While playing around with matrices as part of Liner Algebra learning I encountered a weird issue where numpy and pytorch produces incorrect determinant of a matrix, however tensorflow gives the expected result.

In [110]: a = np.array([[15, 3], [10, 2]], dtype=np.float32)

In [111]: np.linalg.det(a)
Out[111]: 1.6653345e-15)

To the naked eye, the determinant of this matrix is 0, even though the type of the matrix is float32 there is no floating point rounding off needed here, so why does numpy rounds these numbers off and give non-zero determinant.

The same thing happens with pytorch as well

In [114]: at = torch.tensor([[15, 3], [10, 2]], dtype=torch.float32)

In [115]: torch.linalg.det(at)
Out[115]: tensor(-2.6822e-06)

However tensorflow works as expected

In [117]: af = tf.Variable([[15, 3], [10, 2]], dtype=tf.float32)

In [118]: tf.linalg.det(af)
Out[118]: <tf.Tensor: shape=(), dtype=float32, numpy=0.0>

There is more weirdness here

In [121]: a = np.array([[15, 3], [20, 4]], dtype=np.float32)
In [122]: np.linalg.det(a)
Out[122]: 0.0

In [119]: at = torch.tensor([[15, 3], [20, 4]], dtype=torch.float32)
In [120]: torch.linalg.det(at)
Out[120]: tensor(-5.3644e-06)

In [123]: af = tf.Variable([[15, 3], [20, 4]], dtype=tf.float32)
In [124]: tf.linalg.det(af)
Out[124]: <tf.Tensor: shape=(), dtype=float32, numpy=-0.0>

So out of the three libraries tensorflow seems to be the most stable, so can someone please explain this behavior or is this an anomaly? I understand this can cause issues with bunch of different matrix operations.