Syncopation is a number about the metre

Syncopation is usually defined by pointing: a note where the beat is not, an accent in the wrong place, an emphasis displaced. All of those are descriptions of an effect, and none of them is a quantity.

Longuet-Higgins and Lee gave it one in 1984, and the definition is short enough to state in a sentence. Every position in a bar has a metrical weight — zero on the downbeat, one lower at each level down the subdivision tree. A note on a weak position followed by a rest on a stronger one is a syncopation, and it costs the difference between the two weights.

That definition does one thing this site cares about above all: it turns an adjective into a number, and the number is checkable against cases where everybody already agrees. It also inherits a dependency that the previous rung spent its whole length on: a metrical weight is defined relative to a bar line, and the bar line is exactly the quantity the induction rules cannot find.

It agrees with the ear on the easy cases #

A pattern with an onset on every beat and nothing between scores zero. A pattern with an onset on every offbeat and nothing on any beat scores 4 in a bar of eight.

That is the sanity check passed, and it is worth passing before anything more interesting is asked of the measure. A measure of syncopation that scored the hymn tune above zero, or the offbeat pattern at zero, would be measuring something else.

And then it stops being a property of the rhythm #

Read the son clave from step one and it scores 4. Read the same five onsets from step three — the same intervals, in the same order, with nothing moved — and it scores 2. From step two, 7. From step four, 8.

Four to one, between the most and the least syncopated reading of one unchanged pattern. That is not a rounding difference or a boundary effect; it is the whole dynamic range of the measure.

On a circle, where a rhythm has no beginning, those four rows are one object seen from four places. The pattern’s own structure is unchanged by a rotation; what changes is the grid it is being held against, which is not part of the pattern at all.

So syncopation is not a property of a rhythm. It is a property of a rhythm together with a metre, and every published syncopation figure for a named pattern is quietly carrying a metre that somebody chose — usually the transcriber, usually by writing the bar line where the tradition puts it.

Which is fine, until the metre is inferred #

That would be a footnote if the metre always arrived from outside. Sometimes it does: a notated score has bar lines, a dance has feet, a repertoire has a convention.

But this site also has a machine that infers metres from patterns, and combining the two produces a problem that is easy to miss. Run the preference rules over four patterns, and compare the phase they choose against the phase that scores lowest for syncopation:

Three of four agree exactly. Four patterns is not a sample, though, so the same comparison run over every sixteen-step pattern with three to eight onsets — 16,368 of them:

They agree well above chance and much less often than three times in four. A blind guess between four phases would agree 25 per cent of the time, so the two computations are measuring related things; but they disagree on more than four patterns in ten, and they disagree most on the sparse patterns — 48.6 per cent at three onsets against 60.5 at eight — which are the patterns syncopation is an interesting quantity for.

That figure hides something, and separating it out recovers the essay’s own four-case result exactly. The preference rules tie on 38 per cent of these patterns, and on a tie “the rules choose” means the list order chose. Setting those aside and counting only the patterns where the rules pick a phase outright, the two agree 74.9 per cent of the time — three in four, which is what the four hand-picked cases gave.

So both readings are true and they answer different questions. When the rules commit, they agree with the syncopation minimum three times in four. Across all patterns, they agree well under two times in three, because on nearly two patterns in five the rules do not commit at all. The first is a statement about two measures; the second is a statement about one measure and a tie-break.

The rules and the syncopation measure are not the same computation — one counts onsets on strong positions and the other prices notes against following rests — and on the patterns where the first has an opinion they mostly pick the same reading, because both are answering a version of where does this pattern agree with a grid.

The consequence is a trap. Measuring the syncopation of a pattern against a metre the same pattern was used to infer reads close to its own floor, because induction is selecting for agreement and syncopation is measuring disagreement. The number comes out low, and it comes out low by construction rather than because the music is not syncopated.

