1: \documentclass[11pt]{article} % pour LaTeX2e
2: \usepackage{a4wide} % pour LaTeX2e
3: \usepackage{latexsym} % pour LaTeX2e
4: \usepackage{graphicx}
5: \usepackage{eufrak}
6:
7: \newfont{\bbold}{msbm10 scaled \magstep1}
8: \newtheorem{Theorem}{Theorem}%[section]
9: \newtheorem{Proposition}[Theorem]{Proposition}
10: \newtheorem{Lemma}[Theorem]{Lemma}
11: \newtheorem{Definition}[Theorem]{Definition}
12: \newcounter{exampleno}
13: \setcounter{exampleno}{0}
14: \newenvironment{Example}{
15: \refstepcounter{exampleno}%
16: \medbreak\bigskip\noindent%
17: {\sc Example~\theexampleno.} }%
18: {\nobreak{\hfill$\Box$}\medbreak}
19: \newcommand{\stir}[2]
20: {{{#1} \brace {#2}}}
21: \newcommand{\cfrac}[2]{\frac{\displaystyle#1}{\displaystyle#2}}
22:
23: \newtheorem{Remark}[Theorem]{Remark}
24: \newtheorem{Corollary}[Theorem]{Corollary}
25: \newtheorem{Property}[Theorem]{Property}
26:
27: \newtheorem{Remarks}[Theorem]{Remarks}
28: \newtheorem{Question}[Theorem]{Question}
29:
30: \def\Proof{{\noindent \bf Proof. }}
31: \def\SkProof{{\noindent \bf Sketch of Proof. }}
32: \newcommand{\Ref}[1]{(\ref{#1})}
33: \def\qed{$\hfill{\vrule height 3pt width 5pt depth 2pt}$ \bigskip}
34: \def\qee{$\hfill{\Box}$}
35: \def\mb#1{\makebox[5 mm]{#1}} % Pour les dessins de Cyril pstricks
36: \def\tilde{\widetilde}
37:
38: \newcommand{\gf}{generating function}
39: \newcommand{\gfs}{generating functions}
40: \newcommand{\beq}{\begin{equation}}
41: \newcommand{\eeq}{\end{equation}}
42: \newcommand{\al}{\alpha}
43: \newcommand{\be}{\beta}
44: \newcommand{\ga}{\gamma}
45:
46:
47: \newcommand{\ns}{\mbox{\bbold N}}
48: \newcommand{\zs}{\mbox{\bbold Z}}
49: \newcommand{\qs}{\mbox{\bbold Q}}
50: \newcommand{\rs}{\mbox{\bbold R}}
51: \newcommand{\cs}{\mbox{\bbold C}}
52:
53:
54:
55: \newcommand{\gm}[1]{\textbf{#1}}
56: \newcommand{\gc}[1]{\footnote{\textbf{Mireille : #1}}}
57:
58:
59: \title{\bf Generating functions\\ for generating trees}
60:
61: \author{
62: Cyril Banderier \ \ \ \ \ \ \ \and Mireille Bousquet-M\'elou\thanks
63: {Corresponding author.}
64: \ \ \ \ \ \ \ \and Alain Denise
65: \and Philippe Flajolet \and
66: Dani\`ele Gardy \and Dominique Gouyou-Beauchamps
67: }
68: \date{}
69:
70:
71: \begin{document}
72: \maketitle
73: \noindent
74: {\em
75: This article corresponds, up to minor typo corrections,
76: to the article submitted to Discrete Mathematics (Elsevier) in Nov. 1999,
77: and published in its vol. 246(1-3), March 2002, pp. 29-55.
78: This work supplants a
79: preliminary draft ``On Generating Functions of Generating Trees''
80: dated Nov. 1998 which appeared in the Proceedings of the FPSAC'99 Barcelona Conference.}
81:
82: \newcommand{\Sn}{{\EuFrak S}}
83: \def\abstractfr{\if@twocolumn
84: \section*{R\'esum\'e}
85: \else \small
86: \begin{center}
87: {\bf R\'esum\'e\vspace{-.5em}\vspace{0pt}}
88: \end{center}
89: \quotation
90: \fi}
91: \def\endabstractfr{\if@twocolumn\else\endquotation\fi}
92:
93: \def\abstractangl{\if@twocolumn
94: \section*{Abstract}
95: \else \small
96: \begin{center}
97: {\bf Abstract\vspace{-.5em}\vspace{0pt}}
98: \end{center}
99: \quotation
100: \fi}
101: \def\endabstractangl{\if@twocolumn\else\endquotation\fi}
102:
103:
104: \begin{abstractangl}
105: Certain families of
106: combinatorial objects admit recursive descriptions in terms of
107: generating trees: each node of the tree corresponds to an object,
108: and the branch leading to the node encodes
109: the choices made in the construction of the object. Generating trees
110: lead to a fast
111: computation of enumeration sequences (sometimes, to explicit
112: formulae as well) and provide efficient random generation algorithms.
113: We investigate the links between the structural
114: properties of the rewriting rules defining such trees and the rationality,
115: algebraicity, or transcendence of the corresponding generating function.
116: \end{abstractangl}
117:
118: \section{Introduction}
119:
120: Only the simplest combinatorial structures --- like binary strings,
121: permutations, or pure involutions ({\em i.e.},
122: involutions with no fixed point) --- admit product
123: decompositions. In that case, the set $\Omega_n$ of objects of size~$n$
124: is isomorphic to a product set:
125: $\Omega_n\cong[1,e_1]\times[1,e_2]\times\cdots\times[1,e_n]$.
126: Two properties result from such a strong decomposability property:
127: $(i)$~enumeration is easy, since the cardinality of $\Omega_n$ is
128: $e_1e_2\cdots e_n$;
129: $(ii)$~random generation is efficient since it reduces to a sequence
130: of random independent draws from intervals.
131: A simple infinite tree, called a {\em uniform generating
132: tree\/} is determined by the $e_i$: the root has
133: degree $e_1$,
134: each of its $e_1$ descendants has degree~$e_2$, and so on.
135: This tree describes the sequence of all possible choices and
136: the objects of size~$n$
137: are then in natural correspondence with the branches of length~$n$,
138: or equivalently with the nodes of generation~$n$ in the tree. The
139: generating tree
140: is thus fully described by its root degree
141: ($e_1$) and by rewriting rules, here of the special form,
142: \[
143: (e_i)\leadsto (e_{i+1})\,(e_{i+1})\,\cdots (e_{i+1}) \equiv (e_{i+1})^{e_i},
144: \]
145: where the power notation is used to express repetitions. For
146: instance binary strings,
147: permutations, and pure involutions are determined by
148: \[
149: \begin{array}{ll}
150: {\cal S}\,: & [(2),~(2)\leadsto (2)\,(2)] \\
151: {\cal P}\,: &[(1),~\{(k)\leadsto (k+1)^{k}\}_{k\ge1}]\\
152: {\cal I}\,:&[(1),~\{(2k-1)\leadsto (2k+1)^{2k-1}\}_{k\ge1}].
153: \end{array}
154: \]
155:
156: A powerful generalization of this idea consists in
157: considering unconstrained {\em generating trees\/} where any set of rules
158: \begin{equation}\label{i1}
159: \Sigma=[(s_0),~\left\{(k)\leadsto
160: (e_{1,k})\,(e_{2,k})\,\cdots\,(e_{k,k})\right\}]
161: \end{equation}
162: is allowed. Here, the {\em axiom\/} $(s_0)$ specifies the degree of
163: the root,
164: while the {\em productions\/} $e_{i,k}$ list the degrees of the $k$ descendants
165: of a node labeled $k$.
166: Following Barcucci, Del~Lungo, Pergola and
167: Pinzani, we call $\Sigma$ an
168: {\em ECO-system\/} (ECO stands for ``Enumerating
169: Combinatorial Objects'').
170: Obviously, much more leeway is available and there is hope
171: to describe a much wider class of structures than those corresponding
172: to product forms and uniform generating trees.
173:
174:
175: The idea of generating trees
176: has surfaced
177: occasionally in the literature. West introduced it in the context of
178: enumeration of permutations with forbidden subsequences~\cite{West95,West96};
179: this idea has been further exploited in closely related
180: problems~\cite{BaDePePiFPSAC98,BaDePePi,DuGiGu98,DuGiWe96}. A major
181: contribution in this area is due to Barcucci, Del~Lungo, Pergola, and
182: Pinzani~\cite{ECOArticle,BaDePePi_2} who
183: showed that a fairly large number of
184: classical combinatorial structures can be described
185: by generating trees.
186:
187: A form equivalent to generating trees is well worth noting at this
188: stage. Consider the {\em walks\/} on the integer half-line that start
189: at~point $(s_0)$ and such that the only allowable transitions are
190: those specified by~$\Sigma$ {(the steps corresponding to
191: transitions with multiplicities being labeled)}. Then, the walks of
192: length~$n$
193: are in bijective correspondence with the nodes of generation $n$ in the
194: tree.
195: These walks are constrained by the
196: consistency requirement of trees, namely, that the number of outgoing
197: edges from point $k$ on the half-line has to be exactly~$k$.
198:
199: \begin{Example} {\em $123$-avoiding permutations}\label{Catalan123Example}\\
200: The method of ``local expansion'' sometimes gives good results in
201: the enumeration of permutations avoiding specified patterns.
202: Consider for example the set $\Sn _n(123)$ of permutations of length
203: $n$ that {\em avoid the pattern\/} $123$:
204: there exist no integers $i < j< k$ such that $\sigma(i) <
205: \sigma(j) < \sigma(k)$. For instance, $\sigma=4213$ belongs to
206: $\Sn _4(123)$ but $\sigma=1324$ does not, as $\sigma(1) < \sigma(3)
207: < \sigma(4)$.
208:
209: Observe that if $\tau \in \Sn _{n+1}(123)$, then the permutation $\sigma$
210: obtained by erasing the entry $n+1$ from $\tau$ belongs to $ \Sn
211: _{n}(123)$.
212: Conversely, for every
213: $\sigma \in \Sn _{n}(123)$, insert the value $n+1$ in each place that gives an element of $\Sn _{n+1}(123)$
214: (this is the local expansion). For example, the permutation
215: $\sigma=213$ gives $4213$, $2413$ and $2143$, by insertion of $4$ in
216: first, second and third place respectively. The permutation $2134$,
217: resulting {from} the insertion of $4$ in the last place, does not belong
218: to $\Sn _4(123)$. This process can be described by a
219: tree whose nodes are the permutations avoiding $123$:
220: the root is $1$, and the children of any node $\sigma$ are the
221: permutations derived as above. Figure~\ref{pinzani-arbre1}(a) presents
222: the first four levels of this tree.
223:
224: Let us now label the nodes by their number of children: we obtain the
225: tree of Figure~\ref{pinzani-arbre1}(b). It can be proved that the
226: $k$ children of any node labeled $k$ are labeled respectively
227: $k+1,2,3,\ldots,k$ {(see \cite{West95})}. Thus the tree we have
228: constructed is the
229: generating tree obtained from the following rewriting rules:
230: $$
231: [ (2),\ \{(k) \leadsto (2)(3)\ldots(k-1)(k)(k+1) \}_{k\ge 2}].
232: $$
233: The interpretation of {this system} in terms of paths implies that $123$-avoiding permutations
234: are equinumerous with ``walks with returns'' on the half-line,
235: themselves isomorphic to {\L}ukasiewicz codes of plane trees (see,
236: {\em e.g.}, \cite[p.~31--35]{stanley-vol2}).
237: We thus recover a classic result \cite{Knuth1}:
238: $123$-avoiding permutations are counted by Catalan numbers; more precisely,
239: $|\Sn_n(123)|={2n \choose n}/(n+1)$.
240: \end{Example}
241:
242: \begin{figure}[ht]
243: \begin{center}
244: \includegraphics{FigCatalanTree}
245: \end{center}
246: \caption{The generating tree of $123$-avoiding permutations.
247: (a)~Nodes labeled by the permutations.
248: (b)~Nodes labeled by the numbers of children.}
249: \label{pinzani-arbre1}
250: \smallskip
251: \noindent\hrule
252: \end{figure}
253:
254:
255:
256: We shall see below that (certain) generating trees correspond to
257: enumeration sequences of relatively low computational
258: complexity and provide fast
259: random generation algorithms. Hence, there is an obvious interest in
260: delineating as precisely as
261: possible which combinatorial classes admit a generating tree
262: specification. Generating functions condense structural
263: information in a simple analytic entity. We can thus wonder
264: what kind of \gf \ can be obtained through generating trees. More
265: precisely, we study in this paper the
266: connections between the {\em structural\/} properties of the rewriting rules
267: and the {\em algebraic\/} properties of the corresponding generating function.
