On the Transcendence of Formal Power Series
JeanPaul Allouche
L.R.I., Université ParisSud
Algorithms Seminar
December 1, 1997
[summary by Philippe Flajolet]
A properly typeset version of this document is available in
postscript and in
pdf.
If some fonts do not look right on your screen, this might be
fixed by configuring your browser (see the documentation here).
1 Introduction
Algebraicity of generating functions (gf's) is of interest in combinatorial
analysis as it is a sure sign of strong structural properties.
For instance, any (unambiguous) contextfree model leads to
algebraic generating functions; in particular
generating functions of simple families of trees and
random walks (defined by a finite set of node degrees or jumps) are
algebraic. In another context, the algebraic character of the gf's
associated with
2dimensional directed animals in percolation theory points to a
wealth of puzzling combinatorial bijections; see [7] for a
specific illustration.
Conversely, a transcendence result for the gf of a combinatorial class
C means a sort of ``structural complexity lower bound'' on C. For instance, elements of C cannot be encoded by an
unambiguous contextfree grammar. Accordingly, if C already
admits contextfree descriptions, all such descriptions must be inherently
ambiguous.
Methods for establishing the transcendence of generating functions
fall broadly into two categories.

Arithmetic methods are based on numbertheoretic
properties of coefficients.
The most famous criterion in this range is Eisenstein's criterion:
If a series of Q[[z]] is algebraic, then the denominators
of its coefficients contain only finitely many primes.
For instance, f(z)=exp(z) is
transcendental ``because'' its coefficients
f_{n}=1/n! have denominators that contain infinitely many
primes (by Euclid's theorem!).
 Analytic methods are based on the presence of a
transcendental element in a local behaviour, usually taken at a
singular point.
In this perspective, f(z)=exp(z) is
transcendental ``because'' its growth is too fast at infinity, a fact
incompatible with the fact that an algebraic function is locally
described by a Puiseux series (i.e.,
a series involving fractional powers).
The analytic approach is reviewed in [6].
The talk focuses on the arithmetic method, and more specifically on
the following powerful
approach [2, 3, 4, 10].
Principle 1 If f(z)=å_{n} f_{n} z^{n} has integer coefficients
and is algebraic over Q(z), then its reduction
(f(z)mod p):=å (f_{n} mod p)
z^{n} is algebraic over F_{p}(z).
Principle 2 For a series g(z)=å g_{n} z^{n} over a
finite field F_{p},
the following three properties are equivalent:

(i) the correspondence n® g_{n} is computable by a
finite automaton
that inputs the basep representation of n (``the g_{n} are automatic'');
 (ii) the infinite word (g_{0},g_{1},...) is generated by
a regular (length homogeneous) substitution;
 (iii) g(z) is algebraic over F_{p}(z).
This is the classical ``ChristolKamaeMendès FranceRauzy
Theorem'' [4, 5], the equivalence between (i) and (ii)
being due to Cobham in 1972.
For instance, the Catalan gf,
f(z)= 

=z+z^{2}+2z^{3}+5z^{4}+14z^{5}+42z^{6}+132z^{7}+429z^{8}+···

has a reduction modulo 2
g(z)=z+z^{2}+z^{4}+z^{8}+···
where the coefficient g_{n} is 1 exactly when n=2^{r}. Thus the
coefficient sequence is
computable by a finite automaton from the binary representation of
the index n. It is also generated starting from the letter a
by the regular substitution
a® a1, 1® 10, 0® 00.
2 Primitive words
An example originally due to Petersen serves to illustrate nicely the
methods just introduced. Say that a word over some alphabet is
primitive if it is not a ``power'', that is, the repetition of
a shorter pattern. Thus abbab is primitive while abbabbabb is not.
Let m³2 be the alphabet cardinality, W(z)=(1mz)^{1} the gf
of all words, and P(z) the gf of primitive words. Then, since each
word has a ``root'', one has
W(z)=P(z)+P(z^{2})+P(z^{3})+···,
so that, with µ(n) the Moebius function,
P(z)= 

µ(d) W(z^{d}),
P_{n}= 

µ(d) m^{n/d}.

In particular, the reduction modulo m yields


=µ(n)+A· m º µ (n) mod m.

