Oscar Javier Hernandez

Gaussian Integrals

The following formulas are useful for derivations involving Gaussian integrals.

General integrals

\[\int\limits_{-\infty}^{\infty} dx \ \text{Exp}\left( -ax^2+bx+c \right) = \sqrt{\frac{\pi}{a}} \text{Exp}\left( \frac{b^2}{4a}+c \right)\] \[\int\limits_{-\infty}^{\infty} dx \ \frac{1}{\sqrt{2\pi\sigma^2}}\text{Exp}\left( -\frac{1}{2\sigma^2}\left(x-\mu \right)^2+\lambda x \right) = \text{Exp}\left( \frac{\sigma^2}{2}\lambda^2+\lambda \mu \right)\]

Multivariate Gaussian Integrals

\[\int dx_1 \ldots dx_n \text{Exp}\left( -\frac{1}{2} \vec{x}^T A \vec{x} + \vec{J}^T \cdot \vec{x} \right) = \sqrt{\frac{(2\pi)^n}{|\Sigma|}} \text{Exp}\left( -\frac{1}{2} \vec{J}^{T} A^{-1} \vec{J} \right)\] \[\int dx_1 \ldots dx_n \int dk_1 \ldots dk_n \text{Exp}\left( -i\vec{k}^T A \vec{x} + i\vec{k}^T \vec{J}_1 ++ i\vec{J}_2^T \vec{x} \right) \propto \text{Exp}\left( \vec{J}_2^{T} A^{-1} \vec{J}_1 \right)\]

Representations of Delta Function

\[\delta(x) = \frac{1}{2\pi} \int\limits_{-\infty}^{\infty} dk \ e^{ikx}\]

In three dimensions

\[\delta(\vec{x}) = \frac{1}{(2\pi)^3} \int\limits_{-\infty}^{\infty} d^3k \ e^{i\vec{k}\cdot \vec{x}}.\]

The following property of the delta function is also very useful

\[\delta(y-x) = \delta(x'-x+y-x')\\ = \sum_{n=0}^{\infty} \frac{(y-x')^n}{n!} \left( \frac{\partial}{\partial x'} \right) \delta(x'-x) \\ = \sum_{n=0}^{\infty} \frac{(y-x')^n}{n!} \left( - \frac{\partial}{\partial x} \right) \delta(x'-x)\]

Products of Gaussians

\[\int dx_1 \frac{1}{\sqrt{2\pi \sigma^2}} \text{Exp}\left(\frac{1}{2\sigma^2}\left(x_2-b x_1-u_2 \right)^2 \right) \frac{1}{\sqrt{2\pi \sigma^2}} \text{Exp}\left( \frac{1}{2\sigma^2}\left(x_1 -b x_0 -u_1 \right)^2 \right) = \frac{1}{\sqrt{2\pi\sigma^2(1+b^2)}}\text{Exp}\left( -\frac{(x_2-b^2 x_0 -b u_1-u_2)^2}{2\sigma^2(1+b^2)} \right)\] \[\int dx_1 \int dx_2 \frac{1}{\sqrt{2\pi \sigma^2}} \text{Exp}\left(\frac{1}{2\sigma^2}\left(x_3-b x_2-u_3 \right)^2 \right) \frac{1}{\sqrt{2\pi \sigma^2}} \text{Exp}\left(\frac{1}{2\sigma^2}\left(x_2-b x_1-u_2 \right)^2 \right) \frac{1}{\sqrt{2\pi \sigma^2}} \text{Exp}\left( \frac{1}{2\sigma^2}\left(x_1 -b x_0 -u_1 \right)^2 \right) =\\ \frac{1}{\sqrt{2\pi \sigma^2(1+b^2+b^4)}}\text{Exp}\left( -\frac{(x_3-b^3 x_0-b^2 u_1 -b u_2-u_3)^2}{2\sigma^2(1+b^2+b^4)} \right)\]

Bivariate Gaussian

The following notation is useful for the case of the Bivariate Gaussian distribution

\[\vec{\mu} = \begin{bmatrix} \mu_1 \\ \mu_2 \end{bmatrix}, \\ \Sigma = \begin{bmatrix} \sigma^2_1 & \rho \sigma_1 \sigma_2 \\ \rho \sigma_1 \sigma_2 & \sigma_2^2 \end{bmatrix}\\ \vec{x} = \begin{bmatrix} x_1 \\ x_2 \end{bmatrix}, \\\]

where \(\vec{\mu}\) is the vector of mean values, \(\vec{x}\) is the vector of variables and \(\Sigma\) is the covariance matrix. The joint probability distribution of \(x_1\) and \(x_2\) is

\[P(x_1,x_2) = \frac{1}{\sqrt{(2\pi)^2|\Sigma|}} \text{Exp}\left[ \left(\vec{\mu}-\vec{x}\right)^{T}\Sigma^{-1} \left(\vec{\mu}-\vec{x} \right) \right]\]

Conditional Probabilities

Here we look at the conditional probability distribution of the multivariate Gaussian distributions. For the bivariate Gaussian the conditional probability is

\[P(x_1| x_2) = \frac{1}{\sqrt{2\pi(1-\rho^2)\sigma^2_1}} \text{Exp}\left( -\frac{1}{2(1-\rho^2)\sigma^2_1}\left(x_1 -\mu_1 - \frac{\sigma_1}{\sigma_2}\rho(x_2-\mu_2) \right)^2 \right),\]

while the general case is

\[P(\vec{x} |\vec{y}) = \frac{1}{\sqrt{(2\pi)^2|\Sigma|}} \text{Exp}\left[ \left(\hat{\mu}-\vec{x}\right)^{T}\hat{\Sigma}^{-1} \left(\hat{\mu}-\vec{x} \right) \right]\]

where we have that

\[\hat{\mu} = \vec{\mu}_1+\Sigma_{12}\Sigma^{-1}_{22}(\vec{y}-\vec{\mu}_2)\]

and

\[\hat{\Sigma} = {\bf \Sigma}_{11} - {\bf \Sigma}_{12}{\bf \Sigma}^{-1}_{22}{\bf \Sigma}_{21}\]

with

\[\hat{\mu} = \begin{bmatrix} \mu_1 \\ \mu_2 \end{bmatrix}, \\ {\bf \Sigma} = \begin{bmatrix} {\bf \Sigma_{11}} & {\bf \Sigma_{12}} \\ {\bf \Sigma_{21}} & {\bf \Sigma_{22}} \end{bmatrix}\\ \hat{x} = \begin{bmatrix} \vec{x} \\ \vec{y} \end{bmatrix},\]