Will Tebbutt

Thanks for looking at that @devmotion . I agree with your assessment of what should be implemented. > At the interface level, I'd lean towards having a k==k test, and...

Possibly -- I still don't fully understand why they're not already equal to be honest. I also don't really understand why `SEKernel() == SEKernel()` holds, because our SEKernels contain a...

> Maybe a specialised kind of equality could be implemented for category 3 later, if needed, and stick to cases 1 and 2 for now? I would be happy to...

Good catch. Maybe point 1 is better phrased as a necessary condition: k1 and k2 are necessarily unequal if they are instances of different (mutable) structs. If k1 and k2...

> if k1==k2 then isequal(k1(x,y), k2(x,y)) (and similarly for kernelmatrix, etc.) This seems like an excellent definition. This will clearly only hold approximately for things like `5.0 * SEKernel()` vs...

Fantastic. How urgent is this for you? We can implement this in this package now and recurse into Distances when the issue is fixed if you need this soon.

A separate option would be to be a little less ambitious with this initial implementation, and not implement a new matrix type, and just provide an interface for the operation...

> I have no idea why the tests would fail - something about constant allocation heuristics in Zygote? Is it clear whether they're happening in areas where the modifications introduced...

Sorry for not getting around to reviewing this. > I've renamed KernelTensorSum to KernelIndependentSum since in my opinion it better informs what it is. I could also revert this commit,...

That's fair. I agree that I'm not entirely sure that the use of the term is correct, but I still think my preference is consistency + an issue suggesting that...