fix: LogGaussianPrior declares its own (0, inf) support

`LogGaussianPrior`'s support is `(0, inf)` -- `log_prior_from_value` returns
`-inf` for `value <= 0` -- but the prior reported `(-inf, inf)`, because
`Prior.__getattr__` delegated to a `TransformedMessage` whose limits default to
`+/-inf` and were never set. `ClipperPriorBox` worked around it with an
`isinstance` switch (PyAutoFit#1477, follow-up 3); every other consumer of
`lower_limit` was simply told the wrong thing.

Declare the support on the prior itself, and give `Prior` a general strictness
contract so a consumer can tell an exclusive bound from an inclusive one without
a type switch:

- `Prior.lower_limit_strict` / `upper_limit_strict`, class attributes defaulting
  to `False`; `True` for `LogGaussianPrior`'s lower bound.
- `LogGaussianPrior` sets `lower_limit = 0.0` / `upper_limit = inf` in `__init__`.
- `Prior.limits` derives from `lower_limit`/`upper_limit` instead of returning a
  hardcoded `(-inf, inf)`, so the two notions cannot disagree by construction.
- `ClipperPriorBox._limits_from_model` reads the strictness flags; its
  `isinstance` block, import and workaround docstrings are retired.

The limits go on the prior rather than on the `TransformedMessage` deliberately.
On the message they would be dropped by `with_base`, `copy`, `project` and
`__call__`, and would change what `MeanField.lower_limit` and
`LaplaceOptimiser(check_limits=True)` feed to `OptimisationState.valid` -- a live
EP behaviour change well outside this fix.

Verified before/after over 37 behavioural probes: identifiers (prior and
search-with-clipper) byte-identical, so no output directory re-keys;
`log_prior_from_value` pointwise identical; the nested-sampler unit-cube mapping
identical; clipper bounds, projections and masks identical across 3 clipper
configurations x 3 input vectors.

One downstream-visible change: prior passing on a LogGaussian parameter with no
config `Limits` entry previously produced `TruncatedGaussian(-inf, inf)`, whose
`value_for(0.001)` was `-0.545` and whose `log_prior_from_value(-1.0)` was a
finite `-8.0`. It is now lower-bounded at `0.0`.

Tests: the reported support, the strictness flags, regression pins for
`log_prior_from_value` / the unit-cube mapping / the identifier, and a
parametrised property over every prior family asserting that a prior's reported
support matches its actual support -- the general form of this bug, which fails
on the parent commit for `LogGaussianPrior` alone.

Closes #1526

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AXEqAUkXK7UeXw5jGh2i4q

Author: claude

autofit/mapper/prior/abstract.py

@@ -21,6 +21,20 @@ class Prior(Variable, ABC, ArithmeticMixin):
 
     _ids = itertools.count()
 
+    #: Whether the support *excludes* the corresponding limit itself, i.e. whether
+    #: the bound is ``value > limit`` rather than ``value >= limit``.
+    #:
+    #: Most priors are inclusive at their limits: ``UniformPrior.log_prior_from_value``
+    #: is finite exactly on the bound. ``LogGaussianPrior`` is not — its support is the
+    #: open ``(0, inf)`` — and a consumer that clips onto an exclusive bound lands on a
+    #: point of zero density. These flags let such a consumer tell the two apart without
+    #: a per-type ``isinstance`` switch (see ``non_linear.clipper.ClipperPriorBox``).
+    #:
+    #: Class attributes, deliberately: they are a property of the prior *family*, never
+    #: of an instance, so they stay out of ``__dict__`` and out of the identifier.
+    lower_limit_strict = False
+    upper_limit_strict = False
+
     def __init__(self, message, id_=None):
         """
         An object used to mappers a unit value to an attribute value for a specific
@@ -347,10 +361,16 @@ def name_of_class(cls) -> str:
     def limits(self) -> Tuple[float, float]:
         """The (lower, upper) bounds of this prior.
 
