NativeNaNMath

Alternative approach to NaNMath.jl, by using the functions available in Julia Base instead of the Libm ones.

It should be (almost) drop-in replacement for NaNMath.jl.

The following functions are not exported but defined:

• Logarithmic:
  • log(x),log1p(x),log2(x),log10(x),log1p(x)
  • log(x,base)
• Trigonometric:
  • sin(x),cos(x),tan(x),cot(x),sec(x),csc(x)
  • sind(x),cosd(x),tand(x),cotd(x),secd(x),cscd(x)
  • sinpi(x),cospi(x)
  • sincos(x),sincosd(x),sincospi(x)
  • asin(x),acos(x),asec(x),acsc(x)
  • asind(x),acosd(x),asecd(x),acscd(x)
• Hyperbolic:
  • acosh(x),asech(x),atanh(x),acoth(x)
• sqrt(x)
• pow(x,y)
• min(x),max(x)

The package only exports a single function: skipnan(itr) that works in the same way that skipmissing(itr):

x = collect(1.0:10.0)
x[end] = NaN
xn = skipnan(x)
sum(xn) #45

The package uses the nan(::Type{<:Real}) function to obtain an always valid NaN. on types that aren't capable of holding NaNs, (like all integers), it will return a promoted type that can hold NaNs (Float64 for Int8,Int16,Int32,Int64, BigFloat for BigInt). on Rationals, nan(Rational{T}) = nan(T). This function should satisfy isnan(nan(T))

It defaults to zero(x)/zero(x).

• instead of providing NaN-compatible sum, maximum, minimum, etc. It provides a nan-skipping iterator. it can reproduce almost all functionality, except some corner cases:
  • sum(skipnan[NaN]) is 0.0 instead of NaN, because collect(skipnan([NaN])) = Float64[] and sum(Float64[]) == 0.0
  • median(skipnan([NaN]) is not defined. same reason that with sum Other than that, skipnan expands the NaN functionality to any reducing operator.
• pow(x::Integer,y::Integer) will always promote to a float type.
• NativeNaNMath.f(x) where x is not a Real number will always default to Base.f(x). This is useful because automatic differenciation and custom number types can use this package without overloading anything.