This is by design, because the inplace operators are intended to operate on the same chunk of memory as the input - but the broadcast result would not find in that memory.

Having said that, it might be reasonable to allow a1d += a2d to decay into a1d = a1d + a2d, especially given a discussion in #9019

In your case, the fix is simple - just allocate the arrays to the correct size first - at a guess, that would be:

mean = np.zeros(data.shape[1:])
M2 = np.zeros(data.shape[1:])

@mratsim Can you give an example of input that triggers this error? When I run your function on NumPy arrays it works properly:

In [19]: online_variance([np.arange(3.0), -np.arange(3.0)])
Out[19]: array([ 0.,  2.,  8.])

In [20]: online_variance([np.array(1.0), np.array(2.0)])
Out[20]: 0.5

I've found out the issue, I was actually initializing my variables with an empty array due to a mistake.

Screenshot: "os.path.join" missing at initialization:

That was catched by the += but was slient otherwise:

So that can be closed

While investigating, I uncovered a precedence issue with += however. I will raise an issue about that.