CONFERENCES
-
Søren Asmussen (Lund University - Sweden)
-
David Balding (Imperial College - UK)
-
Pierre Collet (CNRS, École Polytechnique - France)
-
Luiz Renato Fontes (Universidade de São Paulo - Brazil)
-
Giambattista Giacomin (Université Paris 7 - France)
-
Ricardo Lima (CNRS, Marseille - France)
-
Servet Martínez (Universidad de Chile - Chile)
-
Yuval Peres (University of California, Berkeley - USA)
-
Rinaldo Schinazi (Université d'Aix Marseille - France)
Søren Asmussen (Lund University - Sweden)
Rare events come up in a variety of applied areas like
insurance risk, reliability, telecommunications and statistical physics.
For example in data transmission problems, one may
be interested in the bit loss rate which is typically
of magnitude $p=10^{-9}$ or smaller. To simulate just
a single bit on its way through the system may be time consuming,
and since to obtain a reasonable precision using straightforward simulation
one needs to simulate a number replications (here bits) somewhat
larger than $p^{-1}$, more sophisticated methods are called upon.
The present talk gives a survey of such methods and open problems
in the area.
The most established method is importance sampling, simulating using
a changed probability measure $Q$ rather than the given one $P$.
The problem is the choice of $Q$. Taking $Q$ as the distribution $Q_A$
given
the rare event $A$ gives variance 0 but is not feasible since the estimator
involves the Radon-Nikodym derivative $dP/dQ_A=P(A)$ on $A$ which
is not available (one simulates because $P(A)$ is unknown!). Instead,
one would try to make $Q$ look as much like $Q_A$ as possible, and
for this purpose large deviations techniques have become an established
tool.
In particular, exponential change of measure plays an important role when
the underlying distributions are light-tailed.
More recently, the large deviations approach has, however, turned out to have severe limitations, as will be demonstrated from problems involving processes with boundaries and distributions with heavy tails.
David Balding (Imperial College - UK)
A fundamental problem in statistical genetics is modelling the complex
correlation structure of datasets that results from the fact that different
individuals have different amounts of shared ancestry. Tree-based models
are natural, but statistical inference under such models is challenging. I
will review computational statistical methods in this application, including
approaches based on Markov chain Monte Carlo and importance sampling, as
well as more computationally efficient approximate methods. I will discuss
applications in human population genetics, including population-based
methods for disease gene mapping.
Pierre Collet (CNRS, École Polytechnique - France)
Concentration phenomena is a common fact in statistical
mechanics connected to large deviation theory. It states
that some observables do not deviate appreciably from
their average in some adequate limit. Recently some new
versions of this effect was discovered for independent
random variables and more general observables leading to
important statistical consequences. Extensions were proven
for non independent random variables. We will discuss the
result for expanding maps of the interval and present some
statistical consequences.
Luiz Renato Fontes (Universidade de São Paulo - Brazil)
Arratia, and later Tóth and Werner, constructed random processes that formally correspond to coalescing one-dimensional Brownian motions starting from every space-time point. We extend their work by constructing and characterizing what we call the Brownian web as a random variable taking values in an appropriate (metric) space whose points are (compact) sets of paths. This leads to general convergence criteria and, in particular, to convergence in distribution of coalescing random walks in the scaling limit to the Brownian web.
Giambattista Giacomin (Université Paris 7 - France)
I will present results aimed at understanding
in detail the repulsion phenomena that arise when
random interfaces are forced to cohabit with
hard walls and/or with other interfaces.
What is observed is the result of an entropy-energy
competition that in the context
of Gaussian effective models (harmonic crystals
or lattice free fields) can be handled with precision.
I will first focus on the results available
in the most basic situation of one interface and
one flat wall, stressing in particular
the dependence of the phenomenon on the dimension
of the space and the relevance of the results
in investigating other interface phenomena like wetting.
I will then analyse the repulsion effects
in presence of a rough wall and for models
of interfaces which are not allowed to intersect
with each other.
Ricardo Lima (CNRS, Marseille - France)
Complexity and entropy are two main tools in characterizing
the properties of symbolic words. The analysis of their
asymptotics has an old and reach story. Instead, we shall
review some properties and constraints of these functions,
when applied to finite lenght sequences. The recent interest
on the subject is related with the use of symbolic dynamics
in the study of dynamical systems as well as with the analysis of
biological sequences. This is why we shall also try to describe
some possible implications for such practical purposes.
Servet Martínez (Universidad de Chile - Chile)
We introduce algebraic automata and describe its dynamics.
We discuss the mathematical tools involved in this study
as well as the limit behavior of larger classes of automata.
Yuval Peres (University of California, Berkeley - USA)
In joint work with A. Dembo, J. Rosen and
O. Zeitouni, we proved two conjectures about simple
random walk in two dimensions:
The first, due to Erdos and Taylor (1960), involves the
number of visits to the most visited lattice site in the
first $n$ steps of the walk. The second, due to Aldous (1989),
concerns the number of steps it takes a simple random
walk to cover all points of the $n$ by $n$ lattice torus.
I will describe the analytical tool behind these proofs,
a multiscale refinement of the "Second moment method",
motivated by probability on trees. I will also present
a different refinement, used in work with I. Benjamini,
H. Kesten and O. Schramm to analyze the uniform spanning
forest in dimensions greater than four.
Rinaldo Schinazi (Université d'Aix Marseille - France)
We introduce a spatial stochastic model for the spread of
tuberculosis and HIV. We have three parameters: the size of the social
cluster for each individual and the infection rates within and outside the
social cluster. We show that when the infection rate from outside the
cluster is low (this is presumably the case for tuberculosis and HIV) then
an epidemic is possible only if the typical social cluster and the within
infection rate are large enough.
Contact
Contact us at: epb6@ime.usp.br.
|