Allan variance is a measure of the stability of a signal as a function of averaging time. It is defined as half the expected value of the squared difference between two consecutive, equally long averages of the signal. Developed for characterizing frequency fluctuations in clocks and oscillators, it is also used to analyze drift and noise in precision sensors and other time series. Unlike an ordinary variance calculated about a sample-wide mean, Allan variance can remain well defined for important noise processes whose low-frequency fluctuations make conventional variance unsuitable. Its square root is called Allan deviation.
Definition
For a continuous signal , define its average over an interval of duration by
The Allan variance at averaging time is
This definition assumes that the expected squared difference is finite and does not depend on the interval index . It does not require itself to be a stationary process: some nonstationary signals have differences of adjacent averages with time-independent statistics.
The Allan deviation is
Allan variance has the square of the signal's units; Allan deviation has the same units as the signal. For fractional frequency fluctuations, both are dimensionless.
The dependence on is essential. A signal may be stable over seconds but unstable over hours, so a single Allan deviation value without its averaging time is incomplete.
Frequency and time-error representations
For an oscillator with nominal frequency , fractional frequency deviation is
If the oscillator phase is written as
its time error relative to the nominal oscillator can be defined by
Under this convention,
Thus, the mean fractional frequency over an interval is
Substituting into the definition gives
Allan variance therefore measures a second finite difference of time error, or equivalently a first difference of averaged frequency. The derivative relationship between time error and frequency explains why descriptions of oscillator noise must specify whether they refer to phase, time error, or frequency.
A constant frequency offset contributes nothing to Allan variance. It changes the oscillator's accuracy but not this measure of its stability. By contrast, a linear frequency drift contributes
Why it differs from conventional variance
Ordinary variance measures fluctuations around a mean. Allan variance measures changes between successive interval averages.
If the adjacent averages have equal variance and their difference has zero mean, then
Consequently, Allan variance equals the variance of an interval average when successive averages are uncorrelated. Otherwise, the covariance term matters. It is not generally interchangeable with ordinary variance or the standard error of a mean.
The differencing operation suppresses very slow fluctuations. This makes Allan variance useful for processes such as flicker frequency noise and random-walk frequency noise, for which conventional estimates of variance may depend strongly on the observation duration or lack a finite stationary interpretation.
This advantage has limits. Sufficiently strong low-frequency noise can also make Allan variance diverge, while deterministic drift can dominate its value at long averaging times.
Estimation from sampled data
Suppose contiguous measurements represent average frequency over sampling intervals of duration . For averaging time , where is a positive integer, define
An overlapping Allan variance estimator is
Each term compares two adjacent, nonoverlapping blocks of length , but consecutive terms share data. “Overlapping” therefore describes the reuse of observations across comparisons, not overlap between the two blocks within one comparison.
A nonoverlapping estimator instead partitions the record into disjoint blocks and computes
Overlapping estimation generally improves statistical precision by using more of the available record. However, its terms are correlated, so their count is not a count of independent observations.
The interpretation of the input data also matters. Instantaneous samples, interval averages, and outputs from instruments with internal weighting are not automatically equivalent. Measurement bandwidth, filtering, and gaps between measurements can change the estimated stability.
Spectral interpretation and noise types
For a process with one-sided fractional-frequency power spectral density , and whenever the integral converges,
This expression follows from the averaging-and-differencing filter and its Fourier transform. It shows that Allan variance measures noise through a frequency-dependent weighting function, rather than simply integrating all noise power equally.
A common model represents oscillator noise as a sum of power-law components,
For a single ideal power law extending over all positive frequencies, the Allan-variance integral converges for
Within this range,
Important oscillator-noise regimes are:
| Noise type | Fractional-frequency spectrum | Typical Allan-deviation dependence |
|---|---|---|
| White phase modulation | Approximately , with bandwidth dependence | |
| Flicker phase modulation | Approximately , with a logarithmic bandwidth-related correction | |
| White frequency modulation | ||
| Flicker frequency modulation | Constant | |
| Random-walk frequency modulation | ||
| Linear deterministic frequency drift | Not a stationary noise spectrum |
The phase-modulation cases require a finite high-frequency cutoff; they cannot be treated as ideal power laws extending to infinite frequency. Their similar Allan-deviation slopes also make them difficult to distinguish using ordinary Allan variance alone.
For white frequency noise with one-sided spectral density ,
For flicker frequency noise with ,
The latter produces a horizontal region on an Allan-deviation plot. Such a region is consistent with flicker frequency noise, but its appearance alone does not uniquely identify the underlying mechanism.
Interpretation and statistical uncertainty
Allan deviation is usually plotted against averaging time on logarithmic axes. Different slopes can reveal changes in the dominant noise process. For example, a falling curve at short averaging times followed by a rising curve at long times can indicate that averaging initially reduces fast noise but eventually exposes slow fluctuations or drift.
Several qualifications are important:
- Long averaging times have limited statistical support. As approaches a substantial fraction of the record duration, few effectively independent comparisons remain.
- Overlapping estimates remain correlated. More summands improve data use but do not provide an equivalent increase in degrees of freedom.
- Noise classification is not unique. Mixtures of noise, periodic disturbances, frequency steps, and environmental changes can produce misleading slopes.
- Drift removal changes the result. Subtracting a fitted trend separates one modelled component from the residual noise; it does not describe the unprocessed signal.
- Gaps and instrument weighting matter. Dead time and undocumented internal averaging can change the statistic being measured.
Confidence intervals are often constructed using a chi-squared distribution approximation with an effective number of degrees of freedom. That number depends on the noise model, averaging factor, estimator, and record length. The raw number of squared differences generally cannot be substituted for it.
Related measures
Several related statistics address particular limitations:
Modified Allan variance introduces additional averaging. It can distinguish white from flicker phase modulation more effectively than ordinary Allan variance, provided the sampling and bandwidth conditions support that distinction.
Time deviation expresses a related stability measure in units of time. With modified Allan deviation ,
Hadamard variance uses second differences of averaged frequency. It rejects linear frequency drift and remains finite for some low-frequency noise processes for which Allan variance does not.
Total variance and related estimators use extensions of the available record to improve statistical confidence, particularly at long averaging times, subject to their underlying assumptions.
These measures are not interchangeable: their filtering properties and responses to drift differ.
History and applications
David W. Allan introduced the statistic in 1966 in the study of atomic frequency standards. It became a central tool in time-and-frequency metrology because oscillator fluctuations often include low-frequency noise that conventional variance does not adequately characterize.
Its principal applications include comparing clocks and oscillators, evaluating frequency-transfer links, and characterizing measurement-system stability. It is also used with inertial sensors, optical measurements, and spectroscopic instruments to examine how fluctuations change with averaging time.
Outside frequency metrology, the same mathematical definition can be applied directly to a measured quantity. The resulting Allan deviation then retains that quantity's units. For example, analysis of a sensor's rate output yields a stability curve in rate units, whereas analysis of integrated sensor output requires a different interpretation.
Allan variance measures variability, not absolute correctness. A constant calibration error is invisible to it, and a highly stable instrument can still be inaccurate. Assessing overall performance therefore also requires calibration, reference comparisons, and an appropriate account of measurement uncertainty.