unexpected results for float32 array (or is expected?).

[prajwel]

I noticed some unexpected results while running the find_transform function. Please see the script below that reproduces it.

import numpy as np
import astroalign as aa

test_data = np.array([[1819.4231 ,  946.6505 ],
                     [2267.2012 , 3106.64   ],
                     [3726.2542 , 2355.1924 ],
                     [2535.0833 , 1680.7942 ],
                     [2837.4094 , 2319.1921 ],
                     [3046.8765 , 3802.3745 ],
                     [3125.3823 , 2531.9395 ],
                     [2563.118  , 2538.3152 ],
                     [1323.7314 , 3177.112  ]])

np.set_printoptions(suppress=True,
   formatter={'float':'{:.5f}'.format})

test_data = test_data.astype(np.float64)
matrix64, _ = aa.find_transform(test_data, test_data)

test_data = test_data.astype(np.float32)
matrix32, _ = aa.find_transform(test_data, test_data)

print(matrix64.params)
print()
print(matrix32.params)

The output

[[1.00000 -0.00000 0.00000]
 [-0.00000 1.00000 0.00000]
 [0.00000 0.00000 1.00000]]

[[1.00000 0.00000 0.00074]
 [0.00000 1.00000 0.00070]
 [0.00000 0.00000 1.00000]]

In some tests, matrix32.params gets even worse. Is this a known behaviour if we use single precision values?