Design interface for missing data in Kronecker models about celerite2

[dfm]

Sorry - I wrote it pretty vaguely!

So if I understand @ericagol's proposal for the missing data algorithm, we just need to remove all the rows of a, U, and V that correspond to the missing data (P is actually trickier in that form - more on that later!). So if you only observe one band at each time, this would be equivalent to just multiplying each row of U and V by the alpha value for the band observed at that time and multiplying that row of a by the square (modulo the diag entries, but you know what I mean!). So in its simplest form, Alpha would be an N-vector where each entry is the alpha parameter for the band at time n. The reason why I said NxJ is that if you add KronTerms that is equivalent to allowing different alphas for different Js and we might want to support that.

Is this at all clearer? Hard to explain without a whiteboard perhaps when it's all a little muddled in my brain.

Also: the reason why the P matrix is harder when masking is because most of the rows of P are 1, I think, but we need the row corresponding to the first observed band of P to be evaluated as exp(-c (t_n - t_{n-1})) (currently it's the first band, which will change if that band gets masked). This is trivial in the "one band per time" case as described above.

[dfm]

Sure that would be great! Next week is getting a bit overloaded already. Perhaps some time on Nov 3? Y'all vote by mail in Washington right?

[tagordon]

@dfm I'm trying to understand your suggestion for the low-rank version. What determines the entries in Alpha?

[ericagol]

@dfm Perhaps we should meet about this; I haven't thought about this much, and so could use an explainer.

[ericagol]

@dfm Voted!

[tagordon]

Election day works for me!