`LogitNormal` gets no default transform, unlike other unit-interval distributions

Description

pm.LogitNormal is supported on the interval (0, 1), but unlike the other distributions on that support it is not assigned a default transform. As a result, NUTS operates on the constrained scale, proposes values outside the support, and a model using it fails at initialization.

Reproducible example

import pymc as pm

with pm.Model() as model:
    mu = pm.Normal("mu", 0, 2)
    sigma = pm.HalfNormal("sigma", sigma=1)
    xs = pm.LogitNormal("xs", mu=mu, sigma=sigma, shape=13)
    pm.Binomial("ks", n=[129] * 13, p=xs, observed=[4] * 13)
    idata = pm.sample()

Result:

pymc.exceptions.SamplingError: Initial evaluation of model at starting point failed!
Starting values:
{'mu': array(-0.58882144), 'sigma_log__': array(0.91378887), 'xs': array([ 0.19462499,  0.39542672,  1.05327918, -0.03184312,  0.82003634,
       -0.61738645,  1.26254703,  0.69715246,  0.60729211,  1.12854095,
        0.72072905, -0.44765787, -0.48220885])}
Logp initial evaluation results:
{'mu': np.float64(-1.66), 'sigma': np.float64(-2.42), 'xs': np.float64(-inf), 'ks': np.float64(-inf)}

Note that sigma appears as sigma_log__, because it did get its transform, while xs did not.

Note the negative entries in the starting value for xs. jitter+adapt_diag perturbs the initial point on the constrained scale, so with 13 components at least one lands outside (0, 1) essentially every time. I tried 6 seeds and it failed on all 6.

Comparison with other distributions on (0, 1)

for name, make in [
    ("Beta",        lambda: pm.Beta("v", 2, 2)),
    ("Kumaraswamy", lambda: pm.Kumaraswamy("v", 2, 2)),
    ("Uniform(0,1)", lambda: pm.Uniform("v", 0, 1)),
    ("LogitNormal", lambda: pm.LogitNormal("v", 0, 1)),
]:
    with pm.Model() as m:
        make()
    print(name, m.rvs_to_transforms[m.free_RVs[0]])

LogitNormal is the only one without a transform, and the only one that is not a subclass of a continuous distribution base class. It appears to be implemented as a wrapper that returns a CustomDist, which is presumably why it does not pick up the default transform that UnitContinuous provides.

For contrast, HalfNormal — also a constrained distribution — does get its LogTransform automatically, so the inconsistency is specific to this distribution rather than general.

Workaround

from pymc.distributions.transforms import logodds

xs = pm.LogitNormal("xs", mu=mu, sigma=sigma, shape=13,
                    default_transform=logodds)

With this, the model initializes and samples normally.

(For the hierarchical model above, a non-centered parameterization — xs = invlogit(mu + sigma * us) with us ~ Normal(0, 1) — both avoids the issue and samples better, but that is a modeling choice and does not address the inconsistency in the distribution itself.)

Question

Should LogitNormal be implemented as a proper UnitContinuous distribution, or at minimum default to the logodds transform? As it stands the failure is hard to diagnose: the error points at the starting values rather than at the missing transform, and nothing in the distribution's documentation suggests it behaves differently from Beta.

Happy to open a PR if there is a preferred approach.

Versions

• PyMC 6.3.1
• PyTensor 3.3.0
• Python 3.12.13
• Linux