In this article we investigate how to simulate probability concepts in computers. We will also discuss its statistical applications.

When Analytic Expression of the Probability Distribution is Known

Posterior Sampling

In Bayesian data analysis, we often need to sample from the posterior distribution. In this case, we often know the analytic expression of the probability distribution only up to constant.

Diffusion Model

Nowadays, many real-world distributions are no longer modelled by traditional parameterized model such as. Instead, diffusion models are widely used. In this case, we can no longer obtain an analytical expression of the probability distribution.

Posterior Sampling via Bayesian Filtering

We are going to solve the following linear inverse problem:

\[\by = \bA\bx +\bn,\quad \bx\in\RR^D,\by\in\RR^d,\bA\in\RR^{d\times D},\bn\sim\cN(0,\sigma^2\bI).\]

The goal is to recover $\bx$ from the incomplete measurement $\by$. Often $D>d$ so this problem is under-determined. To narrow down the solution space, we usually requires additional information about $\bx$. A common assumption, which is statistical in nature, is the prior knowledge of the distribution over $\bx$, represented as $p(\bx)$. Given this, we can traverse the solution space of the linear inverse problem by sampling from the posterior distribution $p(\bx|\by)$ via Bayes’ rule

\[p(\bx|\by)\propto p(\bx)p(\by|\bx),\quad p(\by|\bx)\sim \cN(\by|\bA\bx,\sigma^2\bI).\]

Assume we can only assess to $p(\bx)$ via diffusion model.

We have the forward process

\[q(\bx_{1:N}|\bx_0)=\prod_{k=1}^N q(\bx_i|\bx_{k-1}),\quad q(\bx_k|\bx_{k-1})=\cN(a_k\bx_{k-1},b_k^2\bI )\]

The backward process

\[p_{\btheta}(x_{k-1}|x_k)=\cN ()\]

We can construct