How much does sleep rem min have to change before it stops being noise? The threshold from 650 nights

Before you can know whether your REM sleep has genuinely shifted, you need a threshold — the minimum distance the average must travel to clear the background noise. Across 650 nights of a single body, that threshold is 11.48% when you accumulate 28 nights of data, and 8.12% when you accumulate 56. Below those percentages, the movement you see in your app is part of the natural variability of any given night — not a signal yet.

The starting noise

REM sleep does not hold steady from one night to the next. In the dataset behind these figures, the coefficient of variation is 21.27%. That metric measures how much a variable fluctuates against itself over time, expressed as a percentage of its own median. When the coefficient of variation is 21.27%, the natural night-to-night spread is equivalent to more than one fifth of the median — before any protocol or deliberate change has entered the picture.

The median REM sleep across these 650 nights is 101 minutes per night. But on any given night, with no changes to routine, the reading can fall well above or below that center. The spread around 101 minutes is larger than that of most other variables tracked in the same body. That does not make the data useless — it changes how the data should be read.

That variability is not an equipment flaw. REM sleep responds to dozens of factors a wearable does not control: ambient light, room temperature, alcohol, acute stress, hormonal cycles, prior-day activity. The result is that tonight's number carries no reliable prediction for tomorrow. Reading the app night to night is like trying to detect a climate trend by checking whether it rains today.

The autocorrelation (rho) of REM sleep across these 650 nights is 0.03 — effectively zero. Yesterday's value gives you no information about today's. Each night is drawn from scratch, with no inertia and no short-term memory in the system. For a difference between two averages to be real rather than another lucky draw, the threshold must be high enough to absorb what noise can produce on its own.

The real-world coefficient of variation (RCV) is 58.96%. That figure captures variability at every scale — night to night, week to week, month to month. Even when averaged across a full week, REM sleep remains the noisiest sleep variable among the 31 metrics in the dataset. That context matters before reading any single number in the app as a finding.

The window and its threshold

Averaging more nights cancels more noise. The table below comes directly from the source row ruido.medido.json#sleep_rem_min:

Measurement window	Signal threshold
28 nights	11.48%
56 nights	8.12%
84 nights	6.63%
168 nights	4.69%

Source: ruido.medido.json#sleep_rem_min · 650 nights · single body

These thresholds are not conventions drawn from the published literature — they are the result of calculating the actual variability of this subject and fixing the point at which a shift in the average exceeds what noise alone can produce. The source is the field dataset: 18483 measurements across 650 nights recorded in a single body with a single device.

Each longer window buys precision at the cost of time. If waiting is not an option, the 28-night window is the entry point: it requires a shift of 11.48% before signal can be declared. If the commitment extends longer, 168 nights bring that threshold down to 4.69% — smaller signals become readable when the average is longer and the noise has been smoothed out.

Applying the threshold to your own data

The reference median in this dataset is 101 minutes of REM per night. That belongs to a single subject, measured with a single device, during a specific period of life. Yours may differ, and the thresholds always apply against your own historical baseline — not against this dataset's.

The steps are the same for any body: take the average of your last 28 nights of REM sleep. Compare it to your average for the 28 nights before that. If the difference clears 11.48% of your historical median, the movement is above the noise floor. If it falls short, natural variability can explain it on its own — and that is an answer, not a failure.

To see the signal sooner without waiting for more nights, the only lever is extending the window. At 56 nights the threshold drops to 8.12%; at 84 nights, to 6.63%; at 168 nights, to 4.69%. A longer window means a cleaner signal, at the cost of more nights of recording before the result is legible.

What does not change: the need for your own history. Without your baseline, this dataset's threshold says nothing about your body.

The limit

This threshold describes the statistical dispersion of REM sleep in a single subject, recorded with a single device, over a specific period of time. It is not a universal constant. What this data does not say:

It does not compare bodies. The 11.48% at 28 nights is the signal threshold for this dataset. Your natural variability may be higher or lower. Only your own record can establish your threshold.

It does not identify causes. A threshold is a reading rule, not an explanation. Knowing your average cleared 11.48% confirms that something shifted — it does not name what. That requires a protocol with a defined intervention, a before and an after, and the threshold as the reading criterion.

There is no control group. These 650 nights are a single-subject record, not a controlled trial. There is no parallel group without an intervention. The variability is measured; its cause remains open.

The device is not the phenomenon. The REM minutes a wearable reports are an estimate built from accelerometry and heart-rate data. Physiological REM is defined by EEG. The instrument tracks a proxy, not the direct state.

The question this piece leaves open: which protocol shifts the 28-night average above 11.48% in a sustained way, and how long a record is needed before that confirms. The dataset holds a related file; a sealed answer does not exist yet.

---

Mental Work — all protocols · related file: magnesium-for-sleep · open science — the dataset