How to implement an analyzing polyphase cosine modulated filter bank over a synthesis?

Question

Sound processing

Hello, dsp is a hobby for me, and most likely I need to understand the theory better, nevertheless I will ask.
There is a task to write an analyzer for this synthesizing filter bank ( https ://github.com/FFmpeg/FFmpeg/blob/master/libav...) . In ATRAC3plus we have a decomposition into 16 bands. //github.com/Voskrese/libmpc-msvc/blob/maste... in the case of an arbitrary number of bands and an arbitrary length of the prototype - received emissions on a rectangular signal

and some aliasing from 0hz to a frequency multiple of the doubled section frequency (2756.25 = 2 * 22050 / 16)

For frequencies 1000 hz and 1378.125 hz the picture is as follows:

Although the decomposition itself is more or less true.
It could be attributed to a non-perfect reconstruction. However, ffmpeg decodes the signal without such distortions after encoding by the original atrac3plus encoder.
I would be grateful for advice on how to localize the problem, or links to what to read on the implementation of pqf in the code.