M7. The ± and Correlated Readings
What It Is
The ± beside a figure is its standard error: an estimate, from the take itself, of how far the figure would likely move if the same stretch were heard again under the same conditions.
It is precision, not accuracy. A wrong lift angle or an uncorrected sound card clock moves the figure without widening the ±.
How It Is Worked Out
The ± comes from the short readings inside the span, in time order: the 2 s windows for amplitude and beat error, and every other 10 s rate window for the rate.
d = each reading less the one before it
s = 1.4826 × median distance of d from its median ÷ √2
ρ = lag-one correlation of the readings, after a straight line is
taken off, held between 0 and 0.99
n_eff = n × (1 − ρ) ÷ (1 + ρ)
± = 1.2533 × s ÷ √(1 − ρ) ÷ √n_eff
In words:
- The scatter,
s, comes from the differences between neighbouring readings, and a median rather than a mean, so one wild window doesn’t inflate it. The constants turn it into a standard deviation for normal noise. - The correlation,
ρ, is how much each reading resembles the next. Watch readings do: on the project’s test recordings, neighbouring 2 s amplitude windows correlate by anywhere from 0.01 to 0.90 (the Math Standards Guide, section 12). - The effective count,
n_eff, counts correlated readings as fewer independent ones. At ρ = 0.5, 30 readings are worth 10. - The ± is the standard error of a median of
n_effreadings.
Over several stretches between cuts (positions, breaks, masks), the
differences d are taken within each stretch only, never across a cut.
For amplitude and beat error, a second term is added: how far the
stretches’ own medians spread, past what their noise explains.
τ² = largest of 0 and (variance of the stretches' medians
− mean of their ±²)
term = 1.2533 × τ ÷ √k (k: the stretches of 3 readings or more)
± = √(within² + term²)
The live rate’s ± is read from the 10 s windows since the last break, up to five minutes, then scaled to the Average Over time. It needs about 30 s of unbroken beats. With fewer than 3 readings there is no ±.
Why This Way
- Neighbour differences ignore a slow, steady drift: a watch winding down by a steady 1° a minute doesn’t widen the amplitude’s ±.
- Allowing for correlation: treating correlated readings as independent gives a ± too small. On a seeded test at ρ 0, 0.5 and 0.9, the share of repeats within ±1 standard error stays between 55% and 85% (68% expected).
- The spread between stretches: a watch reading 270° in one stretch and 250° in the next leaves the take’s median uncertain by more than either stretch’s noise. Stretches at one level add nothing on average.
- One function for every figure, in the app, the command line and the MCP server, so the same take gives the same ± everywhere.
When You’d Use It
- Comparing two readings: two figures whose ± ranges overlap widely aren’t shown to differ. Positions, before and after a change, two takes.
- Choosing Average Over: a longer span gives a smaller ±.
- Deciding how many decimals matter: a rate of +4.2 ± 0.8 s/d isn’t known to the tenth.
What It Helps Find
- A noisy setup: a ± much wider than usual for the same watch and span points to the sound, the contact or the stand.
- A watch that wanders: a slow swing or cycle raises ρ and so widens the ±.
What It Can’t Tell You
- Accuracy. The lift angle and the clock are caveats, not part of the ±.
- The next take. A watch can read differently after rewinding, a knock or an hour. The ± is about this stretch only.
- Memory of another shape. The model allows for correlation from one reading to the next. A cycle or a level shift is allowed for only as far as it shows in that one value, so the ± can still be too small.
- A guarantee. It is an estimate under a model of the noise, not a bound on the error.
The project’s precision notes give the model, its sources and the tests behind it.
Previous: M6. Averaging, Windows and Cuts · Next: M8. Positions Test Figures
This page comes from the app's user guide at v1.0.1. See it on GitHub.