Normalized finite distributions #
Real weights make expectations finite sums. Unlike avg, a normalized distribution
cannot exist on an empty type. Uniform expectations retain their old empty-type convention.
An observable bounded by 1 that is strictly below 1 at a point of positive weight has
expectation strictly below 1 (it replaces Moran.expect_lt_one and Voter.expect_lt_one).
Monotonicity of probability: a larger event has a larger probability. A set
inclusion h : A ⊆ B can be passed directly for the events (· ∈ A) and (· ∈ B).
prob computed with any decidability instance for the event.
Events #
Events that agree pointwise have the same probability.
A sure event has probability one.
An impossible event has probability zero.
Inclusion–exclusion for two events, through their complements.
Markov's inequality: P(c ≤ X) ≤ 𝔼X / c for X ≥ 0 and c > 0. The uniform case is
avg_markov.
Markov's inequality on the square (the core of Chebyshev's inequality):
P(c ≤ |Z|) ≤ 𝔼[Z²] / c² for c > 0.
Uniform distribution, available precisely when the sample type is nonempty.
Equations
- Dynamics.Distribution.uniform α = { weight := fun (x : α) => (↑(Fintype.card α))⁻¹, nonneg := ⋯, sum_one := ⋯ }
Instances For
Point mass.
Equations
Instances For
Transport a distribution along a function, summing over its fibers.
Equations
Instances For
A point of positive weight is carried to its image with positive weight (it replaces
Moran.map_weight_pos).
Independent product of a finite family of distributions.
Equations
- Dynamics.Distribution.independent p = { weight := fun (x : ι → α) => ∏ i : ι, (p i).weight (x i), nonneg := ⋯, sum_one := ⋯ }