{\sqrt2\Gamma(\alpha)}
3532: % n^{\alpha + \tfrac12} + r(n),
3533: % \end{equation*}
3534: % where
3535: % \begin{equation*}
3536: % r(n) =
3537: % \begin{cases}
3538: % O(n) & 1/2 < \alpha < 1 \\
3539: % O(n^{1+\epsilon}) & \alpha=1 \\
3540: % O(n^\alpha) & \alpha > 1.
3541: % \end{cases}
3542: % \end{equation*}
3543: %
3544: % Finally, when $\alpha = 1/2$,
3545: % \begin{align*}
3546: % \tau(z) \odot T^2(z) & = -\frac{2\sqrt2}{\Gamma(-1/2)} \Li_{1,0}(ez) +
3547: % c + O(|1-ez|^{1/2}) \\
3548: % & = \sqrt{\frac{2}{\pi}} L(ez) + c +
3549: % O(|1-ez|^{1/2}).
3550: % \end{align*}
3551: % Continuing as in the case $\alpha > 1/2$, we get
3552: % \begin{equation*}
3553: % f(z) = \frac{1}{2\sqrt\pi} (1-ez)^{-1/2} L(ez) +
3554: % O(|1-ez|^{-1/2}),
3555: % \end{equation*}
3556: % and by singularity analysis and Stirling's approximation,
3557: % \begin{equation*}
3558: % f_n = \frac{1}{\sqrt{2\pi}} n \log{n} + O(n).
3559: % \end{equation*}
3560: %
3561: %
3562: % \medskip\noindent
3563: % \underline{The case $t_n=\log{n}$}
3564: % \medskip
3565: %
3566: % When $t_n = \log{n}$, we have $\tau(z) = \Li_{0,1}(z) =
3567: % O(|1-z|^{-1-\epsilon})$ for any $\epsilon > 0$ by~\eqref{eq:3}. Then
3568: % \begin{equation*}
3569: % \tau(z) \odot T^2(z) = c + O(|1-ez|^{\tfrac12 - \epsilon}),
3570: % \end{equation*}
3571: % and in the same way for the case $\alpha < 1/2$, we get
3572: % \begin{equation*}
3573: % f(z) = \frac{\kappa}{2\sqrt2} (1-ez)^{-1/2} + O(|1-ez|^{-\epsilon}),
3574: % \end{equation*}
3575: % which implies
3576: % \begin{equation*}
3577: % f_n = \frac{\kappa}{2} n + O(n^{\tfrac12 + \epsilon}),
3578: % \end{equation*}
3579: % with $\kappa$ defined as at~\eqref{eq:16}.
3580: % \end{proof}
3581:
3582: It is of interest to compare our approach
3583: to that of Knuth and Pittel~\cite{KnPi89}.
3584: These authors use what is fundamentally a ``repertoire approach'',
3585: based on the transforms of two types of tolls, the Dirac
3586: tolls~$\delta_{mn}$ and another family related to
3587: ``tree polynomials''. Their methods
3588: do not clearly appear to be extendible to
3589: the extraction of sublinear terms in asymptotic expansions.
3590: At the same time, their developments require appreciably more
3591: involved and perhaps less transparent computations.
3592:
3593: %% The similarity between the behavior of the uniform Catalan model and
3594: %% the union--find model is striking.
3595:
3596:
3597: % We remark that the results of Section~\ref{sec:cayl-tree-recurr} were
3598: % previously derived, by different techniques and considerably more
3599: % involved computations, as part of Theorem~1 of~\cite{KnPi89}.
3600: %
3601:
3602:
3603: \section{Perspectives}\label{sec:appl-high-moments}
3604:
3605:
3606:
3607:
3608: In this concluding section, we discuss at a fairly informal
3609: and abstract level
3610: applications of the extended
3611: singularity analysis toolkit
3612: developed in the present paper in two further directions:
3613: the determination of higher-order moments for our
3614: basic models, and the treatment
3615: of tree recurrences which are more complex than the ones present in our
3616: lead examples.
3617: (Some of our examples below may accordingly involve
3618: nonbinary tree models.)
3619:
3620:
3621:
3622:
3623:
3624: \subsection{Higher moments and limit distributions}
3625: Let us return to the general framework of Section~\ref{DandC-sec}.
