Merge pull request #257 from quantified-uncertainty/issue250

Improvements to `Means_test.res` (and `expectErrorToBeBounded`)

packages/squiggle-lang/__tests__/Distributions/Invariants/Means_test.res

@@ -1,15 +1,64 @@
 /*
 This is the most basic file in our invariants family of tests. 
 
-See document in https://github.com/quantified-uncertainty/squiggle/pull/238 for details
+Validate that the addition of means equals the mean of the addition, similar for subtraction and multiplication. 
 
-Note: digits parameter should be higher than -4. 
+Details in https://deploy-preview-251--squiggle-documentation.netlify.app/docs/internal/invariants/
+
+Note: epsilon of 1e3 means the invariants are, in general, not being satisfied. 
 */
 
 open Jest
 open Expect
 open TestHelpers
 
+module Internals = {
+  let epsilon = 5e1
+
+  let mean = GenericDist_Types.Constructors.UsingDists.mean
+
+  let expectImpossiblePath: string => assertion = algebraicOp =>
+    `${algebraicOp} has`->expect->toEqual("failed")
+
+  let distributions = list{
+    normalMake(4e0, 1e0),
+    betaMake(2e0, 4e0),
+    exponentialMake(1.234e0),
+    uniformMake(7e0, 1e1),
+    // cauchyMake(1e0, 1e0),
+    lognormalMake(2e0, 1e0),
+    triangularMake(1e0, 1e1, 5e1),
+    Ok(floatMake(1e1)),
+  }
+  let pairsOfDifferentDistributions = E.L.combinations2(distributions)
+
+  let runMean: DistributionTypes.genericDist => float = dist => {
+    dist->mean->run->toFloat->E.O2.toExn("Shouldn't see this because we trust testcase input")
+  }
+
+  let testOperationMean = (
+    distOp: (
+      DistributionTypes.genericDist,
+      DistributionTypes.genericDist,
+    ) => result<DistributionTypes.genericDist, DistributionTypes.error>,
+    description: string,
+    floatOp: (float, float) => float,
+    dist1': SymbolicDistTypes.symbolicDist,
+    dist2': SymbolicDistTypes.symbolicDist,
+    ~epsilon: float,
+  ) => {
+    let dist1 = dist1'->DistributionTypes.Symbolic
+    let dist2 = dist2'->DistributionTypes.Symbolic
+    let received =
+      distOp(dist1, dist2)->E.R2.fmap(mean)->E.R2.fmap(run)->E.R2.fmap(toFloat)->E.R.toExn
+    let expected = floatOp(runMean(dist1), runMean(dist2))
+    switch received {
+    | None => expectImpossiblePath(description)
+    | Some(x) => expectErrorToBeBounded(x, expected, ~epsilon)
+    }
+  }
+}
+
 let {
   algebraicAdd,
   algebraicMultiply,
@@ -26,115 +75,82 @@ let algebraicSubtract = algebraicSubtract(~env)
 let algebraicLogarithm = algebraicLogarithm(~env)
 let algebraicPower = algebraicPower(~env)
 
-describe("Mean", () => {
-  let digits = -4
+let {testOperationMean, distributions, pairsOfDifferentDistributions, epsilon} = module(Internals)
 
-  let mean = GenericDist_Types.Constructors.UsingDists.mean
+describe("Means are invariant", () => {
+  describe("for addition", () => {
+    let testAdditionMean = testOperationMean(algebraicAdd, "algebraicAdd", \"+.", ~epsilon)
 
-  let runMean: result<DistributionTypes.genericDist, DistributionTypes.error> => float = distR => {
-    distR
-    ->E.R2.fmap(mean)
-    ->E.R2.fmap(run)
-    ->E.R2.fmap(toFloat)
-    ->E.R.toExn
-    ->E.O2.toExn("Shouldn't see this because we trust testcase input")
-  }
+    testAll("with two of the same distribution", distributions, dist => {
+      E.R.liftM2(testAdditionMean, dist, dist)->E.R.toExn
+    })
 
