I have read a couple of places that it is possible to do a $2D$ (or $3D$) maximum likelihood fit, but I can't seem to understand how this would work. Suppose I'm considering a probability distribution function depending on $2$ observable variables, $x$ and $y$, given a multi-dimensional set of variable parameters, $\alpha, \beta, \gamma$, ..., PDF($x_i,y_i$|$\alpha, \beta, \gamma, \ldots$).

I would say that I have a good understanding of the ideas behind a maximum likelihood fit, but for some reason I cannot wrap my head around a multi-dimensional maximum likelihood fit.

How do I understand how a $2D$ maximum likelihood fit to determine the best estimate of the parameters $\alpha, \beta, \gamma$ works?

  • 1
    $\begingroup$ Everything about it is the same as unidimensional case. What is unclear for you? $\endgroup$
    – Tim
    Nov 28 '18 at 8:06

Maximum likelihood is about learning

$$\hat{\theta} \in \{\arg\max_{\theta \in \Theta}\sum_{i=1}^n \ln f(x_i|\theta)\}$$

Notice that $\Theta$ need not be of just one dimension and it can be of multiple dimension.

The whole idea is to optimize a quantity and techniques from calculus or numerical optimization might be involved. For smooth cases, we typically begins by exploring the stationary point of the objective function.

  • $\begingroup$ Likelihood is not mean but product. I guess you forgot log? The 1/n constant is not a part of the definition. $\endgroup$
    – Tim
    Nov 28 '18 at 8:08
  • $\begingroup$ I must admit, I still do not see where 2D fit comes in? I understand that 𝚯, the parameters we want to estimate, can be of several dimensions. But we would still do a fit around a single parameter at a time, right? And then possibly re-fit to the remaining parameters. However, what happens if f does not only depend on x, but also y? $\endgroup$
    – NKH
    Nov 28 '18 at 14:01
  • $\begingroup$ We solve them simultaneously rather than one parameter at a time. $\endgroup$ Nov 28 '18 at 14:04
  • $\begingroup$ here is an example where the parameters are $\mu$ and $\sigma$. In that example, while the derivative with respect to $\mu$ can be solved independently, we should view the two first order constraints as a system of equations that we want to satisfy. $\endgroup$ Nov 28 '18 at 14:16
  • $\begingroup$ here , a more complicated example where MLE for the parameters of beta distributed is staed in equation $(2.3)$ and $(2.4)$ and numerical method is used to find them. $\endgroup$ Nov 28 '18 at 14:25

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.