3626: There, the random cost~$X_n$ is related to costs~$X_{K_n}$
3627: and~$\widetilde{X}_{n-a-K_n}$
3628: by the fundamental recursion~(\ref{eq2}). Raising both members
3629: of~(\ref{eq2}) to some integral power~$s$ yields
3630: \begin{equation}\label{hmom1}
3631: X_n^s=X_{K_n}^s+\widetilde{X}_{n-a-K_n}^s+
3632: \sum_{{s_1+s_2+s_3=s\atop s_2,s_3\not=s}}\binom{s}{s_1,s_2,s_3}
3633: t_n^{s_1}X_{K_n}^{s_2}\widetilde{X}_{n-a-K_n}^{s_3},
3634: \end{equation}
3635: where we have made use of the multinomial expansion and
3636: have isolated
3637: the two $s$th
3638: powers. Take expectations with respect to the
3639: model~$\mathfrak{M}_n$ and set~$\mu_n^{(s)}:=\E(X_n^s)$. The recursion
3640: on $s$th moments becomes, thanks to independence of the~$X$
3641: and~$\widetilde{X}$ sequences on the right in~(\ref{hmom1}),
3642: \begin{equation}\label{hmom2}
3643: \mu_n^{(s)}=\sum_k p_{n,k}\left(\mu_k^{(s)}+\mu_{n-a-k}^{(s)}\right)+r_n^{(s)},
3644: \end{equation}
3645: where
3646: \[
3647: r_n^{(s)}:=\sum_{{s_1+s_2+s_3=s\atop
3648: s_2,s_3\not=s}}\binom{s}{s_1,s_2,s_3}
3649: t_n^{s_1}\sum_k p_{n,k}\mu_k^{(s_2)}\mu_{n-a-k}^{(s_3)}.
3650: \]
3651: This calculation shows that the sequence of~$s$th moments for any
3652: fixed~$s$ satisfies the same type of recurrence as the first moments,
3653: save for a more complicated toll $(r_n^{(s)}$) that
3654: involves moments of the smaller orders $0,1,\ldots,s-1$. Define the
3655: normalized generating functions
3656: \[
3657: \mu^{(s)}(z):=\sum_n \mu_n^{(s)}\omega_n z^n,
3658: \qquad
3659: r^{(s)}(z):=\sum_n r_n^{(s)}\omega_n z^n,
3660: \]
3661: with the normalization sequence $\omega_n\equiv1$ for binary search
3662: trees and $\omega_n=C_n$ for uniform binary trees. Then
3663: the relation~(\ref{hmom2}) is solved in terms of generating functions
3664: by an~$\cal L$-transform as
3665: \begin{equation}\label{hmom3}
3666: \mu^{(s)}(z)=\cal L\left(r^{(s)}\right),
3667: \quad
3668: r^{(s)}(z)=\!\!\!\!\!\!\!\sum_{{s_1+s_2+s_3=s\atop
3669: s_2,s_3\not=s}}\binom{s}{s_1,s_2,s_3} \tau^{\odot s_1}(z) \odot
3670: Q(\mu^{(s_2)}(z),\mu^{(s_3)}(z)),
3671: \end{equation}
3672: where
3673: $Q$ is for the binary search tree model and Catalan model,
3674: respectively,
3675: \begin{equation}\label{hmom4}
3676: Q^{\operatorname{BST}}(a(z),b(z))=\int_0^z a(t) b(t)\, dt,
3677: \qquad
3678: Q^{\operatorname{Cat}}(a(z),b(z))=za(z)b(z).
3679: \end{equation}
3680: The~$\cal L$ transform is given in~(\ref{Lbst}) and~(\ref{Lcat}) for
3681: the respective cases; the case of the union--find tree model
3682: [where~$\omega_n=n^{n-1}/n!$ is used for $\mu^{(s)}$ and
3683: $\omega_n' = n^{n-2}(n-1)/n!$ is used for $r^{(s)}$] is similar
3684: but more complicated---see~(\ref{Lcay}) for~$\cal L$,
3685: while
3686: \[
3687: Q^{UF}(a(z),b(z))=\frac12 a(z)b(z).
