search

dendrograms / astrodendro

Generate a dendrogram from a dataset
https://dendrograms.readthedocs.io/
Other
37 stars 36 forks source link

Correctly handle non-binary splitting #99

Open astrofrog opened 10 years ago

ChrisBeaumont commented 10 years ago

Can you elaborate on this? If I understand correctly the data structures themselves handle non-binary splits, so are you referring to a "statistically smart" way of creating non-binary splits when considering things like noise?

astrofrog commented 10 years ago

Sorry for the lack of detail - yes, I want to make sure it works correctly when noise is present. I thought it would do already, but in an email a while back, you said it might not, so I created the issue just to not forget to look into it.

At the moment, why would it not work with noise if delta is set to a value larger than the noise?

ChrisBeaumont commented 10 years ago

I'm not sure if I'm thinking of the same discussion, but we talked about trying to use non-binary splits more intelligently when >2 structures merge at an intensity difference < noise. Right now, I think we only use non-binary splits when >2 structures are merged by exactly one pixel, regardless of noise.

The edge case that confuses me is this one: image

That is, there are several non-binary mergers you would consider, since individual mergers happen within the noise. But the overall height difference from top to bottom is significant. Where should the splits go? There might be a good answer to this question, but I haven't thought of it yet.

astrofrog commented 10 years ago

Ah right yes, mergers can at the moment be arbitrarily close in flux. Nice illustration of the problem! :) I'll think about it a little. It's not clear that there is a unique solution in any case.