-        Returns (-inf, inf) by default. Subclasses with finite bounds
-        (e.g. UniformPrior) override this.
+        Derived from ``lower_limit`` / ``upper_limit`` rather than stated
+        separately, so the two cannot disagree. They used to: this property
+        returned a hardcoded ``(-inf, inf)`` for every prior that did not
+        override it, which was right for ``GaussianPrior`` and wrong for
+        ``LogGaussianPrior``, whose support is ``(0, inf)``.
+
+        Use ``lower_limit_strict`` / ``upper_limit_strict`` to tell whether the
+        support includes the bounds returned here.
         """
-        return (float("-inf"), float("inf"))
+        return (float(self.lower_limit), float(self.upper_limit))
 
     def gaussian_prior_model_for_arguments(self, arguments):
         """Look up this prior in an arguments dict and return the mapped value.

autofit/mapper/prior/log_gaussian.py

@@ -12,6 +12,10 @@ class LogGaussianPrior(Prior):
     __identifier_fields__ = ("mean", "sigma")
     __database_args__ = ("mean", "sigma", "id_")
 
+    #: The support is the *open* interval ``(0, inf)``: ``log_prior_from_value``
+    #: returns ``-inf`` at ``0`` itself, so a consumer must stay strictly above it.
+    lower_limit_strict = True
+
     def __init__(
         self,
         mean: float,
@@ -51,6 +55,15 @@ def __init__(
         self.mean = mean
         self.sigma = sigma
 
+        # Declared on the prior, not on the message below. `TransformedMessage`
+        # defaults its limits to +/-inf and derives its `_support` separately, so
+        # a prior that left them to delegation reported (-inf, inf) for a strictly
+        # positive parameter -- the bug this shadowing fixes. Keeping the message's
+        # own limits untouched keeps the EP/Laplace machinery, which reads them,
+        # behaving exactly as before.
+        self.lower_limit = 0.0
+        self.upper_limit = float("inf")
+
         message = TransformedMessage(
             NormalMessage(mean, sigma),
             log_transform,

autofit/non_linear/clipper.py

@@ -99,8 +99,6 @@
 
 import numpy as np
 
-from autofit.mapper.prior.log_gaussian import LogGaussianPrior
-
 logger = logging.getLogger(__name__)
 
 
@@ -225,7 +223,9 @@ class ClipperPriorBox(AbstractClipper):
       ``(0, inf)``) — inset by an *absolute* ``strict_epsilon``. A relative margin
       is identically zero here, there being no finite width, so it would clip
       exactly onto ``0.0``, where the support is strict and ``log_prior`` is
-      ``-inf``. Only an absolute nudge lands strictly inside.
+      ``-inf``. Only an absolute nudge lands strictly inside. Which bounds are
+      exclusive is read off the prior's ``lower_limit_strict`` /
+      ``upper_limit_strict``, never inferred from its type.
 
     Parameters
     ----------
@@ -249,23 +249,20 @@ def _limits_from_model(self, model):
         The raw ``(lower, upper, lower_strict, upper_strict)`` arrays for ``model``,
         in physical parameter order.
 
-        Limits are read off ``prior.lower_limit`` / ``prior.upper_limit``, which
-        resolves for **every** prior type without a type switch: ``Prior.__getattr__``
-        delegates to the prior's message, and ``AbstractMessage`` defaults both to
-        ``±inf``. ``UniformPrior`` and ``LogUniformPrior`` shadow them with their
-        own attributes, ``TruncatedGaussianPrior`` picks up real limits from
-        ``TruncatedNormalMessage``, and ``GaussianPrior`` correctly falls through
-        to ``±inf``.
-
-        ``LogGaussianPrior`` is the one prior that read gets *wrong*, and silently.
-        Its message is a ``TransformedMessage``, which defaults its limits to
-        ``±inf`` and is never passed any — yet
-        ``LogGaussianPrior.log_prior_from_value`` returns ``-inf`` for
-        ``value <= 0``. Left uncorrected, the clipper would report that coordinate
-        as unbounded and fail to protect precisely the mechanism it exists to fix,
-        so its ``(0, inf)`` support is declared here. Declaring it on the prior
-        itself is the cleaner fix, but it would change a class the EP machinery and
-        the nested samplers also read, so it is deliberately kept local.
+        Limits are read off ``prior.lower_limit`` / ``prior.upper_limit``, and
+        strictness off ``prior.lower_limit_strict`` / ``prior.upper_limit_strict``.
+        Both resolve for **every** prior type without a type switch: ``UniformPrior``
+        and ``LogUniformPrior`` shadow the limits with their own attributes,
+        ``LogGaussianPrior`` declares its ``(0, inf)`` support and flags the lower
+        bound strict, ``TruncatedGaussianPrior`` picks up real limits from
+        ``TruncatedNormalMessage``, and ``GaussianPrior`` falls through
+        ``Prior.__getattr__`` to its message's ``±inf``.
+
+        This clipper used to declare ``LogGaussianPrior``'s support itself, because
+        that prior reported ``(-inf, inf)`` while ``log_prior_from_value`` returned
+        ``-inf`` for ``value <= 0``. That was a workaround for a defect in the prior,
+        fixed in PyAutoFit#1526: any consumer of ``lower_limit`` — not just this one —
+        was being told a strictly positive parameter could go negative.
 