The census puts a size on the trap, and it is bigger than the four cases suggested. If the two always agreed, the bias would be total and the measured syncopation would be the minimum over phases every time. They agree three quarters of the time when the rules commit, so the trap is sprung on three patterns in four — often enough to be systematic, and not often enough for the resulting numbers to look obviously wrong. That combination is what makes it a trap rather than a visible defect: a corpus measured this way would come out with a syncopation distribution biased low by a construction that is right most of the time.

And the two-in-five tie rate is a second, quieter version of the same problem. On a tie the reported metre is the transcriber’s, so the syncopation measured against it is measured against the tradition’s own bar line — which is the right answer for the wrong reason, and indistinguishable in the output from the inferred case.

Anything that wants syncopation as an independent quantity has to take its metre from somewhere the pattern did not supply.

There is one cheap repair available and it is worth naming because it costs nothing. A measure computed at the phase the rules choose is biased; a measure computed at every phase and reported as a range is not, and the range is the honest object — the son clave is a pattern that scores between 2 and 8, and which end of that a performance is at is a fact about the metre it is being played in rather than about the clave. Reporting the range costs one loop and it makes the dependency visible instead of hiding it inside a chosen number.

The available somewheres are the ones the phase rung listed and none of them is in the onset list: the bass, the rate at which the chords change, the notation, and whatever metre was already running. That is not a weakness of the syncopation measure — it is a statement of what kind of quantity it is.

What the measure counts, and what it does not #

The definition prices a note against the rest that follows it, which is a specific and slightly surprising choice.

A note on a weak position followed immediately by another note costs nothing at all, however weak the position. What syncopates is the silence over the strong position — the beat that is skipped rather than the note that is early.

That gives the measure a property worth noticing: a denser pattern is not more syncopated. Filling in the offbeats between the syncopations removes them, because a note can no longer be followed by a rest on a strong position if there is a note there.

Which is, as it happens, what a rhythm section does. The bass syncopates and the hi-hat fills every subdivision, and the second is not adding syncopation to the first — on this measure it is masking it.

Density is not syncopation, and this is testable #

The claim that filling in the offbeats reduces the measured syncopation is a strong one, and it is the sort of claim that ought to be uncomfortable.

It is at least consistent with the repertoire. A drum part that plays every subdivision — a straight sixteenth hi-hat, a shaker, a ride pattern — is universally described as supporting the groove rather than complicating it, and it does so by occupying exactly the positions that would otherwise be the rests a syncopation is priced against.

It is also consistent with what happens when the density comes off. Strip a busy pattern down to its onsets on weak positions and the result is more disorienting, not less, which is the opposite of what a count-the-offbeats measure would predict.

Where it is uncomfortable is in the tied note, and that is a real limitation rather than a subtlety; it is in what the picture cannot show below.

Syncopation and microtiming are different quantities #

There is a second thing that displaces a note from a beat and it is not this one.

Swing displaces every offbeat by a fixed proportion and groove displaces particular beats by tens of milliseconds. Both are displacement, neither is syncopation, and the measure here cannot see either of them: it takes a list of steps, and a note that is 30 milliseconds late is on the same step it was.

The distinction is worth being firm about because the words are used loosely. Syncopation is a fact about which positions of a grid carry notes. Microtiming is a fact about where the notes are relative to the grid. A pattern can have a great deal of one and none of the other, and the two are measured by different machinery with different units — counts of metrical weight against milliseconds.

The tree is a choice #

Every weight in every figure here comes from a subdivision tree, and the tree is supplied rather than derived.

A bar of eight can be [2, 2, 2] — two halves, each two quarters, each two eighths — or [2, 4], or [4, 2]. Each gives a different set of weights and therefore a different syncopation count for the same pattern.

So the measure has two free inputs — the phase and the tree — and reporting a syncopation figure without both of them is reporting a number that cannot be reproduced.

A syncopation count is never a statement about a pattern alone. It is a statement about a pattern and a tree, and the trees are not interchangeable: 4/4 and 6/8 divide eight and six evenly, while 5/8 as 3+2 has an unequal branch that the weight scheme handles by fiat rather than by definition. Every number in this essay inherits whichever tree it was computed against, which is why none of them can be quoted without one — a syncopation of 2 is a syncopation of 2 in four, and the same pattern read in two scores something else entirely.