3688: \]
3689: As seen in the previous section, these $\cal L$ transforms involve
3690: only integration, differentiation, and ordinary
3691: and Hadamard products---all are operations that preserve the character
3692: of being $\Delta$-regular and admitting complete asymptotic expansions
3693: at the dominant singularity. We then have a general result:
3694:
3695:
3696: \begin{theorem}\label{meta-thm} For any of the binary search tree,
3697: uniform binary tree, or union--find model, and for any integer~$s>0$, the
3698: $s$th moment of the cost function associated to a toll $t_n^{\log}$ or
3699: $t_n^\alpha$ admits a complete descending expansion in powers of~$n$
3700: (possibly with logarithmic terms).
3701: \end{theorem}
3702: \begin{proof}
3703: The proof is simply an induction on the order~$s$
3704: of the moments. We establish by induction
3705: the stronger property that the generating functions
3706: $\mu^{(s)}(z)$ are $\Delta$-regular and admit complete asymptotic
3707: expansions in powers of~$(1-z)$, possibly with logarithmic terms,
3708: after rescaling the singularity to be at~1.
3709: The property is true for~$s=1$
3710: by results of the previous section. If the property is assumed to be
3711: true through order~$s-1$, then the tolls $r^{(s)}(z)$
3712: are $\Delta$-regular and admit of complete
3713: asymptotic expansions at their singularity: this results from
3714: closure theorems
3715: of Sections~\ref{sec:sing-expans-diff} and~\ref{sec:hadam-prod-transf}. Next,
3716: the $\cal L$-transform is applied and, again by closure theorems, the property
3717: of~$r^{(s)}(z)$
3718: is seen to extend to $\mu^{(s)}(z)$. Thus the singular structure
3719: of~$\mu^{(s)}(z)$ is fully characterized.
3720: It then suffices to apply basic
3721: singularity analysis in order to recover
3722: the existence of
3723: full asymptotic expansion of the moments
3724: $\mu_n^{(s)}=\frac{1}{\omega_n}[z^n]\mu^{(s)}(z)$.
3725: \end{proof}
3726:
3727:
3728: The process of extracting moments one after the other has been
3729: nicknamed ``moment pumping'' in the article~\cite{FlPoVi98}, where it
3730: was used to determine the shape of the moments of total displacement
3731: in linear hashing tables. It had been employed earlier by Louchard and
3732: Tak\'acs in order to characterize moments of path length in trees and
3733: of area under excursions~\cite{Louchard84,Takacs91}, in a way largely
3734: similar to what has been described here in more general
3735: terms. In favorable cases, a pattern regarding the asymptotic shape
3736: of moments may emerge. In such cases (possibly centering of the random
3737: variable is required), the limiting distribution of costs becomes
3738: accessible through its moments, thanks to the moment convergence
3739: theorem. Instances are found in the already cited
3740: papers~\cite{FlPoVi98,Louchard84,Takacs91}. Fill's
3741: paper~\cite{Fill96} provides another example (although it is based on
3742: direct recurrence manipulations rather than generating functions) to
3743: the effect that the logarithmic toll~$t^{\log}_n$ gives rise to
3744: asymptotically Gaussian costs under the binary search tree model. Yet
3745: other examples, often based on direct recurrence manipulations, are
3746: provided by the recent independent studies of Hwang and
3747: Neininger~\cite{HwNe02} and of Fill and
3748: Kapur~\cite{FK-transfer}. Clearly, a
3749: ``metatheorem'' similar
3750: to Theorem~\ref{meta-thm} is possible for varieties of increasing
3751: trees in the sense of Bergeron-Flajolet-Salvy~\cite{BeFlSa92}
3752: (generalizing the BST model).
3753: For simply generated families of trees in the
3754: sense of Meir and Moon~\cite{MeMo78} (generalizing the Catalan model),
3755: asymptotics of moments as well as limiting distributions have been
3756: derived by Fill and Kapur~\cite{sgtechreport} as part of a
3757: broader project joint with Svante Janson. The union--find
3758: tree model can be generalized to other families of trees, and the
3759: techniques of the present paper can again be applied; this is the
3760: subject of ongoing research by the authors.
