squiggle/packages/squiggle-lang/__tests__/Distributions/Mixture_test.res

open Jest
open Expect
open TestHelpers

// TODO: use Normal.make (etc.), but preferably after the new validation dispatch is in.
let mkNormal = (mean, stdev) => GenericDist_Types.Symbolic(#Normal({mean: mean, stdev: stdev}))
let mkBeta = (alpha, beta) => GenericDist_Types.Symbolic(#Beta({alpha: alpha, beta: beta}))
let mkExponential = rate => GenericDist_Types.Symbolic(#Exponential({rate: rate}))
let mkUniform = (low, high) => GenericDist_Types.Symbolic(#Uniform({low: low, high: high}))
let mkCauchy = (local, scale) => GenericDist_Types.Symbolic(#Cauchy({local: local, scale: scale}))
let mkLognormal = (mu, sigma) => GenericDist_Types.Symbolic(#Lognormal({mu: mu, sigma: sigma}))

describe("mixture", () => {
  testAll(
    "fair mean of two normal distributions",
    list{(0.0, 1e2), (-1e1, -1e-4), (-1e1, 1e2), (-1e1, 1e1)},
    tup => {
      // should be property
      let (mean1, mean2) = tup
      let meanValue = {
        run(Mixture([(mkNormal(mean1, 9e-1), 0.5), (mkNormal(mean2, 9e-1), 0.5)]))->outputMap(
          FromDist(ToFloat(#Mean)),
        )
      }
      meanValue->unpackFloat->expect->toBeSoCloseTo((mean1 +. mean2) /. 2.0, ~digits=-1)
    },
  )
  testAll(
    "weighted mean of a beta and an exponential",
    // This would not survive property testing, it was easy for me to find cases that NaN'd out.
    list{((128.0, 1.0), 2.0), ((2e-1, 64.0), 16.0), ((1e0, 1e0), 64.0)},
    tup => {
      let ((alpha, beta), rate) = tup
      let betaWeight = 0.25
      let exponentialWeight = 0.75
      let meanValue = {
        run(
          Mixture([(mkBeta(alpha, beta), betaWeight), (mkExponential(rate), exponentialWeight)]),
        )->outputMap(FromDist(ToFloat(#Mean)))
      }
      let betaMean = 1.0 /. (1.0 +. beta /. alpha)
      let exponentialMean = 1.0 /. rate
      meanValue
      ->unpackFloat
      ->expect
      ->toBeSoCloseTo(betaWeight *. betaMean +. exponentialWeight *. exponentialMean, ~digits=-1)
    },
  )
  testAll(
    "weighted mean of lognormal and uniform",
    // Would not survive property tests: very easy to find cases that NaN out.
    list{((-1e2, 1e1), (2e0, 1e0)), ((-1e-16, 1e-16), (1e-8, 1e0)), ((0.0, 1e0), (1e0, 1e-2))},
    tup => {
      let ((low, high), (mu, sigma)) = tup
      let uniformWeight = 0.6
      let lognormalWeight = 0.4
      let meanValue = {
        run(
          Mixture([
            (mkUniform(low, high), uniformWeight),
            (mkLognormal(mu, sigma), lognormalWeight),
          ]),
        )->outputMap(FromDist(ToFloat(#Mean)))
      }
      let uniformMean = (low +. high) /. 2.0
      let lognormalMean = mu +. sigma ** 2.0 /. 2.0
      meanValue
      ->unpackFloat
      ->expect
      ->toBeSoCloseTo(uniformWeight *. uniformMean +. lognormalWeight *. lognormalMean, ~digits=-1)
    },
  )
})