Expression of density function f, through derivatives of measures.

47 Views Asked by At

Lemma: Let $P$ and $\mu$ be probability measures on a measurable space $(\Omega, \mathcal{A})$ with restrictions $P_{m}$ and $\mu_{m}$ to the elements of an increasing sequence of σ-fields $\mathcal{A_{1}}⊂\mathcal{A_{2}}⊂···⊂\mathcal{A}$ that generates $\mathcal{A}$. If $P_{m}<<\mu_{m}$ for every $m$, then $dP_{m}/dμ_{m}→dP^{a}/dμ$, almost surely $[\mu]$, for $P^{a}$ the part of $P$ that is absolutely continuous with respect to $\mu$. Furthermore,this convergence is also in $L1(\mu)$ if and only if $P<<μ$

Additional information given from the proof of the Lemma. It states that if $P<<\mu$ then $dP_{m}/dμ_{m}$ is the conditional expectation $\mathbb{E}_{\mu}[dP/dμ|\mathcal{A}_{\mu}]$ and a uniformly integrable $\mu$-martingale that converges to $dP/dμ$.

Later based on this Lemma and on a partition of a space $X$, $(A_{1},...,A_{k})$, it states that a sequence of densities $p_{m}$ can be defined as $p_{m}=dP_{m}/d\mu_{m},$ where $P$ and $\mu$ follow the conditions of the Lemma, and also that the density of $P$ is given as

$$p_{m}=\sum \frac{P(A_{i})}{\mu(A_{i})}1_{A_{i}}$$

Those are notes from the book "Fundamentals of Nonparametric Bayesian Inference".

The Lemma can be found on "L - Miscellaneous Results" while the statment of the density expression for the partition corresponds to the Theorem 3.16 of Chapter 3.

1

There are 1 best solutions below

1
On

Without any further assumption this is surely false. Consider the RND of standard Gaussian measure w.r.t. Lebesgue measure and the partition $(-\infty, 0], (0,\infty)$. If what you are saying is true then the Guassian density function must be a constant on each of the intervals $(-\infty, 0], (0,\infty)$!.