         The ``strict`` flags mark bounds the support *excludes* (``value > limit``
         rather than ``value >= limit``); only those need the absolute inset.
@@ -275,12 +272,8 @@ def _limits_from_model(self, model):
         for prior in model.priors_ordered_by_id:
             low = float(prior.lower_limit)
             high = float(prior.upper_limit)
-            low_strict = False
-            high_strict = False
-
-            if isinstance(prior, LogGaussianPrior):
-                low = 0.0
-                low_strict = True
+            low_strict = bool(prior.lower_limit_strict)
+            high_strict = bool(prior.upper_limit_strict)
 
             lower.append(low)
             upper.append(high)

test_autofit/mapper/prior/test_log_gaussian.py

@@ -1,5 +1,7 @@
 import pickle
+from copy import copy
 
+import numpy as np
 import pytest
 
 import autofit as af
@@ -30,3 +32,129 @@ def test_pickle(log_gaussian):
 
 def test_identifier(log_gaussian):
     Identifier(log_gaussian)
+
+
+# === PyAutoFit#1526: the prior declares its own (0, inf) support ===
+#
+# The support was always (0, inf) -- ``log_prior_from_value`` returns -inf for
+# ``value <= 0`` -- but the prior reported (-inf, inf), because ``Prior.__getattr__``
+# delegated to a ``TransformedMessage`` whose limits default to +/-inf and were never
+# set. ``ClipperPriorBox`` worked around it with an ``isinstance`` switch; every other
+# consumer of ``lower_limit`` was simply told the wrong thing.
+#
+# The pinned values below are the pre-change ones, measured on the commit before the
+# fix. They are the guarantee that declaring the support moved *what the prior says*
+# and nothing about *what it computes*.
+
+
+def test__reports_its_own_support(log_gaussian):
+    assert log_gaussian.lower_limit == 0.0
+    assert log_gaussian.upper_limit == float("inf")
+    assert log_gaussian.limits == (0.0, float("inf"))
+
+
+def test__lower_bound_is_strict__upper_is_not(log_gaussian):
+    """
+    The support is the *open* (0, inf): ``log_prior_from_value(0.0)`` is -inf, so a
+    consumer that clips onto the reported bound must know to stay strictly above it.
+    """
+    assert log_gaussian.lower_limit_strict is True
+    assert log_gaussian.upper_limit_strict is False
+
+    assert log_gaussian.log_prior_from_value(log_gaussian.lower_limit) == -np.inf
+
+
+def test__other_prior_families_keep_their_limits():
+    """
+    ``Prior.limits`` now derives from ``lower_limit``/``upper_limit`` rather than
+    returning a hardcoded (-inf, inf). GaussianPrior must still be unbounded.
+    """
+    assert af.GaussianPrior(mean=0.0, sigma=1.0).limits == (-np.inf, np.inf)
+    assert af.UniformPrior(lower_limit=0.0, upper_limit=2.0).limits == (0.0, 2.0)
+    assert af.LogUniformPrior(lower_limit=0.01, upper_limit=100.0).limits == (
+        0.01,
+        100.0,
+    )
+
+    for cls in (af.GaussianPrior, af.UniformPrior, af.LogUniformPrior):
+        assert cls.lower_limit_strict is False
+        assert cls.upper_limit_strict is False
+
+