3761:
3762:
3763: \subsection{Differential models}
3764: Many tree recurrences
3765: associated to comparison-based searching
3766: and multidimensional retrieval problems
3767: generalizing binary search trees, once translated into generating
3768: functions, lead to integral equations of the form
3769: \begin{equation}\label{dif0}
3770: {\bf \Phi}[f](z)=t(z),
3771: \end{equation}
3772: where ${\bf \Phi}$ is a linear integral operator involving
3773: coefficients in $\C(z)$, that is, rational function
3774: coefficients. Here, as in our lead examples, $f(z)$
3775: is a generating function of expected costs and $t(z)$
3776: is a toll generating function.
3777: By successive differentiations, this transforms into a
3778: linear differential equation of the form
3779: \begin{equation}\label{dif1}
3780: {\bf \Delta}[f](z)=\widetilde t(z), % \qquad {\bf\Delta}\in \C(z)[\partial_z],
3781: \end{equation}
3782: where $\widetilde t(z)$ is a modified toll generating
3783: function and is an elementary variant of~$t(z)$.
3784: We shall let~$d$ denote the order of the differential equation~(\ref{dif1}).
3785:
3786: The description above
3787: corresponds to the situation already encountered with the binary search tree
3788: recurrence,
3789: representing the easy case of a differential order equal to~1.
3790: Other known cases include the $m$--ary search tree studied by Mahmoud
3791: and Pittel (see the account in~\cite{Mahmoud92}) and
3792: others~\cite{ChHw01,FK-transfer},
3793: quicksort with median-of-sample partitioning
3794: and locally balanced trees~\cite{Knuth98a,SeFl96},
3795: quadtrees~\cite{FlGoPuRo93,HoFl92} as well as
3796: multidimensional search trees also known as $k$-d-trees~\cite{FlPu86}.
3797: (A valuable
3798: survey of a class of problems leading to Euler equations
3799: appears in~\cite{ChHwTs02}.)
3800: For instance, in the case of 2-dimensional quadtrees the operator
3801: is given in~\cite{HoFl92} as
3802: \[
3803: {\bf \Phi}[f](z)=f(z) - 4\int_0^z\left[\int_0^y f(x)\,
3804: \frac{dx}{1-x}\right]\,\frac{dy}{y(1-y)},
3805: \]
3806: which leads to a second order differential equation,
3807: \[
3808: z(1-z)\partial_z^2f(z)+(1-2z)\partial_zf(z)-\frac{4}{1-z} f(z) =
3809: \widetilde t(z),
3810: \]
3811: where $\widetilde t(z)=\partial_z[z(1-z)t'(z)]$.
3812:
3813:
3814:
3815:
3816: The variation-of-constants technique applies to equations of order
3817: greater than~1 as well as to linear systems. It may then be used to
3818: express~$f$ as a linear integral transform involving a set $\{h_j\}$
3819: of solutions to the homogeneous equation ${\bf \Delta} h=0$, as we
3820: know explain. Indeed, let the linear differential
3821: equation~(\ref{dif1}) be put into the form of a system
3822: \begin{equation}\label{difsys}
3823: \partial_z {\bf y}(z) = {\bf A} {\bf y}(z)+ {\bf b}(z),
3824: \end{equation}
3825: where ${\bf y}$ is the $d$--dimensional vector ${\bf
3826: y}=(f,f',\ldots,f^{(d-1)})$, ${\bf A}={\bf A}(z)$ is a $d\times d$
3827: matrix of functions [here, by assumption, all in~$\C(z)$],
3828: and~${\bf b}=(\widetilde{t},\widetilde{t}\,',\ldots,
3829: \widetilde{t}^{(d-1)})$; see~\cite[vol~II,
3830: \S9.3]{Henrici77a} for the reduction. Recall that a fundamental
3831: matrix ${\bf W}$ for the system~(\ref{difsys}) is by definition a
3832: nonsingular~$d\times d$ matrix whose columns each satisfy the
3833: homogenous system $\partial_z {\bf y}(z) = {\bf A} {\bf y}(z)$. Then
3834: the general solution to the inhomogeneous
3835: system~(\ref{difsys}) is, by the classical ``variation-of-constants''
3836: formula,
3837: \begin{equation}\label{vofc}
3838: {\bf y}(z)={\bf W}(z)\cdot{\bf W}^{-1}(z_0)\cdot {\bf y}(z_0)
3839: +{\bf W}(z)\cdot \int_{z_0}^z {\bf W}^{-1}(x)\cdot {\bf b}(x)\, dx;
3840: \end{equation}
3841: see once more~\cite[vol~II]{Henrici77a}.