-  let impossiblePath: string => assertion = algebraicOp =>
-    `${algebraicOp} has`->expect->toEqual("failed")
-
-  let distributions = list{
-    normalMake(0.0, 1e0),
-    betaMake(2e0, 4e0),
-    exponentialMake(1.234e0),
-    uniformMake(7e0, 1e1),
-    // cauchyMake(1e0, 1e0),
-    lognormalMake(1e0, 1e0),
-    triangularMake(1e0, 1e1, 5e1),
-    Ok(floatMake(1e1)),
-  }
-  let combinations = E.L.combinations2(distributions)
-  let zipDistsDists = E.L.zip(distributions, distributions)
-
-  let testOperationMean = (
-      distOp: (DistributionTypes.genericDist, DistributionTypes.genericDist) => result<DistributionTypes.genericDist, DistributionTypes.error>, 
-      description: string, 
-      floatOp: (float, float) => float, 
-      dist1': result<SymbolicDistTypes.symbolicDist, string>, 
-      dist2': result<SymbolicDistTypes.symbolicDist, string>
-    ) => {
-    let dist1 = dist1'->E.R2.fmap(x=>DistributionTypes.Symbolic(x))->E.R2.fmap2(s=>DistributionTypes.Other(s))
-    let dist2 = dist2'->E.R2.fmap(x=>DistributionTypes.Symbolic(x))->E.R2.fmap2(s=>DistributionTypes.Other(s))
-    let received =
-      E.R.liftJoin2(distOp, dist1, dist2)
-      ->E.R2.fmap(mean)
-      ->E.R2.fmap(run)
-      ->E.R2.fmap(toFloat)
-    let expected = floatOp(runMean(dist1), runMean(dist2))
-    switch received {
-    | Error(err) => impossiblePath(description)
-    | Ok(x) =>
-      switch x {
-      | None => impossiblePath(description)
-      | Some(x) => x->expect->toBeSoCloseTo(expected, ~digits)
-      }
-    }
-  }
-
-  describe("addition", () => {
-    let testAdditionMean = testOperationMean(algebraicAdd, "algebraicAdd", \"+.")  
-
-    testAll("homogeneous addition", zipDistsDists, dists => {
+    testAll("with two different distributions", pairsOfDifferentDistributions, dists => {
       let (dist1, dist2) = dists
-      testAdditionMean(dist1, dist2)
+      E.R.liftM2(testAdditionMean, dist1, dist2)->E.R.toExn
     })
 
-    testAll("heterogeneous addition (1)", combinations, dists => {
+    testAll(
+      "with two different distributions in swapped order",
+      pairsOfDifferentDistributions,
+      dists => {
         let (dist1, dist2) = dists
-      testAdditionMean(dist1, dist2)
+        E.R.liftM2(testAdditionMean, dist2, dist1)->E.R.toExn
+      },
+    )
   })
 
-    testAll("heterogeneous addition (commuted of 1 (or; 2))", combinations, dists => {
+  describe("for subtraction", () => {
+    let testSubtractionMean = testOperationMean(
+      algebraicSubtract,
+      "algebraicSubtract",
+      \"-.",
+      ~epsilon,
+    )
+
+    testAll("with two of the same distribution", distributions, dist => {
+      E.R.liftM2(testSubtractionMean, dist, dist)->E.R.toExn
+    })
+
+    testAll("with two different distributions", pairsOfDifferentDistributions, dists => {
       let (dist1, dist2) = dists
-      testAdditionMean(dist2, dist1)
-    })
+      E.R.liftM2(testSubtractionMean, dist1, dist2)->E.R.toExn
     })
 
