Addition of non-uniform priors.

[arm61]

This PR implements the ability to pass arbitrary scipy.stats objects as priors. This means that non-uniform priors can be passed to different FittingBase objects.

However, to achieve this, I have removed the bounds keyword argument, which is an API change; hence, this will be a part of a 2.1.0 release to account for this.

Closes #143

[Copilot]

Pull request overview

This PR replaces parameter bounds with support for arbitrary scipy.stats frozen distributions as priors across FittingBase-derived fitting classes, enabling non-uniform priors (per #143) and updating tests/docs accordingly.

Changes:

• Introduce priors support in FittingBase and propagate it through Arrhenius/VTF and Yeh-Hummer fitting.
• Add/adjust unit tests to cover prior initialization and usage.
• Update Arrhenius tutorial documentation to demonstrate providing priors.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
kinisi/fitting.py	Replaces bounds-based configuration with prior distributions; adds MAP objective and validation.
kinisi/arrhenius.py	Switches Arrhenius/VTF wrappers to pass priors into FittingBase.
kinisi/yeh_hummer.py	Updates YehHummer to accept priors and passes them to FittingBase; adjusts default prior construction.
kinisi/tests/test_fitting.py	Adds new unit tests for FittingBase priors and sampling.
kinisi/tests/test_arrhenius.py	Updates Arrhenius/VTF tests to use priors instead of bounds.
kinisi/tests/test_yeh_hummer.py	Updates YehHummer tests to exercise custom priors and invalid priors.
docs/source/arrhenius_t.ipynb	Updates tutorial to demonstrate priors-based API and explain scipy.stats.uniform.

Comments suppressed due to low confidence (3)

kinisi/tests/test_yeh_hummer.py:67

• Leftover debug print statements in the test will pollute test output and can cause CI noise/failures when output is asserted.

        assert 1e-4 < yh.shear_viscosity.value < 1e-2

    def test_yeh_hummer_mcmc(self):

kinisi/arrhenius.py:161

• VogelFulcherTammann.__init__ still uses a bounds-like type annotation for priors, but FittingBase expects a sequence of frozen scipy.stats distributions. The current annotation is inconsistent with runtime validation.

    def __init__(
        self,
        diffusion,
        priors: tuple[
            tuple[VariableLike, VariableLike], tuple[VariableLike, VariableLike], tuple[VariableLike, VariableLike]
        ]
        | None = None,
    ) -> 'VogelFulcherTammann':

kinisi/fitting.py:204

• max_aposteriori is annotated to return tuple[float] but currently returns None. Returning the fitted parameters keeps the API consistent with the type hint and with max_likelihood.

            self.data_group[name] = result[i] * self.parameter_units[i]

    def mcmc(self, n_samples: int = 1000, n_walkers: int = 32, n_burn: int = 500, n_thin=10) -> None:
        """

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
2031	1923	95%	0%	🟢

New Files

File	Coverage	Status
kinisi/tests/test_fitting.py	100%	🟢
TOTAL	100%	🟢

Modified Files

File	Coverage	Status
kinisi/_init_.py	100%	🟢
kinisi/arrhenius.py	100%	🟢
kinisi/fitting.py	96%	🟢
kinisi/tests/test_arrhenius.py	100%	🟢
kinisi/tests/test_yeh_hummer.py	100%	🟢
kinisi/yeh_hummer.py	88%	🟢
TOTAL	97%	🟢

updated for commit: 08d9b25 by action🐍

[arm61]

@copilot please re-review

[Copilot]

@copilot please re-review

Addressed the remaining unresolved threads in commit Fix docstring, default prior construction, return annotation, and test slope/viscosity labeling:

• yeh_hummer.py docstring — updated to say [D_0, slope] instead of [D_0, viscosity], which matches how priors are actually applied internally.
• Default slope prior bug — removed the erroneous * self._slope_unit multiplication (lines 73-74) that wrapped the plain floats from viscosity_to_slope() back into scipp variables, causing a UnitError when the optimizer called logpdf with plain floats. Also fixed the D_prior scale from the incorrect D_max + 0.8 to 1.2 * D_max so the support is [0.8*D_max, 2.0*D_max] as intended.
• max_likelihood / max_aposteriori — added explicit -> None return annotations since both methods store results as side effects on data_group.
• Test — fixed the second prior comment from "viscosity prior" to "slope prior", and updated the assertion to compare the fitted slope value against the slope prior bounds (instead of comparing shear_viscosity against a slope-unit prior range, which was semantically wrong).

[PythonFZ]

unrelated to the PR, but could be interesting to you https://github.com/ofek/hatch-vcs ( very easy to add PythonFZ/zndraw-joblib#2 ) and auto-updates the version based on git tags + you get the commit hash if you are working with a dev version

[arm61]

Happy to welcome a PR that implemented this!

[PythonFZ]

Looks great. The only thing worth mentioning is, are we fine with removing the units previously available in the bounds? Are there any scenarios where this could cause issues?

[arm61]

Yeah, the units thing is a bit annoying. I had an idea to write a wrapper for script.stats objects that did the unit work, but I think it is a bit of a big job...

[PythonFZ]

The wrapper is probably out of scope and the addition is quite valuable to tweak sampling from e.g. log-space distributed samples. LGTM 👍