Index required for efficient random access for trees

[jeromekelleher]

Current TreeSequence.at() and other methods to seek to a particular tree work by iterating along the trees. This isn't very efficient, if we want to get a tree in the middle of the sequence.

Can we seek efficiently using the current indexes, or do we need to build some sort of interval tree?

[jeromekelleher]

Related: #24 and #4

[grahamgower]

The intervals are non-overlapping, so an interval tree should be unnecessary---just binary search the sorted list of start coordinates.

[jeromekelleher]

They are overlapping though, aren't they? We should be able to get them from the existing indexes though, by binary searching the in-order and out-order indexes...

[jeromekelleher]

See also #685

[jeromekelleher]

I've thought about this again, and no, I can't see anyway of doing this efficiently without a different type of index. The indexes we have tell us the order in which edges go in and out in a left-to-right traversal, basically by sorting on the left and right coordinates. Since the edges for a given tree are not adjacent within these lists in general, then binary searching for a given position will only give you the edges that start/end closest to that point. You can have long edges (say) spanning the entire interval which are not inserted/removed anywhere near this point along the genome.

So, we will need some sort of data structure that'll support interval overlap queries if we're going to implement this (and edge finding in general, #685) efficiently.

[benjeffery]

They are overlapping though, aren't they? We should be able to get them from the existing indexes though, by binary searching the in-order and out-order indexes...

To do this you'd need to find the first in edge i that has i.right > target, then you also find the first out edge o that has o.right > i.left. Then you could continue normal tree iteration from that point.
I'm not sure this is so simple as binary search though as in is not sorted by right and vice-versa.

[benjeffery]

Heh, seems we wrote at the same time.

[grahamgower]

Oh, yes. Sorry for the misunderstanding. So you possibly want nested containment lists. https://academic.oup.com/bioinformatics/article/23/11/1386/199545

[jeromekelleher]

To do this you'd need to find the first in edge i that has i.right > target, then you also find the first out edge o that has o.right > i.left. Then you could continue normal tree iteration from that point.

That's can't be true, can it? Suppose we have one edge (0, L, a, b) at spans the entire TS. This appears at the start of the edges_in list and the end of the edges_out list. Then, suppose we have lots of trees, and we want to produce the one at L / 2. Searching for L / 2 won't bring us anywhere near this edge in either the edges_in or edges_out liest.

[benjeffery]

Yes, long edges make this the same as normal iteration.

[benjeffery]

This looks promising: https://github.com/biocore-ntnu/ncls shame the C is so tied to the python. There is also this https://github.com/databio/AIList/ although it is GPL :(