Issues with the dot-tilde (i.e. .~) syntax

[yebai]

• Error with .~ and rand  #405
• Possibly confusing .~ meaning #435
• .~ seems to give incorrect answers #28
• Support for variables with changing size with dot_tilde #700

Below is copied from #710 (comment)

I lean towards removing the dot syntax. The benefits of using dot syntax against a loop or fillarray/product_distribution are relatively marginal compared to its complexity and lack of robustness.

[yebai]

cc @torfjelde @penelopeysm @mhauru @willtebbutt

[willtebbutt]

I would also be in favour of removing the .~ syntax -- could you remind me what the benefits of it are? I vaguely remember there being a couple of things that it can be used to do, that someone mentioned being trickier with fillarray / product_distribution.

[mhauru]

Might poll our userbase on Slack, see if anyone is desperate to keep .~. I would be very glad if we could get rid of it, but don't want to break user code lightly.

[torfjelde]

I would also be in favour of removing the .~ syntax -- could you remind me what the benefits of it are? I vaguely remember there being a couple of things that it can be used to do, that someone mentioned being trickier with fillarray / product_distribution.

There are "semantic" differences.

x ~ filldist(Normal(), 2)

treats x as a multivariate random variable of length 2, while

x .~ Normal()

treats x as a vector of 2 univariate random variables. This is effectively the difference between a multivariate random variable vs. a vector of IID variables.

[yebai]

Here is the idea I mentioned briefly during the meeting. We lower

x .~ Normal()

to

x = tilde_assume!!(__context__, IIDDistribution(Normal()), __varinfo__)

Then, we overload tilde_assume!! for IIDDistribution to store each RV x[i] separately in __varinfo__. This allows users to individually apply model operations like condition/fix on x[i].

We could make this work for other distribution sugars by overloading the tilde_assume!! (also tilde_observe!!) interfaces for new types like IIDDistribution.

EDIT: I am happy to remove the .~ feature if we can't find a good, robust workaround.