Histogram dropping values when dealing with signed zero

[mtanneau]

I encountered this issue through UnicodePlots (JuliaPlots/UnicodePlots.jl#229), but I believe the root cause lies in StatsBase.

It appears that histogram will drop some entries when the left-end of the data is a signed zero -0.0.
From the original issue

julia> histogram([0.0, 1.0])
              ┌                                        ┐ 
   [0.0, 0.5) ┤█████████████████████████████████████  1  
   [0.5, 1.0) ┤  0                                       
   [1.0, 1.5) ┤█████████████████████████████████████  1  
              └                                        ┘ 
                               Frequency

julia> histogram([-0.0, 1.0])
              ┌                                        ┐ 
   [0.0, 0.5) ┤  0                                         # where did that zero go?
   [0.5, 1.0) ┤  0                                       
   [1.0, 1.5) ┤█████████████████████████████████████  1  
              └                                        ┘ 
                               Frequency

julia> histogram([0.0, -0.0])
              ┌                                        ┐ 
   [0.0, 1.0) ┤█████████████████████████████████████  1    # only one value?
              └                                        ┘ 
                               Frequency

On the one hand, I can understand that -0.0 would technically fall in a separate bin than +0.0, because of floating-point ordering.
However, it was surprising to see those entries dropped from the total count.

[t-bltg]

I think this will be more representative in StatsBase:

julia> using StatsBase
julia> fit(Histogram, [-0.0, 1.0])
Histogram{Int64, 1, Tuple{StepRangeLen{Float64, Base.TwicePrecision{Float64}, Base.TwicePrecision{Float64}}}}
edges:
  0.0:0.5:1.5
weights: [0, 0, 1]
closed: left
isdensity: false

julia> fit(Histogram, [-0.0, 1.0], closed=:right)
Histogram{Int64, 1, Tuple{StepRangeLen{Float64, Base.TwicePrecision{Float64}, Base.TwicePrecision{Float64}}}}
edges:
  -0.5:0.5:1.0
weights: [1, 0, 1]
closed: right
isdensity: false

[nalimilan]

Good catch. Even beyond the bug, it seems that treating both zeros exactly the same makes more sense for histograms. See #768.