What's so special about `array`s?

I’m working on a new video about S3 objects in R and class inheritance / hierarchy. One of my favorite functions for exploring objecs and their structure is unclass().

The documentation states

unclass returns (a copy of) its argument with its class attribute removed. (It is not allowed for objects which cannot be copied, namely environments and external pointers.)

So, unclass() should remove the class of any and all objects. For the sake clarity of this post I’m going to make a helper function.

class_unclass <- function(x) class(unclass(x))

This works for factors which are just integer vectors which have the attribute levels

class_unclass(factor())

[1] "integer"

and data.frames are lists (also technically a vector just not atomic but instead recursive) with attributes row.names, and names.

class_unclass(data.frame())

[1] "list"

But when we get down to matrix another poser type (just like data.frame and factor pretending to be something they’re not actually), we get something different.

class_unclass(matrix())

[1] "matrix" "array" 

Well, why the heck is that? What about it makes it so special? Let’s explore this a bit more.

The two things that make a matrix are the classes c("matrix", "array") and the dim attribute which specifies the dimensions. Matrixes are two dimensional arrays, by the way!

attributes(matrix())

$dim
[1] 1 1

What is weird is that you can make a matrix just by adding the dim attribute to a vector.

m <- structure(integer(1), dim = c(1, 1))
class_unclass(m)

[1] "matrix" "array" 

We didn’t even specify the class. Why does this happen?

And when we remove the dim attribute….

attr(m, "dim") <- NULL
class_unclass(m)

[1] "integer"

we get an integer vector. This differs from the behavior of other similar types. Recall that factors are integer vectors with an attribute of levels that is a character vector.

So let’s try something here. Let’s create a factor from scratch.

structure(integer(), levels = character(), class = "factor")

factor()
Levels: 

Now, if the behavior is similar to matrix or array we would expect that by omitting the class attribute R should reconstruct it.

structure(integer(), levels = character())

integer(0)
attr(,"levels")
character(0)

Nope! Would you look at that!

@yjunechoe pointed me to some excerpts from Deep R Programming, a book I wish I had read yesterday. It refers to these attributes as “special attributes”. The author, Marek Gagolewski (author of stringi, by the way), makes note of this special behavior of matrix but leaves it at that.

To me, this is a fundamental inconsistency in the language. Either all poser types (data.frame, matrix, factor, etc) should be automatically created if their special attributes are set on the appropriate type or not at all. What justification is there for only matrix having a special behavior in unclass()?

To me, this warrants a bug report for unclass(). Based on the documentation, unclass() should always remove the class attribute from an object but it fails to do so for arrays and matrixes.

Looking deeper!

With some further exploratory help of June we can see the internal representation of these objects.

Let’s create an object with a dim attribute and a custom class. Suprisingly, the custom class is respected and matrix and array aren’t slapped on it.

x <- structure(integer(1), dim = c(1, 1), class = "meep")
class(x)

[1] "meep"

If we look at the internal representation using .Internal(inspect(x)) we get to see some of the C goodies.

.Internal(inspect(x))

@1310a0d28 13 INTSXP g0c1 [OBJ,REF(2),ATT] (len=1, tl=0) 0
ATTRIB:
  @131097930 02 LISTSXP g0c0 [REF(1)] 
    TAG: @156014080 01 SYMSXP g1c0 [MARK,REF(2750),LCK,gp=0x4000] "dim" (has value)
    @1310a0cf0 13 INTSXP g0c1 [REF(65535)] (len=2, tl=0) 1,1
    TAG: @1560141d0 01 SYMSXP g1c0 [MARK,REF(33684),LCK,gp=0x6000] "class" (has value)
    @131092e08 16 STRSXP g0c1 [REF(65535)] (len=1, tl=0)
      @1462b34a8 09 CHARSXP g0c1 [REF(2),gp=0x60] [ASCII] [cached] "meep"

Just look to the right hand side of these gibberish. See that the dim has a value of 1,1 and class has a value of meep. There is no matrix or array or nothing.

Now we remove the meep class and check again.

class_unclass(x)

[1] "matrix" "array" 

Boom matrix and array. But if we look at the internals…

.Internal(inspect(unclass(x)))

@130620d60 13 INTSXP g0c1 [ATT] (len=1, tl=0) 0
ATTRIB:
  @1305255b0 02 LISTSXP g0c0 [REF(1)] 
    TAG: @156014080 01 SYMSXP g1c0 [MARK,REF(2754),LCK,gp=0x4000] "dim" (has value)
    @1310a0cf0 13 INTSXP g0c1 [REF(65535)] (len=2, tl=0) 1,1

THEY AREN’T THERE!!!!!!