Merge pull request #39 from bgctw/dev2

support fitting ScaledLogNormal of LogNormal(-x)

docs/src/lognormal.md

@@ -45,6 +45,33 @@ d = fit_mean_relerror(LogNormal, 2, 0.2)
 (true, true)
 ```
 
+## ScaledLogNormal: LogNormal(-x)
+To support the use case of a distribution of strictly negative values, the
+fitting of a mirrored LogNormal on `-x` is supported.
+
+There is a type-alias 
+`ScaledLogNormal = LocationScale{T, Continuous, LogNormal{T}} where T`,
+denoting a scaled and shifted LogNormal distribution.
+
+There are fitting function dispatched by this type that fit
+such a mirrored distribution.
+
+```jldoctest; output = false, setup = :(using DistributionFits)
+d = fit_mean_Σ(ScaledLogNormal, -1, log(1.1))
+(mean(d), σstar(d)) .≈ (-1.0, 1.1)
+# output
+(true, true)
+```
+
+```jldoctest; output = false, setup = :(using DistributionFits)
+d = fit(ScaledLogNormal, -1.0, @qp_ll(-1.32), Val(:mode))
+(mode(d), quantile(d, 0.025)) .≈ (-1.0, -1.32)
+# output
+(true, true)
+```
+Note the usage of lower quantile for the mirrored distribution, here.
+
+
 ## Detailed API
 
 ```@docs

ext/DistributionFitsOptimExt.jl

@@ -11,7 +11,7 @@ function DistributionFits.optimize(f, ::OptimOptimizer, lower, upper)
 end
 
 function __init__()
-    @info "DistributionFits: setting OptimOptimizer"
+    #@info "DistributionFits: setting OptimOptimizer"
     #DistributionFits.set_optimizer(DistributionFitsOptimExt.OptimOptimizer())
     DistributionFits.set_optimizer(OptimOptimizer())
 end

inst/scratch.jl

@@ -0,0 +1 @@
+

src/DistributionFits.jl

@@ -36,7 +36,7 @@ export
 #   optimize, set_optimizer  
 
 # LogNormal
-export AbstractΣstar, Σstar, σstar
+export AbstractΣstar, Σstar, σstar, ScaledLogNormal
 
 # LogitNormal  
 export fit_mode_flat, shifloNormal

src/univariate/continuous/lognormal.jl

@@ -166,3 +166,32 @@ function fit_mean_relerror(::Type{LogNormal}, mean, relerror)
     σ = sqrt(log(w))
     LogNormal(μ, σ)
 end
+
+
+#---- support LogNormal(-x) of negative values ------------
+const ScaledLogNormal{T} = LocationScale{T, Continuous, LogNormal{T}} where T
+
+σstar(d::ScaledLogNormal) = exp(params(d.ρ)[2])
+
+
+function fit_mean_Σ(::Type{ScaledLogNormal}, mean::T1, σ::T2) where {T1 <: Real,T2 <: Real}
+    _T = promote_type(T1, T2)
+    fit_mean_Σ(ScaledLogNormal{_T}, mean, σ)
+end
+function fit_mean_Σ(d::Type{ScaledLogNormal{T}}, mean::Real, σ::Real) where T
+    mean < 0 && return(-1 * fit_mean_Σ(LogNormal{T}, -mean, σ))
+    1 * fit_mean_Σ(LogNormal{T}, mean, σ)
+end
+
+function fit_mode_quantile(::Type{ScaledLogNormal}, mode::T, qp::QuantilePoint) where T<:Real
+    fit_mode_quantile(ScaledLogNormal{T}, mode, qp)
+end
+function fit_mode_quantile(::Type{ScaledLogNormal{T}}, mode::Real, qp::QuantilePoint) where T
+    if mode < 0 
+        return(-1 * fit_mode_quantile(LogNormal{T}, -mode, QuantilePoint(-qp.q,1-qp.p)))
+        #return(-1 * fit_mode_quantile(LogNormal{T}, -mode, qp))
+    end
+    1 * fit_mode_quantile(LogNormal{T}, mode, qp)
+end
+
+

test/univariate/continuous/lognormal.jl

@@ -54,3 +54,23 @@ end;
     @test mode(d) ≈ 0.5
     @test quantile(d, 0.95) ≈ 0.9
 end;
+
+@testset "fitting ScaledLogNormal" begin
+    d2 = fit_mean_Σ(ScaledLogNormal, -1, log(1.1))
+    @test d2 isa ScaledLogNormal
+    @test σstar(d2) == 1.1
+
+    d2 = fit_mean_Σ(ScaledLogNormal, 1.0, log(1.1))
+    @test d2 isa ScaledLogNormal
+    @test mean(d2) == 1.0
+    @test σstar(d2) == 1.1
+
+    # take care to specify qp_ll here, its lower than mode
+    d3 = fit(ScaledLogNormal, -1.0, @qp_ll(-1.32), Val(:mode))
+    @test d3 isa ScaledLogNormal
+    @test mode(d3) == -1.0
+    @test quantile(d3, 0.025) ≈ -1.32
+
+    d3 = fit(ScaledLogNormal, 1.0, @qp_uu(1.32), Val(:mode))
+    @test d3 isa ScaledLogNormal
+end;