According to Bishop, the author from "Statistical Pattern Recognition", we can optimize the hyperparameters of a Gaussian process by maximizing the likelihood function

$$p(\textbf{t}|\theta),$$ where $\textbf{t}$ denotes the target vectors $(t_1, .. ,t_N)$ of the corresponding input values $x_1, ..., x_N$ and $\theta$ the hyperparameters.

He then claims, that the log likelihood function is given by the standard form for a multivariate Gaussian distribution:

$$\ln{p(\textbf{t}|\theta}) = -\frac{1}{2}\ln{|C_N|}-\frac{1}{2} \textbf{t}^T C_N^{-1}\textbf{t}-\frac{N}{2}\ln{2\pi}$$

Considering the multivariate Gaussian distribution for $X = [X_1, ...., X_N]^T$ is given by: (https://cs229.stanford.edu/section/gaussians.pdf)

$$p(x|\mu,\Sigma) = \frac{1}{(2\pi)^{\frac{n}{2}}|\Sigma|^{\frac{1}{2}} }\exp\left(-\frac{1}{2}(x-\mu)\Sigma^{-1}(x-\mu)\right)$$

To me, the first equation is derived by simply the $\log$ of the second equation (keep in mind that the mean is zero in the Gaussian Process described by Bishop) without taking the product into account, which would be $\prod_N{\ln{p(\textbf{t}|\theta})}$. I am not sure, what I am missing here. As far as I know, taking the log of a gaussian distribution is not enough to maximize a wanted parameter.

To be more precise, would the MLE of equation 2 not be:

$$\prod_K{p(x|\mu,\Sigma) = \frac{1}{(2\pi)^{\frac{kn}{2}}|\Sigma|^{\frac{k}{2}} }\exp\left(-\frac{1}{2}\sum_K(x-\mu)\Sigma^{-1}(x-\mu)\right)}$$

which is not equal to the first equation, after taking the log. Notice the $k$ in the denumerator and the sum of the term in the exp term. Thus, I would expect something like:

$$\ln{p(\textbf{t}|\theta}) = -\frac{K}{2}\ln{|C_N|}-\frac{1}{2} \sum \textbf{t}^T C_N^{-1}\textbf{t}-\frac{KN}{2}\ln{2\pi}$$

This assumption is further verified by the accepted answer in this post: Maximum Likelihood Estimators - Multivariate Gaussian

For me, the first equation is just the log of a gaussian multivariate normal distribution with zero mean, not the MLE.


1 Answer 1


$$\log_k\left(\prod x_i\right)=\sum\log_k\left(x_i\right)$$


$$ \log_k\left(k^x\right)=x $$


$$ \log_k\left(\prod k^{x_i}\right) =\sum x_i $$

You’re taking the base-$e$ logarithm of a product of $e^x$ expressions, turning the base-$e$ logarithm of the product into the sum of the exponents.

  • $\begingroup$ Please see my edit, I think I phrased my question poorly $\endgroup$
    – kklaw
    Aug 14, 2022 at 15:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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