Inconsistencies in HARK.distribution: DiscreteDistributionLabeled, combine_indep_dstns, expected

[sbenthall]

There's something funky about how a few of our distribution operations work together.

The following code calls on several features in HARK.distribution:

• DiscreteDistributionLabeled
• combine_indep_dstns
• expected

and breaks unexpectedly.

from HARK.distribution import expected, Normal
from HARK.model import discretized_shock_dstn
import numpy as np

shocks = {
    "gamma": Normal(mu= 0, sigma = 1)
}

approx = {
    "gamma" : {"N" : 7}
}

ds = discretized_shock_dstn(shocks, approx)

discretized_shock_dstn creates a DiscreteDistributionLabeled class for each shock passed in (as a dictionary), and then combines them. See here: https://github.com/econ-ark/HARK/blob/master/HARK/model.py#L44

Then the following code breaks:

def sample_f(shocks):    
    return shocks["gamma"] * np.array([2,4])

expected(func = sample_f, dist = ds)

This produces an error:

Cell In[18], line 3, in sample_f(shocks)
      1 def sample_f(shocks):
----> 3     return shocks["gamma"] * np.array([2,4])

ValueError: operands could not be broadcast together with shapes (7,) (2,) 

The issue being that the full grid of the discretized distribution is being passed (in a dictionary, 'labeled' form) to the function sample_f, instead of one value being passed in at a time. This doesn't typically raise an error, but it does when the function includes an array operation.

I believe these different functions in HARK.distribution were written by different people, so many it's a matter of harmonizing them.

[sbenthall]

Oh, this looks like it's because of the difference between expected and calc_expectations.
expected wants the function passed to it to take... the whole vector of probability atoms?

It says "This function should take the full array of distribution values
and return either arrays of arbitrary shape or scalars."

https://github.com/econ-ark/HARK/blob/master/HARK/distribution.py#L2430

That doesn't make sense to me... the values of either the probabilities or the realized values isn't enough to compute the expected value, and there isn't an argument for both.

I see this is widely used in the HARK models but the docs don't make sense to me. Can somebody explain?