Wrong handling of term ordering leads to incorrect differentiation

[e-dorigatti]

Hello, differentiation in the latest version sometimes gives incorrect results compared to the earlier version:

import formulaic
print(formulaic.__version__)

f = formulaic.Formula("x1 + x2 + x3 + x1:x2 + {x3**2}")
print(f)

for var in ['x1', 'x2', 'x3']:
    print(var, f.differentiate(var, use_sympy=True))

Here are the outputs:

1.1.0
1 + x1 + x2 + x3 + x3 ** 2 + x1:x2
x1 0 + 1 + 0 + 0 + 0 + x2
x2 0 + 0 + 1 + 0 + 0 + x1
x3 0 + 0 + 0 + 1 + 0 + (2*x3)

and

1.0.2
1 + x1 + x2 + x3 + x3 ** 2 + x1:x2
x1 root:
    [0, 1, 0, 0, 0, x2]
x2 root:
    [0, 0, 1, 0, 0, x1]
x3 root:
    [0, 0, 0, 1, (2*x3), 0]

Seems like the issue is related to the default ordering of the terms, as the correct results can be obtained by explicitly disabling it:

f = formulaic.Formula("x1 + x2 + x3 + x1:x2 + {x3**2}")
f.ordering = formulaic.formula.OrderingMethod.NONE

[matthewwardrop]

Hi @e-dorigatti ! Thanks for surfacing this! I'll make sure it gets fixed soon.

I didn't realise anyone other than myself had ever shown an interest in the differentiation features. How do you use them in practice?

[e-dorigatti]

Hi @matthewwardrop thank you for your prompt answer! We are using formulas in our bofire package to express the model for an experimental design, and use the differentiation features to optimize the design with respect to some optimality criteria. In practice we use this to plan experiments :)

[matthewwardrop]

Nice! Maybe I should upgrade them from experimental status then :).

[matthewwardrop]

This has been fixed and released as 1.1.1.