M12. Isochronism and Rate against Amplitude
What It Is
A balance is isochronous when its period doesn’t depend on how far it swings. A real one isn’t quite: the hairspring, the escapement’s impulse and the balance’s poise all make the period move a little with the amplitude. As the mainspring lets down, the amplitude falls, so a watch’s rate moves through the day with it.
The Isochronism pane (chapter 19) measures that: the rate at each amplitude, the slope of rate against amplitude, and where the rate changes least. The Positions Test’s full wind against 24 hours (chapter 20, M8) measures it another way: the change in each position between two states of wind.
How It Is Worked Out
Pairs. Each 10 s rate reading is paired with an amplitude for the same span. A rate reading is the slope of a line fitted through its beats, and such a fit weighs the beats near the middle of the span most: 6u(1 − u) at the fraction u of the way through. So the paired amplitude is the mean of the reading’s 2 s windows, each weighted the same way:
weight of a window from u₀ to u₁ = F(u₁) − F(u₀), F(u) = 3u² − 2u³
five windows: 0.104, 0.248, 0.296, 0.248, 0.104
A short event inside a reading then moves the amplitude it is paired with as far as it moves the rate.
Steps and curve. The pairs are sorted into 2° steps of amplitude between the 2nd and 98th percentile, each step with at least 10 pairs, and at least 3 steps. Each step gives its median rate and its quartiles. A quadratic is fitted through the medians, each weighted by its pairs:
rate ≈ c₀ + c₁ (A − Ā) + c₂ (A − Ā)²
turn at A = Ā − c₁ ÷ (2 c₂) shown when inside the middle 80%
The slope. A straight line through every pair, by least squares, after leaving out any rate or amplitude reading over 6σ from its 5-minute running median (none when more than 3% would go), as the Steadiness views do. But the amplitude readings are noisy, and noise in the quantity along the bottom always pulls a fitted slope towards 0. The slope is corrected for it (errors in variables, Fuller 1987):
σ_w² = (1.4826 × MAD of neighbouring window differences ÷ √2)²
noise = σ_w² × Σ w² (about 0.232 σ_w² for five windows)
λ = 1 − noise ÷ variance of the paired amplitudes
slope = least-squares slope ÷ λ
λ is the reliability: the share of the amplitudes’ spread that is real. The slope is withheld when λ is under 0.5, and the pane says the amplitude readings’ own noise is half or more of their variance. The ± allows for correlated readings, as in M7.
The walk. One point per 5 minutes: the median rate of its 10 s readings and the median amplitude of its 2 s windows, each read on its own. A 5 minutes with under half of either is left out. Past 250 points (a take over about 20.8 hours) neighbouring 5 minutes are merged, 10, 15 … minutes a point: each series the mean of the joined medians, weighted by their counts.
Why This Way
- Medians in steps: a few wild readings don’t move a step’s rate.
- The slope through every reading: none is smoothed away, so a short swing of amplitude still counts.
- The noise correction: without it a noisy amplitude would make any watch look more isochronous than it is.
- Weights matching the rate’s fit: a median or a plain mean of the windows moves less than the rate does in a short event, and biases the slope by about 10% on a rub of 8 s in 10 (the project’s series notes).
- The walk: the curve throws time away; the walk keeps it, so a step or a return shows.
When You’d Use This
- On a run-down: a watch fully wound and left to run for hours covers the amplitudes it meets in a day.
- After adjusting the hairspring, the balance or the escapement: compare the slope and the turn with a take from before.
- Judging a positions test: a rate difference between positions comes partly from their different amplitudes.
What It Helps Find
- How much the rate moves through a day’s wind: the slope times the fall in amplitude.
- The amplitude where the watch is steadiest: the turn.
- A change in the watch between two takes, with the same slope read the same way.
What It Can’t Tell You
- Cause. The rate and the amplitude went together; temperature, position or anything else that changed over the take moves with them.
- The shape beyond the range read. The curve is fitted only over the amplitudes the take covered.
- A guarantee. The noise correction assumes the window noise is independent from window to window.
Previous: M11. Time and Frequency Maps · Contents
This page comes from the app's user guide at v1.0.1. See it on GitHub.