Makoto YUI

But as described in the paper, `delete` can be implemented with a careful implementation. No `merge` feature (merging small leaves and branch shrinking) since my original motivation was append-only usecase....

see https://gist.github.com/myui/e2424bcd2b11741f68581a6c80b5e438

Thank you for the patch. I would like to make the core as simple as possible for maintenancebility. I think at least peekMinimum can more simply be implemented using [IndexQuery](https://github.com/myui/btree4j/blob/master/src/main/java/btree4j/indexer/IndexQuery.java)....

Bulk delete requires node shrinking (garbage collection). WIP for it.

By modifying algorithms described in the original paper, adam-hd worked fine https://github.com/apache/incubator-hivemall/blob/master/core/src/main/java/hivemall/optimizer/Optimizer.java#L674 This thesis (multiplicatative hypergradient descent) helped. https://github.com/damaru2/convergence_analysis_hypergradient_descent/blob/master/dissertation_hypergradients.pdf Negative learning rate can be seen in the original experiments but...

@akaniklaus learning rates monotonically increased certain condition because `∇f(θ_{t-1})・∇f(θ_{t-2})` will usually become greater than 0.

Backtracking make sense.

Doesn't `f` string simply work? ```python column1 = None client.query(f"INSERT INTO table (column1) VALUES ({column1 or 'NULL'})") ```

@CodeDream-Mo BTreeIndexDup allows multiple values for a key. https://github.com/myui/btree4j/blob/master/src/test/java/btree4j/BTreeIndexTest.java#L128 To get values for BTreeIndexDup, search() can be used. Values for a key is `not ordered`. ```java BTreeIndexDup btree = new...

Close as answered. Please reopen if you have additional questions.