```suggestion - `DEMetropolisZ`, an improved variant of `DEMetropolis` brings better parallelization and higher efficiency with fewer chains with a slower initial convergence. This implementation is experimental. See [#3784](https://github.com/pymc-devs/pymc3/pull/3784) for more info. ```

Shouldn't be "DEMetropolisZ can achieve the same effective sample sizes with less chains."

same as my other comment

I think this import is not used. Also could we follow the convention of using alias when importing the libraries like `import numpy as np`, I think this is the style we follow in most notebooks.

Need link.

```suggestion - `DEMetropolisZ`, an improved variant of `DEMetropolis` brings better parallelization and high efficiency with few chains with a slower initial convergence. This implementation is experimental. See [#3784](https://github.com/pymc-devs/pymc3/pull/3784) for more info. ```

Being picky here. This test would be easier to read if written like this ``` tune =300 tune_drop_fraction=.85 draws=200 step = DEMetropolisZ(tune_drop_fraction=tune_drop_fraction) trace = sample( tune=tune, draws=draws, step=step, cores=1, chains=1, discard_tuned_samples=False ) assert len(step._history) == (tune - tune * tune_drop_fraction) + draws ``` That way there's less mental steps to correlate where the step_history expectation is coming from. Same for trace

Same comment here about parameterization It's worth parameterizing this so in the vent multiple tests fail its easier to figure out which ones are failing. Right now earlier test failures will hide the failures of the subsequent ones

It's worth parameterizing this so in the vent multiple tests fail its easier to figure out which ones are failing. Right now earlier test failures will hide the failures of the subsequent ones

What is this pass doing?

Can this be made a standard triple quotes comment? That way it shows up with `help()` and doc discovery tools, such as the ones in ides

What's 2.38? Can a comment be left for the magic number?

This is very picky; Would be easier to read if written `if tune not in {None, 'scaling', 'lambda'}:` because then it reads like english instead of like a program. Not implying that a change is required

Could we use a more informative name?

what about `iz1, iz2 = np.random.choice(range(it+1), size=2, replace=False)`

Is there any concern for memory for large num_step? It would be great if we can clear out `self._history` if user are calling from `pm.sample`

Should `self._it` and `self._history` also be reset after tuning?

need an Experimental warning.