+@pytest.mark.parametrize(
+    "value, expected",
+    [
+        (-3.0, -np.inf),
+        (-1e-09, -np.inf),
+        (0.0, -np.inf),
+        (1e-12, -204.83588563011642),
+        (0.001, -8.892033838618836),
+        (0.1, 0.14164835190717184),
+        (0.5, 0.3396055360729164),
+        (1.0, -0.04733727810650888),
+        (2.0, -0.7185718164978876),
+        (10.0, -3.3735407249713125),
+        (1000.0, -19.437601069254285),
+    ],
+)
+def test__log_prior_from_value_is_unchanged(value, expected):
+    """
+    The density was always correct; only the reported limits were wrong. Pinned
+    against the pre-change values either side of zero.
+    """
+    prior = af.LogGaussianPrior(mean=0.4, sigma=1.3)
+    assert prior.log_prior_from_value(value) == pytest.approx(expected, rel=1e-12)
+
+
+@pytest.mark.parametrize(
+    "unit, expected",
+    [
+        (1e-09, 0.0006129978595719644),
+        (0.0001, 0.011858368820420245),
+        (0.01, 0.07249394511581292),
+        (0.1, 0.2819523947584061),
+        (0.25, 0.6207439038545902),
+        (0.5, 1.4918246976412703),
+        (0.75, 3.5852803622761034),
+        (0.9, 7.893321602745905),
+        (0.99, 30.69968015862632),
+        (0.999999999, 3630.585196715403),
+    ],
+)
+def test__unit_cube_mapping_is_unchanged(unit, expected):
+    """
+    The nested samplers work in unit-cube coordinates and map through the prior. The
+    limits live on the prior while the mapping lives on the message stack, which this
+    change does not touch -- so no stored nested-sampling result shifts.
+    """
+    prior = af.LogGaussianPrior(mean=0.4, sigma=1.3)
+    assert prior.value_for(unit) == pytest.approx(expected, rel=1e-12)
+    assert prior.unit_value_for(prior.value_for(unit)) == pytest.approx(unit, rel=1e-6)
+
+
+def test__identifier_is_unchanged():
+    """
+    If the declared limits fed the identifier, every existing output directory would
+    re-key and its stored results would be orphaned. ``__identifier_fields__`` is
+    ("mean", "sigma"), and the limits are instance/class attributes outside it, so the
+    hash is untouched. Pinned to the pre-change value.
+    """
+    prior = af.LogGaussianPrior(mean=0.4, sigma=1.3)
+    assert str(Identifier(prior)) == "34cb61ade6bafa6050229e8b6b390235"
+
+
+def test__declared_support_survives_copy_pickle_and_projection(log_gaussian):
+    """
+    The limits are derived in ``__init__`` rather than carried as parameters, which is
+    what makes them survive every path that rebuilds the prior -- including the JAX
+    pytree round-trip, where only (mean, sigma, id) are flattened.
+    """
+    assert copy(log_gaussian).lower_limit == 0.0
+    assert pickle.loads(pickle.dumps(log_gaussian)).lower_limit == 0.0
+
+    samples = np.exp(np.random.default_rng(0).normal(1.0, 2.0, 500))
+    projected = log_gaussian.project(samples, np.zeros(500))
+    assert projected.lower_limit == 0.0
+
+    rebuilt = af.LogGaussianPrior.tree_unflatten((), log_gaussian.tree_flatten()[0])
+    assert rebuilt.lower_limit == 0.0