268:
269:
270: We shall prove several conjectures that were presented to
271: us by Pinzani and his coauthors in March 1998.
272: Our main results can be roughly described as follows.
273: \begin{itemize}
274: \item[---] {\em Rational systems.} Systems satisfying strong
275: regularity conditions
276: lead to rational generating functions (Section~\ref{SectionRational}).
277: This covers systems that
278: have a finite number of allowed degrees, as well as
279: systems like $(2.a)$, $(2.b)$, $(2.c)$ and $(2.d)$
280: below where the labels are constant except for
281: a fixed number of labels that depend linearly and uniformly on $k$.
282: \item[---] {\em Algebraic systems.}
283: Systems of a {\em factorial\/} form,
284: {\em i.e.}, where a finite modification of the set $\{1,\ldots,k\}$ is
285: reachable
286: from~$k$, lead to algebraic generating functions
287: (Section~\ref{SectionAlgebraic}). This includes in particular
288: cases $(2.f)$ and $(2.g)$.
289: \item[---] {\em Transcendental systems.} One possible reason for a
290: system to give a transcendental series is the
291: fact that its coefficients grow too fast, so that its radius of convergence is
292: zero. This is the case for System~$(2.h)$ below.
293: Transcendental generating functions are also associated with systems
294: that are too ``irregular''. An example is System~$(2.e)$.
295: We shall also discuss the
296: holonomy of transcendental systems (Section~\ref{SectionTranscendant}).
297: \end{itemize}
298:
299:
300:
301: \newpage
302: \begin{Example} \label{SystemExamples}
303: {\em A zoo of rewriting systems}\\
304: Here is a list of
305: examples recurring throughout this paper.
306: \begin{flushright}
307: \begin{tabular}{ll}
308: $ [(3), \{(k) \leadsto (3)^{k-3} (k+1) (k+2) (k+9)\}]$
309: & $(2.a)$ \\
310: $ [(3), \{(k) \leadsto (3)^{k-1} (3k+6)\}]$
311: & $(2.b)$ \\
312: $ [(2), \{(k) \leadsto (2)^{k-2} (2+(k \bmod 2)) (k+1)\}]$
313: & $(2.c)$ \\
314: $ [(2), \{(k) \leadsto (2)^{k-2} (3-(k \bmod 2)) (k+1)\}]$
315: & $(2.d)$ \\
316: $ [(2), \{(k) \leadsto (2)^{k-2} (3 - [\exists p\!:\!k=2^p]) (k+1)\}]$
317: \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ & $(2.e)$ \\
318: $ [(2), \{(k) \leadsto (2)(3)\ldots(k-1)(k)(k+1)\}]$
319: & $(2.f)$ \\
320: $ [(1), \{(k) \leadsto (1)(2)\ldots(k-1) (k+1)\}]$
321: & $(2.g)$ \\
322: $ [(2), \{(k) \leadsto (2)(3) (k+2)^{k-2}\}]$ &$(2.h)$ \\
323: \end{tabular}
324: \end{flushright}
325: (In $(2.e)$, we make use of Iverson's brackets: $[P]$ equals $1$ if $P$ is
326: true, $0$ otherwise.)
327: \end{Example}
328:
329: \paragraph{Notations.}
330: {}From now on, we adopt functional notations for rewriting rules:
331: systems will be of the form
332: $$
333: [(s_0), \quad \{(k) \leadsto (e_{1}(k))\; (e_{2}(k)) \ldots (e_{k}(k))\}]
334: $$
335: where $s_0$ is a constant and each $e_i$ is a function of
336: $k$. Moreover, we assume that % $s_0$ and the $e_i$'s are such that
337: all the values appearing in the generating tree are positive: {each
338: node has at least one descendant.}
339:
340: In the generating tree, let $f_n$ be the number of nodes at level $n$
341: and $s_n$ the sum of the labels of these nodes. By convention, the
342: root is at level $0$, so that $f_0=1$. In terms of walks, $f_n$ is the
343: number of walks of length $n$.
344: The generating function associated {with} the system is
345: $$
346: F(z) = \sum_{n \geq 0} f_n z^n.
347: $$
348: Remark that $s_n=f_{n+1}$, and that the sequence $(f_n)_n$ is
349: nondecreasing.
350:
351: Now let $f_{n,k}$ be the number of nodes at level $n$ having
352: label $k$ (or the number of walks of length $n$ ending at position
353: $k$). The following generating functions will be also of interest:
354: $$
355: F(z,u) = \sum_{n,k \geq 0} f_{n,k} z^n u^k
356: \qquad \hbox{and} \ \
357: F_k(z) = \sum_{n \geq 0} f_{n,k} z^n.
358: $$
359: We have $F(z)=F(z,1)=\sum_{k\ge 1}F_k(z)$.
360: Furthermore, the $F_k$'s satisfy the relation
361: \begin{equation} \label{equationF_k}
362: F_k(z) = [k=s_0] + z \sum_{j \geq 1} \pi_{j,k} F_j(z),
363: \end{equation}
364: where $\pi_{j,k} = |\{i \le j: e_i(j)=k \}|$ denotes the number of one-step
365: transitions from $j$ to $k$. This is equivalent to the following
366: recurrence for the numbers $f_{n,k}$,
367: \begin{equation}\label{rec1}
368: {f_{0,k}=[k=s_0]} \qquad
369: \hbox{and} \ \ f_{n+1,k}=\sum_{j \geq 1} \pi_{j,k} f_{n,j},
370: \end{equation}
371: that results from tracing all the paths that lead to $k$ in $n+1$
372: steps.
373:
374: \paragraph{Counting and random generation.}
375: The recurrence~(\ref{rec1})
376: gives rise to an algorithm that computes the successive rows of the
377: matrix $(f_{n,k})$ by ``forward propagation'': to compute the
378: $(n+1)$th row, propagate the contribution $f_{n,j}$ to
379: $f_{n+1,e_i(j)}$ for all pairs $(i,j)$ such that $i\le j$.
380: Assume the system is {\em linearly bounded\/}: this means that the
381: labels of the nodes that can be reached in $m$ steps are bounded by
382: a linear function of~$m$. (All the systems given in
383: Example~\ref{SystemExamples}, except for $(2.b)$, are
384: linearly bounded; more generally, systems where forward jumps are bounded by a
385: constant are linearly bounded.)
386: Clearly, the forward propagation algorithm
387: provides a counting algorithm of arithmetic complexity that is at most
388: cubic.
389:
390: For a linearly bounded system, uniform random generation can also be achieved
391: in polynomial
392: time, as shown in \cite{BaDePe99}. We present here the general
393: principle.
394:
395: Let $g_{n,k}$ be the number of walks of length~$n$ that start from
396: label~$k$.
397: These numbers
398: are determined by the recurrence
399: $g_{n,k}=\sum_i g_{n-1,e_i(k)},$
400: that traces all the possible continuations of a path given its initial
401: step. Obviously, $f_n=g_{n,s_0}$, with $s_0$ the axiom of the system.
402: As above, the $g_{n,k}$
403: can be determined in time $O(n^3)$ and~$O(n^2)$ storage.
404: Random generation is then achieved as follows:
405: In order to generate a walk of length $n$ starting from state~$k$,
406: pick up a transition $i$ with probability
407: $g_{n-1,e_i(k)}/g_{n,k}$,
408: and generate recursively a walk of length~$n-1$ starting from state
409: $e_i(k)$.
410: The cost of a
411: single random generation is then $O(n^2)$ if a sequential search is
412: used over the $O(n)$ possibilities of each of the $n$ random drawings;
413: the time complexity goes down to $O(n\log n)$ if binary search is
414: used, but at the expense of an increase in storage complexity of
415: $O(n^3)$ (arising from $O(n^2)$
416: arrays of size $O(n)$ that binary search requires).
417:
418:
419: \section{Rational systems}
420: \label{SectionRational}
421:
422: We give in this section three main criteria (and a variation on one of them)
423: implying that the generating
424: function of a given ECO-system is rational.
425:
426:
427:
428: Our first and simplest criterion applies to systems in which
429: the functions $e_i$ are uniformly bounded.
430: \begin{Proposition}\label{obvious}
431: If finitely many labels
432: appear in the tree, then $F(z)$ is rational.
433: \end{Proposition}
434:
435: \Proof
436: Only a finite number of $F_k$'s are nonzero, and they are related by
437: linear equations like Equation~(\ref{equationF_k}) above.
438: \qed
439:
440: \begin{Example} \label{FibonacciExample}
441: {\em The Fibonacci numbers} \\
442: The system $[(1),\, \{(k) \leadsto (k)^{k-1}
443: ((k \bmod 2)+1)\}]$ can be also written as $[(1),\, \{(1)
444: \leadsto (2),\, (2) \leadsto (1) (2)\}]$. Hence the only labels that
445: occur in the tree are $1$ and $2$. Eq.~\Ref{equationF_k} gives $F_1(z)=1+zF_2(z)$
446: and $F_2(z)=z(F_1(z)+F_2(z))$. Finally,
447: $$F(z)=\frac{1}{1-z-z^2}=\sum_{n\ge 0} f_n z^n= 1+z+2z^2+3z^3+5z^4+\cdots,$$
448: the well-known Fibonacci generating function.
449: \end{Example}
450:
451:
452: None of the systems of Example~\ref{SystemExamples} satisfy the
453: assumptions of Proposition~\ref{obvious}. However, the following criterion can
454: be applied to systems $(2.a)$ and~$(2.b)$.
455:
456: \begin{Proposition} \label{Prop1Rule}
457: Let $\sigma(k)=e_1(k)+e_2(k)+\cdots +e_k(k)$.
458: If $\sigma$ is an affine function of $k$, say $\sigma(k)=\alpha k+\beta$, then
459: the series $F(z)$ is rational. More precisely:
460: $$
461: F(z) \ =\ {1 + (s_0-\alpha)z \over 1-\alpha z-\beta z^2}.
462: $$
463: \end{Proposition}
464: \Proof
465: Let $n \geq 0$ and let $k_1, k_2, \ldots k_{f_n}$ denote the labels of
466: the $f_n$ nodes at level $n$. Then
467: \begin{eqnarray*}
468: f_{n+2} \ =\ s_{n+1} &=& (\alpha k_1+\beta) + (\alpha k_2+\beta)
469: + \cdots + (\alpha k_{f_n}+\beta) \\
470: &=& \alpha s_n + \beta f_n \ =\ \alpha f_{n+1} + \beta f_n.
471: \end{eqnarray*}
472: We know that $f_0=1$ and $f_1=s_0$. The result follows.
473: \qed
474:
475:
476:
477: \begin{Example} \label{FiboBisectionExample}
478: {\em Bisection of Fibonacci sequence} \\
479: The system $[(2),\,\{(k)
480: \leadsto (2)^{k-1} (k+1)\}]$ gives $F(z)={1-z \over
481: 1-3z+z^2}=1+2z+5z^2+\cdots$,
482: the generating function for Fibonacci numbers of even index.
483: (Changing the axiom to $(s_0)=(3)$ leads to the other half of the Fibonacci
484: sequence.)
485: Some other systems, like
486: $$\begin{array}{l}
487: {[(2),\,\{(k) \leadsto (1)^{k-1} (2k)\}]}, \\
488: {[(2),\,\{(k) \leadsto (2)^{k-2} (3-(k \bmod 2)) (k + (k \bmod 2))\}]} ,\\
489: {[(2),\,\{(k) \leadsto (2)^{k-2} (3-[k \mbox{ is prime}]) (k + [k \mbox{ is prime}])\}],}
490: \end{array}$$
491: lead to the same function $F(z)$ since $\sigma(k)=3k-1$
492: and $s_0=2$. However, the generating trees are
493: different, as are the bivariate functions $F(z,u)$.
494: \end{Example}
495:
496:
497: \begin{Example} \label{ParodoxExample}
498: {\em Prime numbers and rational \gfs}\\
499: Amazingly, it is possible to construct a
500: generating tree whose set of labels is the set of prime
501: numbers but that has a rational generating function $F(z)$.
502: This is a bit unexpected, as prime numbers are usually thought ``too
503: irregular'' to be associated with rational \gf s.
