Sampling the soft weights of hard-core boosting #
Votes are ordinary Boolean words. The number of correct votes, the threshold, and the dyadic
rate determine a natural-number numerator. Saturating subtraction implements both clipped
regions of the real weight curve; neither real arithmetic nor rejection sampling is needed.
isPPTOn_sampleWeight certifies the whole program through the shared programming combinators.
sampleWeight_density_deviation estimates the source-averaged density with a Hoeffding bound,
using a natural success counter and independent source examples.
Each call draws a fresh coin. An implementation of a random set must additionally cache answers to repeated membership queries; this program alone is not such a stateful oracle.
Main definitions #
Cslib.Crypto.Pseudoentropy.Boosting.weightNumerator: clipped dyadic weights using naturals.Cslib.Crypto.Pseudoentropy.Boosting.sampleWeight: a fresh exact weight coin.
Main results #
Cslib.Crypto.Pseudoentropy.Boosting.eval_sampleWeight_true: exact agreement with the weight.Cslib.Crypto.Pseudoentropy.Boosting.sampleWeight_density_deviation: density concentration.Cslib.Crypto.Pseudoentropy.Boosting.isPPTOn_sampleWeight: the uniform strict PPT certificate.
References #
Holenstein, Section 2.2, Figure 1 and Claims 2.6–2.8. Adapted from
Crypto/Computational/Pseudoentropy/Boosting/Sampling.lean on the cslib crypto branch at
5ff69ca93402e7c947208dc56396d6d86203ccd8, with typed tidy-machine certificates.
The numerator of a clipped linear weight, computed entirely with natural numbers.
The excess of correct over incorrect votes is 2 * votes.count truth - votes.length in
integer arithmetic; the inner natural subtraction keeps only its excess over the threshold.
Instances For
Every generated numerator is within its denominator, including malformed parameter choices.
Saturating natural arithmetic exactly implements the clipped real-valued weight curve.
The weight's natural numerator is polynomial time in its complete encoded input.
The program's acceptance probability is exactly the soft weight used in the potential proof.
Sampling a source example and then its weight coin accepts with exactly the soft density.
Repeated source examples and fresh weight coins estimate the soft density. The executable counter uses natural numbers; the empirical fraction occurs only in its specification.
Runtime-dependent parameters, votes, and labels compose to a uniform strict PPT sampler.