added note and assert that sbvm conc < 10k (#3412)

pyro/distributions/sine_bivariate_von_mises.py

@@ -55,6 +55,8 @@ class SineBivariateVonMises(TorchDistribution):
     .. note:: In the context of :class:`~pyro.infer.SVI`, this distribution can be used as a likelihood but not for
         latent variables.
 
+    .. note:: Normalization remains accurate up to concentrations of 10,000.
+
     ** References: **
       1. Probabilistic model for two dependent circular variables Singh, H., Hnizdo, V., and Demchuck, E. (2002)
       2. Protein Bioinformatics and Mixtures of Bivariate von Mises Distributions for Angular Data,
@@ -108,6 +110,16 @@ def __init__(
         ) = broadcast_all(
             phi_loc, psi_loc, phi_concentration, psi_concentration, correlation
         )
+
+        max_conc = torch.maximum(
+            torch.max(phi_concentration), torch.max(psi_concentration)
+        )
+        assrt_hstr = (
+            "Normalization of SineBiviateVonMises is inaccurate for"
+            f"current max concentration ({max_conc} > 10,000)."
+        )
+        assert max_conc <= torch.tensor(10_000.0), assrt_hstr
+
         self.phi_loc = phi_loc
         self.psi_loc = psi_loc
         self.phi_concentration = phi_concentration

tests/distributions/test_sine_bivariate_von_mises.py

@@ -132,8 +132,15 @@ def guide(data):
         assert_equal(expected[k].squeeze(), actual.squeeze(), 9e-2)
 
 
-@pytest.mark.parametrize("conc", [1.0, 10.0, 1000.0, 10000.0])
+@pytest.mark.parametrize("conc", [1.0, 10.0, 1000.0, 10_000.0, 10_001.0])
 def test_sine_bivariate_von_mises_norm(conc):
+    if conc > 10_000.0:
+        try:
+            dist = SineBivariateVonMises(0, 0, conc, conc, 0.0)
+            pytest.fail()
+        except AssertionError:
+            return
+
     dist = SineBivariateVonMises(0, 0, conc, conc, 0.0)
     num_samples = 500
     x = torch.linspace(-torch.pi, torch.pi, num_samples)