504: For $n\ge 1$, let $p_n$ denote the $n$th prime; hence $(p_1, p_2, p_3,
505: \ldots) = (2,3,5,\ldots)$. Assume for the moment that the Goldbach
506: conjecture is true: every even number larger than $3$ is the sum of
507: two primes. Remember that, according to Bertrand's postulate,
508: $p_{n+1}< 2p_n$ for all $n$ (see, {\em e.g.}, \cite[p.~140]{Ribenboim}).
509:
510: For $n \ge 1$, the number $2p_n-p_{n+1}+3$ is an even number larger
511: than $3$. Let $q_n$ and $r_n$ be two primes such that
512: $2p_n-p_{n+1}+3=q_n+r_n$. In particular, $q_1=r_1=2$. Consider the
513: system
514: $$[(2), \{ (p_n)\leadsto (p_{n+1}) (q_n) (r_n) (2)^{p_n-3}\}].$$
515: It satisfies the criterion of Proposition~\ref{Prop1Rule}, with
516: $\sigma(k)=4k-3$. Hence, the \gf \ of the associated generating tree
517: is
518: $$F(z)=\frac{1-2z}{1-4z+3z^2}=\frac{1}{2} \left[
519: \frac{1}{1-z}+\frac{1}{1-3z}\right].$$
520: Consequently, the number of nodes at level $n$ is simply {$f_n=(1+3^n)/2$.}
521: This can be checked on the first few levels of the tree drawn in
522: Figure~\ref{goldbach}.
523:
524: \begin{figure}[ht]
525: \begin{center}
526: \includegraphics{goldbach}
527: \end{center}
528: \caption{A generating tree with prime labels and rational \gf .}
529: \label{goldbach}
530: \smallskip
531: \end{figure}
532:
533: Now, one can object that the Goldbach conjecture is not proved;
534: however, it is known that every even number is the sum of at most six
535: primes~\cite{Ramare}, and a similar example can be constructed using
536: this result.
537:
538: \end{Example}
539:
540:
541:
542: \medskip
543: Proposition~\ref{Prop1Rule} can be adapted to apply to systems that
544: ``almost'' satisfy the criterion of Proposition~\ref{Prop1Rule}, like
545: System~$(2.c)$ {or~$(2.d)$.}
546: Let us consider a system of the form
547: \begin{eqnarray*}
548: (s_0), & (k) \leadsto e_1^{[0]}(k), \ldots , e_k^{[0]}(k) \ \ \hbox{ if }
549: k \hbox{ is even},\\
550: & (k) \leadsto e_1^{[1]}(k), \ldots , e_k^{[1]}(k) \ \ \hbox{ if }
551: k \hbox{ is odd}.
552: \end{eqnarray*}
553: Assume, moreover, that:
554:
555: $(i)$~the corresponding functions $\sigma_0$ and $\sigma_1$ are
556: affine and have the same leading coefficient $\alpha$, say $\sigma_0(k)=\alpha
557: k+\beta_0$ and $\sigma_1(k)=\alpha k+\beta_1$;
558:
559: $(ii)$~exactly $m$
560: odd labels occur in the right-hand side of each rule, for some $m\ge 0$.
561: \begin{Proposition} \label{Prop2Rules}
562: If a system satisfies properties {\em ({\it i})} and {\em ({\it ii})}
563: above, then
564: $$
565: F(z) = {1 + (s_0-\alpha)z + (s_1-\alpha s_0-\beta_0)z^2
566: \over
567: 1 - \alpha z - \beta_0z^2 - m(\beta_1-\beta_0)z^3}.
568: $$
569: Of course, if $\beta_0=\beta_1$, we recover the \gf \ of
570: Proposition~{\em \ref{Prop1Rule}. }
571: \end{Proposition}
572: \Proof The proof is similar to that of
573: Proposition~\ref{Prop1Rule}. The only new ingredient is the
574: fact that, for
575: $n\ge 1$, the number of nodes of odd label at level $n$ is
576: $mf_{n-1}$.
577: \qed
578:
579:
580: System~$(2.c)$ satisfies properties
581: ({\it i}) and ({\it ii}) above with
582: $\alpha=3$, $\beta_0=-1$, $\beta_1=0$, $m=1$, $s_0=2$ and $s_1=5$.
583: Consequently, its generating
584: function is $F(z)= { 1-z \over 1-3z+z^2-z^3}$.
585: System~$(2.d)$, although very close to $(2.c)$,
586: does not satisfy property ({\it ii}) above, so that
587: Proposition~\ref{Prop2Rules} does not apply.
588: {However, another minor variation on the argument of
589: Proposition~\ref{Prop1Rule}, based on the fact that the number $o_n$ of
590: odd labels at level $n$ satisfies $o_n=2(f_{n-1}-o_{n-1})$, proves the
591: rationality of $F(z)$.}
592:
593: {Alternatively, rationality follows from the last
594: criterion of this section, which is of a different nature.}
595: We consider systems $[(s_0),\,
596: \{(k) \leadsto (e_1(k)) (e_2(k)) \ldots (e_{k}(k))\}]$ that can be
597: written as
598: \begin{equation}
599: [(s_0),\,
600: \{(k) \leadsto (c_1(k)) (c_2(k)) \ldots (c_{k-m}(k))
601: (k+a_1)
602: (k+a_2)\ldots
603: (k+a_m)\}]
604: \label{lineaire1}
605: \end{equation}
606: where $1\le a_1\le a_2\le \cdots\le a_m$ and
607: the functions $c_i$ are uniformly bounded.
608: Let $C =\max_{i,k} \{s_0, c_i(k)\}$.
609:
610:
611: \begin{Proposition}
612: Consider the system {\em (\ref{lineaire1})}, and let
613: $\pi_{j,k} = |\{i\le j : e_i(j)=k \}|$. If all the series
614: $$
615: \sum_{j \ge 1} \pi_{j,k} \ t^j
616: $$
617: for $k \le C$ are rational, then so is the series $F(z)$.
618: \end{Proposition}
619:
620: \Proof
621: We form an infinite system of equations defining the series $F_k(z)$ by writing
622: Eq.~(\ref{equationF_k}) for all $k\ge 1$. In particular, for
623: $k> C$, we obtain
624: $$F_k(z)=z\sum_{\ell=1}^m F_{k-a_\ell}(z),$$
625: with $F_j(z)=0$ if $j\le 0$.
626: This part of the system is easy to solve in terms of $F_1, \ldots
627: ,F_C$. Indeed, for $k\in \zs$:
628: \begin{equation}
629: \label{recFP}
630: F_k(z) = \sum_{i=1}^{C}P_{i,k}(z) F_i(z)
631: \end{equation}
632: where the $P_{i,k}$ are polynomials in $z$
633: defined by the following recurrence: for all $i\le C$,
634: \begin{equation}
635: P_{i,k}(z) = \left\{
636: \begin{array}{ll}
637: 0 & \hbox{\rm if } k \le 0, \\
638: {[k=i]} & \hbox{\rm if } 0<k \le C ,\\
639: \displaystyle
640: z \sum_{\ell=1}^m{ P_{i,k-a_\ell}(z)} &
641: \mbox{\rm if } k >C. \\
642: \end{array}\right.
643: \label{Pik}
644: \end{equation}
645: \medskip
646: Using (\ref{recFP}), we find
647: $$F(z)=\sum_{k\ge 1}F_k(z)= \sum_{i=1}^C\left[ F_i(z) \sum_{k \ge 1}
648: P_{i,k}(z) \right].$$
649: According to (\ref{Pik}), for all $i\le C$, the series
650: $\sum_{k \ge 1} P_{i,k}(z)t^k$ is a rational
651: function of $z$ and $t$, of denominator
652: $1-z\sum_\ell t^{a_\ell}$. At $t=1$, it is rational in $z$.
653: Hence, to prove the rationality of $F(z)$, it suffices to prove the
654: rationality of the $F_i(z)$, for $i\le C$.
655:
656: Let us go back to the $C$ first equations of our system; {using
657: (\ref{recFP})}, we find, for $k\le C$:
658: $$
659: F_k(z) = [k=s_0] + z \sum_{i=1}^C\left[ F_i(z)\sum_{j \ge 1}P_{i,j}(z)\pi_{j,k}\right].
660: $$
661: Again, $\sum_{j \ge 1}P_{i,j}(z)\pi_{j,k}t^j$ is a
662: rational function of $z$ and $t$ (the Hadamard product of two rational series
663: is rational). Thus the series $F_k(z)$, for $k\le C$, satisfy a linear
664: system with rational coefficients: they are rational themselves, as
665: well as $F(z)$.
666: \qed
667:
668:
669:
670:
671: Examples~$(2.a)$, $(2.c)$, $(2.d)$ and $(2.e)$
672: have the form (\ref{lineaire1}). The above proposition implies that
673: the first three have a rational generating function. System~$(2.e)$
674: will be discussed in Section~\ref{SectionTranscendant}, and proved to
675: have a transcendental \gf .
676:
677: \section{Factorial walks and algebraic systems}
678: \label{SectionAlgebraic}
679:
680: In this section, we consider systems that are of a {\em factorial
681: form\/}. By this, we mean informally that the set of
682: successors of $(k)$ is a finite modification of the integer interval
683: $\{1,2,\ldots,k\}$. As was
684: detailed in the introduction, ECO-systems can be rephrased
685: in terms of walks over the integer half-line. We
686: thus consider the
687: problem of enumerating walks
688: over the integer half-line such that the set of allowed moves from
689: point~$k$ is a finite modification of the
690: integer interval
691: $[0,k]$. We shall mostly study modifications around the point $k$
692: (although some examples where the interval is modified around $0$ as
693: well are given at the end of the section).
694: Precisely, a {\em factorial walk\/} is
695: defined by a finite (multi)set $A \subset \zs$ and a finite set $B \subset \ns^+$, where $\ns ^+=\{1,2,3, \dots\}$,
696: specifying respectively the {\em
697: allowed supplementary jumps\/} (possibly labeled)
698: and the {\em forbidden backward jumps\/}. In
699: other words, the possible moves from $k$ are given by the rule:
700: \begin{equation}\label{alg1}
701: (k)\leadsto [0,k-1] \setminus (k-B) \ \cup \ (k+A).
702: \end{equation}
703: Observe that these walk models are not necessarily ECO-systems, first
704: because we allow labels to be zero -- but a simple translation can
705: take us back to a model with positive labels -- and second because we
706: do not require $(k)$ to have exactly $k$ successors.
707:
708: We say that an ECO-system is factorial if a shift of indices
709: transforms it into a factorial walk. Hence the rules of a factorial
710: ECO-system are of the form
711: \[
712: (k+r)\leadsto [r, k+r-1] \setminus (k+r-B) \ \cup \ (k+r+A),
713: \]
714: that is,
715: \beq \label{eco-factorial}
716: (k)\leadsto
717: [r,k-1] \setminus(k-B) \ \cup \ (k+A) \ \ \ \hbox{ for } k \ge r \ge 1.
718: \eeq
719: The \gf \ $F(z)$ for
720: such an ECO-system, taken with axiom $(s_0)$, equals the \gf \ for the
721: walk model \Ref{alg1}, taken with axiom $(s_0-r)$. However, remember
722: that
723: the rewriting rules defining a generating tree have to obey the
724: additional condition that a node labeled $k$ has exactly $k$
725: successors. Taking $k=r$ in \Ref{eco-factorial}, this implies that $r=|A|$. Taking
726: $k>r+\max B$, this implies that $r+|B|=|A|$, so that finally $B =
727: \emptyset$. Hence, strictly speaking, either
728: one has a ``fake'' factorial ECO-system (that is some of its initial
729: rules are not of the factorial type),
730: either one has a ``real'' factorial ECO-system and then it is given by rules of the form
731: \[
732: (k)\leadsto
733: [r,k-1] \ \cup \ (k+A) \ \ \ \hbox{ for } k \ge r \ge 1,
734: \]
735: where $A$ is a multiset of integers of cardinality $r$.
736: For instance, Systems~$(2.f)$ and $(2.g)$ are factorial.
737: We shall prove that
738: all factorial walks have an algebraic generating function.
739: The result naturally applies to factorial ECO-systems.
740:
741: We consider again the generating function
742: $F(z,u)=\sum_{n,k\geq 0}f_{n,k} z^n u^k$, where
743: $f_{n,k}$ is the number of walks of length $n$ ending at point
744: $k$. We also denote by $F_k(z)$ the coefficient of $u^k$ in this series,
745: and by $f_n(u)$ the coefficient of $z^n$.
