What bearophile means is something like this:

struct Tensor(T, size_t nDim)
{
     T[nDim] dimensions;
     T[] values;

    this(...) etc.
}

so Tensor!(int,3) is a 3D perfectly cubic 'matrix' if ints, with for example dimensions 3*5*8.
The only thing CT-defined is the rank (scalar: 0, vector: 1, etc), the number of dimensions. The dimensions themselves are RT values. The total size is the product of these values. It can be calculated during construction.

What you lose is the CT checking that can be done for multiplication or additions if all dimensions are exposed in the type.
I don't know, would that be an possible addition to Phobos? I remember many people here creating matrix libraries, so someone probably made this already.


Concerning arrays of arrays of arrays (or ranges), a useful template is one that gives the number of []'s, the rank:

template rank(T)
{
  static if (!isArray!T)  // or (isInputRange!T)
    enum size_t rank = 0;
  else
    enum size_t rank = 1 + rank!(ElementType!T); // continue recursion
}

As you can see, it considers a simple scalar type to have rank 0. Maybe that's not what you want.
This allows you to put simple template constraints, like  if(rank!T == 2) ...

I played with ranges of ranges a few month ago. Maybe some functions there could interest you:

http://svn.dsource.org/projects/dranges/trunk/dranges/docs/rangeofranges.html
http://www.dsource.org/projects/dranges/browser/trunk/dranges/rangeofranges.d

Mapping a range of ranges, creating them by tensorial product, etc. But the docs are still incomplete, and the global module is due a clean-up, I'm afraid.


Philippe