Past

The Brownian snake (with lifetime given by a normalized

Brownian excursion) arises as a natural limit when studying random trees. This

may be used in both directions, i.e. to obtain asymptotic results for random

trees in terms of the Brownian snake, or, conversely, to deduce properties of

the Brownian snake from asymptotic properties of random trees. The arguments

are based on Aldous' theory of the continuum random tree.

I will discuss two such situations:

1. The Wiener index of random trees converges, after

suitable scaling, to the integral (=mean position) of the head of the Brownian

snake. This enables us to calculate the moments of this integral.

2. A branching random walk on a random tree converges, after

suitable scaling, to the Brownian snake, provided the distribution of the

increments does not have too large tails. For i.i.d increments Y with mean 0,

a necessary and sufficient condition is that the tails are o(y^{-4}); in

particular, a finite fourth moment is enough, but weaker moment conditions are

not.

The celebrated Levy-Khintchine formula provides us an explicit

structure of Levy processes on $R^d$. In this talk I shall present a

structure result for quasi-regular semi-Dirichlet forms, i.e., for

those semi-Dirichlet forms which are associated with right processes

on general state spaces. The result is regarded as an extension of

Levy-Khintchine formula in semi-Dirichlet forms setting. It can also

be regarded as an extension of Beurling-Deny formula which is up to

now available only for symmetric Dirichlet forms.

Consider approximating a set of discretely defined values $f_{1}, \ldots , f_{m}$ say at $x=x_{1}, x_{2}, \ldots, x_{m}$, with a chosen approximating form. Given prior knowledge that noise is present and that some might be outliers, a standard least squares approach based on $l_{2}$ norm of the error $\epsilon$ may well provide poor estimates. We instead consider a least squares approach based on a modified measure of the form $\tilde{\epsilon} = \epsilon (1+c^{2}\epsilon^{2})^{-\frac{1}{2}}$, where $c$ is a constant to be fixed.

\\

The choice of the constant $c$ in this estimator has a significant effect on the performance of the estimator both in terms of its algorithmic convergence to a solution and its ability to cope effectively with outliers. Given a prior estimate of the likely standard deviation of the noise in the data, we wish to determine a value of $c$ such that the estimator behaves like a robust estimator when outliers are present but like a least squares estimator otherwise.

\\

We describe approaches to determining suitable values of $c$ and illustrate their effectiveness on approximation with polynomial and radial basis functions. We also describe algorithms for computing the estimates based on an iteratively weighted linear least squares scheme.

In Clarendon Lab