746: The first ingredient of the proof is {a} linear operator $M$, acting on
747: formal power series in $u$, that encodes the possible moves. More
748: precisely, for all $n\ge 0$, {we will have:}
749: \[
750: M[f_n](u)=f_{n+1}(u).
751: \]
752: The operator $M$ is constructed step by step as follows.
753: \begin{itemize}
754: \item[---]
755: The set of moves from $k$ to all the positions $0, 1,
756: \dots, k-1$ is described by the operator $L_0$ that maps $u^k$ to
757: $u^0+u^1+\cdots+u^{k-1} =(1-u^k)/(1-u)$.
758: As $L_0$ is a linear operator, we have, for any series $g(u)$:
759: \[
760: L_0[g](u)=\frac{g(1)-g(u)}{1-u}.
761: \]
762: \item[---]
763: The fact that transitions near $k$ are modified, with those of type
764: $k+\al$ (with $\al \in A$)
765: allowed and those of type $k-\be $ (with $\be \in B $)
766: forbidden, is expressed by a Laurent polynomial
767: $$
768: P(u)=\sum_{k=-b}^a p_k u^k=A(u)-B(u)\quad\hbox{with}\quad
769: A(u)=\sum_{\al \in A} u^{\al} \ \ \ \hbox{and }\ \ \
770: B(u)=\sum_{\be \in B} u^{-\be}.
771: $$
772: The degree of $P$ is $a:= \max A$, the largest forward jump
773: and $b:=\max(0,-B,-A)$ is largest forbidden backward jump or the largest
774: supplementary backward jumps
775: (we take $b=0$ if the set $B$ is empty).
776:
777: The operator
778: $$L[g](u):=L_0[g](u) + P(u)g(u)$$
779: describes the extension of a walk by one step.
780: \item[---] Finally, the operator $M$ is given by
781: \[
782: M[g](u)=L[g](u)-\{u^{<0}\} L[g](u),
783: \]
784: where $\{u^{<0}\}h(u)$ is the sum of all the monomials in
785: $h(u)$ having a
786: negative exponent. Hence $M$ is nothing but $L$ stripped of the negative
787: exponent monomials, which correspond to walks ending on the nonpositive
788: half-line. Observe that, for any series $g(u)$, the only part of
789: $L[g](u)$ that is likely to contain monomials with negative exponents
790: is $P(u)g(u)$. Consequently,
791: \[
792: M[g](u)=L[g](u)-\{u^{<0}\}[P(u)g(u)]
793: \]
794: and if $g(u)=\sum_k g_k u^k$, then
795: \begin{equation}
796: \{u^{<0}\}[P(u)g(u)]
797: =\sum_{i=1}^{b} \sum_{k=0}^{i-1}g_k p_{-i} \, u^{k-i}
798: =\sum_{k=0}^{b-1} g_k r_k(u). \label{<0-simple}
799: \end{equation}
800: \end{itemize}
801: Assume for simplicity that the
802: initial point of the walk is~$0$;
803: other cases follow the same argument.
804: The linear relation $f_{n+1}(u)=M[f_n](u)$, together with
805: $f_0(u)=1$, yields
806: \begin{eqnarray*}
807: F(z,u) &=&1+zM[F](z,u)\\
808: &=& 1+ z \left( \frac{F(z,1)-F(z,u)}{1-u}+ P(u) F(z,u) +
809: \{u^{<0}\} [P(u)F(z,u)] \right).
810: \end{eqnarray*}
811: Thanks to \Ref{<0-simple}, we can write
812: $$\{u^{<0}\} [P(u)F(z,u)]=
813: \sum_{k=0}^{b-1} r_k(u) F_k(z),$$
814: where $r_k(u)$ is a Laurent polynomials (defined by Equation~\ref{<0-simple})
815: whose exponents belong to $[k-b, -1]$.
816: Thus, $F(z,u)$ satisfies the following functional equation:
817: \begin{equation} \label{e1} F(z,u) \left(1+\frac{z}{1-u}-z P(u)\right)=
818: 1 +\frac{zF(z,1) }{1-u}
819: +z \sum_{k=0}^{b-1} r_k(u) F_k(z). \end{equation}
820:
821: \medskip
822: Let us take an example. The moves
823: \[
824: (k)\leadsto (0)(1)\cdots(k-5) (k-3)(k-1)(k)(k+7)(k+9),
825: \]
826: lead to $A(u)=u^0+u^7+u^9$ and $B(u)=u^{-4}+u^{-2}$. Moreover,
827: $$\{u^{<0}\} [B(u)F(z,u)]=(u^{-2}+u^{-4})F_0(z)
828: +(u^{-1}+u^{-3})F_1(z)+u^{-2}F_2(z)
829: +u^{-1} F_3(z),$$
830: so that the functional equation defining $F(z,u)$ is
831: $$ F(z,u) \left(1+\frac{z}{1-u}-z (1+u^7+u^9-u^{-4}-u^{-2})\right)=
832: \hskip 50mm {} $$
833: $$1 +\frac{zF(z,1) }{1-u} +
834: z(u^{-2}+u^{-4})F_0(z)
835: +z(u^{-1}+u^{-3})F_1(z)+zu^{-2}F_2(z)
836: +zu^{-1} F_3(z).$$
837:
838: \bigskip
839:
840: The second ingredient of the proof, sometimes called the
841: {\em kernel method\/}, seems to belong to the
842: ``mathematical folklore'' since the 1970's. It has been used in
843: various combinatorial problems~\cite{CoRi,Knuth1,Odlyzko}
844: and in probabilities~\cite{FaIa}. See also~\cite{Bo98,BoPe98,Pe98} for
845: more recent and systematic applications.
846: This method consists in cancelling
847: the left-hand side of
848: the fundamental functional equation~(\ref{e1})
849: by coupling $z$ and $u$, so that the coefficient of
850: the (unknown) quantity $F(z,u)$ is zero. This constraint defines $u$ as one of
851: the branches of an algebraic
852: function of $z$. Each branch that can be
853: substituted analytically into the functional equation yields a linear
854: relation between the unknown series $F(z,1)$ and
855: $F_k(z)$, $0\le k <b$. If enough branches can be
856: substituted analytically,
857: we obtain a system of linear equations, whose solution gives
858: $F(z,1)$ and the $F_k(z)$ as algebraic
859: functions. {}From there, an expression for $F(z,u)$ also results in the
860: form of a bivariate algebraic function.
861:
862: Let us multiply Eq.~(\ref{e1}) by $u^b(1-u)$ to obtain an equation
863: with polynomial coefficients {(remind that we take $b=0$ if the set $B$ of
864: forbidden backward steps is empty)}. The new equation reads
865: $K(z,u)F(z,u)=R(z,u)$, where $K(z,u)$ is the {\em kernel\/} of the equation:
866: \begin{eqnarray}
867: K(z,u)&=&
868: u^b(1-u)\left(1+\frac{z}{1-u}-z P(u)\right),\nonumber\\
869: &=& u^b(1-u)+zu^b-z(1-u)\sum_{\al \in A} u^{\al+b}+z(1-u)\sum_{\be \in
870: B} u^{b-\be}.\label{kern}
871: \end{eqnarray}
872: This polynomial has degree $a+b+1$ in $u$, and hence, admits $a+b+1$
873: solutions, which are algebraic functions of $z$.
874: The classical theory of algebraic functions
875: and the Newton polygon construction enable us to expand the
876: solutions near any point as Puiseux series
877: (that is, series involving fractional exponents; see~\cite{Dieudonne}).
878: The $a+b+1$ solutions, expanded around 0, can be classified as follows:
879: \begin{itemize}
880: \item[---] the ``unit'' branch, denoted by $u_0$, {is a power
881: series in $z$ with constant term $1$;}
882: \item[---] $b$ ``small'' branches, denoted by $u_1, \dots, u_{b}$,
883: {are power series in $z^{1/b}$ whose first nonzero term is $\zeta z^{1/b}$,
884: with $\zeta ^b+1=0$;}
885: \item[---] $a$ ``large'' branches, denoted by $v_1, \dots, v_a$,
886: {are Laurent series in $z^{1/a}$ whose first nonzero term is $\zeta
887: z^{-1/a}$, with $\zeta ^a+1=0$.}
888: \end{itemize}
889: {In particular}, all the roots are distinct. {(It is not difficult
890: to check ``by hand'' the existence of these solutions: for instance,
891: plugging $z=t^b$ and $u=tw(t)$ in $K(z,u)=0$ confirms the existence of
892: the $b$ small branches.)}
893: Note that there are exactly $b+1$ finite branches: the unit
894: branch $u_0$ and the $b$ small branches $u_1,\ldots,u_b$.
895: {As $F(z,u)$ is a series in $z$ with \em{polynomial}} {coefficients
896: in $u$, these $b+1$ series $u_i$, having no negative exponents, can be
897: substituted for $u$ in $F(z,u)$. More specifically, let us replace $u$
898: by $u_i$ in~(\ref{e1}): the right-hand side of the equation vanishes,
899: giving a linear equation relating the $b+1$ unknown series $F(z,1)$ and
900: $ F_k(z)$, $0\le k <b$. Hence the $b+1$ finite branches give a set of
901: $b+1$ linear equations relating the $b+1$ unknown series. One could solve directly this
902: system, but the following argument is more elegant.}
903:
904:
905: The right-hand side of~(\ref{e1}), once multiplied by $u^b(1-u)$, is
906: $$R(z,u)= u^b (1-u) \left(1+\frac{z}{1-u} F(z,1)
907: +z \sum_{k=0}^{b-1} r_k(u) F_k(z)\right).$$
908: By construction, it is a {\em polynomial\/} in~$u$ of degree $b+1$ and
909: leading coefficient $-1$. {Hence, it admits $b+1$ roots, which
910: depend on $z$.} Replacing $u$ by the series $u_0, u_1, \dots, u_b$ in
911: Eq.~\Ref{e1} shows that these series are exactly the $b+1$
912: roots of $R$, so that
913: $$R(z,u)=-\prod_{i=0}^b (u-u_i).$$
914: Let $p_a:=[u^a]P(u)$
915: be the multiplicity of the largest forward jump.
916: Then the coefficient of $u^{a+b+1}$ in $K(z,u)$ is $p_az$,
917: and we can write
918: $$K(z,u)= p_a z \prod_{i=0}^{b} (u-u_i)\prod_{i=1}^{a} (u-v_i).$$
919: Finally, as $K(z,u)F(z,u)=R(z,u)$, we obtain
920: \begin{equation}\label{biv}
921: F(z,u) =\frac{-\prod_{i=0}^{b} (u-u_i)}
922: {u^b(1-u)+zu^b-zu^b(1-u)P(u)}
923: =-\frac{1}{ p_a z \prod_{i=1}^{a} (u-v_i)}
924: .
925: \end{equation}
926: We have thus proved the following result.
927: \begin{Proposition}\label{thm-algebraic}
928: The \gf \ $F(z,u)$ for factorial walks defined by {\em \Ref{alg1}} and
929: starting from $0$ is algebraic; it is given by {\em \Ref{biv}}, where
930: $u_0,\ldots,u_b$ (resp. $v_1, \ldots , v_a$)
931: are the finite (resp. infinite) solutions at~$z=0$ of the equation
932: $K(z,u)=0$ and the kernel $K$ is defined by~{\rm
933: (\ref{kern}).}
934: In particular, the generating function for all walks, irrespective of
935: their endpoint, is
936: \[
937: F(z,1)=-\frac{1}{z}\prod_{i=0}^b (1-u_i),
938: \]
939: and the \gf \ for {\em excursions}, i.e., walks ending at $0$, is, for $b<0$:
940: \[
941: F(z,0)=\frac{(-1)^{b}}{z}\prod_{i=0}^b u_i,
942: \]
943: (for $b=0$, the relation becomes $F(z,0)=\frac{(-1)^{b}}{1+z-p_0 z}\prod_{i=0}^b u_i$.)
944: \end{Proposition}
945: {These results
946: could be derived by a detour via multivariate linear recurrences,
947: and the present treatment is closely related to~\cite{BoPe98,Pe98};
948: however, our results were obtained independently in March 1998
949: \cite{Ba}. }
950:
951: The asymptotic behaviour of the number of $n$-step
952: walks can be established via
953: singularity analysis or saddle point methods. The series
954: $u_i$ have ``in general'' a square root singularity, yielding
955: an asymptotic behaviour of the form
956: $A \mu^n n^{-3/2}.$
957: We plan to develop this study in a forthcoming paper.
