Sarah Lutteropp

> Is this (especially the denominator) how you meant it? Nevermind, I just noticed my mistake: The denominator has to be the cardinality of the set union, not the sum...

Yes, it is now randomly placed. There are not many arc removal moves to consider. After all, we can remove at most O(number of reticulations) arcs, but the number of...

@celinescornavacca This sounds like a great approach! :) But why stop at a single pair of ml trees then? Can't we do the entire network this way? And then have...

@celinescornavacca While I do like the approach you proposed, I do not see how it helps with the "which arc insertion should we take when moving from k reticulations to...

Lol. Turns out blob optimization is not the way to go with this new likelihood definition we are using. Instead, each node in the network needs to store 2^(number_of_reticulations_in_subnetwork_rooted_at_the_node) clv...

But even if memory usage ends up being an issue in the future, the blobs per-se do not make much sense anymore. We could simply arbitrarily choose some on-the-fly nodes...

RIP blobs and gray-code stuff... well, they still make sense for the "wrong" likelihood definition. Maybe we can use the wrong one as a heuristic during horizontal search, to speed...

Even more CLV saving potential, but not worth the extra effort and dirty case distinctions needed: 

Okay, actually this extra clv/computation saving is easier to implement than I thought. I'm confident that I can do it without much extra effort :slightly_smiling_face: 

In order to avoid pointer craziness, I will modify pll_update_partials in my libpll2 fork: Instead of giving it a long list of operations, I will always just give a single...