Compact trace soundness #
The detailed step-trace results are transported to compact traces of general-affinity draws, with
Proof.Traces.outputGivenTrace as the fiber: the Markov version normalizedOutputGivenTrace
factors the joint laws, and almost surely it agrees with outputGivenTrace on the source, while
on the target outputGivenTrace is the Dirac mass at the target output.
The compact replay as the fiber #
Proof.Traces.outputGivenTrace normalized to a Markov kernel, as an s-finite kernel.
Equations
- Determinize.Proof.Traces.normalizedKernel source = { kernel := Determinize.Proof.StepTraces.normalizedOutputGivenTrace source, sfinite := ⋯ }
Instances For
The normalized compact replay read through the retained draws of a detailed trace.
Equations
Instances For
On the detailed joint laws, the normalized compact replay is a sound fiber.
On the compact joint laws, the normalized compact replay is a sound fiber.
Almost every terminating target trace gives the source replay mass one.
Replaying the target along its own compact trace returns its output.
The factorization with its almost-sure identifications #
Trace soundness for an expectation-affinity source, as a factorization by the normalized
compact replay together with the identifications that give the public statement: almost
surely the source fiber is outputGivenTrace itself, and the target replay is the Dirac mass
at the target output.
Apply expectation-affinity soundness using silent subtyping for general-affinity programs.
Some trace factorization exists: the input of the corollaries.
A measure composed with a kernel is the bind that pairs each point with its draw.
Determinization returns the canonical replay mean on each terminating trace.
Replay factorization and its conditional-mean property, used to identify the conditional laws.