Replaced sigma argument with noise in MarginalSparse.marginal_likelihood

[fonnesbeck]

Currently, MarginalSparse.marginal_likelihood uses a sigma argument, while Marginal.marginal_likelihood uses noise. Changed to noise to make it consistent, and deprecated sigma.

[twiecki]

Wouldn't it be more consistent to have everything be sigma instead?

[fonnesbeck]

Depends on what we are being consistent with. The rest of the GP module uses noise, the normal distribution uses sd. It should at least be the same within the class of models.

[bwengals]

The logic for the different names was that in the general case the noise doesn't have to be diagonal, you can either give an arbitrary covariance func or just sigma to it. But for MarginalSparse it needs to be white. sd instead of sigma makes sense though. I'm ok with changing the name either way.

[fonnesbeck]

I think the difference can be made clear in the docstrings, rather than in the naming. It felt odd having to switch the keyword argument when changing from a full to a sparse GP. noise is kind of nice because, though more colloquial than sd or sigma, it is clear that it is referring to observation noise, rather than process error. However, a case can be made for sigma and sd to make them consistent with either MvNormal or Normal respectively, though it seems that scalars are generally passed, which would favor using sd. Thoughts?

[junpenglao]

I think the difference can be made clear in the docstrings, rather than in the naming.

Second this. @bwengals' explanation should definitely go to doc/docstring.

[bwengals]

Everything OK with me