Minorisation and Coupling in probability theory.

32 Views Asked by At

In probability theory, when studying the convergence of a stochastic process to equilibrium minorisation conditions can be exploited (see for instance Assumption 2.1. of https://www.sciencedirect.com/science/article/pii/S0304414902001503?ref=pdf_download&fr=RR-2&rr=7e1ea215b82bb2e7).

What is the intuitive reason that such a condition can provide convergence to equilibrium, I have heard in passing that it is related to coupling techniques but can not see how.

1

There are 1 best solutions below

0
On

Not sure if this helps for the above specific paper, but for the Discrete case the reasoning is much easier to explain.

For any kernel $K$ acting from $\mathcal{X} \to \mathcal{Y}$, we can decompose it as convex combination of two Kernels as $$ K = \alpha 1 \pi^T + (1 − \alpha)M $$ where $\alpha = \sum_{y \in \mathcal{Y}} \min_{x \in \mathcal{X}} K(y|x) $ is the Doeblin coefficient of Kernel $K$ and $\pi(y) = \min_{x \in \mathcal{X}} K(y|x)/\alpha $.

Now, whenever a measure $\mu$ is acted upon by a kernel $K$, there is essentially an 'erasure of the $\alpha$-th fraction of information' in $\mu$ due to the kernel $1 \pi^T $ in the above decomposition. This idea is basically at the heart of many convergence results.