958:
959: \begin{Example}\label{CatalnExample}{\em Catalan numbers}\\
960: This is the simplest factorial walk, $(k)\leadsto
961: (0)(1)\dots(k)(k+1)$, which corresponds to the ECO-system $(2.f)$.
962: The operator $M$ is given by
963: $$M[f](u)=\frac{f(1)-f(u)}{1-u}+(1+u)f(u).$$
964: The kernel is $K(z,u)=1-u+z-z(1-u)(1+u)=1-u+zu^2$, hence
965: $u_0(z)=\frac{1-\sqrt{1-4z}}{2z}$, so that
966: $$F(z,1)=-\frac{1-u_0}{z}=\frac{1-2z-\sqrt{1-4z}}{2z^2}=
967: \sum_{n\ge 1} {2n \choose n}\frac{z^{n-1}}{n+1},$$
968: the generating function of the Catalan numbers
969: (sequence {\bf M1459}\footnote{%
970: The numbers {\bf Mxxxx} are identifiers of the
971: sequences in {\em The Encyclopedia of Integer Sequences}~\cite{SlPl95}.
972: }%
973: ).
974: This result could be expected, given the obvious relation
975: between these walks and {\L}ukasie\-wicz codes.
976: \end{Example}
977:
978: \begin{Example}\label{MotzkinExample}{\em Motzkin numbers}\\
979: This example, due to Pinzani and his co-authors, is derived from the
980: previous one by forbidding ``forward'' jumps of length zero. The rule is
981: then
982: \[
983: (k)\leadsto (0)\cdots(k-1)(k+1).
984: \]
985: The operator $M$ is
986: $$M[f](u)=\frac{f(1)-f(u)}{1-u}+uf(u).$$
987: The kernel is $K(z,u)=1-u+z-zu(1-u)=1+z-u(1+z)+zu^2$, leading to
988: \begin{eqnarray*}
989: F(z,1)&=&\frac{1-z-\sqrt{1-2z-3z^2}}{2z^2}=1+z+2z^2+4z^3+9z^4+21z^5+O(z^6),
990: \end{eqnarray*}
991: the generating function for Motzkin numbers (sequence {\bf
992: M1184}).
993: \end{Example}
994:
995:
996: \begin{Example}\label{SchroederExample}{\em Schr\"oder numbers}\\
997: This example is also due to the Florentine group.
998: The rule is $(k) \leadsto (0)\dots (k-1)(k) (k+1)^2$.
999: {}From Proposition~\ref{thm-algebraic}, we derive
1000: $$F(z,1)=\frac{1-3z-\sqrt{1-6z+z^2}}{4z^2}=1+3z+11z^2+45z^3+197z^4+O(z^5).$$
1001: The coefficients are the Schr\"oder numbers ({\bf M2898}: Schr\"oder's
1002: second problem). We give in Table 1 at the end of the paper a
1003: generalization of Catalan and Schr\"oder numbers, corresponding
1004: to the rule $(k) \leadsto (0)\dots (k-1)(k) (k+1)^m$. {This generalized
1005: rule has recently been shown to describe a set of permutations
1006: avoiding certain patterns \cite{Darla}.}
1007: \end{Example}
1008:
1009: The above examples were all quadratic.
1010: However, it is clear from our treatment that algebraic functions of
1011: arbitrary degree can be obtained: it suffices that the set of
1012: ``exceptions'' around $k$ have a span greater than~1.
1013: Let us start with a family of ECO-systems where supplementary forward jumps of
1014: length larger than one are allowed.
1015: \begin{Example}\label{TernaryExample}{\em Ternary trees, dissections of a
1016: polygon, and $m$-ary trees}\\
1017: The ECO-system with axiom $(s_0)=(3)$ and rule \[(k)\leadsto
1018: (3)(4)\cdots(k)(k+1)(k+2)\] is equivalent to the walk \[(k)\leadsto
1019: (0) (1)\cdots(k)(k+1)(k+2).\]
1020: The kernel is $K(z,u)=1-u+zu^3$, and the \gf \
1021: $$F(z,1)=\sum_{n\ge1} {3n \choose n} \frac{z^{n-1}}{2n+1}$$
1022: counts ternary trees ({\bf M2926}).
1023:
1024:
1025:
1026: More generally, the system with axiom $(m)$
1027: and rewriting rules
1028: \[
1029: (k)\leadsto (m)\cdots(k)(k+1)(k+2)\cdots(k+m-1)
1030: \]
1031: yields the $m$-Catalan numbers, ${mn \choose n}/((m-1)n+1)$, that count $m$-ary
1032: trees. The kernel is $1-u+zu^m$ and the generating function $F(z,1)$
1033: satisfies
1034: $F(z,1)=(1+zF(z,1))^m$.
1035: In particular, the $4$-Catalan numbers ${4n \choose n}/(3n+1)$
1036: appear in \cite{SlPl95} (sequence {\bf M3587}) and count dissections of a
1037: polygon.
1038:
1039: \end{Example}
1040:
1041:
1042: In the above examples, all backward jumps are allowed. In
1043: other words, each of these examples corresponds to an ECO-system.
1044: Let us now give an example where backward jumps of length $1$ are
1045: forbidden.
1046:
1047: \begin{Example}
1048: {}\\
1049: Consider the following modification of the Motzkin rule:
1050: \[
1051: (k)\leadsto (0)\cdots(k-2)(k+1).
1052: \]
1053: The kernel is now $K(z,u)=u(1-u)+zu-z(1-u)(u^2-1)$, and, according to
1054: \Ref{biv}, the series $F(z)=F(z,1)$ is given by $F(z)=1/[z(v_1-1)]$, where
1055: $v_1$ satisfies $K(z,v_1)=0$ and is infinite at $z=0$. Denoting
1056: $G=zF(z)$, we find that the algebraic equation defining $G$ is:
1057: $$G=z \ \frac{1+2G+G^2+G^3}{1+G}.$$
1058:
1059: \end{Example}
1060:
1061:
1062:
1063: So far, we have only dealt with walks for which the set of allowed
1064: moves was obtained by modifying the interval $[0,k]$ around
1065: $k$. One can also modify this interval around $0$: we shall see -- in
1066: examples -- that the \gf \ remains algebraic. However, it is
1067: interesting to note that in these examples, the kernel method does not
1068: {immediately} provide enough equations between the ``unknown
1069: functions'' to solve the functional equation.
1070:
1071: Let us first explain how we modify the interval $[0,k]$ around
1072: $0$. The walks we wish to count are still specified by a multiset $A$
1073: of allowed supplementary jumps and a set $B$ of forbidden backward
1074: jumps. But, in addition, we forbid backward jumps to end up in $C$,
1075: where $C$ is a given finite subset of $\ns$. In other
1076: words, the possible moves from $k$ are given by the rule
1077: $$(k)\leadsto [0, k-1] \setminus(C \cup (k-B)) \ \cup \ (k+A).$$
1078: Again, we can write a functional equation defining $F(z,u)$:
1079: \beq \label{eq-funct-C}
1080: F(z,u)=1+z\left(
1081: \frac{F(z,1)-F(z,u)}{1-u}+P(u)F(z,u)+\sum_{k=0}^{b-1}r_k(u) F_k(z)
1082: -\sum_{\ga \in C}u^\ga G_\ga (z)\right),
1083: \eeq
1084: where, as above,
1085: $$P(u)=\sum_{\al \in A} u^\al - \sum_{\be \in B} u^{-\be}
1086: \ \ \ \hbox{ and } \ \ \ r_k(u)=\sum_{\be >k, \ \be \in B}u^{k-\be},$$
1087: the new terms in the equations being
1088: $$G_\ga(z)=F(z,1)-\sum_{k=0}^{\ga} F_k(z) - \sum_{\be \in B} F_{\be +
1089: \ga}(z).$$
1090: Observe that the first three terms are the same as in the case
1091: $C=\emptyset$.
1092: The equation, once multiplied by $u^b(1-u)$, reads $K(z,u)F(z,u)=R(z,u)$
1093: where $K(z,u)$ is given by \Ref{kern} and
1094: $$R(z,u)=u^b(1-u)\left(1+
1095: \frac{zF(z,1)}{1-u}+z\sum_{k=0}^{b-1}r_k(u) F_k(z)
1096: -z\sum_{\ga \in C}u^\ga G_\ga (z)\right).$$
1097: The kernel is not modified by the introduction of $C$. As above, it
1098: has degree $a+b+1$ in $u$, and admits $b+1$ finite
1099: roots $u_0, \ldots , u_b$ around $z=0$. However, $R(z,u)$ now involves
1100: $b+1+|C|$ unknown functions, namely $F(z,1)$, the $F_k(z)$, $0\le k <b$
1101: and the $G_\ga(z)$, $\ga \in C$. The degree of $R$ in $u$ is no longer
1102: $b+1$ but $b+c+1$, where $c=\max C$.
1103: The $b+1$ roots of $K$ that can be substituted for $u$ in
1104: Eq.~\Ref{eq-funct-C}
1105: provide $b+1$ linear equations between the $b+|C|+1$ unknown
1106: functions. Additional equations will be obtained by extracting the
1107: coefficient of $u^j$ from Eq.~\Ref{eq-funct-C}, for some values of $j$. In
1108: general, we have:
1109: \beq \label{Fj}
1110: F_j(z)=[j=0]+ z\sum_{\al \in A} F_{j-\al}(z)+z[j \not \in C]
1111: \left(F(z,1)-\sum_{k=0}^jF_k(z)-\sum_{\be \in B} F_{j+\be}(z)\right).
1112: \eeq
1113: It is possible to construct a finite subset $S\subset \ns$ such
1114: that the combination of the $b+1$ equations obtained via the kernel
1115: method and the equations \Ref{Fj} written for $j \in S$ determines all unknown
1116: functions as algebraic functions of $z$ -- more precisely, as rational
1117: functions of $z$ and the roots $u_0, \ldots , u_b$ of the
1118: kernel. However, this is a long
1119: development, and these classes of walks play a marginal role in the
1120: context of ECO-systems. For these reasons, we shall merely give two
1121: examples. {The details
1122: on the general procedure for constructing the set $S$
1123: can be found in \cite{mbm}}.
1124:
1125: \begin{Example}
1126: {}\\
1127: This example is obtained by modifying the Motzkin rule of
1128: Example~\ref{MotzkinExample} around {the point} $0$. Take $A=C=\{1\}$ and
1129: $B=\emptyset$. The rewriting rule is
1130: $$(k) \leadsto (0)(2)(3) \cdots (k-1)(k+1).$$
1131: The functional equation reads
1132: \beq \label{C-ex1}
1133: (1-u+z-zu(1-u))F(z,u)=1-u+zF(z,1)-zu(1-u)G_1(z),
1134: \eeq
1135: with $G_1(z)=F(z,1)-F_0(z)-F_1(z)$. The kernel has a {\em unique\/}
1136: finite root at $z=0$:
1137: $$u_0=\frac{1+z-\sqrt{1-2z-3z^2}}{2z},$$
1138: {whereas the right-hand side of Eq.~\Ref{C-ex1} contains {\em two\/}
1139: unknown functions.}
1140: Writing Eq.~\Ref{Fj} for $j=0$ and $j=1$ yields
1141: $$F_0(z)=1+z(F(z,1)-F_0(z))\ \ \ \hbox{ and } \ \ \ F_1(z)=zF_0(z).$$
1142: These two equations allow us to express $F_0$ and $F_1$, and hence
1143: $G_1$, in terms of $F(z,1)$:
1144: $$G_1(z)=(1-z)F(z,1)-1.$$
1145: {This equation relates the two unknown functions of Eq.~\Ref{C-ex1}.}
1146: We replace $G_1(z)$ by the above expression in \Ref{C-ex1}, so that
1147: only one unknown function, namely $F(z,1)$, is left. The kernel method
1148: finally gives:
1149: $$F(z,1)=\frac{3-3z^2-2z^3-(1+z)\sqrt{1-2z-3z^2}}{2(1-z-z^2+z^3+z^4)}
1150: =1+z+2z^2+3z^3+6z^4+12z^5+O(z^6).$$
1151: \end{Example}
1152:
1153: \begin{Example}
1154: {}\\
1155: Let us choose $A=\{1\}$, $B=\{2\}$ et $C=\{2\}$.
