[Docs] get_fantasy_model - are posterior covariances computed from scratch or using efficient cache updates?

[fteufel]

📚 Which caches are updated efficiently in get_fantasy_model

Hi,

I'm interested in adding fantasy observations to a GP and computing the posterior covariance.
The docs are pretty clear that get_fantasy_model should be used because it is more efficient by using linear time cache updates (https://docs.gpytorch.ai/en/stable/models.html#gpytorch.models.ExactGP.get_fantasy_model).

I understand from the source that the mean cache is updated efficiently. But when it comes to the covariance cache, I can't directly tell from the source whether there is a low rank update happening ((https://github.com/cornellius-gp/gpytorch/blob/master/gpytorch/models/exact_prediction_strategies.py#L212)

It would be great to know if the covar caches are updated efficiently using get_fantasy_strategy - maybe it's somewhere in LinearOperator and I just don't see it directly?

I've found an old discussion in #177 but am not sure what has made it into the current gpytorch version. Also, is it still accurate that only LOVE makes use of cached results when computing covariances? Is there support for the efficient update introduced in Jiang 2020 (https://arxiv.org/pdf/2006.15779.pdf) ?

Thanks!

[fteufel]

Turns out there are efficient updates in place, they are just not working currently #2494