# Multivariate gaussian vs univariate gaussian

What is the intuition behind the change in formula from the univariate gaussian to the multivariate gaussian? Why are the determinant and $(2\pi)^{n/2}$ added into the equation?

Thanks

• Linear algebra says there is an orthogonal transformation of $x$ that diagonalizes $\Sigma$. Among other things, such transformations do not change volumes. Thus, you can (and always should) think of the multivariate formula as really being the product of $n$ versions of the first formula, each with its own variance $\sigma_i^2$. Your question then comes down to asking why the determinant is the product of the $\sigma_i^2$. – whuber Dec 31 '17 at 16:14

You are attempting to generalize; instead, particularize from the multivariate case to the univariate case.

The $n$-variate normal density is, as you say, $$f(\mathbf x; \mathbf \mu, \Sigma) = \frac{1}{(2\pi)^{n/2}|\det \Sigma|^{1/2}}\exp\big(-\frac 12 (\mathbf x-\mathbf \mu)^T\Sigma^{-1}(\mathbf x-\mathbf \mu)\big) \tag{1}$$ where $\mathbf x$ is a column vector of length $n$, and $\Sigma$ is the $n\times n$ covariance matrix with variances down the main diagonal. Note that $\frac{1}{(2\pi)^{n/2}|\det \Sigma|^{1/2}}$ is a constant that makes the $n$-dimensional integral of the density $(1)$ have value $1$. Now, for the case $n=1$, the $1\times 1$ covariance matrix is just $[\sigma^2]$ with determinant $\sigma^2$ and inverse matrix $[\sigma^{-2}]$. So, getting rid of the distinction between $1\times 1$ matrices and ordinary scalars, the univariate normal distribution is obtained as \begin{align} f(x; \mu, \sigma^2) &= \frac{1}{(2\pi)^{1/2}(\sigma^2)^{1/2}}\exp\big(-\frac 12 (x-\mu)\sigma^{-2}(x-\mu)\big)\\ &= \frac{1}{\sigma(2\pi)^{1/2}}\exp\Bigr(-\frac 12 \left(\frac{x-\mu}{\sigma}\right)^2\Bigr)\tag{2} \end{align} exactly as you have it.

In both cases, everything before the exponential is simply a normalizing constant that makes the expression integrate to one. This is needed to have a proper probability distribution.

For example, in the univariate case:

$$\int_{-\infty}^\infty \exp \left (-\frac{(x-\mu)^2}{2 \sigma^2} \right ) = \sigma \sqrt{2 \pi}$$

So, multiplying by $\frac{1}{\sigma \sqrt{2 \pi}}$ makes $p(x; \mu, \sigma^2)$ integrate to one.

I like to think about how the gaussian distribution is constructed from the "inside". The middle of the equation, or the $(x-\mu)^2$, is a unit deviance, a function that satisfies:

$$d(y; y) = 0$$

and

$$d(y; \mu) > 0$$ There is actually a big group of distributions constructed from unit deviances (see The Theory of Dispersion Models). Anyway, if we plot $(x-\mu)^2$, with $\mu = 0$ for example, varying $x$, we have this:

But this doesn't look like a distribution, right? What is the first step to change this plot? Put a $-$ before the equation, or $-(x-\mu)^2$, which turns the plot upside down:

Ok, but we want the distribution to give us plausible values. We can smooth the tails of the equation with $exp(-(1/2\sigma^2)(x-\mu)^2)$. The plot for this, considering $\sigma^2 = 1$ is:

And now we "just" need to scale the equation in a way that it produces values between 0 and 1 and integrates 1 (I used quotes there because finding a good normalizing constant is not always an easy task). Then, we end up with: $$f(x) = \frac{exp(-(1/2\sigma^2)(x-\mu)^2)}{\sqrt{(2 \pi \sigma^2}}$$

Plotting this, we have:

This is not a very formal explanation (the way I did it in here), but a good one for understanding the Gaussian distribution. The multivariate case is then a generalization of what I explained above.