What can adjacent frequency averages reveal that an ordinary standard deviation or a clock offset cannot?
Allan deviation measures how neighboring frequency averages differ over a chosen averaging time. A constant frequency offset cancels in the calculation.
Source research checked . Teaching calculations are ideal models.
A clock can run at a very stable but incorrect frequency, or have a correct long-term average while fluctuating substantially over short intervals. Stability and accuracy describe different properties. Allan deviation compares neighboring frequency averages over a chosen averaging time, making it useful for studying time-scale-dependent fluctuations. It does not directly report the offset from an ideal reference. A complete clock comparison should therefore show the stability statistic together with the reference, systematic corrections and uncertainty that establish accuracy. One curve cannot replace the other evidence.
For equally spaced nonoverlapping fractional-frequency averages y_k of duration τ, the basic Allan variance estimate is one half of the mean squared difference between adjacent averages. Taking its square root gives Allan deviation. The differencing removes a constant frequency offset, so adding the same bias to every y_k leaves the statistic unchanged. That feature is useful for examining fluctuations, but it also makes clear why a small Allan deviation cannot certify correct frequency. The time interval τ is part of the quantity, not an optional label on the plot.
Changing τ changes which fluctuations are emphasized. For an ideal white-frequency-noise model, averaging can produce a familiar inverse-square-root trend. Other noise processes can flatten or raise the curve. Deterministic drift can also make long-time differences grow. It is unsafe to extrapolate a short-time slope indefinitely or assume all points on a log-log plot represent independent evidence. Longer averaging intervals leave fewer effective comparisons in a finite record, and overlapping estimators share data. Their confidence depends on record length, noise type and estimator choice.
Data preparation can change the conclusion. Removing an estimated drift may be appropriate for a stated residual-stability question, but the original drift remains relevant to an uncorrected clock. Gaps, outliers and instrument dead time also affect interpretation. Any preprocessing should be described and, where useful, shown alongside the unprocessed result. A clock comparison should specify whether the data are time deviations, phase measurements or fractional-frequency averages, because converting between them requires conventions and can amplify or suppress different noise components.
A small synthetic record illustrates both the statistic and its limits. An alternating sequence can have a nonzero one-step Allan deviation yet average to identical values in pairs. Its two-step estimate can then be zero, even though the original sequence plainly fluctuates. This does not prove excellent long-term clock behavior. It reveals a particular periodic pattern and an extremely limited sample. A credible instrument characterization needs enough independent information across the claimed time scales and should look for periodic, environmental or servo-related structure that a single summary can hide.
Dead time and irregular sampling require special care because they change which frequency fluctuations are observed. A convenient sequence index is not automatically a uniform time axis. Before interpreting a slope as a named noise process, retain the acquisition times and ensure that the chosen Allan-family statistic matches the sampling procedure.
An Allan-deviation curve gains meaning from the data-acquisition and noise model behind it. A stable offset disappears from adjacent differences, while fluctuations on selected time scales remain. Short records provide few effectively independent long-time comparisons, so the smoothness of a plotted curve does not guarantee a precise estimate at its longest averaging times. Reporting the record length, sampling gaps, drift treatment and uncertainty prevents the display from suggesting more evidence than was collected. Stability can guide averaging and control-loop choices. Absolute agreement with a frequency reference and the systematic correction budget remain separate questions, even when the stability curve reaches a very small value.
Follow the mechanism
- Convert the measured record into a declared phase or fractional-frequency representation with known sampling and gaps.
- Form adjacent averages at a specified τ and calculate the chosen Allan-family estimator.
- Evaluate record length, noise assumptions, overlap and preprocessing when attaching uncertainty or fitting a slope.
- Combine stability with reference and systematic information for the actual holdover, comparison or timing task.
Nonoverlapping Allan variance of four synthetic fractional-frequency averages
Inputs: y: [1e-12, 2e-12, 1e-12, 2e-12]; tau s: 1; constant bias: 5e-11
Allan_variance=sum((y[k+1]-y[k])^2)/(2*(M-1))
| Calculated quantity | Synthetic result |
|---|---|
| Allan variance | 5e-25 |
| Allan deviation | 7.0710678e-13 |
| two second single difference estimate | 0 |
Synthetic four-sample record: y=[1,2,1,2]×10⁻¹², with each value an average over τ=1 s. The three adjacent differences have magnitude 10⁻¹², so Allan variance is 0.5×10⁻²⁴ and Allan deviation is 7.0711×10⁻¹³. Add a constant bias of 5×10⁻¹¹ to all four values; the result is unchanged. If the original record is averaged into two nonoverlapping 2-s blocks, both block means are 1.5×10⁻¹² and the single available difference is zero. That zero is a property of this tiny alternating sequence, not a reliable long-term stability estimate or an accuracy certificate.
A distinction to keep
The tempting conclusion: A tiny Allan deviation proves the clock tells the correct time.
What follows instead: The statistic can ignore a constant frequency bias and does not by itself establish reference accuracy or accumulated time error.
| Quantity or decision | What it means |
|---|---|
| Stability | How frequency or time fluctuates over specified intervals. |
| Accuracy or systematic uncertainty | How the output relates to the intended reference after justified corrections. |
| Holdover behavior | How error accumulates during a period without an external correction, depending on both offset and noise. |
Does a constant frequency bias affect Allan deviation?
It cancels from adjacent differences. The clock can therefore have small Allan deviation while accumulating substantial time error from an uncorrected offset.
Why does averaging time belong in the result?
The statistic compares fluctuations on that time scale. Different noise processes behave differently as τ changes.
Are overlapping Allan points independent?
They share samples and can have substantial dependence. Confidence intervals and fits should account for the estimator and noise model.
Should drift be removed?
Only for a clearly stated analysis purpose. Report that preprocessing and retain the original drift when it matters to the clock application.
Before interpreting a result
- Sampling, data type and estimator convention.
- Record duration, effective comparisons and noise-model-aware intervals.
- Preprocessing plus independent reference and systematic-error information.
Allan deviation describes fluctuations on a specified time scale. Accuracy and holdover require the reference, offsets and physical uncertainty that differencing alone cannot reveal.
Sources and model boundaries
- Handbook of Frequency Stability Analysis ↗
Frequency stability analysis uses defined sampling, Allan-family statistics and noise-model-aware reporting. The four-sample sequence teaches a statistic and is not sufficient to characterize a real oscillator.