Reparametrization in Expectation Sampling

The expectation value of a function $f(z)$ over a Guassian distribution $\mathscr N(z;\mu, \sigma)$ is equivalent to the expectation value of $f()$ a Gaussian distribution $\mathscr N(z;\mu=0, \sigma=1)$, i.e.,

$$ {\mathbb E}_{\mathscr N(z; \mu, \sigma)} \left[ f(z) \right] = {\mathbb E}_{\mathscr N(z; 0, 1)} \left[ f() \right] $$

where

$$ \mathscr N = \frac{1}{\sqrt{2\pi\sigma^2}} \exp \left( -\frac{(z-\mu)^2}{2\sigma^2}\right). $$

$$ \begin{align} {\mathbb E}_{\mathscr N(z; \mu, \sigma)} \left[ f(z) \right] &= \int \mathrm d z \frac{1}{\sqrt{2\pi\sigma^2}}\exp \left( -\frac{(z-\mu)^2}{2\sigma^2}\right) f(z) \\ &= \int \mathrm dz \frac{1}{\sigma} \frac{1}{\sqrt{2\pi}} \exp \left( -\frac{1}{2} \left(\frac{z-\mu}{\sigma}\right)^2 \right) f(z) \\ &= \int \mathrm d \left( \sigma z' + \mu \right) \frac{1}{\sigma} \frac{1}{\sqrt{2\pi}} \exp \left( -\frac{1}{2} z'^2 \right) f(\sigma z' + \mu) \\ &= \int \mathrm d z' \frac{1}{\sqrt{2\pi}}\exp \left( -\frac{1}{2} z'^2 \right) f(\sigma z' + \mu) \\ &= \int \mathrm d z' \mathscr N(z'; \mu=0, \sigma=1) f(\sigma z' + \mu) \\ &= {\mathbb E}_{\mathscr N(z'; \mu=0, \sigma=1)} \left[ f(\sigma z' + \mu) \right] \end{align} $$

Define a field $\phi = {\mu, \sigma}$, the derivatives of the reparametrization

$$ \begin{align} \frac{\partial}{\partial \phi} {\mathbb E}_{\mathscr N(z; \mu, \sigma)} \left[ f(z) \right] &= \frac{\partial}{\partial \phi} {\mathbb E}_{\mathscr N(z'; \mu=0, \sigma=1)} \left[ f(\sigma z' + \mu) \right] \\ &= {\mathbb E}_{\mathscr N(z'; \mu=0, \sigma=1)} \left[ \frac{\partial}{\partial \phi} f(\sigma z' + \mu) \right] . \end{align} $$

This is a great result. The derivatives are passed onto the function $f$. We do not need to deal with the distribution.

In fact, we have a much general formalism for such properties.

Planted: by ;

LM (2021). 'Reparametrization in Expectation Sampling', Datumorphism, 01 April. Available at: https://datumorphism.leima.is/cards/statistics/reparametrization-expectation-sampling/.