Nice summary of the concerns with 'retreats' from computational intractability! I'm in the approximation camp. Would love to get Iris van Rooij 💭 et al's take on our sampling hypothesis as approximation - I don't think it fits the limited inputs requirement. osf.io/q9pjb/

@LizBonawitz Iris van Rooij 💭 Yes I'm in the approximation camp too :-) .
I just wish I spent less time on Twitter and more time on reading!

JoJaSciPo @LizBonawitz I spend too much time here too :)

Fwiw, I have had so much to say on the topic of approximation and intractability already (see also: sciencedirect.com/science/articl…) that it is hard to fit it all in tweets.

JoJaSciPo @LizBonawitz Given @LizBonawitz invitation, I now plan a blog to bring some ideas together to make them more accessible. Short answer is, ‘approximation is not a panacea’, see also: sciencedirect.com/science/articl…), but intuitions are strong (even if wrong).

Iris van Rooij 💭 JoJaSciPo @LizBonawitz You might be interested in this paper and the work of my PhD student Nils Donselaar on parameterized complexity of approximate Bayesian inference.

sciencedirect.com/science/articl…

Iris van Rooij 💭 JoJaSciPo @LizBonawitz tl;dr of this paper: only for approximating marginals (not posteriors) the number of samples needed for a guaranteed approximation quality is dependent only on the quality, for posteriors also on properties of the underlying distribution.