test_autofit/mapper/prior/test_prior_properties.py

@@ -271,3 +271,83 @@ def test__from_mode_matches_mean_and_variance(cls, mean, variance):
     message = cls.from_mode(np.asarray(mean), np.asarray(variance))
     assert float(message.mean) == pytest.approx(mean, rel=1e-6)
     assert float(message.variance) == pytest.approx(variance, rel=1e-6)
+
+
+# === P6: the reported support matches the actual support ===
+
+
+def interior_probes(prior):
+    """
+    Points that lie strictly inside the prior's *reported* ``limits``.
+
+    Every one of them must have finite log prior. If a prior reports a box wider
+    than its true support, some point in here falls in the gap and the density
+    there is -inf -- which is precisely the shape of PyAutoFit#1526, where
+    ``LogGaussianPrior`` reported (-inf, inf) for a support of (0, inf).
+    """
+    lo, hi = prior.limits
+    probes = []
+
+    if np.isfinite(lo) and np.isfinite(hi):
+        width = hi - lo
+        probes += [lo + 0.25 * width, lo + 0.5 * width, lo + 0.75 * width]
+        probes += [lo + 1e-9 * width, hi - 1e-9 * width]
+    elif np.isfinite(lo):
+        probes += [lo + 1e-9, lo + 1.0, lo + 1e3]
+    elif np.isfinite(hi):
+        probes += [hi - 1e-9, hi - 1.0, hi - 1e3]
+    else:
+        probes += [-1e6, -1.0, 0.0, 1.0, 1e3]
+
+    return probes
+
+
+@pytest.mark.parametrize("prior", all_priors(), ids=prior_id)
+def test__log_prior_is_finite_everywhere_inside_the_reported_limits(prior):
+    """
+    The general form of PyAutoFit#1526. A prior that reports a bound it does not
+    actually have hands every consumer a licence to evaluate where the density is
+    zero -- and for a strictly positive parameter, ``log(0)`` or a division by it
+    is the failure that follows.
+    """
+    for value in interior_probes(prior):
+        assert np.isfinite(
+            prior.log_prior_from_value(value)
+        ), f"{prior} reports limits {prior.limits} but log_prior({value}) is not finite"
+
+
+@pytest.mark.parametrize("prior", all_priors(), ids=prior_id)
+def test__log_prior_is_minus_inf_outside_the_reported_limits(prior):
+    """
+    The converse: the reported box must not be *narrower* than the support either.
+    """
+    lo, hi = prior.limits
+
+    if np.isfinite(lo):
+        assert prior.log_prior_from_value(lo - 1.0) == -np.inf
+    if np.isfinite(hi):
+        assert prior.log_prior_from_value(hi + 1.0) == -np.inf
+
+
+@pytest.mark.parametrize("prior", all_priors(), ids=prior_id)
+def test__strictness_flags_agree_with_the_density_at_the_bounds(prior):
+    """
+    A bound flagged strict must have zero density on it; a bound not flagged strict
+    must have finite density on it. This is what lets a consumer clip onto a bound
+    without a per-type switch -- ``ClipperPriorBox`` insets only the strict ones.
+    """
+    lo, hi = prior.limits
+
+    if np.isfinite(lo):
+        at_lower = prior.log_prior_from_value(lo)
+        if prior.lower_limit_strict:
+            assert at_lower == -np.inf
+        else:
+            assert np.isfinite(at_lower)
+
+    if np.isfinite(hi):
+        at_upper = prior.log_prior_from_value(hi)
+        if prior.upper_limit_strict:
+            assert at_upper == -np.inf
+        else:
+            assert np.isfinite(at_upper)

test_autofit/non_linear/test_clipper.py

@@ -69,16 +69,24 @@ def test__gaussian_prior__passes_through_as_infinite(self):
         assert not np.isnan(lower).any()
         assert not np.isnan(upper).any()
 
-    def test__log_gaussian_prior__lower_bound_is_declared_by_the_clipper(self):
+    def test__log_gaussian_prior__lower_bound_is_declared_by_the_prior(self):
         """
-        LogGaussianPrior reports (-inf, inf) because its TransformedMessage is never
-        given limits, yet ``log_prior_from_value`` is -inf for value <= 0. The
-        clipper declares the real (0, inf) support, strictly, so the projected value
+        LogGaussianPrior declares its own (0, inf) support and flags the lower bound
+        strict, so the clipper reads it like any other prior and the projected value
         lands *above* zero rather than on it.
+
+        This test used to assert ``prior.lower_limit == -np.inf`` and the clipper
+        supplied the real bound itself via an ``isinstance`` switch. That was a
+        workaround for PyAutoFit#1526: the prior was telling every consumer, not just
+        this one, that a strictly positive parameter could go negative. The bounds
+        assertions below are unchanged by the fix — that equivalence is the point.
         """
         prior = af.LogGaussianPrior(mean=0.0, sigma=1.0)
 
-        assert prior.lower_limit == -np.inf
+        assert prior.lower_limit == 0.0
+        assert prior.upper_limit == np.inf
+        assert prior.lower_limit_strict is True
+        assert prior.upper_limit_strict is False
 
         model = _model(alpha=prior)
         lower, upper = ClipperPriorBox(strict_epsilon=1.0e-12).bounds_from_model(