Thus, the problem is reduced to showing that µ(n) is the
coefficient sequence of a transcendental series.
Now, by a theorem a Cobham, if a sequence has an algebraic gf over a finite
field, and if it assumes some fixed value with a limit density d,
then
d is a rational number. (Think of the characterization by
finite automata.) But, here, µ(n)=1 whenever n is squarefree, an
event whose density is 6/p^{2}. The transcendence of å_{n}
µ(n)z^{n} then follows from the irrationality of p.
Reduction modulo m thus provides a proof of the fact that the
language of all primitive words cannot be an unambiguous context free
language.
In the analytic perspective, transcendence results from the fact that
P(z) has infinitely many poles inside the unit circle. Such poles,
at points m^{1/r}exp(2ikp/r), arise from W(z) and the
Moebius inversion formula for P(z).
3 Stanley's conjecture
In his fundamental paper of 1980 on Dfinite series,
Stanley [9] conjectured that the binomial series
is transcendental for any integers t³2. Of course, we have
B_{1}(z)=1/(14z)^{1/2}.
In the case of even t, B_{t} is clearly transcendental given the presence of
logarithmic elements induced by the asymptotic form of coefficients,
In addition B_{2} is also known to be an elliptic integral.
The case of odd t is harder. An analytic proof was
suggested by Flajolet [6] in 1987 and an algebraic proof
was given by Woodcock and Sharif [10] in 1989.
The proof of [10] consists in reducing first B_{t}(z) modulo
a prime p. The resulting series is algebraic,
since a theorem of Furstenberg states that algebraic functions over
finite fields are closed under Hadamard (termwise) products. (This property
is also clear from the characterization by finite automata.)
However, by means of arguments from algebraic number
theory, Woodcock and Sharif are able to estimate the degree of
(B_{t}(z)mod p)
over F_{p}(z) and deduce that there exists an infinity of special
prime
values of p for
which this degree grows without bound. This in turn implies
the transcendence of B_{t}(z).
In contrast,
from the analytic standpoint,
it is the examination of the Puiseux expansion
of B_{t}(z) near its singularity z=4^{t} that leads to
the transcendence result via the arithmetic transcendence of the
number p.
4 Miscellaneous examples
There are a great many cases where reduction modulo a prime leads to
transcendence results for generating functions. Here are a few examples.
In [6], the language {a^{n}bv_{1}a^{n}v_{2}} was shown to
be inherently ambiguous through transcendence of
since poles accumulate near 1/2. Alternatively, simple manipulations
show that, modulo 2, the transcendence of S(z) is equivalent to the
transcendence of the divisor series
The latter form is transcendental over F_{2}(z) since,
upon reduction modulo 2, it is the
indicator series of squares, and squares are known not to be automatic
(Minsky).
A similar process applies to the Goldstine language whose gf involves
the theta function
Q(z)=å_{n³0} z^{n(n+1)/2}, and to the partition series
P(z)=Õ (1z^{n})^{1} whose logarithmic derivative is closely
related to divisor functions.
An amusing example due to Allouche, Betrema, and Shallit is the
``Bourbaki definition of integers''
Ø, {Ø}, {Ø,{Ø}}, {Ø,{Ø},{Ø,{Ø}}}, ...,
which, upon binary encoding, leads to the nonregular substitution
[a® aab, b® b]. The associated infinite
word (interpret a as 0, b
as 1) has a gf that is transcendental, being
related to the series
that also shows up in a formal language example of [6].
5 Lucas sequences
The talk concludes with a description of some recent results of
Allouche, GouyouBeauchamps, and Skordev [1]. Lucas showed that
where the m_{j},n_{j} are the digits of m,n in base p
for prime p.
More generally,
following [8],
define a pLucas sequence (p prime) by the property
a_{pn+j}º a_{n} a_{j} mod p.
For instance, the Apéry numbers
are pLucas. Then, Allouche et alii
characterize the
strong property for a sequence to be simultaneously algebraic
(automatic) over Q and
pLucas for all large enough p. In essence, the only
possibility for such a sequence is
to be, up to normalization, the sequence of values of the Legendre
polynomials at some rational point. In other words, the corresponding
gf F(z) is of the form
A particular case is the central binomial coefficient ().
From Lucas' property
and this characterization, a new proof of Stanley's conjecture can be
deduced. There are also interesting extensions to Hadamard products of
series involving (), (), etc.
References
 [1]

Allouche (J.P.), GouyouBeauchamps (D.), and Skordev (G.). 
Transcendence of binomial and Lucas's formal power series. 
Preprint, December 1997.
 [2]

Allouche (JeanPaul). 
Note on an article of H. Sharif and C. F. Woodcock:
``Algebraic functions over a field of positive characteristic and
Hadamard products''. Séminaire de Théorie des Nombres de Bordeaux.
Série 2, vol. 1, n°1, 1989, pp. 163187.
 [3]

Allouche (JeanPaul). 
Finite automata and arithmetic. In Séminaire Lotharingien de
Combinatoire (Gerolfingen, 1993), pp. 118. 
Univ. Louis Pasteur, Strasbourg, 1993.
 [4]

Christol (G.). 
Ensembles presquepériodiques kreconnaissables. Theoretical Computer Science, vol. 9, 1979, pp. 141145.
 [5]

Christol (G.), Kamae (T.), Mendès France (M.), and Rauzy (G.). 
Suites algébriques, automates et substitutions. Bulletin de
la Société Mathématique de France, vol. 108,
1980, pp. 401419.
 [6]

Flajolet (P.). 
Analytic models and ambiguity of contextfree languages. Theoretical Computer Science, vol. 49, 1987, pp. 283309.
 [7]

GouyouBeauchamps (D.) and Viennot (G.). 
Equivalence of the twodimensional directed animal problem to a
onedimensional path problem. Advances in Applied Mathematics, vol. 9,
n°3, 1988, pp. 334357.
 [8]

McIntosh (Richard J.). 
A generalization of a congruential property of Lucas. The
American Mathematical Monthly, vol. 99, n°3, 1992,
pp. 231238.
 [9]

Stanley (R. P.). 
Differentiably finite power series. European Journal of
Combinatorics, vol. 1, 1980, pp. 175188.
 [10]

Woodcock (Christopher F.) and Sharif (Habib). 
On the transcendence of certain series. Journal of Algebra,
vol. 121, n°2, 1989, pp. 364369.
This document was translated from L^{A}T_{E}X by
H^{E}V^{E}A.