The closely related 1-modifiers Fold (´) and Insert (˝) apply a dyadic operand function 𝔽 repeatedly between elements or major cells of 𝕩. Neither is quite like the APL2-style Reduce operator (/ or ⌿ in APL), although I sometimes use the term "reduction" to mean either Fold or Insert. There are a bunch of other names like "accumulate" and "aggregate" for this class of calculations—I don't know which is best but I know "catamorphism" is worst.
A distinguishing feature of APL-family reductions is that they don't use an initial value, and try to derive an "identity element" for 𝔽 if the argument array is empty. BQN retains this capability but also allows the programmer to supply an initial value as 𝕨.
As its glyph suggests, Fold is slightly simpler than Insert. The argument 𝕩 must always be a list, and Fold applies 𝔽 between elements—always two at a time—of the list to yield a single result value. In this sense, 𝔽´ removes a layer of depth from 𝕩, although it's not necessarily true that the depth of 𝔽´𝕩 is less than that of 𝕩 because the function 𝔽 might increase depth.
+´ 2‿4‿3‿1 10 +´ ⟨2‿4, 3‿1⟩ ⟨ 5 5 ⟩
Any function can be used as an operand. You can find the largest number out of an entire list with Maximum (⌈), or the smallest with Minimum (⌊).
⌈´ 2‿4‿3‿1 4 ⌊´ 2‿4‿3‿1 1 ×´ 2‿4‿3‿1 # Product as well 24
The logic function And (∧) tests if all elements of a boolean list are 1, while Or (∨) tests if any are 1.
∧´ 1‿1‿0 0 ∨´ 1‿1‿0 1
Folding over a list of length 1 never calls the operand function: it returns the lone element unchanged.
↗️!´ ⟨⎊⟩ ⎊
Folding over a list of two values applies 𝔽 once, since 𝔽 is always called on two arguments. But what about zero values? Should 𝔽 be applied minus one times? Sort of. BQN checks to see if it knows an identity value for the operand function, and returns that, never calling the function. This works for the arithmetic functions we showed above, always returning a single number.
+´ ⟨⟩ # Add nothing up, get zero 0 ⌈´ ⟨⟩ # The smallest number ¯∞ ∧´ ⟨⟩ # All the elements in the list are true… 1
Here's the full list of identity values Fold has to support.
| Id | Fn | Fn | Id |
|---|---|---|---|
0 |
+ |
- |
0 |
1 |
× |
÷ |
1 |
1 |
⋆ |
¬ |
1 |
∞ |
⌊ |
⌈ |
¯∞ |
0 |
∨ |
∧ |
1 |
0 |
≠ |
= |
1 |
0 |
> |
≥ |
1 |
The functions we've shown so far are associative (ignoring floating point imprecision), meaning it's equally valid to combine elements of the argument list in any order. But it can be useful to fold using a non-associative function. In this case you must know that Fold performs a right fold, starting from the end of the array and working towards the beginning.
↗️⋈´ "abcd" ⟨ 'a' ⟨ 'b' "cd" ⟩ ⟩ 'a' ⋈ 'b' ⋈ 'c' ⋈ 'd' # Expanded form ⟨ 'a' ⟨ 'b' "cd" ⟩ ⟩
Pair (⋈) as an operand shows the structure nicely. This fold first pairs the final two characters 'c' and 'd', then pairs 'b' with that and so on. This matches BQN's right-to-left order of evaluation. More declaratively we might say that each character is paired with the result of folding over everything to its right.
BQN doesn't provide a left Fold (` is Scan). However, you can fold from the left by reversing (⌽) the argument list and also reversing (˜) the operand function's argument order.
⋈˜´ ⌽ "abcd" ⟨ ⟨ "ab" 'c' ⟩ 'd' ⟩
One consequence of this ordering is that folding with Minus (-) gives an alternating sum, where the first value is added, the second subtracted, the third added, and so on. Similarly, ÷ gives an alternating product, with some elements multiplied and some divided.
-´ 30‿1‿20‿2‿10 57
And the operand +⟜÷ is a quick way to compute a continued fraction's value from a list of numbers. Here are a few terms from the continued fraction for e.
+⟜÷´ 2‿1‿2‿1‿1‿4‿1‿1 2.7183098591549295
When 𝔽 isn't just an arithmetic primitive, folding with no initial element can be dangerous. Even if you know 𝕩 isn't empty, saving you from an "Identity not found" error, the case with only one element can easily violate expectations. Here's a somewhat silly example of a function meant to merge elements of the argument into a single list (∾⥊¨ is a much better way to do this):
∾○⥊´ ⟨2‿4≍6‿8,"abcd",0⟩ ⟨ 2 4 6 8 'a' 'b' 'c' 'd' 0 ⟩ ∾○⥊´ ⟨2‿4≍6‿8,"abcd"⟩ ⟨ 2 4 6 8 'a' 'b' 'c' 'd' ⟩ ∾○⥊´ ⟨2‿4≍6‿8⟩ ┌─ ╵ 2 4 6 8 ┘
The result always has rank 1, until the one-element case, when ∾○⥊ is never applied and can't deshape anything. Using Fold with lots of complex operands and no initial element can make a program fragile.
