Add analytical results as member functions

[kyleabeauchamp]

I changed the test systems to be classes, so that we can add member functions that calculate various analytical (or experimental) properties that will be useful for unit testing.

PS: we should also include docstrings with references and some equations, if necessary. The point here is for the member functions for calculating analytical properties to be self-documenting.

[kyleabeauchamp]

PS: at some point we might want to create an abstract base class for all these test system classes

[jchodera]

PS: at some point we might want to create an abstract base class for all these test system classes

Agreed. How about a TestSystem object with some methods for computing analytical properties?

Which properties would we want to compute? I could imagine

• potential energy: mean and standard deviation
• volume (for NPT systems like IdealGas): mean and standard deviation

Are others useful?

[jchodera]

Should we create a single method getAnalyticalProperties() that returns a dict of analytical properties, or should we have separate get* functions such as getPotentialEnergyMean, getPotentialEnergyStddev, that return None if this property is unavailable?

If we use a dict, what naming convention is best? PotentialEnergyMean, or potential_energy_mean, or a double-dict with property['potential energy']['mean'] and property['potential energy']['stddev'], etc?

[kyleabeauchamp]

So I think we should just have one function per analytical property.

A member function can be accessed in the same manner as a dict object, but has the added benefit of having a docstring attached to it. To me, that's a huge benefit, as it allows us to attaching provenance and derivations directly to the data.

[kyleabeauchamp]

There are other ways we could handle this, obviously, but to me this is the most obvious one.

[jchodera]

Good point. We'll effectively be using introspection to see what analytical results are available.

[jchodera]

Is there any advantage to having a base class that implements all possible analytical result methods and returns a NotImplementedException for all of them? Or is introspection the more Pythonic way to do this?

[kyleabeauchamp]

I don't see any major advantage to that.

To me, one main benefit of having an abstract base class is to teach new developers how to create their own test systems--by listing the necessary member functions, we define a clear interface for future test additions.

[kyleabeauchamp]

PS: the base class would also be at the top of the module and say "This is how you create a new test..."