StatsBase.jl

StatsBase.jl copied to clipboard

fit(Histogram, X) fails when X contains values x for which iszero(x) == false but isinf(1 / x) == true.

Example:

julia> fit(Histogram, [2e-314; zeros(100)])
ERROR: InexactError: Int64(NaN)
Stacktrace:
  [1] Int64
    @ ./float.jl:723 [inlined]
  [2] histrange(lo::Float64, hi::Float64, n::Int64, closed::Symbol)
    @ StatsBase ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:99
  [3] histrange(v::Vector{Float64}, n::Int64, closed::Symbol)
    @ StatsBase ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:39
  [4] #142
    @ ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:103 [inlined]
  [5] map
    @ ./tuple.jl:232 [inlined]
  [6] histrange
    @ ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:102 [inlined]
  [7] #fit#158
    @ ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:332 [inlined]
  [8] #fit#151
    @ ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:276 [inlined]
  [9] fit(::Type{Histogram{Int64, N, E} where {N, E}}, v::Vector{Float64})
    @ StatsBase ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:276
 [10] #fit#161
    @ ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:383 [inlined]
 [11] fit(::Type{Histogram}, args::Vector{Float64})
    @ StatsBase ~/.julia/packages/StatsBase/pJqvO/src/hist.jl:383
 [12] top-level scope
    @ REPL[66]:1

julia> iszero(2e-314)
false

julia> isinf(1 / 2e-314)
true

Huh, weird edge case. Perhaps worth noting that 2e-314 is subnormal, i.e. it's less than floatmin(), and using floatmin() here works as expected. This isn't my area of expertise; perhaps @simonbyrne might have some sagely guidance here?