Isn't the magic kernel simply a 2× bilinear downscale filter? For example, if you use 2× bilinear downscaling in Tensorflow (as opposed to bilinear downsampling which would end up being just a box filter) using tf.image.resize it uses the magic kernel [1,3,3,1].

There is an article [0] that points this out, although later it goes into recommending odd filters, which insidiously break many applications where you would want to use downscaling by shifting the image by a fraction of a pixel.

[0] https://cbloomrants.blogspot.com/2011/03/03-24-11-image-filt...