Distribution for "joint multi-variate gaussian distribution" (joint MVG):

$$f_{X}(x) = \frac{1}{(2\pi)^{n/2}\prod \limits_{i=0}^{n}\sigma_i} ~~\text{exp}\bigg[-\frac{1}{2} \sum \limits_{i=1}^{n}\Big(\frac{x_i\mu_i}{\sigma_i}\Big)^2 \bigg]\tag{2}$$

Distribution for "multi-variate gaussian distribution" (MVG):

$$f_{\mathbf{x}}(x) = \frac{1}{(2\pi)^{n/2}~|\text{det }C|^{1/2} }~~\text{exp}\bigg[ -\frac{1}{2} (\mathbf{x}-\mathbf{\mu})^T C^{-1} (\mathbf{x}-\mu)\bigg]\tag{1}$$


$[n\times1] ~~~\mathbf{x} \Rightarrow \text{sample vector}$

$[n\times1]~~~\mathbf{\mu} \Rightarrow \text{mean vector}$

$[n\times n]~~~ C \Rightarrow \text{covariance matrix (symmetric positively defined matix)}$

What's the difference between the two?


There's no difference in formula:

Try setting the $C=\text{diag}(\sigma_1^2, \sigma_2^2,\ldots,\sigma_n^2)$ in $(1)$ to see how $(1)$ simplifies to $(2)$. So, your formula $(2)$ is just a special case of $(1)$.

In the diagonal case, the elements of the multivariate distribution are jointly Gaussian, which is much stronger than them all just being Gaussian distributed.

Joint Gaussian distribution has a name of its own (unlike other multivariate distributions with identically distributed elements) because "joint gaussian multivariate random varaible" has the very interesting feature that simple MVG doesn't have: that uncorrelatedness implies independence. This is actually something that makes it useful in communications engineering (because rotation and normalization doesn't change joint gaussianness and doesn't introduce interdependencies between components - so, in an IQ receiver, the noise on I and Q are still independent, even after phase correction, for example).


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.