The left argument, if given, is the initial right argument to 𝔽. This puts 𝕨 on the same level as an element of 𝕩, so it doesn't need to be enclosed and will usually have one smaller depth. For ∾○⥊´ the natural starting element is the empty list.
⟨⟩ ∾○⥊´ ⟨2‿4≍6‿8⟩ ⟨ 2 4 6 8 ⟩
With a non-empty 𝕨 we can see it's placed at the end of the result list, because it's passed to 𝔽 before any elements of 𝕩.
"end" ∾○⥊´ ⟨"start","middle"⟩ "startmiddleend"
Folding with 𝕨 never needs to come up with an identity value, and the number of function applications is exactly the length of 𝕩. A function P can be applied to each element of 𝕩 before operating using 𝕨P⊸F´𝕩, which is equivalent to 𝕨 F´ P¨𝕩 except for the order in which F and P are invoked (if they have side effects).
"STOP" ⌽⊸∾´ "ABCDE"‿"012"‿"abcd" "EDCBA210dcbaSTOP"
Fold only works on lists. What if you want to, say, sum the columns of a table?
↗️⊢ tab ← (2+↕5) |⌜ 9+↕3 ┌─ ╵ 1 0 1 0 1 2 1 2 3 4 0 1 3 4 5 ┘ +˝ tab ⟨ 9 7 12 ⟩
The Insert (˝) modifier will do this for you. And because it works on the leading axis of the argument, Insert can be applied to axes other than the first with Rank. Sum each row (second axis) with Cells (˘), for example.
+˝˘ tab ⟨ 2 3 6 5 12 ⟩
This case is tricky, because +´˘ tab yields the same result but is actually unsound—if tab contains arrays then they will be merged together at the end. Remember: if you want to reduce along one axis of an array and get an array of results out, you should use Insert (possibly adding Each to work on elements instead of cells; see APL2 reduction below).
A function with Insert 𝔽˝ is nearly equivalent to 𝔽´<˘ (and both fail on unit arguments, because there's no axis to apply along). Besides being more convenient, 𝔽˝ is a little safer because it takes the argument shape into account when returning an identity value:
+´<˘ 0‿4⥊0 0 +˝ 0‿4⥊0 ⟨ 0 0 0 0 ⟩
Just like Fold, Insert allows an initial element for the left argument, so that you don't need to rely on the interpreter knowing the identity (in Fold terms it's {𝕨𝔽´<˘𝕩}). Below, we see that 𝔽 is called on the last major cell and 𝕨, then the next-to-last major cell and so on. This makes ≠𝕩 calls, while there would be 1-˜≠𝕩 without the initial value.
"id" ⋈˝ ["row0 ","row1 ","row2 "] ┌─ · "row0 " ⟨ "row1 " ⟨ "row2 " "id" ⟩ ⟩ ┘
One trick involving Insert is ∾˝, which uses Join to to merge the first two axes of 𝕩 into one long axis. It even works on empty arrays, because BQN knows that there's only one result shape that makes sense (unlike ∾´⟨⟩, where many results sometimes fit but none of them always fit).
⊢ let ← ("AHW"-'A') +⌜ "aA" +⌜ ↕4 ┌─ ╎"abcd ABCD ·hijk HIJK ·wxyz WXYZ" ┘ ∾˝ let ┌─ ╵"abcd ABCD hijk HIJK wxyz WXYZ" ┘ ≢ ∾˝ ↕3‿2‿4 ⟨ 6 4 ⟩ ≢ ∾˝ ↕0‿2‿4 # The identity is an empty cell ⟨ 0 4 ⟩
As a historical note, Insert is named after J's adverb /, which comes from SHARP APL's ⌿, reduce-down. In the original APL, only arithmetic reductions were defined, and nested arrays didn't exist—arrays were either all characters or all numbers. SHARP extended them by splitting the array into cells as we've shown. However, there's another interpretation, which is what you'll find in mainstream APLs today…
The function ⍪⌿ in Dyalog APL gives very different results from BQN's ∾˝. Instead of combining the cells like we see above, APL applies the function on pairs of elements much like Fold. The difference is that, because reduction happens only along one axis but an array might have other axes, there can be multiple values in the result, so that it will always be an array like the argument. BQN can perform this operation as well: ⍪⌿ is written ∾¨˝ in BQN (but please use <˘⍉ instead).
tab ┌─ ╵ 1 0 1 0 1 2 1 2 3 4 0 1 3 4 5 ┘ ∾¨˝ tab ⟨ ⟨ 1 0 1 4 3 ⟩ ⟨ 0 1 2 0 4 ⟩ ⟨ 1 2 3 1 5 ⟩ ⟩
This kind of reduction has an interesting property not found in ´ or ˝: it always removes exactly one axis, so that the result's shape is 𝕩's major cell shape. When applied to a later axis using Rank or Cells, it removes that axis instead.
≢ ∾¨˝ ↕4‿2‿3 # Reduce out the first axis ⟨ 2 3 ⟩ ≢ ∾¨˝˘ ↕4‿2‿3 # Reduce out the second ⟨ 4 3 ⟩
When the operand is an arithmetic function, say ⌊, APL2-style reduction is no different from Insert: ⌊¨˝ is the same as ⌊˝, because ⌊¨ and ⌊ are the same on arrays. That means that Insert with an arithmetic operand also has this axis-removing property.