ADAM is known to have better convergence bounds than other methods. Theoretical bounds may not explain the full story of why a method works well, but it is how mathematicians reason about it. I'm only blaming them for not sharing those relevant parts of what we already know.

My bigger pain point even is how they choose to allocate their space: the theorem statements for the most relevant results are missing, the proofs for the more interesting theorems are just citations, while the proofs for basic and arguably tangentially relevant lemmata from eg probability theory take up pages and pages.