Reject invalid replay-grid log likelihoods

src/pyrecest/filters/replay_grid_likelihood.py

@@ -254,9 +254,10 @@ def adaptive_position_proposal_probability(
 
 
 def grid_proposal_weights(log_likelihood) -> np.ndarray:
-    """Convert finite grid log likelihoods to normalized proposal weights."""
+    """Convert valid grid log likelihoods to normalized proposal weights."""
 
     values = np.asarray(log_likelihood, dtype=float)
+    _validate_grid_log_likelihood_values(values)
     finite = np.isfinite(values)
     if not np.any(finite):
         raise ValueError("all grid log-likelihoods are non-finite")
@@ -356,9 +357,15 @@ def _coerce_grid_values(values, expected_size: int) -> np.ndarray:
     values = np.asarray(values, dtype=float)
     if values.shape != (expected_size,):
         raise ValueError(f"log_likelihood must have shape ({expected_size},)")
+    _validate_grid_log_likelihood_values(values)
     return values
 
 
+def _validate_grid_log_likelihood_values(values: np.ndarray) -> None:
+    if np.any(np.isnan(values) | np.isposinf(values)):
+        raise ValueError("log_likelihood must contain finite values or -np.inf")
+
+
 def _linear_rectilinear_grid_values(
     positions: np.ndarray,
     values: np.ndarray,
@@ -433,8 +440,8 @@ def _nearest_grid_values(
 
     indices = _nearest_bin_indices(positions[finite_positions], bin_tree)
     nearest_values = values[indices]
-    # A non-finite grid value represents zero or undefined likelihood. Borrowing
-    # the minimum finite value from another bin would create spurious mass.
+    # A -inf grid value represents zero likelihood. Borrowing the minimum finite
+    # value from another bin would create spurious mass.
     nearest_values = np.where(
         np.isfinite(nearest_values), nearest_values, float(log_zero)
     )

tests/filters/test_replay_grid_log_likelihood_validation.py

@@ -0,0 +1,47 @@
+import unittest
+
+import numpy as np
+
+# pylint: disable=no-name-in-module
+from pyrecest.filters import grid_proposal_weights, replay_grid_log_likelihood_values
+
+
+class TestReplayGridLogLikelihoodValidation(unittest.TestCase):
+    def test_rejects_nan_and_positive_infinity(self):
+        bin_centers = np.asarray([[0.0], [1.0]])
+        positions = np.asarray([[0.0]])
+
+        for invalid_value in (np.nan, np.inf):
+            log_likelihood = np.asarray([0.0, invalid_value])
+            with self.subTest(invalid_value=invalid_value, api="lookup"):
+                with self.assertRaisesRegex(
+                    ValueError, "finite values or -np.inf"
+                ):
+                    replay_grid_log_likelihood_values(
+                        positions,
+                        log_likelihood,
+                        bin_centers,
+                        interpolation="nearest",
+                    )
+            with self.subTest(invalid_value=invalid_value, api="proposal"):
+                with self.assertRaisesRegex(
+                    ValueError, "finite values or -np.inf"
+                ):
+                    grid_proposal_weights(log_likelihood)
+
+    def test_negative_infinity_remains_zero_likelihood(self):
+        bin_centers = np.asarray([[0.0], [1.0]])
+        log_likelihood = np.asarray([0.0, -np.inf])
+
+        np.testing.assert_allclose(grid_proposal_weights(log_likelihood), [1.0, 0.0])
+        looked_up = replay_grid_log_likelihood_values(
+            np.asarray([[1.0]]),
+            log_likelihood,
+            bin_centers,
+            interpolation="nearest",
+        )
+        self.assertTrue(np.isneginf(looked_up[0]))
+
+
+if __name__ == "__main__":
+    unittest.main()