Why does standard statistical variance fail for oscillators?

Standard variance requires the 'mean' (average) of the data set to be stable. If you measure a clock that has long-term drift (like a quartz crystal slowly aging), the average frequency keeps changing. Because the mean is moving, standard statistical variance math breaks down and approaches infinity the longer you measure. It becomes completely useless for characterizing the stability of the clock.

How does Allan Variance fix the math?

Instead of comparing every single data point to a global average, Allan Variance simply compares a block of time to the *very next* block of time. It calculates the difference between consecutive measurements (the first difference). Because it only looks at adjacent time periods, it effectively ignores the slow, long-term drift that ruins standard variance, allowing engineers to cleanly measure the clock's stability.

What is the difference between Allan Variance and Allan Deviation?

They are mathematically identical concepts, just like Variance and Standard Deviation in normal statistics. Allan Variance is the squared value. Allan Deviation (ADEV) is simply the square root of the Allan Variance. Engineers almost exclusively use Allan Deviation in datasheets because its units directly represent 'Fractional Frequency' (e.g., parts per million or parts per billion), making it intuitive to read.

System Performance

Allan Variance

An engineer needs to buy a highly stable oscillator for a new GPS satellite. They want to know how much the clock's frequency wanders over time. If they measure the frequency thousands of times and calculate the standard statistical variance, the result is useless—the math explodes because the crystal's frequency slowly, continuously drifts as it ages. Standard statistics cannot handle a moving target. Instead, the engineer looks at the datasheet's Allan Variance plot. This graph shows the stability of the clock based on the "averaging time" (τ). For a 1-second interval, the Allan Deviation might be 1x10^-11, meaning the clock only wanders by a tiny fraction of a Hertz. But if the interval is expanded to 10,000 seconds, the graph shows the deviation rising rapidly due to thermal drift and aging. By using Allan Variance, the engineer can exactly characterize the stability of the clock across both short-term jitter and long-term drift.

Category: System Performance

Domain: Time Domain (unlike Phase Noise)

Primary Use: Characterizing atomic clocks and precision oscillators

The Allan Deviation Curve (Log-Log Plot)

Averaging Time (τ)	Dominant Noise Source	Curve Trajectory
Short (e.g., 0.01 seconds)	White Phase Noise / White Frequency Noise	Slope goes DOWN as time increases (Averaging helps)
Medium (e.g., 100 seconds)	Flicker Noise (1/f) / The "Floor"	Flat / Horizontal (Maximum stability point)
Long (e.g., 10,000 seconds)	Random Walk / Temperature Drift / Aging	Slope goes UP as time increases (Drift dominates)

Two-Sample Variance Math:
Instead of comparing data to a grand mean, Allan Variance compares adjacent measurements. If y_i is the average fractional frequency over a time period τ, and y_i+1 is the average over the *next* time period τ:
σ_y²(τ) = 1 / [2(N-1)] · Σ (y_i+1 - y_i)²
Because it only subtracts the *next* value from the *current* value, a slow, steady drift (where y is slowly rising over years) is mathematically canceled out, preventing the variance from exploding to infinity.

Fractional Frequency:
Allan Deviation does not output raw Hertz. It outputs fractional frequency (Δf / f). An Allan Deviation of 1x10^-9 for a 10 MHz oscillator means the clock is stable to within 0.01 Hz over that specific time interval.

Common Questions

Frequently Asked Questions

How is this different from Phase Noise?

They measure the exact same instability, just in different domains. Phase noise measures stability in the *Frequency Domain* (how the power spreads out around the carrier frequency). Allan Variance measures stability in the *Time Domain* (how the time between consecutive ticks wanders). You can actually mathematically convert an Allan Variance plot directly into a Phase Noise plot (though it is computationally heavy).

Why does the Allan curve go down and then back up?

It shows the difference between random noise and physical drift. At very short intervals, random white noise dominates. If you average the data for 1 second instead of 0.1 seconds, the random noise cancels itself out, and the stability improves (the line goes down). However, if you keep averaging out to 1 hour, the physical temperature of the room will have changed. The crystal expanded, and the frequency drifted. The drift now overwhelms the averaging, and the stability gets worse (the line goes up). The lowest point on the curve is the ultimate stability floor of the clock.

What is Modified Allan Variance (MVAR)?

Standard Allan Variance has a flaw: it cannot distinguish between White Phase Noise and Flicker Phase Noise (they both look the same on the graph). Engineers created Modified Allan Variance (MVAR), which adds an extra layer of averaging to the math. This slightly alters the slope of the graph, allowing engineers to visually separate the different types of underlying physics causing the noise.

Related Terms

← All-Digital PLL AM-AM Conversion →

System Performance

Allan Deviation Drift Calculator

Input your oscillator's base frequency and its published Allan Deviation (ADEV) for a given averaging time (τ). Instantly calculate the absolute time error (in nanoseconds) your system will accumulate over that interval, critical for GPS and telecom synchronization.

Calculate Accumulated Time Error