Scalar extension is a way to apply a function with a scalar argument when an array of a particular non-empty shape would be expected. The scalar is extended to this shape by treating it as an array with each element equal to the scalar's only element. This is equivalent to reshaping the scalar to fit the desired shape.
History and terminology
The concept of scalar extension has been around since APL\360. An example which extends the scalar
2 × 1 2 3 4 2 4 6 8
A Programming Language describes the above computation as a "scalar multiple" but does not generalise it to arbitrary scalar functions, so it's unclear when scalar extension as a unified concept was adopted in Iverson notation.
The word "extension" applies to scalar extension in two ways: first, a function is extended by making a case which would have been a RANK ERROR into a valid application. Second, the application works by conceptually extending the scalar to function as though it were an array of higher rank.
Two arrays are said to conform if they have the same shape or at least one can be extended (it is a scalar, or, in langauges with singleton extension, has exactly one element). A pair of conforming arrays defines a single shape which describes how their elements are paired: if neither is a scalar, it is their shared shape; if one is a scalar, it is the other's shape; if both are scalars, it is the empty vector,
The term "scalar extension" is sometimes used to refer to the practice of allowing a scalar when a higher rank is expected. The scalar is treated as an array of the expected minimum rank whose shape is a vector of 1s (that is, a singleton). For example,
8⍴'a' both produce an 8-element array even though the shape of an array is always a vector and so cannot be equal to
8. This type of extension, which differs from ordinary scalar extension in that there is no expected shape and only an expected rank, has also been present since APL\360.
(1 1⍴5) + 10 20 15 25 ⍴ (1 1⍴5) + 10 20 2
In this case addition accepts a singleton, and discards its shape. If two singletons are used as arguments, they are still considered to conform; the shape of the result is taken from the argument with higher rank.
Singleton extension can sometimes conflict with other extensions, an issue which does not occur with scalar extension. If a function is extended to allow a shorter vector argument to be extended (perhaps by padding with 0), but it also supports singleton extension, then there is a conflict with length-1 vectors.
Extension in the Rank operator
The Rank operator uses a generalization of scalar extension to pair cells. A function called with rank 0 exhibits ordinary scalar extension: it acts like a scalar function. A function with higher rank extends not scalars (arrays with empty shape) but arrays whose frame is empty. An empty frame implies there is only one cell, and it has a scalar-like array structure. This cell can be extended by reusing it for every function call.
1 2 3 4 * 2 1 4 9 16
⍴ 2 ¯3 / 7 1 8⍴⍳56 7 5 8
In languages which allow a vector left argument to Rotate, the behavior with a scalar left argument follows from scalar extension. In the following example a length-2 vector could be used to rotate each row by a different amount. A scalar rotates both rows by the same amount.
3⌽2 6⍴'extendscalar' endext larsca
|APL features |
|Built-ins||Primitive function ∙ Primitive operator ∙ Quad name|
|Array model||Shape ∙ Rank ∙ Depth ∙ Bound ∙ Index ∙ Axis ∙ Ravel ∙ Ravel order ∙ Element ∙ Scalar ∙ Vector ∙ Matrix ∙ Simple scalar ∙ Simple array ∙ Nested array ∙ Box ∙ Cell ∙ Major cell ∙ Subarray ∙ Empty array ∙ Prototype|
|Concepts and paradigms||Leading axis theory ∙ Scalar extension ∙ Conformability ∙ Scalar function ∙ Glyph ∙ Identity element ∙ Total array ordering|
|Errors||LIMIT ERROR ∙ RANK ERROR ∙ SYNTAX ERROR|