NaN definition for Integer types?

[Júlio Hoffimann]

Hi,

Is there any definition of NaN for Integer types? Let's say I want to create a matrix with some unspecified entries, that is entries marked special. I am using NaN for Float64 but would like to make it work with other types too. Suggestions?

-Júlio

[Tamas Papp]

Every bit pattern in a Signed or Unsigned represents an integer,
there is no representation for a NaN like in eg Float64 etc. You
would have to use Nullable types, see
https://julia.readthedocs.org/en/latest/manual/types/#nullable-types-representing-missing-values

Best,

Tamas

On Fri, Jul 24 2015, Júlio Hoffimann <julio.h...@gmail.com>
wrote:

[Erik Schnetter]

On Fri, Jul 24, 2015 at 1:09 PM, Júlio Hoffimann <julio.h...@gmail.com> wrote:

Hi,

Is there any definition of NaN for Integer types? Let's say I want to create a matrix with some unspecified entries, that is entries marked special. I am using NaN for Float64 but would like to make it work with other types too. Suggestions?

There is no pre-defined integer nan equivalent.

Option 1 (fast but unsafe): Pick a special int value that is rarely used, e.g. typemin(Int).

Option 2 (works with all types, including Float64, and is safer): Use a nullable type. This might complicate your code and may slightly increase your memory usage.

-erik

--

Erik Schnetter <schn...@gmail.com> http://www.perimeterinstitute.ca/personal/eschnetter/

[Júlio Hoffimann]

Thank you all, that is what I thought, I will stick with plain Float64 for now.

-Júlio

[Scott Jones]

Maybe you could make your own 64-bit bitstype, which restricted values to integer values in the range that will fit in a Float64 (around -2^53 to 2^53).  Then you could be sure you didn't lose any information (give an InexactError if a value doesn't fit), and still allow for a NaN value, and fairly efficient computation.

[Stefan Karpinski]

This is definitely possible, but integer operations will be a bit slower since you'll have to do something like this:

bitstype 64 NullableInt

+(x::NullableInt, y::NullableInt) = ifelse(

        (reinterpret(Int,x) !== typemin(Int)) & (reinterpret(Int,y) !== typemin(Int)),

        reinterpret(NullableInt, reinterpret(Int,x) + reinterpret(Int,y)),

        reinterpret(NullableInt, typemin(Int)),

    )

Fortunately, this is not as bad as it looks, but it's still a lot more work than just doing an integer add instruction.

[Júlio Hoffimann]

Thank you all, I'll try to figure out a good solution for my use case.

-Júlio

[Sisyphuss]

By the way, is it a v0.4 feature?

[Stefan Karpinski]

Yes, but available with Compat too.

[David Gold]

If you end up trying the `Nullable` route, you might find our in-progress `NullableArrays` package helpful: https://github.com/johnmyleswhite/NullableArrays.jl. Its designed to extend the functionality of the `AbstractArray` interface to a container type specialized to handle `Nullable` entries. I'm currently working on this for the JSoC, so there is a point person for issues and functionality requests.

Note, though, that it's being developed for 0.4.