Heuristic doubling period trick by preprocessing pitch correlation values?

[YangWangGit]

I think the trick in this surrounding "for" code block is to suppress insignificant, possible doubling period. Is there any more in-depth motivation, explanation, or paper mentioning this? Thanks. https://github.com/xiph/LPCNet/blob/2159b4f85d9839d9627f1843a1cb5ea298c63159/src/lpcnet_enc.c#L597