This package defines a Composition{D} type representing a D-part composition as defined by
Aitchison 1986. In Aitchison's geometry,
the D-simplex together with addition (a.k.a. pertubation) and scalar multiplication
(a.k.a. scaling) form a vector space, and important properties hold:
- Scaling invariance
- Pertubation invariance
- Permutation invariance
- Subcompositional coherence
In practice, this means that one can operate on compositional data (i.e. vectors whose entries represent parts of a total) without destroying the ratios of the parts.
Get the latest stable release with Julia's package manager:
] add CoDaCompositions are static vectors with named parts:
julia> using CoDa
julia> c = Composition(CO₂=2.0, CH₄=0.1, N₂O=0.3)
3-part composition
┌ ┐
CO₂ ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 2.0
CH₄ ┤■■ 0.1
N₂O ┤■■■■■ 0.3
└ ┘
julia> CoDa.parts(c)
(:CO₂, :CH₄, :N₂O)
julia> CoDa.components(c)
3-element StaticArrays.SVector{3, Union{Missing, Float64}} with indices SOneTo(3):
2.0
0.1
0.3
julia> c.CO₂
2.0Default names are added otherwise:
julia> c = Composition(1.0, 0.1, 0.1)
3-part composition
┌ ┐
w1 ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 1.0
w2 ┤■■■■ 0.1
w3 ┤■■■■ 0.1
└ ┘ and serve for internal compile-time checks.
Compositions can be added, subtracted, negated, and multiplied by scalars. Other operations are also defined including dot product, induced norm, and distance:
julia> cₒ = Composition(CO₂=1.0, CH₄=0.1, N₂O=0.1)
3-part composition
┌ ┐
CO₂ ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 1.0
CH₄ ┤■■■■ 0.1
N₂O ┤■■■■ 0.1
└ ┘
julia> -cₒ
3-part composition
┌ ┐
CO₂ ┤■■ 0.047619047619047616
CH₄ ┤■■■■■■■■■■■■■■■■■■■ 0.47619047619047616
N₂O ┤■■■■■■■■■■■■■■■■■■■ 0.47619047619047616
└ ┘
julia> 0.5c
3-part composition
┌ ┐
CO₂ ┤■■■■■■■■■■■■■■■■■■■■ 0.6207690197922022
CH₄ ┤■■■■ 0.13880817265812764
N₂O ┤■■■■■■■■ 0.24042280754967013
└ ┘
julia> c - cₒ
3-part composition
┌ ┐
CO₂ ┤■■■■■■■■■■■■■■■■■■■■■■■ 0.3333333333333333
CH₄ ┤■■■■■■■■■■■■ 0.16666666666666666
N₂O ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 0.5
└ ┘
julia> c ⋅ cₒ
3.7554028908352994
julia> norm(c)
2.1432393747688687
julia> aitchison(c, cₒ) # Aitchison distance
0.7856640352007868More complex functions can be defined in terms of these
operations. For example, the function below defines the
composition line passing through cₒ in the direction of c:
julia> f(λ) = cₒ + λ*c
f (generic function with 1 method)Finally, two compositions are considered to be equal when their closure is approximately equal:
julia> c == c
true
julia> c == cₒ
falseCurrently, the following log-ratio transformations are implemented:
julia> alr(c)
2-element StaticArrays.SArray{Tuple{2},Float64,1,2} with indices SOneTo(2):
1.8971199848858813
-1.0986122886681096
julia> clr(c)
3-element StaticArrays.SArray{Tuple{3},Float64,1,3} with indices SOneTo(3):
1.6309507528132907
-1.3647815207407001
-0.2661692320725906
julia> ilr(c)
2-element StaticArrays.SArray{Tuple{2},Float64,1,2} with indices SOneTo(2):
-2.1183026052494185
-0.3259894019031434and their inverses alrinv, clrinv and ilrinv.
The transforms for tables are defined in the TableTransforms.jl
package, they are: Closure, Remainder, ALR, CLR, ILR.
These transforms are functors that can be used as follows:
julia> table |> ILR()It is often useful to compose D columns of a table into D-part compositions. The
package provides a CoDaArray type that implements the Julia array interface and the
Tables.jl interface. We recommend using the function compose(table, cols) to construct
such arrays:
julia> table = (a=[1,2,3], b=[4,5,6], c=[7,8,9])
(a = [1, 2, 3], b = [4, 5, 6], c = [7, 8, 9])
julia> ctable = compose(table, (:a,:b))
(c = [7, 8, 9], CODA = Composition{2, (:a, :b)}[1.000 : 4.000, 2.000 : 5.000, 3.000 : 6.000])
julia> ctable.CODA[1]
2-part composition
┌ ┐
a ┤■■■■■■■■■ 1.0
b ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 4.0
└ ┘ D-part compositions can be created at random from a Dirichlet distribution:
julia> rand(Composition{3})
3-part composition
┌ ┐
w1 ┤■■■■■■■■■■■■■■■■■ 0.39938229705106565
w2 ┤■■■■■■ 0.1491859823748656
w3 ┤■■■■■■■■■■■■■■■■■■■ 0.45143172057406883
└ ┘Separate packages are available for plotting compositional data:
- Relative variation biplots: Biplots.jl
- Ternary diagrams (Makie.jl) TernaryDiagrams.jl
- Ternary diagrams (Plots.jl) TernaryPlots.jl
This package is heavily influenced by Aitchison's monograph:
- Aitchison, J. 1986. The Statistical Analysis of Compositional Data
and by other textbooks:
- den Boogaart, K. & Tolosana-Delgado. 2011. Analyzing Compositional Data with R
- Pawlowsky-Glahn et al. 2015. Modeling and Analysis of Compositional Data
- Pawlowsky-Glahn, V. & Buccianti, A. 2011. Compositional Data Analysis - Theory and Applications
The unicode display of composition objects can be obtained with the following code:
using UnicodePlots
using CoDa
function Base.show(io::IO, mime::MIME"text/plain",
c::Composition{D,PARTS}) where {D,PARTS}
w = CoDa.components(c)
x = Vector{Float64}()
p = Vector{Symbol}()
m = Vector{Symbol}()
for i in 1:D
if ismissing(w[i])
push!(m, PARTS[i])
else
push!(p, PARTS[i])
push!(x, w[i])
end
end
plt = barplot(p, x, title="$D-part composition")
isempty(m) || annotate!(plt, :t, "missing: $(join(m,", "))")
show(io, mime, plt)
endThe code is not added to the CoDa.jl package itself because the UnicodePlots.jl package has become a very heavy dependency, see UnicodePlots/issues/291.