Solves #136 (#137)

* Change a convert method, addressin Issue 136

This implements an idea of David to solve Issue 136; it does
solve it, but a test is now failing.

* Prevent case when rationalize(x) is zero, for non-zero x

* Add tests for intervals with small values

src/intervals/conversion.jl

@@ -46,7 +46,12 @@ function convert{T<:AbstractFloat, S<:Real}(::Type{Interval{T}}, x::S)
 end
 
 function convert{T<:AbstractFloat}(::Type{Interval{T}}, x::Float64)
-    convert(Interval{T}, rationalize(x))
+    II = convert(Interval{T}, rationalize(x))
+    # This prevents that rationalize(x) returns a zero when x is very small
+    if x != zero(x) && II == zero(Interval{T})
+        II = Interval{T}(string(x))
+    end
+    II
 end
 
 function convert{T<:AbstractFloat}(::Type{Interval{T}}, x::Interval)

test/interval_tests/numeric.jl

@@ -34,6 +34,9 @@ facts("Numeric tests") do
     @fact Interval(0.0, 1.0)/Interval(0.0,1.0) --> Interval(0.0, Inf)
     @fact Interval(-1.0, 1.0)/Interval(0.0,1.0) --> entireinterval(c)
     @fact Interval(-1.0, 1.0)/Interval(-1.0,1.0) --> entireinterval(c)
+    a = @interval(1.e-20)
+    @fact a --> Interval(1.0e-20, 1.0000000000000001e-20)
+    @fact diam(a) --> eps(1.e-20)
 end
 
 facts("Power tests") do