Allan and Modified Allan Variance

• Time-domain links between Allan and Modified Allan variances are statistical descriptors that quantify frequency stability and noise performance in oscillators and timing systems.
• They employ finite-difference operators and kernel-weighted integral transforms to bridge time-domain measurements with power spectral density noise models.
• Practical applications include fiber-optic frequency transfer and clock stability analysis, where accurate variance estimation ensures robust performance evaluation.

Time-domain links between the Allan variance and Modified Allan variance serve as the basis for the quantitative assessment of frequency stability and noise characterization in time and frequency metrology, frequency transfer links, oscillators, and electronic timing systems. The Allan variance (AVAR) and Modified Allan variance (MVAR) encapsulate the temporal evolution of frequency fluctuations, enabling a rigorous connection between time-domain statistical descriptors and frequency-domain noise models through explicit integral transforms and kernel structures. These variances are foundational for both empirical data evaluation and analytic conversion between the PSD and integration-time stabilized metrics, especially in the context of practical fiber links, wireless systems, and high-performance oscillators (Marchi et al., 2023, Rubiola et al., 2022, Georgakaki et al., 2012, Calosso et al., 2015, Calosso et al., 2015).

1. Mathematical Formalism: Allan and Modified Allan Variance

The single-sided Allan variance for fractional frequency error $y(t)$ is formally defined as

$$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$

where $\tau$ is the integration (averaging) time. In practice, for $N$ frequency samples $y_k$ with gate time $\tau_0$, the estimator is

$$\sigma^2_y(m\tau_0) = \frac{1}{2(N-2m+1)} \sum_{k=1}^{N-2m+1} \left[ \bar{y}_{k+m} - \bar{y}_k \right]^2$$

where $m = \tau/\tau_0$ and $\bar{y}_k$ is the local average over an interval of length $m$.

The Modified Allan variance introduces an additional moving average, yielding a fourth-order filter kernel:

$$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$0

and, in discrete data:

$$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$1

where $$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$2 are phase samples. Both AVAR and MVAR are quadratic forms in time-domain finite-difference operators and possess explicit Parseval-type representations linking them to single-sided PSDs (Marchi et al., 2023, Rubiola et al., 2022, Georgakaki et al., 2012).

2. Time–Frequency Domain Links: Kernels and Inversion

The direct connection between the time-domain variances and the frequency-domain PSD is via kernel-weighted integrals:

• For AVAR:

$$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$3

• For MVAR:

$$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$4

The kernel $$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$5 implements a finite-difference operator and acts as a weighting function, attenuating high-frequency noise in the AVAR, while $$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$6 provides even stronger filtering for MVAR (Marchi et al., 2023, Rubiola et al., 2022, Georgakaki et al., 2012).

The inversion problem (recovering $$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$7 from a time-domain variance profile) is unambiguously solvable only for a single power-law noise regime; for piecewise or composite processes, explicit reconstruction algorithms produce an approximate, piecewise power-law PSD consistent with the measured $$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$8 within set intervals. The mapping between time-domain slopes $$\sigma^2_y(\tau) = \frac{1}{2} \left\langle \left[ y(t + \tau) - y(t) \right]^2 \right\rangle$$9 and frequency-domain slopes $\tau$0 is given by

$\tau$1

where $\tau$2 is a kernel integral dependent on the AVAR or MVAR filter (Marchi et al., 2023).

3. Noise Process Identification and Power-Law Mapping

Each noise process gives rise to a characteristic slope in $\tau$3: | Noise Process | $\tau$4 | $\tau$5 scaling | $\tau$6 scaling | |-------------------|-------------|----------------------------------|--------------------------------------| | White PM | $\tau$7 | $\tau$8 | $\tau$9 | | Flicker PM | $N$0 | $N$1 | $N$2 | | White FM | $N$3 | $N$4 | $N$5 | | Flicker FM | $N$6 | const (floor) | $N$7 | | Random walk FM | $N$8 | $N$9 | $y_k$0 |

The Modified Allan variance distinguishes phase-modulated noise types (white PM, flicker PM) more effectively due to its enhanced high-frequency rejection. Explicit conversion formulas, such as $y_k$1 and coefficient maps via kernel integrals, enable analytic translation between measured Allan profiles and underlying power-law PSDs (Marchi et al., 2023, Georgakaki et al., 2012, Rubiola et al., 2022).