-  describe("subtraction", () => {
-    let testSubtractionMean = testOperationMean(algebraicSubtract, "algebraicSubtract", \"-.")
-
-    testAll("homogeneous subtraction", zipDistsDists, dists => {
+    testAll(
+      "with two different distributions in swapped order",
+      pairsOfDifferentDistributions,
+      dists => {
         let (dist1, dist2) = dists
-      testSubtractionMean(dist1, dist2)
+        E.R.liftM2(testSubtractionMean, dist2, dist1)->E.R.toExn
+      },
+    )
   })
 
-    testAll("heterogeneous subtraction (1)", combinations, dists => {
+  describe("for multiplication", () => {
+    let testMultiplicationMean = testOperationMean(
+      algebraicMultiply,
+      "algebraicMultiply",
+      \"*.",
+      ~epsilon,
+    )
+
+    testAll("with two of the same distribution", distributions, dist => {
+      E.R.liftM2(testMultiplicationMean, dist, dist)->E.R.toExn
+    })
+
+    testAll("with two different distributions", pairsOfDifferentDistributions, dists => {
       let (dist1, dist2) = dists
-      testSubtractionMean(dist1, dist2)
+      E.R.liftM2(testMultiplicationMean, dist1, dist2)->E.R.toExn
     })
 
-    testAll("heterogeneous subtraction (commuted of 1 (or; 2))", combinations, dists => {
+    testAll(
+      "with two different distributions in swapped order",
+      pairsOfDifferentDistributions,
+      dists => {
         let (dist1, dist2) = dists
-      testSubtractionMean(dist2, dist1)
-    })
-  })
-
-  describe("multiplication", () => {
-    let testMultiplicationMean = testOperationMean(algebraicMultiply, "algebraicMultiply", \"*.")
-
-    testAll("homogeneous subtraction", zipDistsDists, dists => {
-      let (dist1, dist2) = dists
-      testMultiplicationMean(dist1, dist2)
-    })
-
-    testAll("heterogeneoous subtraction (1)", combinations, dists => {
-      let (dist1, dist2) = dists
-      testMultiplicationMean(dist1, dist2)
-    })
-
-    testAll("heterogeneoous subtraction (commuted of 1 (or; 2))", combinations, dists => {
-      let (dist1, dist2) = dists
-      testMultiplicationMean(dist2, dist1)
-    })
+        E.R.liftM2(testMultiplicationMean, dist2, dist1)->E.R.toExn
+      },
+    )
   })
 })

packages/squiggle-lang/__tests__/TestHelpers.res

@@ -1,6 +1,25 @@
 open Jest
 open Expect
 
+/*
+This encodes the expression for percent error
+The test says "the percent error of received against expected is bounded by epsilon"
+
+However, the semantics are degraded by catching some numerical instability: 
+when expected is too small, the return of this function might blow up to infinity. 
+So we capture that by taking the max of abs(expected) against a 1. 
+
+A sanity check of this function would be welcome, in general it is a better way of approaching 
+squiggle-lang tests than toBeSoCloseTo. 
+*/
+let expectErrorToBeBounded = (received, expected, ~epsilon) => {
+  let distance = Js.Math.abs_float(received -. expected)
+  let expectedAbs = Js.Math.abs_float(expected)
+  let normalizingDenom = Js.Math.max_float(expectedAbs, 1e0)
+  let error = distance /. normalizingDenom
+  error->expect->toBeLessThan(epsilon)
+}
+
 let makeTest = (~only=false, str, item1, item2) =>
   only
     ? Only.test(str, () => expect(item1)->toEqual(item2))

packages/squiggle-lang/src/rescript/ReducerInterface/ReducerInterface_GenericDistribution.res

@@ -117,7 +117,8 @@ module Helpers = {
         | Error(err) => GenDistError(ArgumentError(err))
         }
       }
-    | Some(EvDistribution(b)) => switch parseDistributionArray(args) {
+    | Some(EvDistribution(b)) =>
+      switch parseDistributionArray(args) {
       | Ok(distributions) => mixtureWithDefaultWeights(distributions)
       | Error(err) => GenDistError(ArgumentError(err))
       }