- Mode
-
Whether a draw is kept random or replaced by its mean; affinity in the Lean development
(Affinity).
- E
- Replaced by its mean in the determinized program.
- G
- Kept random; in one run, both programs draw the same value.
- Sample site
- A place in the program that draws from a distribution.
- Symbolic state
-
The proof's symbolic semantics (paper, section 5): σ binds a symbol to each E draw so
far, and the program computes with the symbols. The simulator computes it; it is not
part of the Lean pipeline (its definitions in Lean).
- G trace
-
The values of a run's G draws, in order (Trace).
- Returned, rejected, failed
-
A run returns a value, is rejected by
observe, or fails: division by zero,
a parameter outside a distribution's domain, or the step limit.
- Exact values
-
The mean, the variance and the probability of returning, as fractions, from a program's
finite model: its states and their exact transition probabilities (Model). A program has one if its random draws are coin flips and finite discrete draws, and
its states stay within Lean's limits. The simulator computes them with its port of
Lean's finite models; Lean's command-line tool computes them with a certificate that its
kernel checks (checkStatistics).
- Additive mode
-
How the exact values' models are explored, as by default in Lean's command-line tool: a
loop's pending outer sums become rewards on the model's edges (RewardModel.Model), so that a loop like
x + f(n - 1) reaches finitely many states.
- Markov chain
-
A program's finite model with each step of probability 1 contracted, shown below the
distributions with "Show the Markov chains": a node is where a run draws, returns, is
rejected or diverges. A source branches at its E and its G draws; its determinized
program only at the G draws, as the E draws are replaced by their means.
- Variance-reduction factor
- The source's variance divided by the determinized program's, over returned runs.