4. Implementation: Algorithms and Practical Guidelines

Given a measured or modeled AVAR profile consisting of $y_k$2 contiguous power-law regimes, an algorithm constructs a matching piecewise PSD through:

• Reordering of slopes/amplitudes;
• Calculation of mapping kernels $y_k$3;
• Determination of break frequencies $y_k$4 by enforcing continuity $y_k$5;
• Assembly of the composite $y_k$6;
• Optional self-validation via re-integration to reconstruct $y_k$7 (Marchi et al., 2023).

These methods extend to the MVAR by antisymmetrizing the kernel and replacing the AVAR kernel with the appropriate sixth-order filter. Limitations include loss of precision for strongly curved $y_k$8 dependence and requirements for high-frequency cut-offs when $y_k$9 or $\tau_0$0 (Marchi et al., 2023).

5. Applications in Frequency Transfer and Metrology

Time-domain variance analysis is indispensable for characterizing the stability and spectral properties of fiber-optic links, atomic clocks, and satellite navigation systems. Real-world applications include:

• Astronomical/geodetic time series evaluation, where AVAR and its modifications (WADEV, WMADEV) handle unequal weights, multidimensional data, reveal spectral regimes, and quantify fractal noise properties (Malkin, 2011);
• Fiber-link data analysis, where anti-alias filtering enables unbiased AVAR-based estimation of long-term stability without recourse to MVAR, provided all high-frequency noise is properly filtered before decimation or analysis (Calosso et al., 2015, Calosso et al., 2015);
• Empirical separation of link-limited and clock-limited instability in optical frequency dissemination, essential for establishing true frequency-transfer uncertainty floors (Calosso et al., 2015);
• Digital phase measurement acquisition, where sine-fitting techniques connect time-domain noise extraction and AVAR estimation directly at the ADC digitization level (DeVoe, 2017).

6. Extensions and Variants: Weighted and Parabolic Variance

The parabolic variance (PVAR/Omega-type, "parADEV") and lambda-type (MVAR-related) measures generalize the classical framework, yielding wavelet-form time-domain variances with tailored weightings optimal for different noise mechanisms:

• PVAR uses a parabolic weight and achieves higher degrees of freedom with faster convergence and reduced measurement time for a fixed confidence interval compared to MVAR, without sacrificing analytic tractability for arbitrary (including non-integer) power-law noise processes (Vernotte et al., 2020).
• Weighted AVAR (WADEV) and multidimensional forms address outlier robustness and vector-valued time series, prominent in geodetic and radio-astronomy data (Malkin, 2011).
• Selection of weighting and variance type is determined by the dominant regime—parabolic for white phase noise, rectangular for white frequency noise—according to explicit minimization criteria for type-A uncertainty (Benkler et al., 2015).

7. Limitations and Validation

Reverse mapping from AVAR to PSD is only exact in single-regime (pure power law) cases; in practical, multi-slope scenarios, the piecewise construction reflects a best-fit approximation rather than a unique inversion. Strong curvature or discontinuities in $\tau_0$1 may introduce artifacts, especially near transitions. Self-validation by reconstructing the original variance from the piecewise PSD is recommended for algorithmic reliability. High-$\tau_0$2 and high-$\tau_0$3 behavior should be treated with care, enforcing cutoffs as needed for convergence (Marchi et al., 2023, Rubiola et al., 2022).

---

In summary, the rigorous time-domain links between Allan (and modified Allan) variance and power spectral density form the basis for quantitative oscillator, link, and frequency standard analysis, enabling physical insight, numerical conversion, and robust practical assessment across a wide range of real and simulated systems (Marchi et al., 2023, Georgakaki et al., 2012, Malkin, 2011, Rubiola et al., 2022, Vernotte et al., 2020, Calosso et al., 2015, Benkler et al., 2015).