Merge pull request rapidsai#4683 from rapidsai/branch-22.04

[gpuCI] Forward-merge branch-22.04 to branch-22.06 [skip gpuci]

python/cuml/neighbors/kernel_density.py

@@ -377,7 +377,7 @@ def sample(self, n_samples=1, random_state=None):
 
         supported_kernels = ["gaussian", "tophat"]
         if (self.kernel not in supported_kernels
-                or self.metric != "euclidian"):
+                or self.metric != "euclidean"):
             raise NotImplementedError(
                 "Only {} kernels, and the euclidean"
                 " metric are supported.".format(supported_kernels))

python/cuml/test/test_kernel_density.py

@@ -102,12 +102,18 @@ def test_kernel_density(arrays, kernel, metric, bandwidth):
         assert np.allclose(np.exp(as_type("numpy", cuml_prob_test)),
                            ref_prob_test, rtol=tol, atol=tol, equal_nan=True)
 
-    if kernel in ["gaussian", "tophat"] and metric == "euclidian":
+    if kernel in ["gaussian", "tophat"] and metric == "euclidean":
         sample = kde.sample(100, random_state=32).get()
-        nearest = skl_pairwise_distances(sample, X, metric=metric)
+        nearest = skl_pairwise_distances(sample, X_np, metric=metric)
         nearest = nearest.min(axis=1)
         if kernel == "gaussian":
-            assert np.all(nearest < 5 * bandwidth)
+            from scipy.stats import chi
+            # The euclidean distance of each sample from its cluster
+            # follows a chi distribution (not squared) with DoF=dimension
+            # and scale = bandwidth
+            # Fail the test if the largest observed distance
+            # is vanishingly unlikely
+            assert chi.sf(nearest.max(), X.shape[1], scale=bandwidth) > 1e-8
         elif kernel == "tophat":
             assert np.all(nearest <= bandwidth)
     else: