cov(Vector,Vector,dims=1) fails

[PaulSoderlind]

cov(rand(5),rand(5),dims=1) fails, while both cov(rand(5),rand(5)) and cov(rand(5),rand(5,1),dims=1) work.

Why bother? Well, to be consistent with the other functions in Statistics. For instance, std(rand(5),dims=1) gives a 1x1 vector. It is surely a convenience for users to be able to interchange Nx1 matrices and N-vectors in such basic calculations.

The quick fix is to add a dummy dims argument to the current method (in Statistics.jl) as cov(x::AbstractVector, y::AbstractVector; dims::Int=1, corrected::Bool=true) =. This would produce a Number. A better fix might be produce a 1x1 vector to be consistent with std and other functions.

[PaulSoderlind]

I have spent some time with the cov() and cor() functions in Statistics - and I get the impression that the cov() function could use an overhaul. For instance, it is several times slower than cor(). However, it is a somewhat heavy machinery.

In the short run, the issue with the missing cov(Vector,Vector,dims=1) method, and similarly for cor(), can be solved by the crude approach suggested below. The performance seems to be within 1% of the existing code, and attempts to rely on dispatching on dims (via the covm()) do not improve.

function cov(x::AbstractVector, y::AbstractVector; dims=:, corrected::Bool=true)
    result = covm(x, mean(x), y, mean(y); corrected=corrected)
    if dims === (:)
        return result
    elseif isa(dims,Int) && dims >= 1
        dims == 1 ? result2 = fill(result,1,1) : result2 = fill(NaN,length(x),length(y))
        return result2
    else
        error("wrong value of dims")
    end
end

function cor(x::AbstractVector, y::AbstractVector; dims=:)
    result = corm(x, mean(x), y, mean(y))
    if dims === (:)
        return result
    elseif isa(dims,Int) && dims >= 1
        dims == 1 ? result2 = fill(result,1,1) : result2 = fill(NaN,length(x),length(y))
        return result2
    else
        error("wrong value of dims")
    end
end