1156: The rewriting rule is now:
1157: $$(k) \rightarrow (0)(1)(3)(4)(5) \ldots (k-3)(k-1)(k+1).$$
1158: The functional equation reads
1159: $$\left[u^2(1-u)+zu^2-zu^3(1-u)+z(1-u)\right]F(z,u)\hskip 8cm $$
1160: \beq \hskip 2cm =u^2(1-u)+
1161: zu^2F(z,1)+z(1-u)\left[F_0(z)+uF_1(z)\right] -zu^4(1-u) G_2(z)
1162: ,\label{eq2}\eeq
1163: with $G_2(z)=F(z,1)-F_0(z)-F_1(z)-F_2(z)-F_4(z)$.
1164: Only three roots, $u_0, u_1, u_2$ can be substituted for $u$ in the
1165: kernel, while the right-hand side of the equation contains four
1166: unknown functions, $F(z,1), F_0(z), F_1(z)$ and $G_2(z)$.
1167: Writing \Ref{Fj} for $j=0,1$ and $2$ yields
1168: $$\begin{array}{ll}
1169: F_0(z)& =1+z\left[ F(z,1)-F_0(z)-F_2(z)\right], \\
1170: F_1(z) &= zF_0(z) + z\left[ F(z,1)-F_0(z)-F_1(z)-F_3(z)\right],\\
1171: F_2(z)&=zF_1(z).
1172: \end{array}$$
1173: The second equation is not of much use but, by combining the first and
1174: third one, we find
1175: $$F_0(z) =\frac{1+z\left[ F(z,1)-zF_1(z)\right]}{1+z}.$$
1176: Replacing $F_0(z)$ by this expression in \Ref{eq2} gives:
1177: $$\left[u^2(1-u)+zu^2-zu^3(1-u)+z(1-u)\right]F(z,u)
1178: =u^2(1-u)+\frac{z(1-u)}{1+z}\hskip 4cm $$
1179: \beq +zF(z,1) \left[
1180: u^2+\frac{z(1-u)}{1+z}\right] +z(1-u)F_1(z) \left[
1181: u-\frac{z^2}{1+z}\right] -zu^4(1-u)G_2(z) .\label{eq2bis}\eeq
1182: We are left with three unknown functions, related by three linear
1183: equations obtained by cancelling the kernel. Solving these equations
1184: would give $F(z,1)$ as an enormous rational function of $z$, $u_0,u_1$
1185: and $u_2$, symmetric in the $u_i$. This implies that $F(z,1)$ can also
1186: be written as a rational function of $z$ and $v\equiv v_1$, the
1187: {fourth and last}
1188: root of the kernel. In particular, $F(z,1)$ is algebraic of degree at
1189: most $4$.
1190:
1191: In order to obtain directly an expression of $F(z,1)$ in terms of $z$
1192: and $v$, we can proceed as follows. Let $R'(z,u)$ denote the
1193: right-hand side of Eq.~\Ref{eq2bis}. Then $R'(z,u)$ is a polynomial in
1194: $u$ of degree $5$, and three of its roots are $u_0, u_1,
1195: u_2$. Consequently, as a polynomial in $u$,
1196: the kernel $K(z,u)$ divides $(u-v)R'(z,u)$.
1197:
1198: Let us evaluate $(u-v)R'(z,u)$ modulo $K(z,u)$: we obtain a
1199: polynomial of degree $3$ in $u$, whose coefficients depend on $z, v,
1200: F(z,1), F_1(z)$ and $G_2(z)$. This polynomial has to be zero: this
1201: gives a system of four (dependent) equations relating the three unknown
1202: functions $F(z,1), F_1(z)$ and $G_2(z)$. Solving the first three of
1203: these equations yields
1204: \begin{eqnarray*}
1205: F(z,1)&=&\frac {1+z+z^2-(z+1)zv+(z+1)zv^2-z^2v^3}
1206: {1-z^2-z(1-z^2)v+ z^3v^3}\\
1207: &=& 1+z+2z^2+3z^3+6z^4+11z^5+23z^6+47z^7+101x^8+O(z^9).\\
1208: \end{eqnarray*}
1209: Eliminating $v$ between this expression and $K(z,v)=0$ gives
1210: a quartic equation satisfied by $F(z,1)$.
1211: \end{Example}
1212: \section{Transcendental systems}
1213: \label{SectionTranscendant}
1214:
1215: \subsection{Transcendence}
1216: The radius of convergence of an algebraic series is always positive.
1217: Hence, one possible reason for a system to give a transcendental series is the
1218: fact that its coefficients grow too fast, so that its radius of convergence is
1219: zero. This is the case for
1220: System~$(2.h)$, as proved by the following proposition.
1221:
1222:
1223:
1224:
1225: \begin{Proposition} \label{zero1}
1226: Let $b$ be a nonnegative integer.
1227: For $k \ge 1$, let $m(k)=|\{i: \ e_i(k) \ge k-b\}|$.
1228: Assume that:
1229:
1230: $1.$ for all $k$, there exists a forward
1231: jump from~$k$ (i.e., $e_i(k)>k$ for some $i$),
1232:
1233: $2.$ the sequence $(m(k))_k$ is nondecreasing and tends to infinity.
1234:
1235: \noindent Then the {(ordinary)} generating function of the system has radius of
1236: convergence $0$.
1237: \end{Proposition}
1238: \Proof
1239: Let $s_0$ be the axiom of the system.
1240: {Let us denote by $h_n$ the product} $m(s_0+b)m(s_0+2b)\cdots m(s_0+nb)$. Let us
1241: prove that the generating tree contains at least $h_n$ nodes at
1242: level $n(b+1)$.
1243: At level $nb$, take a node $v$ labeled $k$, with $k \ge
1244: s_0+nb$.
1245: Such a node exists thanks to the first assumption.
1246: By definition of $m(k)$, this
1247: node $v$ has $m(k)$ sons whose label is at least $k-b$. As $m$
1248: is non decreasing, $v$ has at least $m(s_0+nb)$ sons of label at least
1249: $s_0+(n-1)b$. Iterating this procedure shows that, at level
1250: $nb+i$, at least $m(s_0+(n-i+1)b)\cdots m(s_0+nb)$ descendants
1251: of $v$ have a label larger than or equal to $s_0+(n-i)b$, for $0<
1252: i \le n $. In particular, for $i=n$, we obtain at level $n(b+1)$ at
1253: least $h_n$ descendants of $v$ whose label is at least $s_0$.
1254:
1255: Hence $f_{n(b+1)} \ge h_n$. But as $h_{n}/h_{n-1} = m(s_0+nb)$
1256: goes to infinity with $n$, the series $\sum_n h_n z^{n(b+1)}$ has radius of
1257: convergence $0$, and the same is true for $F(z)=\sum_n f_nz^n$.
1258: \qed
1259:
1260: In particular, this proposition implies that {\em the \gf \ of any
1261: ECO-system in which the length of backward jumps is bounded has
1262: radius of convergence $0$.} Many examples of this type will be given
1263: in the next subsection, in which we shall study whether the
1264: corresponding \gf \ is holonomic or not.
1265: The following example, in which backward jumps are not bounded,
1266: was suggested by Nantel Bergeron.
1267:
1268: \begin{Example}
1269: {\em A fake factorial walk} \\
1270: Consider the system with axiom $(1)$ and rewriting rules
1271: $\{(k) \leadsto (2)(4) \cdots (2k)\}$. Proposition~\ref{zero1}
1272: applies with $b=0$ and $m(k)=1+ \lfloor k /2 \rfloor$. Note that
1273: the radius of convergence of $F(z)$ is zero
1274: although {\em all\/} the functions $e_i$ are bounded, and indeed
1275: constant: $e_i(k)=2i$ for all $k \ge i$.
1276: The series $F(z)$ is of course
1277: transcendental. Note, however,
1278: that $F(z,u)$ satisfies a functional equation that is at first sight
1279: reminiscent of the equations studied in Section \ref{SectionAlgebraic}:
1280: $$F(z,u)=u+ zu^2 \ \frac{F(z,1)-F(z,u^2)}{1-u^2}.$$
1281: \end{Example}
1282:
1283: The following example shows that Proposition~\ref{zero1} is not far
1284: from optimal: an ECO-system in which all
1285: functions $e_i$ grow linearly can have a finite radius of convergence.
1286:
1287:
1288: \begin{Example}
1289: {}\\
1290: The system with axiom $(1)$ and rules $(k) \leadsto (\lceil k/2
1291: \rceil)^{k-1}(k+1)$ leads to a \gf \ with a positive radius of
1292: convergence.\\
1293: Let us start from the recursion defining the numbers $f_{n,k}$. We
1294: have $f_{0,1}=1$ and for $n \ge 1$,
1295: $$f_{n+1,k}=f_{n,k-1}+(2k-1)f_{n,2k}+(2k-2)f_{n,2k-1}.$$
1296: The largest label occurring at level $n$ in the tree is $n+1$. Let us
1297: introduce the numbers $g_{n,k}=f_{n,n-k+1}$, for $k \le n$. The above
1298: recursion can be rewritten as:
1299: \beq g_{n+1,k}=g_{n,k}+(2n-2k+3)g_{n,2k-n-3}+(2n-2k+2)g_{n,2k-n-2}.
1300: \label{g} \eeq
1301: We have $g_{n,k}=0$ for $k<0$. Hence Eq.~\Ref{g} implies that for
1302: $k \ge 0$, the sequence $(g_{n,k})_{n}$ is nondecreasing and reaches a
1303: constant value $g(k)$ as soon as $n\ge 2k-1$ {(see Table~\ref{fgtable}).}
1304:
1305: \begin{table}
1306: $$\begin{array}{l|ccccccc}
1307: n \ k& 1 &2 &3 &4 &5 &6 \\
1308: \hline
1309: 0 &\bf{1} \\
1310: 1 &\bf{0} &\bf{1} \\
1311: 2 & 1 &\bf{0} &\bf{1} \\
1312: 3 &0 &\bf{3} &\bf{0} &\bf{1} \\
1313: 4 &3 &3 &\bf{3} &\bf{0} &\bf{1} \\
1314: 5 &3 &9 &\bf{7} &\bf{3} &\bf{0} &\bf{1} \\
1315: \end{array}
1316: \hskip 2cm
1317: \begin{array}{l|ccccccc}
1318: n \ k& 0 &1 &2 &3 &4 &5 \\
1319: \hline
1320: 0 &\bf{1} \\
1321: 1 &\bf{1} &\bf{0} &\\
1322: 2 &\bf{1} &\bf{0} & 1 \\
1323: 3 &\bf{1} &\bf{0} &\bf{3} &0\\
1324: 4 &\bf{1} &\bf{0} &\bf{3} &3 &3 \\
1325: 5 &\bf{1} &\bf{0} &\bf{3} &\bf{7}&9 &3 \\
1326: \end{array}
1327: $$
1328: \caption{The numbers $f_{n,k}$ and $g_{n,k}$. Observe the convergence
1329: of the coefficients.}
1330: \label{fgtable}
1331: \end{table}
1332: Going back to the number $f_n$ of nodes at level $n$, we have $$f_n
1333: =\sum_{k=0}^n g_{n,k} \le \sum_{k=0}^n g(k).$$
1334: But $$\sum_{n\ge 0}z^n \sum_{k=0}^n g(k) = \frac{1}{1-z}
1335: \sum_{k=0}^n g(k)z^k,$$
1336: and hence it suffices to prove that the \gf \ for the numbers $g(k)$
1337: has a finite radius of convergence, that is, that these numbers grow
1338: at most exponentially.
1339:
1340: Writing \Ref{g} for $n+1=2k-i$, for $1\le i\le k$, we obtain:
1341: $$g_{2k-i,k}=g_{2k-i-1,k}
1342: +(2k-2i+1)g_{2k-i-1,i-2}+(2k-2i)g_{2k-i-1,i-1}.$$
1343: Iterating this formula for $i$ between $1$ and $k$ yields
1344: \begin{eqnarray*}
1345: g(k) &=& g_{2k-1,k}= \sum_{i=1}^k \left[
1346: (2k-2i+1)g_{2k-i-1,i-2}+(2k-2i)g_{2k-i-1,i-1}\right]\\
1347: &\le & \sum_{i=1}^k \left[
1348: (2k-2i+1)g(i-2)+(2k-2i)g(i-1)\right]=
1349: \sum_{i=0}^{k-2} (4k-4i-5)g(i).
