Source lean/Determinize/Spec/Primitives.lean
import Determinize.Spec.Typesimport Mathlib.Probability.Distributions.Betaimport Mathlib.Probability.Distributions.Exponentialimport Mathlib.Probability.Distributions.Gaussian.Realimport Mathlib.Probability.Distributions.Poisson.Basicimport Mathlib.Tactic.DeriveCountable/-!# Primitive distributionsEach primitive has one fiber: at evaluated parameters it is the primitive's law at astochastic site and the Dirac mass at the primitive's mean at a mean site, the formdeterminization leaves behind. Both are the zero measure outside the parameter domain,which makes an off-domain call stuck.-/namespace Determinize.Spec.Paperopen MeasureTheory ProbabilityTheorynoncomputable section/-- Primitive names recorded at G draws. A discrete label records the number ofsupplied probabilities; there is one additional outcome with the remaining probability. -/inductive Op where  | uniform  | gaussian  | poisson  | exponential  | beta  | gamma  | bernoulli  | discrete (arity : Nat)deriving DecidableEq, Repr, Countable/-- Uniform probability measure on a closed interval, including a point interval. -/def uniformMeasure (lower upper : ℝ) : Measure ℝ :=  if _hDomain : lower ≤ upper then    if _hPoint : lower = upper then Measure.dirac lower    else (ENNReal.ofReal (upper - lower))⁻¹ • volume.restrict (Set.Icc lower upper)  else 0/-- `uniform(lower, upper)`: the uniform law on `[lower, upper]` or its mean. -/def uniformFiber (action : DistributionAction) (lower upper : ℝ) : Measure ℝ :=  if lower ≤ upper then    match action with    | .sample _ => uniformMeasure lower upper    | .mean => Measure.dirac ((lower + upper) / 2)  else 0/-- `gaussian(mean, variance)`: the normal law or its mean. -/def gaussianFiber (action : DistributionAction) (mean variance : ℝ) : Measure ℝ :=  if h : 0 ≤ variance then    match action with    | .sample _ => gaussianReal mean ⟨variance, h⟩    | .mean => Measure.dirac mean  else 0/-- `poisson(rate)`: the Poisson law on the naturals, read as reals, or its mean. -/def poissonFiber (action : DistributionAction) (rate : ℝ) : Measure ℝ :=  if h : 0 ≤ rate then    match action with    | .sample _ => (poissonMeasure ⟨rate, h⟩).map (fun value : Nat ↦ (value : ℝ))    | .mean => Measure.dirac rate  else 0/-- `exponential(rate)`: the exponential law or its mean `1 / rate`. -/def exponentialFiber (action : DistributionAction) (rate : ℝ) : Measure ℝ :=  if 0 < rate then    match action with    | .sample _ => expMeasure rate    | .mean => Measure.dirac (1 / rate)  else 0/-- `beta(alpha, beta)`: the beta law or its mean `alpha / (alpha + beta)`. -/def betaFiber (action : DistributionAction) (alpha beta : ℝ) : Measure ℝ :=  if 0 < alpha ∧ 0 < beta then    match action with    | .sample _ => betaMeasure alpha beta    | .mean => Measure.dirac (alpha / (alpha + beta))  else 0/-- `gamma(shape, rate)`: the gamma law or its mean `shape / rate`. -/def gammaFiber (action : DistributionAction) (shape rate : ℝ) : Measure ℝ :=  if 0 < shape ∧ 0 < rate then    match action with    | .sample _ => gammaMeasure shape rate    | .mean => Measure.dirac (shape / rate)  else 0/-- Bernoulli on numeric outcomes zero and one. Invalid probabilities have zero mass. -/noncomputable def bernoulliFiber (action : DistributionAction) (probability : ℝ) : Measure ℝ :=  if 0 ≤ probability ∧ probability ≤ 1 then    match action with    | .sample _ => ENNReal.ofReal (1 - probability) • Measure.dirac 0 +        ENNReal.ofReal probability • Measure.dirac 1    | .mean => Measure.dirac probability  else 0/-- `discrete(p₀, …, pₙ₋₁, *)` has outcomes `0, …, n`. The last probability is`1 - ∑ i, pᵢ`; supplied probabilities must be nonnegative and sum to at most one.The same domain applies at mean sites. -/noncomputable def discreteFiber (action : DistributionAction) (probabilities : List ℝ) :    Measure ℝ :=  if (∀ i : Fin probabilities.length, 0 ≤ probabilities[i]) ∧      ∑ i : Fin probabilities.length, probabilities[i] ≤ 1 then    match action with    | .sample _ =>      (∑ i : Fin probabilities.length,        ENNReal.ofReal probabilities[i] • Measure.dirac ((i : ℕ) : ℝ)) +      ENNReal.ofReal (1 - ∑ i : Fin probabilities.length, probabilities[i]) •        Measure.dirac (probabilities.length : ℝ)    | .mean => Measure.dirac ((probabilities.length : ℝ) +        ∑ i : Fin probabilities.length, (((i : ℕ) : ℝ) - probabilities.length) * probabilities[i])  else 0endend Determinize.Spec.Paper