3842: (The initial conditions at some~$z_0$ are assumed to be known.)
3843: This provides the solution to~(\ref{dif0}) as
3844: \[
3845: f(z)={\cal L}[\widetilde t(z)],
3846: % \qquad
3847: % \hbox{with}\quad
3848: % {\cal L}\in\C[h_1,\ldots,h_m][\partial_z].
3849: \]
3850: with~$\cal L$ a \emph{linear integral transform} that involves
3851: polynomially the elements of a fundamental matrix~${\bf W}$ as well as the
3852: inverse of the Wronskian $\det {\bf W}$.
3853: (The case of Euler equations
3854: is somewhat simpler, as it is fully
3855: explicit~\cite{ChHwTs02,FK-transfer}.)
3856: For instance, the case of 2-dimensional quadtrees leads to
3857: a still explicit form~\cite{HoFl92}, namely,
3858: $
3859: f(z)={\cal L}[\widetilde t(z)]$ where
3860: \[
3861: {\cal L}[e(z)]=\frac{1+2z}{(1-z)^2}
3862: \int_0^z \frac{(1-y)^3}{y(1+2y)^2}\left[
3863: \int_0^y \frac{1+2x}{(1-x)^2} e(x)\, dx\right]\, dy.
3864: \]
3865:
3866: From here onward we suppose for simplicity that $\omega_n \equiv 1$,
3867: so that $f$ is an ordinary generating function, though our discussion
3868: extends readily to more general normalization constants. Call a
3869: system \emph{dominantly regular}\footnote{The term ``dominantly
3870: regular'' evokes the fact that the condition concerns the
3871: \emph{dominant} singularity of the solution function, where the
3872: singularity of the system is of the so-called ``\emph{regular}''
3873: type (first-kind implies regular singularity by a well-known
3874: theorem; see~\cite[vol~II, Theorem~9.4d]{Henrici77a}).} % = Theorem
3875: % 9.4d, p.~111.
3876: if it is singular at~1 (i.e., if the matrix ${\bf A}$ has a pole at~1,
3877: but at no other point in~$|z|\le1$ except possibly~0) and if the pole
3878: of~${\bf A}$ at~1 is simple---the latter case is known as a
3879: singularity ``of the first kind''. All the classical examples listed
3880: above and generalizing binary search trees satisfy this condition. We
3881: then have:
3882: \begin{theorem}\label{genBST-thm}
3883: Let a tree recurrence be expressed by a differential system that is
3884: dominantly regular. Then the expectations of
3885: costs induced by the tolls~$t_n^\alpha$
3886: and~$t_n^{\log}$ admit complete asymptotic expansions in
3887: descending powers of~$n$, possibly with logarithmic terms.
3888: \end{theorem}
3889: \begin{proof}
3890: First, we observe that for any tree recurrence, the cost induced by
3891: an eventually increasing nonnegative toll $t_n \to +\infty$ is at least~$t_n$
3892: (by the very nature of the tree recurrence) and at most $O(nt_n)$
3893: (by induction). Thus, for the tolls under consideration, the
3894: generating function of costs, $f(z)$, has radius of convergence
3895: exactly equal to~1. We also observe that the values of~$f$ and its
3896: derivatives at some point~$z_0$ such that $|z_0|<1$ are well-defined.
3897: %% (they are in fact computable real numbers).
3898: We may adopt for
3899: instance~$z_0=\frac12$ in the variation-of-constants
3900: formula.
3901:
3902: By the classical theory of singularities of the first kind, each of
3903: the column vectors of matrix~${\bf W}$ is analytic for $z$ in a
3904: neighborhood
3905: of~1 slit along the ray $[1,+\infty)$.
3906: There, as $z$ tends to~1, it admits a representation
3907: as a finite combination of terms of the form
3908: \[
3909: (1-z)^{\alpha}L(z)^k R(1-z),
3910: \]
3911: where~$\alpha$ is an algebraic number (a root of the indicial
3912: equation), $k$ an integer, and $R$ is analytic at~0. Thus, each
3913: element of~${\bf W}$ is amenable to singularity analysis
3914: as it is $\Delta$-continuable and admits a
3915: bona fide expansion near~1.
