Corollaries of trace soundness #
Jensen's inequality between the source and target output laws, preservation of expectations
in the extended reals, preservation of the output mass and hence of the expectation conditioned
on acceptance, and the law of total variance along traces with its consequence that
determinization does not increase the variance. All follow from a TraceFactorization
alone: trace by trace, the target output is the mean of the source output fiber.
The two output laws behind MeanOnTraces: the source output is the mixture of the fibers
over the trace law and the target output is the pushforward of the trace law along output.
Under a trace factorization the two output laws have the same mass, the mass of the trace law: every fiber is a probability measure.
Output mass is preserved along the traces.
The expectations conditioned on acceptance agree: the unnormalized mean and the output mass are both preserved, so their quotient is.
Jensen's inequality along the traces.
∫⁻ (a - b)⁺ ≤ ∫⁻ a⁺ when b is nonnegative.
∫⁻ a⁺ ≤ ∫⁻ (a - b)⁺ + ∫⁻ b⁺.
For nonnegative a, b, the extended-real difference ∫⁻ a - ∫⁻ b equals
∫⁻ (a - b)⁺ - ∫⁻ (a - b)⁻ as soon as one of the two integrals is finite.
Extended-real expectations are preserved along the traces.
The second moment about m of a real law with a finite second moment, through its first
two moments and its mass.
Mathlib's variance of a real law with a finite second moment, through its first two
moments and its mass; the law need not be normalized.
The law of total variance along a trace factorization, through the second moments: the target output law has a finite second moment, the fiber variances are integrable over the trace law, and the source second moment is the target second moment plus the mean fiber variance.
The law of total variance along a trace factorization: the fiber variances are integrable
over the trace law and the source output variance is the target output variance plus the mean
fiber variance. Both output laws have the same mass and the same mean, so Mathlib's variance
needs no normalization.
Determinization does not increase the second moment or the variance of the output law.
Extended-real expectation preservation, from operational trace soundness.
Jensen's inequality between the two output laws, from operational trace soundness.
Output mass preservation, from operational trace soundness.
Variance non-increase, from operational trace soundness.
Preservation of the expectation conditioned on acceptance, from operational trace soundness.
The law of total variance along traces, stated with regular conditional distributions.