Patrick Kidger

If my derivation is correct then I think the term needed in the SRK should be: `I_k0 = (t - s) * (H_{s,t} + 0.5 W_{s,t})` But it's also 1am...

Your concerns about tests are reasonable - bugs and tests are definitely the next things to handle. I agree that having to specify levy area to use SRK should be...

@lxuechen This is labelled as Ito-specific: https://github.com/google-research/torchsde/blob/133a74570affed019006ce89bb0135ef220f697a/torchsde/_core/base_sde.py#L53 Is there any part of that which actually relies on that though? It looks like it's just general drift/diffusion operations.

I agree. IIRC both of the SRKs, and NV, are high weak order schemes.

The reason they're there is that several parts of the internal code -- for example `misc.batch_mvp` and `ForwardSDE.dg_ga_jvp` -- implicitly assume there is a single batch dimension. If we can...

So this is a bit weird! In particular because one tends to imagine a `BrownianInterval` being scale-invariant. I've been able to figure out why this happens, but I don't have...

See [the FAQ](https://github.com/google/jaxtyping/blob/main/FAQ.md#what-about-pep-646-and-variadic-generics).

So checking intermediate annotations like this is really the job of a runtime type checker (typeguard/beartype) -- just like how checking the argument/return annotations is already the job of a...

Right, so JAX sits in a lovely spot for applicability of runtime type checking, because Python is only ever being used as a metaprogramming language for XLA. In this context...

Absolutely! I completely agree. So at the moment this is a limitation of the current approach. The checking is performed via `isinstance`, which simply returns `True` or `False`, and it's...