I am not a lawyer, so take what I'm saying here with several large grains of salt.

My understanding is that a subset and a superset would both be okay. I am more certain about subset than superset -- I am not sure that something like Microsoft's J++ would pass muster either. But a subset I think would still be permissible under fair use so long as the goal was to achieve compatibility.

Oracle's argument is that Dalvik is not the JVM, nor is it an attempt to be the JVM. It's a competitor to the JVM that uses parts of the Java API. I don't know if they're right, and I don't know that being right should mean that Google has to pay them money or what have you as a result. I do think that people are overstating the effect that a ruling that Google infringed on Oracle's copyrights would have.

> So are they claiming that it infringes copyright if your implementation is a superset or a subset of the original implementation, but if it is an exact implementation, then it's not copyright infringement?

They're claiming that any of the above would require an API license, and that a license (requiring GPL for the implementation) was granted only to the latter, so a subset would be unauthorized infringement.

But when you start creating an implementation, it ends up being a subset while it is a work in progress. Are you stating that works in progress are infringing, until they become "magically" not infringing at the point where they reach parity?

Does Mono infringe on Microsoft's .NET API because it has always lagged behind the official .NET API, even though the goal of the project is to achieve compatibility?

.NET isn't relevant because Microsoft isn't trying to control it with the same kind of licensing terms that Oracle's using to try to ensure that anybody else's Java implementation adheres to their standards.