The need for an approach to arrays is clear and is acute to many DFDL constituencies.

The first step in any approach to arrays for DFDL is  an XML model for array data and an XSD for describing it. Then DFDL can put properties on this.

I suggest the following model. Consider a 2-d case. This will generalize to N dimensions.

Each axis is named. The array itself is represented as elements, with attributes used to identify the position of the value on each axis conceptually like so:

<a x="5" y="-2">51</a>

That is, you think of each array element as having attributes identifying its position in the array. Of course DFDL allows data to be processed without ever creating elements like that, so this is a conceptual model only, particularly for a dense array.

That element is of an array named 'a', at position x=5, y=-2, having value 51.

The declaration in XSD would be like this:

<element name="a" maxOccurs="unbounded">
   <complexType>
    <extension base="int">
        <simpleContent>
           <attribute name="x">
               <simpleType>
                   <restriction base="int">
                      <maxInclusive value="5"/>
                      <minInclusive value="-5"/>
                    </restriction>
               </simpleType>
           </attribute>
       <attribute name="y">
               <simpleType>
                   <restriction base="int">
                      <maxInclusive value="10"/>
                      <minInclusive value="-10"/>
                    </restriction>
               </simpleType>
           </attribute>
        </simpleContent>
   </extension>
  </complexType>
</element>

Notice how the ranges of the index values are captured in XSD by use of the simple type restriction, and can cover arbitrary sections of the integer space, including negative indices.

DFDL would then provide properties for

1) declaring that 'a' is an array and that 'x' and 'y' are array indices (and therefore do not have values stored anywhere in the data).
2) declaring the storage-order of the array. This can be an ordered list of the dimension names. E.g., "x y" or "y x" depending on which index changes fastest in the storage ordering.

Access to elements would be by XPath expressions like this: ..../a[x='5' and y='-2']. Processors would recognize that x and y are array indices based on DFDL annotations and would thereby recognize predicates involving the indices and treat them specially. For example, we could preclude slicing arrays like this:  ..../a[x='0'] that is, where the 'y' axis is unconstrained.