Flat lossless condensation by the supplied polynomial map #
For every qualifying nonempty flat source, the seeded output is close to a distribution that is uniform on exactly the original number of source points in every seed fiber. The comparison preserves the seed and bounds all tests on the joint seed-output space. Its error is the GUV expansion loss divided by the field cardinality.
For a larger flat source of size at least K, the same error bound holds
against an explicit mixture of seedwise K-flat witnesses, provided
K ≤ (2^r)^m. The original support may exceed the output alphabet size.
This is the flat-source distributional consequence of Guruswami--Umans--Vadhan
(2009), Theorem 3.3. The exponent 2^r is the characteristic-power choice used
by Cheraghchi (2010) for source linearity. The field and modulus remain supplied;
this theorem provides neither their uniform construction nor an encoded
evaluator. General weighted sources require a separate decomposition argument.
Polynomial expansion gives a seed-preserving flat witness at the original support size, with error equal to the normalized degree loss. The witness may depend on the source; the condenser map itself is fixed.
Every flat source above the threshold is close to an explicit mixture of
seedwise K-flat outputs. Only the threshold must fit in the output space.