Less permissive one

tests/test_numerics_arrays.py

@@ -100,7 +100,6 @@ def check_inv(matrix, rtol=None):
     expected[combined_relative_mask] = 0.0
 
 
-
     # Check both directions
     numpy_zeros = (expected == 0.0)
     our_zeros = (result == 0.0)
@@ -109,30 +108,8 @@ def check_inv(matrix, rtol=None):
     # Where numpy has zeros but we don't; no cases require it but it makes sense to do
     result[mask_exact_zeros] = np.where(np.abs(result[mask_exact_zeros]) < thresh, 0.0, result[mask_exact_zeros])
 
-    # Where we have zeros but numpy doesn't  - this is the one we discovered
-    expected[mask_exact_zeros] = np.where(np.abs(expected[mask_exact_zeros]) < thresh, 0.0, expected[mask_exact_zeros])
-
-    # numpy_values_at_our_zeros = expected[our_zeros]
-    # expected[our_zeros] = np.where(
-    #     np.abs(numpy_values_at_our_zeros) < thresh,
-    #     0.0,
-    #     numpy_values_at_our_zeros
-    # )
-
-    # We also need to check against the values we get in the inverse; it is helpful 
-    # to zero out anything too close to "zero" relative to the values used in the matrix
-    # This is very necessary, and was needed when testing on different CPU architectures
-    inv_norm = np.max(np.sum(np.abs(result), axis=1))
-    if cond < 1e10:
-        zero_thresh = 100*thresh
-    elif cond < 1e14:
-        zero_thresh = 1000*thresh
-    else:
-        zero_thresh = 10000*thresh
-    trivial_relative_to_norm = np.where(np.abs(result)/inv_norm < zero_thresh)
-    result[trivial_relative_to_norm] = 0.0
-    trivial_relative_to_norm = np.where(np.abs(expected)/inv_norm < zero_thresh)
-    expected[trivial_relative_to_norm] = 0.0
+    # Where we have zeros but numpy doesn't - this is the one we discovered. Apply the check only to `numpy_zeros`
+    expected[numpy_zeros] = np.where(np.abs(expected[numpy_zeros]) < thresh, 0.0, expected[numpy_zeros])
 
     if rtol is None:
         rtol = get_rtol(matrix)