Would assert_array_almost_equal(..., decimal=5) be the most similar to assert_allclose(..., atol=1.5e-5)?

Would assert_array_almost_equal(..., decimal=5) be the most similar to assert_allclose(..., atol=1.5e-5)?

It would be equivalent, but that is not recommended for the same reason assert_array_almost_equal is not recommended. Comparing with an absolute tolerance only works if one knows the scale of the data (and is able to compensate for it). Often, a casual reader would not know the scale of the data (e.g. here before I printed those values in the failed test). The risk is then of false positives e.g. if the data has a scale of 1e-5 or lower and we use a tolerance of 1.5e-5, the test will always pass without the user realizing that the test doesn't enforce anything.

For that reason, it's better to use rtol instead, except when compared with 0 (cf PEP 485). In our case it means changing atol to rtol, there is a risk of test failure if the threshold is too low, but it's also an opportunity to check that the tests are behaving as expected.

For instance, I do hope that the data we are comparing with in the following tests is larger than 0.1 (or at least of the order of 1.0),

sklearn/decomposition/tests/test_nmf.py
198:        assert_array_almost_equal(A_fit_tr, A_tr, decimal=1)

sklearn/linear_model/tests/test_least_angle.py
356:    assert_array_almost_equal(lars_coef, lasso_coef2, decimal=1)

sklearn/linear_model/tests/test_logistic.py
1636:    assert_array_almost_equal(sgd.coef_, log.coef_, decimal=1)
1652:        assert_array_almost_equal(coefs[0], coefs[1], decimal=1)
1654:        assert_array_almost_equal(coefs[0], coefs[2], decimal=1)
1656:        assert_array_almost_equal(coefs[1], coefs[2], decimal=1)

sklearn/mixture/tests/test_gaussian_mixture.py
1020:        assert_array_almost_equal(gmm.means_, means_s, decimal=1)

sklearn/svm/tests/test_sparse.py
234:    assert_array_almost_equal(clf.coef_, sp_clf.coef_, decimal=1)
235:    assert_array_almost_equal(clf.intercept_, sp_clf.intercept_, decimal=1)

probably, but hard to be sure just by looking at it.

Also related #13977