Alice puts a flat prior on the parameter θ. Bob works with φ = g(θ) and puts a flat prior on that. Their posteriors, translated back to θ, disagree. The Jeffreys prior (proportional to the square root of the Fisher information) is the one choice that gives the same answer whichever parameterisation is used: here the Fisher information of φ is computed numerically from the likelihood, not from the transformation rule, and the two Jeffreys priors agree.
Priors on θ: flat in θ is p(θ) = 1; flat in φ is p(θ) = |g′(θ)| (the change of variables); Jeffreys is p(θ) ∝ √F(θ). All curves are normalised numerically on a fine grid in the unconstrained variable (logit or log); when a prior is improper and the posterior does not integrate, a warning is shown.
code/ch03_statistics.py (fig_3_13); all figures in the code index.