I'm trying to analytically find the gradient the log a multivariate Gaussian density, which is given by

$$ f(x_{1}, \dots, x_{p}) = f(\vec{x}) = \frac{1}{(2\pi)^{p/2}|\Sigma|^{1/2}} \exp \Bigg(-\frac{1}{2}(\vec{x} - \vec{\mu})^{T}\Sigma^{-1}(\vec{x} - \vec{\mu})\Bigg) \sim \mathcal{N}(\vec{\mu}, \Sigma) $$

where $\vec{x}, \vec{\mu} \in \mathbb{R}^{p}$, and $\Sigma \in \mathbb{R}^{p \times p}$.

I want to find the analytical expression for $\nabla f$, where $\nabla$ is the gradient operator defined as

$$ \nabla f(\vec{x}) = \Bigg[ \frac{\partial f(\vec{x})}{\partial x_{1}}, \dots, \frac{\partial f(\vec{x})}{\partial x_{p}} \Bigg] $$

My vector calculus is very rusty and not sure how to take partial derivatives with respect to $x_{1}, \dots, x_{p}$.

Any help is much appreciated!



The relevant part is $$\exp\left(\frac12 \sum_{i,j=1}^p (\Sigma^{-1})_{ij}(x_i-\mu_i)(x_j-\mu_j)\right)$$

so its $k$-th partial derivative is

$$\exp\left(\frac12 \sum_{i,j=1}^p (\Sigma^{-1})_{ij}(x_i-\mu_i)(x_j-\mu_j)\right) \cdot \frac\partial{\partial x_k}\left(\frac12 \sum_{i,j=1}^p (\Sigma^{-1})_{ij}(x_i-\mu_i)(x_j-\mu_j)\right) \\= \exp\left(\frac12 \sum_{i,j=1}^p (\Sigma^{-1})_{ij}(x_i-\mu_i)(x_j-\mu_j)\right)\cdot \frac12 \left(\sum_{i,j=1}^p (\Sigma^{-1})_{ij}\delta_{ik}(x_j-\mu_j) + \sum_{i,j=1}^p (\Sigma^{-1})_{ij}\delta_{jk}(x_i-\mu_i)\right)$$ Since $\Sigma^{-1}$ is symmetric, the two sums are equal so the final result is $$\exp\left(\frac12 \sum_{i,j=1}^p (\Sigma^{-1})_{ij}(x_i-\mu_i)(x_j-\mu_j)\right) \cdot\left(\sum_{i=1}^p (\Sigma^{-1})_{ik}(x_i-\mu_i)\right)$$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.