Test correctness of `eigh` solver in `FastICA`

fastica_test.py

from scipy import linalg
import numpy as np
from sklearn.utils._testing import assert_array_almost_equal
import warnings

def assert_sign_redundant(x,y):
    X, Y = x.copy(), y.copy()
    for A in (X,Y):
        for c in range(A.shape[1]):
            if A[0,c] < 0:
                A[:,c] *= -1
    assert_array_almost_equal(X, Y)

for i in range(100):
    X = np.random.rand(2000, 300)
    K = min(X.shape)
    d, u = linalg.eigh(X.T.dot(X))
    eps = np.finfo(np.double).eps
    degenerate_idx = d < eps
    if np.any(degenerate_idx):
        warnings.warn(
            "There are some small singular values, using "
            "svd_solver = 'svd' might lead to more "
            "accurate results."
        )
    d[degenerate_idx] = eps  # For numerical issues
    d = np.sqrt(d)
    # Reorder based on eigenvalues
    sort_indices = np.argsort(d)[::-1]
    d, u = d[sort_indices], u[sort_indices]
    # Reshape to match SVD
    u = u[:,:K]
    # Reorder to match SVD
    L = np.eye(u.shape[0])[::-1]
    R = np.eye(u.shape[1])[::-1]
    u = L@u@R
    u2, d2 = linalg.svd(X.T, full_matrices=False, check_finite=False)[:2]
    assert_sign_redundant(u,u2)
    assert_array_almost_equal(d,d2)