1350: \end{eqnarray*}
1351: This inequality, together with the fact that $g(0)=1$, implies that
1352: for all $k \ge 0$, $g(k)\le \tilde g(k)$, where the sequence $\tilde
1353: g(k)$ is defined by $\tilde g(0)=1$ and $\tilde g(k) =
1354: \sum_{i=0}^{k-2} (4k-4i-5)\tilde g(i)$ for $k >0$. But the series
1355: $\sum_k \tilde g(k) z^k$ is rational, equal to
1356: $(1-z)^2/(1-2z-2z^2-z^3)$, and has a finite radius of
1357: convergence. Consequently, the numbers $\tilde g(k)$ and $ g(k) $
1358: grow at most exponentially.
1359: \end{Example}
1360:
1361: \medskip
1362: Algebraic generating functions are strongly constrained in their
1363: algebraic structure (by a polynomial equation) as well as in their
1364: analytic structure (in terms of singularities and asymptotic
1365: behaviour). In particular, they have a finite number of
1366: singularities, which are algebraic numbers, and they admit local asymptotic
1367: expansions that involve only rational exponents.
1368: {\em A contrario\/}, a generating function that has infinitely many
1369: singularities ({\em e.g.}, a natural boundary) or that involves a
1370: transcendental element ({\em e.g.}, a logarithm) in a local asymptotic
1371: expansion is by necessity transcendental; see~\cite{Flajolet87}
1372: for a discussion of such transcendence criteria. In the case of
1373: generating trees,
1374: this means that the presence of a condition involving a transcendental
1375: element is expected to lead to a transcendental generating function.
1376: This is the case in the following example.
1377:
1378: \begin{Example}{\em A Fredholm system}\label{FredholmExample}\\
1379: We examine System~$(2.e)$, in which
1380: the rules are irregular at powers of~2:
1381: $$ (s_0)=(2), \ \ \ \ (k) \leadsto (2)^{k-2} (3 - [\exists p\!:\!k=2^p])
1382: (k+1), \ \ \ k\ge 2.$$
1383: This example will involve the Fredholm series
1384: $h(z):=\sum_{p\ge 1} z^{2^p}$, which is well-known to admit the unit
1385: circle as a natural boundary. (This can be seen by way of the
1386: functional equation $h(z)=z^2+h(z^2)$, from which there results that
1387: $h(z)$ is infinite at all iterated square-roots of unity.)
1388: According to Eq.~\Ref{equationF_k}, we have, for $k>3$,
1389: $F_k(z)=zF_{k-1}(z),$
1390: so that $$F_k(z)=z^{k-3}F_3(z) \ \ \ \hbox{ for } k\ge 3.$$
1391: Now, writing Eq.~\Ref{equationF_k} for $k=2$ gives
1392: \begin{eqnarray*}
1393: F_2(z)&=&1+z\sum_{k\ge 3} (k-2)F_k(z) +z \sum_{p\ge 1} F_{2^p}(z)\\
1394: &=& 1+\frac{z}{(1-z)^2} F_3(z) +z F_2(z) + F_3(z) \left(
1395: \frac{h(z)}{z^2}-1\right)\\
1396: &=& 1+zF_2(z) + F_3(z) \left( \frac{z}{(1-z)^2} +
1397: \frac{h(z)}{z^2}-1\right).
1398: \end{eqnarray*}
1399: For $k=3$, we obtain:
1400: \begin{eqnarray*}
1401: F_3(z)&=&zF_2(z)+z\sum_{k\ge 3, \ k \not = 2^p} F_k(z)\\
1402: &=& zF_2(z)+F_3(z)\left( \frac{1}{1-z}-\frac{h(z)}{z^2}\right).
1403: \end{eqnarray*}
1404: Solving for $F_2(z)$ and $F_3(z)$, then summing ($F(z) = F_2(z) +
1405: F_3(z) / (1-z)$), we obtain:
1406: \[
1407: F(z) = \frac { (1-z)^2 h(z)}
1408: { (1-2z) (1-z)^2 h(z) - z^4 }=1+2z+5z^2+14z^3+39z^4+108z^5+O(x^6)
1409: .
1410: \]
1411: The functions $h(z)$ and $F(z)$ are rationally related, so that
1412: $F(z)$ is itself transcendental. The series $h$ has radius $1$, but
1413: the denominator of $F$ vanishes before $z$ reaches $1$ -- actually,
1414: before $z$ reaches $1/2$. Hence the radius of $F$ is the smallest root
1415: of its denominator. Its value is easily
1416: determined numerically and found to be about 0.360102.
1417: \end{Example}
1418:
1419: \subsection{Holonomy}
1420: In the transcendental case, one can also discuss the {\em
1421: holonomic\/} character of the generating function $F(z)$.
1422:
1423:
1424: A series is said to be
1425: {\em holonomic\/}, or {\em D-finite\/} \cite{stanley}, if it satisfies
1426: a linear differential equation with polynomial coefficients in
1427: $z$. Equivalently, its coefficients $f_n$ satisfy a linear recurrence
1428: relation with polynomial coefficients in $n$. Consequently, given a
1429: sequence $f_n$, the ordinary \gf \ $\sum_n f_n z^n$ is holonomic if
1430: and only if the exponential \gf \ $\sum_n f_n z^n /n!$ is holonomic.
1431: The set of holonomic series has nice closure properties: the sum or product of
1432: two of them is still holonomic, and the substitution of an algebraic
1433: series into an holonomic one gives an holonomic series.
1434: Holonomic series include algebraic series, and have a
1435: finite number of singularities.
1436: This implies that Example~\ref{FredholmExample}, for which $F(z)$ has
1437: a natural boundary, is not holonomic.
1438:
1439: We study below five ECO-systems that, at first sight, do not
1440: look to be very
1441: different. In particular, for each of them, forward
1442: and backward jumps are bounded. Consequently, Proposition~\ref{zero1}
1443: implies that the corresponding ordinary \gf \ has radius of convergence
1444: zero. However, we shall see that the first three systems have an
1445: holonomic \gf , while the last two have not. We have no general
1446: criterion that would allow us to distinguish between systems leading
1447: to holonomic generating functions and those leading to
1448: nonholonomic generating functions.
1449:
1450: Among the systems with bounded jumps, those for which $e_i(k)-k$
1451: belongs
1452: to $\{-1,0,1\}$ for all $i\le k$
1453: have a nice property: the \gf \ for the corresponding {\em
1454: excursions\/} (walks starting and ending at level $0$) can be written
1455: as the following continued fraction \cite{flajolet-fracont}:
1456: $$\frac{1}{\displaystyle 1-b_0z-
1457: \frac{\displaystyle a_1c_0z^2}{\displaystyle 1-b_1z-
1458: \frac{\displaystyle a_2c_1z^2}{\displaystyle 1-b_2z-
1459: \frac{\displaystyle a_3c_2z^2}{\cdots}}}},$$
1460: where the coefficients $a_k, b_k$ and $c_k$ are the multiplicities appearing
1461: in the rules, which read $(k) \leadsto (k-1)^{a_k}(k)^{b_k}(k+1)^{c_k}.$
1462:
1463:
1464: \begin{Example} \label{Arrangements}{\em Arrangements}\\
1465: The system $(k) \leadsto (k)(k+1)^{k-1}$ with axiom~$(s_0)=(2)$ generates a
1466: sequence that starts with $1,2,5,16,65, 326$
1467: ({\bf M1497}). It is not hard to see that the
1468: triangular array $f_{n,k+2}$ is given by the arrangement
1469: numbers $k!{n \choose k}$, so that the {\em exponential\/} generating
1470: function (EGF) of
1471: the sequence is $$\tilde F(z,u)
1472: = \sum_{n\ge 0,k\ge 2} f_{n,k}u^k \frac{z^n}{n!}
1473: =\frac{u^2 e^z}{1-uz}.$$ This system satisfies the
1474: conditions of Proposition~\ref{zero1} with $b=0$ and
1475: $m(k)=k$. Accordingly, one has $f_n\sim e\, n!$, so that the {\em ordinary\/}
1476: generating function $F(z)$ has radius of convergence~0 and cannot be
1477: algebraic. However, $\tilde F(z,1)=e^z/(1-z)$ is holonomic, and so is
1478: $F(z)$.
1479: \end{Example}
1480:
1481: \begin{Example} \label{ExampleInvolutions}{\em Involutions and
1482: Hermite polynomials}\\
1483: The system $(k) \leadsto (k-1)^{k-1}(k+1)$ with axiom~$(s_0)=(1)$
1484: generates a sequence that starts with $1,1,2,4, 10,26, 76$ ({\bf
1485: M1221}). These numbers count involutions: more precisely, one easily
1486: derives from the recursion satisfied by the coefficients $f_{n,k}$ that
1487: $f_{n,k}$ is the number of involutions on $n$ points, $k-1$ of which
1488: are fixed. Proposition~\ref{zero1} applies with $b=1$ and $m(k)=k$.
1489:
1490: The corresponding EGF is
1491: \beq \tilde F(z,u)= \sum_{n\ge 0,k\ge 1} f_{n,k}u^k \frac{z^n}{n!}
1492: = u\exp\left (zu+\frac{z^2}{2}\right),\label{inv-egf} \eeq
1493: and its value at $u=1$ is holonomic.
1494:
1495: The polynomials $f_n(u)=\sum_k f_{n,k}u^k$ counting involutions on $n$
1496: points are in fact closely related to the Hermite polynomials, defined by:
1497: $$\sum_{n\geq 0} H_n(x) \frac{t^n}{n!}=\exp\left(xt-\frac{t^2}{2}\right).$$
1498: Indeed, comparing the above identity with \Ref{inv-egf} shows that
1499: $ f_n(u)=u\ i^nH_n(-i u).$
1500: \end{Example}
1501:
1502:
1503:
1504:
1505: \begin{Example}{\em Partial permutations and Laguerre polynomials}
1506: \label{ExampleLaguerre}\\
1507: The rewriting rule $(k)\leadsto (k+1)^{k-1}
1508: (k+2)$, taken with axiom $(2)$, generates a sequence that starts with
1509: $1,2,7,34,209,...$ ({\bf M1795}). From the recursion satisfied by the
1510: coefficients $f_{n,k}$, we derive that $f_{n,n+k}$ is the number of
1511: partial injections of $\{1,2,\ldots , n\}$ into itself in which $k-2$
1512: points are unmatched.
1513: From this, we obtain:
1514: $$\tilde F(z,u)=
1515: \frac{u^2}{1-uz}\exp\left( \frac{u^2z}{1-uz}\right)=u^2\sum_{n\geq 0}
1516: L_n(-u) \frac{(uz)^n}{n!} $$
1517: where $L_n(u)$ is the $n$th Laguerre polynomial.
1518: Again, $\tilde F(z,1)$ is holonomic.
1519: \end{Example}
1520:
1521:
1522: \noindent The next two systems, as announced,
1523: lead to nonholonomic generating functions.
1524:
1525: \begin{Example}\label{ExampleStirling}{\em Set partitions and
1526: Stirling polynomials}\\
1527: Let us consider the system $[(1),(k)\leadsto (k)^{k-1}(k+1)]$. From
1528: the recursion satisfied by the coefficients $f_{n,k}$, we derive that
1529: $f_{n,k+1}$ is equal to the Stirling number of
1530: the second kind $\stir{n}{k}$, which counts partitions of $n$ objects
1531: into $k$ nonempty subsets. The corresponding EGF is
1532: $$\tilde F(z,u) =
1533: u \exp\left(u(\exp{z}-1)\right).$$
1534: At $u=1$, this \gf \ specializes to
1535: $$\tilde
1536: F(z,1)=\exp(\exp(z)-1))=\sum_{n \ge 0} B_n \frac{z^n}{n!}=
1537: 1+z+2\frac{z^2}{2!}+5
1538: \frac{z^3}{3!}+15\frac{z^4}{4!}+52\frac{z^5}{5!}+203\frac{z^6}{6!}+\dots$$
1539: This is the exponential generating function of the Bell
1540: numbers ({\bf M1484}). It is known that $\log B_n =n \log n -n \log
1541: \log n +O(n)$ (see~\cite{Odlyzko}), and this
1542: cannot be the asymptotic behaviour of the logarithm of the
1543: coefficients of an holonomic
1544: series (see~\cite{WiZe85b} for admissible types).
