Promote arguments of norminvcdf and norminvccdf (#132)

* Promote arguments of `norminvcdf` and `norminvccdf`

* Simplify promotions

* Improve clarity of tests

Co-authored-by: Seth Axen <seth.axen@gmail.com>

* Simplify tests

* Add method with `rtol`

* Simplification

Co-authored-by: Seth Axen <seth.axen@gmail.com>

Project.toml

@@ -1,6 +1,6 @@
 name = "StatsFuns"
 uuid = "4c63d2b9-4356-54db-8cca-17b64c39e42c"
-version = "0.9.14"
+version = "0.9.15"
 
 [deps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"

src/distrs/norm.jl

@@ -39,15 +39,15 @@ function normlogpdf(μ::Real, σ::Real, x::Number)
         z = zval(μ, σ, x)
     end
     normlogpdf(z) - log(σ)
-end            
+end
 
 # cdf
 normcdf(z::Number) = erfc(-z * invsqrt2)/2
 function normcdf(μ::Real, σ::Real, x::Number)
     if iszero(σ) && x == μ
         z = zval(zero(μ), σ, one(x))
-    else        
-        z = zval(μ, σ, x)        
+    else
+        z = zval(μ, σ, x)
     end
     normcdf(z)
 end
@@ -56,8 +56,8 @@ normccdf(z::Number) = erfc(z * invsqrt2)/2
 function normccdf(μ::Real, σ::Real, x::Number)
     if iszero(σ) && x == μ
         z = zval(zero(μ), σ, one(x))
-    else        
-        z = zval(μ, σ, x)        
+    else
+        z = zval(μ, σ, x)
     end
     normccdf(z)
 end
@@ -69,8 +69,8 @@ normlogcdf(z::Number) = z < -1.0 ?
 function normlogcdf(μ::Real, σ::Real, x::Number)
     if iszero(σ) && x == μ
         z = zval(zero(μ), σ, one(x))
-    else        
-        z = zval(μ, σ, x)        
+    else
+        z = zval(μ, σ, x)
     end
     normlogcdf(z)
 end
@@ -82,17 +82,21 @@ normlogccdf(z::Number) = z > 1.0 ?
 function normlogccdf(μ::Real, σ::Real, x::Number)
     if iszero(σ) && x == μ
         z = zval(zero(μ), σ, one(x))
-    else        
-        z = zval(μ, σ, x)        
+    else
+        z = zval(μ, σ, x)
     end
     normlogccdf(z)
 end
 
 norminvcdf(p::Real) = -erfcinv(2*p) * sqrt2
-norminvcdf(μ::Real, σ::Real, p::Real) = xval(μ, σ, norminvcdf(p))
+# Promote to ensure that we don't compute erfcinv in low precision and then promote
+norminvcdf(μ::Real, σ::Real, p::Real) = norminvcdf(promote(μ, σ, p)...)
+norminvcdf(μ::T, σ::T, p::T) where {T<:Real} = xval(μ, σ, norminvcdf(p))
 
 norminvccdf(p::Real) = erfcinv(2*p) * sqrt2
-norminvccdf(μ::Real, σ::Real, p::Real) = xval(μ, σ, norminvccdf(p))
+# Promote to ensure that we don't compute erfcinv in low precision and then promote
+norminvccdf(μ::Real, σ::Real, p::Real) = norminvccdf(promote(μ, σ, p)...)
+norminvccdf(μ::T, σ::T, p::T) where {T<:Real} = xval(μ, σ, norminvccdf(p))
 
 # invlogcdf. Fixme! Support more precisions than Float64
 norminvlogcdf(lp::Union{Float16,Float32}) = convert(typeof(lp), _norminvlogcdf_impl(Float64(lp)))

test/rmath.jl

@@ -17,7 +17,15 @@ end
 get_statsfun(fname) = eval(Symbol(fname))
 get_rmathfun(fname) = eval(Meta.parse(string("RFunctions.", fname)))
 
-function rmathcomp(basename, params, X::AbstractArray, rtol=100eps(float(one(eltype(X)))))
+function rmathcomp(basename, params, X::AbstractArray)
+    # compute default tolerance:
+    # has to take into account `params` as well since otherwise e.g. `X::Array{<:Rational}`
+    # always uses a tolerance based on `eps(one(Float64))` even when parameters are of type
+    # Float32
+    rtol = 100 * eps(float(one(promote_type(Base.promote_typeof(params...), eltype(X)))))
+    rmathcomp(basename, params, X, rtol)
+end
+function rmathcomp(basename, params, X::AbstractArray, rtol)
     # tackle pdf specially
     has_pdf = true
     if basename == "srdist"
@@ -264,6 +272,8 @@ end
         ((0.0, 0.0), -6.0:0.1:6.0),
         ((0f0, 1f0), -6f0:0.01f0:6f0),
         ((0.0, 1.0), -6f0:0.01f0:6f0),
+        ((0, 2), -6//1:1//2:6//1),
+        ((0f0, 2f0), -6//1:1//2:6//1),
         # Fail since `SpecialFunctions.erfcx` is not implemented for `Float16`
         #((Float16(0), Float16(1)), -Float16(6):Float16(0.01):Float16(6)),
         #((0f0, 1f0), -Float16(6):Float16(0.01):Float16(6)),