Preserve Gauss-von-Mises single-point autograd

src/pyrecest/distributions/cart_prod/gauss_von_mises_distribution.py

@@ -234,7 +234,7 @@ def pdf(self, xs):
         )
 
         if single_point:
-            return float(p[0])
+            return p[0]
         return array(p)
 
     def mode(self):

tests/distributions/test_gauss_von_mises_pytorch_autograd.py

@@ -0,0 +1,47 @@
+from __future__ import annotations
+
+import pytest
+
+import pyrecest.backend
+from pyrecest.distributions.cart_prod.gauss_von_mises_distribution import (
+    GaussVonMisesDistribution,
+)
+
+torch = pytest.importorskip("torch")
+
+pytestmark = pytest.mark.skipif(
+    pyrecest.backend.__backend_name__ != "pytorch",
+    reason="PyTorch backend regression",
+)
+
+
+def test_single_point_pdf_preserves_pytorch_autograd() -> None:
+    distribution = GaussVonMisesDistribution(
+        mu=2.0,
+        P=1.3,
+        alpha=3.0,
+        beta=0.0,
+        Gamma=0.001,
+        kappa=0.7,
+    )
+    point = torch.tensor(
+        [0.8, 1.4],
+        dtype=torch.float64,
+        requires_grad=True,
+    )
+
+    density = distribution.pdf(point)
+
+    assert torch.is_tensor(density)
+    assert density.ndim == 0
+    assert density.dtype == point.dtype
+    assert density.device == point.device
+    assert density.requires_grad
+    assert torch.isfinite(density)
+    assert density > 0.0
+
+    density.backward()
+
+    assert point.grad is not None
+    assert torch.all(torch.isfinite(point.grad))
+    assert torch.all(point.grad != 0.0)