Inverting calls

Slicing and inversion are confusing and error-prone features that are unfortunately very useful and therefore ubiquitous.


TODO write overview

TODO maybe I should start with slicing, then introduce inversion as a notation shortcut

Slicing is confusing because sliced events are evaluated with less context than it looks like they have in the score.

Inversion is confusing because it means a track shows up twice in the stack. In general, the call order is not the same as what is on the screen. For instance, two inverting calls will wind up inverting until the recursion limit is hit (in fact, the limit exists for that reason).


There are several variants of slicing, all of which are implemented in Derive.Slice.

The original motivation was a "delay" transformer that causes a note to be delayed by a bit. A transformer can either change the environment to encourage the deriver to produce a certain output, or evaluate the deriver and transform the events directly.

So while directly transforming the score events is appropriate in some cases, we still need to be able to transform controls by manipulating the environment.

However, in the normal order of evaluation, controls and pitches are evaluated before the note call since they themselves are transformers, and only change the environment of the note deriver. So a delay on the note is too late, the controls have already been evaluated. In a functional notation pseudo-code, if you put the delay on the note you get dyn [1] (pitch [4c] (>inst (delay (note)))). In score notation this looks like:

    dyn ->      * ->    >inst
    1           4c      delay |

The delay is too late to affect the dynamics or pitch. What we really want is to put the delay outside: dyn [1] (delay (pitch [4c] (>inst (note)))):

    dyn ->      > ->    * ->    >inst
    1           delay   4c      ""

Unfortunately, tracks don't work that way. Only control tracks can modify other tracks, and control tracks generate a single signal that has scope over the entire track below it.

However, tracks could work that way, even control tracks. Given:

    dyn ->      >inst
    1           ""
    .5          ""

You could view this as dyn [1, .5] (>inst (note <> note)) (if <> is the merge operator) where both notes get the [1, .5] signal, but it also could make sense to interpret it as dyn [1] (>inst (note)) <> dyn [.5] (>inst (note)), i.e. the dyn track is sliced up for each note. Which one is appropriate depends on the musical context, so if dyn is a decrescendo curve then the first is correct, but if it is intended to be individual dynamics for each note then the second is correct. This yields to the correct behaviour for a delay: if it's a decrescendo, the delayed note should move to the quieter part of the curve, but if it's individual dynamics, a delayed note should keep its dynamic, and not pick up the dynamic of its successor! So the notation needs to distinguish between the two.

So there are two complementary extensions:

Visually, the transformation looks like this:

    dyn ->      >inst ->        *
    1           a               4c
    .5          b               4d

[1, .5] is intended to be a decrescendo, while 4c and 4d belong to notes a and b respectively. After inversion and slicing we effectively have two blocks:

    dyn ->      * ->            >inst
    1           4c              a

    dyn ->      * ->            >inst
    .5          4d              b

Of course this transformation isn't useful as-is, since we could have written in the second form in the first place (ignoring the slicing into two blocks part). But if we only invert the generator part of the note expression and wrap the transformer part around the sliced children, we can transform this:

    dyn ->      >inst ->        *
    1           delay 1 | a     4c
    .5          delay 2 | b     4d

Into this:

    dyn ->      delay 1 | ->    * ->            >inst
    1                           4c              a

    dyn ->      delay 2 | ->    * ->            >inst
    .5                          4d              b

Here, delay 1 | isn't real tracklang syntax, but represents a transformer wrapped around its children. In functional notation:

    dyn [1, .5] ((delay 1 (pitch [4c] (>inst (a))))
              <> (delay 2 (pitch [4d] (>inst (b)))))

Which is exactly what is needed for delay to work as expected.

Note transformers

Slicing without inversion is called a note transformer, since it's basically a way for a note track to take other notes as arguments. Since the note call isn't inverting itself below the control track children, there has to be another note track below it. E.g.

    >inst ->    >
    tuplet      a
      |         b

In this case, the tuplet call (really named t, defined in Derive.Call.Prelude.Parent) slices events within its duration, so it gets a and b as arguments and is free to do with them as it wishes. This is how ornaments that transform multiple other notes are implemented.

As usual, there are a few more wrinkles:

Note that inverting and non-inverting slicing calls can be combined, for instance:

    >inst ->    > ->    *
    tuplet      ""      4c
      |         ""      4d

tuplet will slice the two events under it and receive the two null note calls as arguments. If it chooses to evaluate them, they will then invert themselves beneath the * pitch track.

old version

This is an old version of the doc, maybe it makes more sense:

Transformers operate on derivers. They can change the environment to encourage the deriver to produce a certain output. For instance, they can modify the pitch to transpose the derived music, or modify the warp to delay it.

Since the controls must be evaluated and in the environment for the notes to pick them up, by the time the note track is being evaluated it's too late to change them. So a "delay" transform on a note may delay the note itself but will not affect the controls. Sometimes this is desired, e.g. if you delay a note during a decrescendo you may not want the decrescendo to move with the note, but sometimes it isn't, e.g. if you delay a note its pitch should be delayed along with it.

So what is needed is for the transform to be placed "above" the note, e.g. (decrescendo (delay (pitch (note)))). This can be written in terms of tracks, but it's awkward, because the delay, which is logically part of the note track, must be split into two tracks, like so:

    dynamics ->     transform ->    pitch ->    note
    1               delay           4c          ""
                    |                           |
    i 0             V                           V

It's not clear what the duration of "delay" means, and the fact that it applies to the "" note two tracks below it. So there's a concept of track inversion, which is an automatic transformation from this:

    note    pitch
    t | n   4c

To this:

    note    pitch   (note)
    t       4c      n
    |               |
    V               V

If there are tracks below the note track, they are sliced horizontally in the range of the note being evaluated, and the generator part of the note is put in a new "track" below them.

Inverting calls check for the presence of subtracks, and if found, do the slicing described above and reinsert themselves below their subtracks, hence inverting the call structure.