Chad Scherrer

> This seems like a cool use case that I admittedly have never thought about before. Thanks! > How is this sampling expected to behave? Sample one distribution of the...

@baggepinnen understands this much better then I do, but I think the issue is how to compute `a unsafe_comparisons(true) [ Info: Unsafe comparisons using the function `Statistics.mean` has been enabled...

It seems that relative to `rand(Bernoulli(0.3))`, `rand($pd)` does 1000x the work in 200x the time, is that right? I'm missing how this shows inefficiency. But it's early here, so maybe...

Are the `isa` calls in `replace_particles` all compiled away? If not, would it be faster to turn each of these into a different method? Or if generated functions would help,...

This looks great! I wonder, how far can "particle semantics" be pushed? For the last `pd` in your example, could `rand(pd)` return a `Particles` by drawing one sample from each...

In a typical MCM use case, say you have some function `f(x::Real)` that returns another `Real`. If you give that function a `Particles`, you might expect it to return another...

This comes up all the time in probabilistic programming. We have uncertainty on our parameter estimates (because they're samples from the posterior distribution), which we propagate through the observation model...

For some cases there are some nice optimizations, e.g. ```julia rand(ParticleDistribution(Normal, m, s)) == m + s * Particles(Normal()) ``` (assuming I got that right)

That's really close! I'd guess the latter might be a little more stable, since it takes advantage of stratified sampling. But I'm not sure about that. You mentioned https://github.com/JuliaArrays/StructArrays.jl, did...

Oh interesting, I thought `Particles()` used systematic sampling by default. Think I'm misunderstanding something