Change linear_operator_test_util.random_postive_definite_matrix to return nicer conditioned matrix by default (a Wishart matrix). Enable tests in oss again.

PiperOrigin-RevId: 294817440
Change-Id: Ib9a976e9c78fba41863639d7c158e2a3b25523e8

tensorflow/python/kernel_tests/linalg/BUILD

@@ -260,7 +260,6 @@ cuda_py_test(
     srcs = ["linear_operator_full_matrix_test.py"],
     shard_count = 5,
     tags = [
-        "no_oss",  # b/149346156
         "noasan",
         "optonly",
     ],

tensorflow/python/ops/linalg/linear_operator_test_util.py

@@ -825,15 +825,23 @@ class NonSquareLinearOperatorDerivedClassTest(LinearOperatorDerivedClassTest):
       return 2
 
 
-def random_positive_definite_matrix(shape, dtype, force_well_conditioned=False):
-  """[batch] positive definite matrix.
+def random_positive_definite_matrix(shape,
+                                    dtype,
+                                    oversampling_ratio=4,
+                                    force_well_conditioned=False):
+  """[batch] positive definite Wisart matrix.
+
+  A Wishart(N, S) matrix is the S sample covariance matrix of an N-variate
+  (standard) Normal random variable.
 
   Args:
     shape:  `TensorShape` or Python list.  Shape of the returned matrix.
     dtype:  `TensorFlow` `dtype` or Python dtype.
-    force_well_conditioned:  Python bool.  If `True`, returned matrix has
-      eigenvalues with modulus in `(1, 4)`.  Otherwise, eigenvalues are
-      chi-squared random variables.
+    oversampling_ratio: S / N in the above.  If S < N, the matrix will be
+      singular (unless `force_well_conditioned is True`).
+    force_well_conditioned:  Python bool.  If `True`, add `1` to the diagonal
+      of the Wishart matrix, then divide by 2, ensuring most eigenvalues are
+      close to 1.
 
   Returns:
     `Tensor` with desired shape and dtype.
@@ -843,11 +851,21 @@ def random_positive_definite_matrix(shape, dtype, force_well_conditioned=False):
     shape = tensor_shape.TensorShape(shape)
     # Matrix must be square.
     shape.dims[-1].assert_is_compatible_with(shape.dims[-2])
+  shape = shape.as_list()
+  n = shape[-2]
+  s = oversampling_ratio * shape[-1]
+  wigner_shape = shape[:-2] + [n, s]
 
   with ops.name_scope("random_positive_definite_matrix"):
-    tril = random_tril_matrix(
-        shape, dtype, force_well_conditioned=force_well_conditioned)
-    return math_ops.matmul(tril, tril, adjoint_b=True)
+    wigner = random_normal(
+        wigner_shape,
+        dtype=dtype,
+        stddev=math_ops.cast(1 / np.sqrt(s), dtype.real_dtype))
+    wishart = math_ops.matmul(wigner, wigner, adjoint_b=True)
+    if force_well_conditioned:
+      wishart += linalg_ops.eye(n, dtype=dtype)
+      wishart /= math_ops.cast(2, dtype)
+    return wishart
 
 
 def random_tril_matrix(shape,