The twelve-step case, where the tree is genuinely disputed #

A bar of twelve can be [2, 3, 2] — a bar of two, each a triple, each divided — or [3, 2, 2], or [2, 2, 3]. Those are 6/8, 3/4 and something else, and for a twelve-step timeline the tradition does not settle which.

For a music organised around simultaneous groupings, a measure that requires a single tree is asking a question the music declines to answer. The number can be computed three ways and the honest report is all three.

What the measure is good for #

None of the above is a reason to discard it, and the site’s habit is to say what a model is for rather than to leave it in pieces.

Its first use is comparative, within one metre. Given a bar line — from the notation, from the dance, from anywhere — the measure ranks patterns, and the ranking is stable and matches judgement. That is a real instrument and it does not care that the absolute number is meaningless.

Its second use is exactly the one this rung is warning about, run deliberately. Compute the syncopation of a pattern under every rotation, and the shape of that profile is a property of the pattern after all — how much the pattern’s character depends on where the bar line goes. The clave’s profile runs 4, 7, 2, 8; a pattern of onsets on every beat runs 0, 4, 0, 4; a pattern of onsets on every step runs 0, 0, 0, 0, which is the flat profile a rhythm has when its bar line genuinely does not matter; and the tresillo’s runs 2, 2, 1, 3, which is nearly flat and is one reason it travels so well.

That is a derived quantity rather than the published one, it is cheap to compute, and it is the honest way to get a number that belongs to the rhythm alone.

Whose music, and whose bar #

The definition was written for tonal Western music with notated bar lines, and it was validated against judgements from listeners raised on it. That is not a criticism — it is what the paper says it is doing — and it is the reason for two of the constraints above.

The tree comes from the time signature, which exists on the page. The phase comes from the bar line, which also exists on the page. Given both, the measure is well-defined, reproducible, and matches intuition on the cases where intuition is firm.

Take the page away, and both inputs have to be inferred from the same pattern being measured. That is the situation in every repertoire this site’s rhythm figures actually draw from — Cuban, Balkan, West African — and it is the situation in which the measure is least reliable and most often quoted.

What the picture cannot show #

It cannot show a syncopation that is not a rest. The definition needs a silence on the strong position. A strong position occupied by a quieter note, or by a note in an inner voice, is not a syncopation on this measure and is one to most listeners.

It cannot show duration. Every onset here is a point. A note that begins on a weak position and is held through the following strong one is the classic tied syncopation of Western notation, and the measure treats it identically to a short note followed by silence — which is right for a drum and wrong for a voice.

It cannot show accent. A pattern played with the offbeats loud and the downbeats soft is more syncopated than the same pattern played flat, and no figure here has a dynamic in it.

It cannot show that syncopation is enjoyable. The measure is monotone: more is more. What listeners actually report is an inverted U — some syncopation is pleasurable and a great deal is disorienting — and the peak sits at a low count. A number that runs to 8 on the clave is not a scale of enjoyment and reading it as one is the standard misuse.

It cannot show the level above the bar. Every count here is inside one bar. A phrase that syncopates against the four-bar group rather than against the beat is doing the same thing one level up, and extending the weight tree upward is straightforward and has been done by nobody in the figures on this page.

And the weights are integers with no justification beyond the tree. Zero, −1, −2, −3 down the levels is a stipulation. Nothing measured says the drop from the downbeat to the half-bar is the same size as the drop from the half-bar to the beat, and every figure here assumes it is.

The ladder from here #

Two rungs of this ladder have now taken the metre as given and asked what can be computed against it. The next asks what happens when the evidence for the metre is taken away — the beat itself silenced, for two whole bars — and the answer is that the rule set does something no listener does.

That failure has a shape, and the shape says exactly what kind of model would be needed instead.

After that, the case this measure cannot represent at all: a bar whose beats are not the same length, where the subdivision tree has an unequal branch and the whole weight scheme has to be stipulated rather than derived.