1545: Hence, $\tilde
1546: F(z,1)$, as well as $F(z,1)$, is nonholonomic.
1547: \end{Example}
1548:
1549: \begin{Example}\label{ExampleBessel}{\em Bessel numbers}\\
1550: We study the system with axiom~(2) and rewriting rules
1551: \beq
1552: (2)\leadsto(2)(3),\ \ \ (k)\leadsto (k-1)(k)^{k-2}(k+1), \ \ k\ge
1553: 3.
1554: \label{bess}
1555: \eeq
1556: We shift the labels by $2$ to obtain a walk model with
1557: axiom $(0)$ and rules
1558: $$ (0)\leadsto (0)(1), \ \ \ (k)\leadsto (k-1)(k)^k(k+1), \ \ k\ge
1559: 1.$$
1560: The corresponding bivariate \gf \ $F(z,u)$
1561: satisfies the functional differential equation
1562: \[
1563: F(z,u)\left(1-z(u+u^{-1})\right)
1564: =1+z(1-u^{-1})F(z,0)+zu\frac{\partial F}{\partial u}(z,u),
1565: \]
1566: which is certainly not obvious to solve.
1567: However, as observed in \cite{flajolet-fracont}, it is easy to obtain
1568: a continued fraction expansion of the excursion \gf :
1569: \[
1570: F(z,0)=1+z+2z^2+4z^3+9z^4+\cdots=\cfrac{1}{1-z-
1571: \cfrac{z^2}{1-z-
1572: \cfrac{z^2}{1-2z-
1573: \cfrac{z^2}{1-3z-
1574: \ddots
1575: }}}}=\frac{1}{1-z-z^2B(z)},
1576: \]
1577: where $B(z)=\sum_n B^*_nz^n=1+z+2z^2+5z^3+14z^4+43z^5+143z^6+\cdots
1578: $
1579: is the generating function of Bessel numbers ({\bf M1462}) and counts
1580: non-overlapping partitions \cite{FlSc90}.
1581: As $F(z,0)$ itself, the series $B(z)$ has radius of convergence
1582: zero. The fast increase of $B^*_n$ entails
1583: \[
1584: [z^n]F(z,0)\sim B_{n-2}^*.
1585: \] {}From~\cite{FlSc90}, we know that $\log B^*_n = n \log n -n \log
1586: \log n +O(n).$
1587: Again, this prevents $F(z,0)$ from being holonomic.
1588:
1589: In order to prove that $F(z,1)$ itself is nonholonomic, we are going
1590: to prove that its coefficients $f_n$ have the same asymptotic
1591: behaviour as the coefficients of $F(z,0)$. Clearly,
1592: $$[z^n]F(z,0) =f_{n,0} \le\sum_k f_{n,k} =f_n.$$
1593: To find an upper bound for $f_n$, we compare the system~\Ref{bess}
1594: (denoted $\Sigma_1$ below) to
1595: the system $\Sigma_2$ with axiom $(2)$ and rule $(k)\leadsto
1596: (k)^{k-1}(k+1)$. This system generates a tree with counting sequence
1597: $g_n$. The form of the rules implies that the (unlabeled) tree associated
1598: with $\Sigma_1$ is a subtree of the tree associated with $\Sigma_2$. Hence
1599: $f_n \le
1600: g_n$. Comparing $\Sigma_2$ to the system studied in the previous example
1601: shows that $g_n$ is the Bell number $ B_{n+1}$, the logarithm of which
1602: is also known to be $n \log n -n \log
1603: \log n +O(n)$ (see~\cite{Odlyzko}). Hence $\log f_n = n \log n -n \log
1604: \log n +O(n),$ and this prevents the series $F(z,1)$ from being holonomic.
1605: \end{Example}
1606:
1607:
1608: \newpage
1609: \noindent {\bf{A small catalog of ECO-systems}}\\
1610: To conclude, we present in Table~2 a small catalog of ECO-systems that
1611: lead to sequences of
1612: combinatorial interest. Several examples are detailed in the paper;
1613: others are due to West \cite{West95,West96} or Barcucci, Del Lungo,
1614: Pergola, Pinzani \cite{ECOArticle,BaDePePiFPSAC98,BaDePePi,BaDePePi_2},
1615: or are folklore. Each of them is an instance of
1616: application of our criteria.
1617:
1618:
1619:
1620: \begin{table}[t]
1621: \begin{small}
1622: \begin{tabular}{|c|l|l|c|c|}
1623: \hline
1624: Axiom & System & Name & Id. & Generating Function \\
1625: \hline
1626: &\bf{Rational OGF} &&& {\bf OGF}\\
1627: $(1)$ & $(k) \leadsto (k)^{k-1} ((k \bmod 2)+1)$ &
1628: Ex. \ref{FibonacciExample}: Fibonacci & M0692 & $\frac{1}{1-z-z^2}$
1629: \makebox(0,12)[0cm]{}
1630: \\
1631: $(2)$ & $(k) \leadsto (2)^{k-1} (k+1)$ & Ex. \ref{FiboBisectionExample}: even Fibonacci & M1439 & $\frac{1-z}{1-3z+z^2}$
1632: \makebox(0,12)[0cm]{}
1633: \\
1634: $(3)$ & $(k) \leadsto (2)^{k-1} (k+1)$ &
1635: Ex. \ref{FiboBisectionExample}: odd Fibonacci & M2741 & $\frac{1}{1-3z+z^2}$
1636: \makebox(0,12)[0cm]{}
1637: \\
1638: &&&&\\
1639: \hline
1640: &\bf{Algebraic OGF} &&& {\bf OGF}\\
1641: $(1)$ & $(k) \leadsto (1) % (2)
1642: \cdots (k-1) (k+1)$ & Ex. \ref{MotzkinExample}: Motzkin numbers & M1184 &
1643: $\frac{1-z-\sqrt{1-2z-3z^2}}{2z^2}$
1644: \makebox(0,12)[0cm]{}
1645: \\
1646: $(2)$ & $(k) \leadsto (2) % (3)
1647: \cdots (k) (k+1)$ & Ex. \ref{CatalnExample}: Catalan numbers & M1459 &
1648: $\frac{1-2z-\sqrt{1-4z}}{2z^2}$
1649: \makebox(0,12)[0cm]{}
1650: \\
1651: $(3)$ & $(k) \leadsto (3) % (4)
1652: \cdots (k) (k+1)^2$ & Ex. \ref{SchroederExample}: Schr\"oder numbers & M2898 &
1653: $\frac{1-3z-\sqrt{1-6z+z^2}}{4z^2}$
1654: \makebox(0,12)[0cm]{}
1655: \\
1656: $(4)$ & $(k) \leadsto (4) % (5)
1657: \cdots (k) (k+1)^3$ &\hskip 20mm --- & M3556 &
1658: $\frac{1-4z-\sqrt{1-8z+4z^2}}{6z^2}$
1659: \makebox(0,12)[0cm]{}
1660: \\
1661: $(m)$ & $(k) \leadsto (m) % (m\!+\!1)
1662: \cdots (k) (k+1)^{m-1}$ &\hskip 20mm --- &--- &
1663: $\frac{1-mz-\sqrt{1-2mz+(m-2)^2z^2}}{2(m-1)z^2}$
1664: \makebox(0,12)[0cm]{}
1665: \\
1666: $(3)$ & $(k)\leadsto (3) % (4)
1667: \cdots % (k) (k+1)
1668: (k+2)$ & Ex. \ref{TernaryExample}: Ternary trees & M2926 &
1669: $F=(1+zF)^3$
1670: \makebox(0,12)[0cm]{}
1671: \\
1672: $(4)$ & $(k)\leadsto (4)% (5)
1673: \cdots % (k+1) (k+2)
1674: (k+3)$ & Ex. \ref{TernaryExample}: Dissections of a polygon & M3587 &
1675: $F=(1+zF)^4$
1676: \\
1677: $(m)$ & $(k)\leadsto (m)\cdots% (k)(k+1)(k+2)\cdots
1678: (k+m-1)$ & Ex. \ref{TernaryExample}: $m$-ary trees & &
1679: $F=(1+zF)^m$
1680: \\
1681: &&&&\\
1682: \hline
1683: &\bf{Holonomic} &&& {\bf EGF}\\
1684: &{\bf transcendental OGF} &&&\\
1685:
1686: $(1)$ & $(k) \leadsto (k+1)^k$ & Permutations & M1675 & $1/(1-z)$ \\
1687:
1688: $(2)$ & $(k) \leadsto (k)(k+1)^{k-1}$ & Ex. \ref{Arrangements}:
1689: Arrangements & M1497 & $e^z/(1-z)$ \\
1690:
1691:
1692: $(1)$ & $(k) \leadsto (k-1)^{k-1} (k+1)$ &
1693: Ex. \ref{ExampleInvolutions}: Involutions & M1221 &
1694: $e^{z+\frac{1}{2}z^2}$\\
1695:
1696:
1697:
1698: $(2)$ & $(k) \leadsto (k+1)^{k-1} (k+2)$ & Ex.~\ref{ExampleLaguerre}:
1699: Partial permutations &
1700: M1795 & $e^{z/(1-z)}/(1-z)$ \\
1701:
1702: $(2)$ & $(k) \leadsto (k-1)^{k-2} (k) (k+1)$ & Switchboard problem & M1461 & $e^{2z+\frac{1}{2}z^2}$\\
1703:
1704:
1705: $(2)$ & $(k) \leadsto (k-1)^{k-2} (k+1)^2$ & Bicolored involutions & M1648 & $e^{2z+z^2}$\\
1706:
1707:
1708: &&&&\\
1709: \hline
1710: &\bf{Nonholonomic OGF} && & {\bf EGF}\\
1711: $(1)$ & $(k) \leadsto (k)^{k-1} (k+1)$ & Ex. \ref{ExampleStirling}:
1712: Bell numbers & M1484 &
1713: $e^{e^z-1}$\\
1714: $(2)$ & $(k) \leadsto (k)^{k-2} (k+1)^2$ & Bicolored partitions & M1662
1715: & $e^{2(e^z-1)}$\\
1716: $(2)$ & $(k)\leadsto (k-1)(k)^{k-2}(k+1)$ & Ex. \ref{ExampleBessel}: Bessel numbers & M1462
1717: & --- \\
1718: \hline
1719: \end{tabular}
1720: \end{small}
1721: \caption{ Some ECO-systems of combinatorial interest.}
1722: \end{table}
1723:
1724: \bigskip\noindent
1725: {\bf Acknowledgements.}
1726: We thank Elisa Pergola and Renzo Pinzani who presented us the problem
1727: we deal with in this paper. We are also very grateful for helpful
1728: discussions with Jean-Paul Allouche.
1729:
1730:
1731:
1732: \newpage
1733:
1734: \bibliographystyle{plain}
1735:
1736: \bibliography{eco}
1737:
1738: \newpage
1739: \bigskip
1740:
1741: \noindent {\em Cyril Banderier, Philippe Flajolet} \\
1742: Projet Algorithmes\\
1743: INRIA Rocquencourt\\
1744: F-78153 Le Chesnay\\
1745: {\sc France}\\
1746: {\tt Cyril.Banderier@inria.fr}\\
1747: {\tt Philippe.Flajolet@inria.fr}
1748: \bigskip
1749:
1750: \noindent {\em Mireille Bousquet-M\'elou}\\
1751: LaBRI, Universit\'e Bordeaux 1\\
1752: 351 cours de la Lib\'eration\\
1753: F-33405 Talence Cedex\\
1754: {\sc France}\\
1755: {\tt bousquet@labri.u-bordeaux.fr}
1756:
1757: \bigskip
1758:
1759: \noindent {\em Alain Denise, Dominique Gouyou-Beauchamps}\\
1760: LRI, B\^atiment 490\\
1761: Universit\'e Paris-Sud XI \\
1762: F-91405 Orsay Cedex\\
1763: {\sc France}\\
1764: {\tt Alain.Denise@lri.fr}\\
1765: {\tt dgb@lri.fr}
1766:
1767: \bigskip
1768:
1769: \noindent {\em Dani\`ele Gardy}\\
1770: Universit\'e de Versailles/Saint-Quentin \\
1771: Laboratoire PRISM\\
1772: 45, avenue des \'Etats-Unis\\
1773: F-78035 Versailles Cedex\\
1774: {\sc France}\\
1775: {\tt Daniele.Gardy@prism.uvsq.fr}
1776:
1777:
1778: \end{document}
1779:
1780:
1781:
1782: