The actual affine construction tolerates deterministic seed leakage on average #
The source supplies the selected affine construction's entropy reserve plus
all (t + 1) * d leakage bits. For arbitrary shifts and tampered base seeds
chosen as functions of the honest base seed, the average conditional output
discrepancy is at most the construction's selected dyadic error. Conditioning
uses exact source masses and normalized completions for impossible leakage
transcripts, with no conditional entropy or security hypothesis from the caller.
This proves the average-error step of Chattopadhyay--Liao,
Extractors for Sum of Two Sources, Lemma 5.3, printed pp.19--20:
https://arxiv.org/abs/2110.12652. The tampered seed retains f y i; the
argument does not replace it by the honest base seed. The subsequent bad-seed
counting argument is a separate finite averaging theorem.
All source-dependent seed leakage is paid before applying the actual selected affine map.