3916:
3917: By formula~(\ref{vofc}), there remains to discuss the elements of
3918: ${\bf W}^{-1}$. By the cofactor rule, the elements of ${\bf W}^{-1}$
3919: involve polynomially the elements of~${\bf W}$ divided by
3920: the Wronskian determinant $\det {\bf W}(z)$. It is a well-known fact
3921: (see {\S9.3} of~\cite[vol~II]{Henrici77a}) that the Wronskian is
3922: expressible in terms of the system alone and one has
3923: \[
3924: [\det {\bf W}(z)]^{-1}= [\det {\bf W}(z_0)]^{-1}\exp
3925: \left(-\int_{z_0}^z \operatorname{tr} {\bf A}(x)\, dx\right).
3926: \]
3927: [Here~$\operatorname{tr}(\cdot)$ denotes the matrix trace operator.]
3928: By the dominant regularity assumption, the trace is here a rational function
3929: with at most a simple pole at~1, so that its integral is either
3930: analytic at~1 or logarithmic. In either case, $(\det {\bf W}(z))^{-1}$
3931: is of singularity analysis type, and so are consequently all the
3932: elements of the inverse of the fundamental matrix~${\bf W}^{-1}(z)$.
3933: By the singular integration and singular differentiation theorems of
3934: Section~\ref{sec:sing-expans-diff}, there results that the integral
3935: transform~(\ref{vofc}) preserves for functions the character of being
3936: amenable to singularity analysis. Since the toll generating functions
3937: are of singularity analysis type, basic singularity analysis is
3938: applicable to~$f(z)$. The result follows.
3939: \end{proof}
3940:
3941: In principle, higher moments will also become accessible to
3942: singularity analysis once the nonlinear integral forms~$Q$
3943: extending $Q^{BST}$ of~(\ref{hmom4}) have been worked out.
3944: We are however not aware of existing research in this direction, despite
3945: the fact that the splitting probabilities are known in a number of cases
3946: (see, e.g.,~\cite{FlGoPuRo93} for quadtrees). There
3947: is interest in these questions, as partly heuristic recent work by
3948: Majumdar and collaborators (see, e.g.,~\cite{DeMa02} for the type of
3949: method employed and succinct developments) indicates the
3950: probable existence of phase transitions in
3951: the number of internal nodes of~$d$-dimensional quadtrees
3952: for large enough~$d$ ($d\ge d_c=9$ is suggested) in a way similar to what is
3953: already well established for the size of~$m$-ary search
3954: trees~\cite{ChHw01,FK-transfer,Mahmoud92}.
3955:
3956:
3957: \smallskip
3958:
3959: As a final note, we'd like to mention digital trees, which were
3960: recognized to be amenable to treatment by \emph{ordinary}
3961: (rather than the more customary exponential) generating functions
3962: in~\cite{FlRi92}. Techniques of the present paper
3963: would most likely be usable in such a context, in particular as
3964: regards tolls of the form~$n^\alpha$ and~$\log n$. A partial
3965: classification of cost functions along these lines has
3966: already been given by Derfel and Vogl in~\cite{DeVo01}.
3967:
3968:
3969:
3970: \medskip
3971: \noindent
3972: \begin{small}
3973: {\bf Acknowledgements.} The second author is grateful to Don Knuth
3974: for a summer invitation to
3975: Stanford in 1988, during which his work on the subject was started.
3976: This work was supported in part by the Future and Emerging
3977: Technologies programme of the EU under contract number
3978: IST-1999-14186 ({\sc Alcom--Ft} Project).
3979:
3980: The first and third authors' research is supported by NSF grant
3981: DMS--9803780 and DMS--0104167, and by The Johns Hopkins University's
3982: Acheson J.~Duncan Fund for the Advancement of Research in Statistics.
3983: \end{small}
3984: %%\JF{What is Flajolet reference~\cite{flajolet}?}
3985: % \bibliographystyle{abbrv}
3986:
3987: %% \malert{Reminder to JAF/NK:\ Check global references to FK Catalan.}
3988: \small
3989:
3990:
3991: \bibliographystyle{acm}
3992: \bibliography{algo,msn,misc}
3993:
3994:
3995: \end{document}
3996:
3997: %%
3998: