Merge pull request matplotlib#25813 from ksunden/npy_prec_robustness

[TST] Adjust tests to be more tolerant to floating point math operations being imprecise

Author: oscargus

lib/matplotlib/tests/test_axes.py

@@ -4415,7 +4415,8 @@ def test_mollweide_forward_inverse_closure():
 
     # set up 1-degree grid in longitude, latitude
     lon = np.linspace(-np.pi, np.pi, 360)
-    lat = np.linspace(-np.pi / 2.0, np.pi / 2.0, 180)
+    # The poles are degenerate and thus sensitive to floating point precision errors
+    lat = np.linspace(-np.pi / 2.0, np.pi / 2.0, 180)[1:-1]
     lon, lat = np.meshgrid(lon, lat)
     ll = np.vstack((lon.flatten(), lat.flatten())).T
 

lib/matplotlib/tests/test_polar.py

@@ -425,13 +425,13 @@ def test_axvline_axvspan_do_not_modify_rlims():
 def test_cursor_precision():
     ax = plt.subplot(projection="polar")
     # Higher radii correspond to higher theta-precisions.
-    assert ax.format_coord(0, 0) == "θ=0π (0°), r=0.000"
+    assert ax.format_coord(0, 0.005) == "θ=0.0π (0°), r=0.005"
     assert ax.format_coord(0, .1) == "θ=0.00π (0°), r=0.100"
     assert ax.format_coord(0, 1) == "θ=0.000π (0.0°), r=1.000"
-    assert ax.format_coord(1, 0) == "θ=0.3π (57°), r=0.000"
+    assert ax.format_coord(1, 0.005) == "θ=0.3π (57°), r=0.005"
     assert ax.format_coord(1, .1) == "θ=0.32π (57°), r=0.100"
     assert ax.format_coord(1, 1) == "θ=0.318π (57.3°), r=1.000"
-    assert ax.format_coord(2, 0) == "θ=0.6π (115°), r=0.000"
+    assert ax.format_coord(2, 0.005) == "θ=0.6π (115°), r=0.005"
     assert ax.format_coord(2, .1) == "θ=0.64π (115°), r=0.100"
     assert ax.format_coord(2, 1) == "θ=0.637π (114.6°), r=1.000"
 

lib/mpl_toolkits/mplot3d/tests/test_axes3d.py

@@ -200,9 +200,9 @@ def test_bar3d_lightsource():
 
     # Testing that the custom 90° lightsource produces different shading on
     # the top facecolors compared to the default, and that those colors are
-    # precisely the colors from the colormap, due to the illumination parallel
-    # to the z-axis.
-    np.testing.assert_array_equal(color, collection._facecolor3d[1::6])
+    # precisely (within floating point rounding errors of 4 ULP) the colors
+    # from the colormap, due to the illumination parallel to the z-axis.
+    np.testing.assert_array_max_ulp(color, collection._facecolor3d[1::6], 4)
 
 
 @mpl3d_image_comparison(['contour3d.png'], style='mpl20')