mirror wgf for constrained params

[eclipse1605]

• particles stay in dual value_vars; particle drift uses primal ∇; noise σ = exp(-½ log_jac_det); ArviZ posterior packed via backward
• dual vs primal grad split: point APIs stay ∇_y; flat/batched/drift apply primal scaling

closes #6

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[yannmclatchie]

Suggested rewording: "The likelihood is still misspecified: a single Gaussian cannot match $P_0$, but a scale mixture of Gaussians can."

Reply via ReviewNB

[yannmclatchie]

Maybe just add one sentence here to the effect of "constrained parameters are automatically detected and the flow uses a mirror-mapped geometry based on the default PyMC [transformations](https://www.pymc.io/projects/docs/en/stable/api/distributions/transforms.html)."

Reply via ReviewNB

[yannmclatchie]

Now we can write here how we will investigate whether the traces respected their constraints throughout the course of the flow. (which I think you have previously written in the cells above)

Reply via ReviewNB

[yannmclatchie]

And here we can write how by identifying a bimodal mixture over the scales the induced predictive distribution also more closely resembles the observed data distribution. If you wanted to take this farther, you could even fit the same model with a Bayes posterior and plot their respective posterior predictive checks: the Bayes posterior will have poor predictive calibration/coverage. (This is not strictly necessary for the purposes of getting the package up and running, but potentially interesting for people who are less familiar with predictively oriented posteriors)

Reply via ReviewNB

[yannmclatchie]

probably no need for this cell or the one titled Scale trajectories, we can just write a single code comment at the top of the cell like # plot particle trajectories and # plot posterior histogram : the point of this notebook isn't to discuss inference output shapes but rather the mirror mapped behaviour (you can maybe add a cell after the trajectories which points out that the particles remain positive; in fact you could even show the y-axis in sqrt or log scale to make this clearer, but I don't know if it's super helpful)

Reply via ReviewNB