Application of the Allan Variance to Time Series Analysis in Astrometry and Geodesy: A Review

This paper reviews the application of the classical Allan variance and its modified versions—specifically weighted, multi-dimensional, and weighted multi-dimensional variants—to analyze noise characteristics in unevenly weighted and multi-dimensional astrometric and geodetic time series.

The Gist

Imagine you are trying to listen to a faint radio station while driving through a city. Sometimes the signal is clear, sometimes it crackles with static, and sometimes a giant truck drives by, causing a massive burst of noise.

This paper is about a special tool called the Allan Variance (AVAR). Think of AVAR as a "noise detective" that helps scientists figure out exactly what kind of static is messing up their measurements.

Here is the breakdown of the paper in plain English, using some everyday analogies.

1. The Problem: Measuring the Unmeasurable

Scientists who study the Earth (Geodesy) and the stars (Astrometry) need to track things very precisely. They track:

• Station positions: How much a GPS tower moves (maybe due to earthquakes or melting ice).
• Star positions: How much a distant radio source wobbles in the sky.
• Earth's rotation: How fast the Earth spins and wobbles.

But their measurements aren't perfect. They have "noise" (errors).

• The Old Way: For a long time, scientists used a simple ruler called the "Standard Deviation" (or WRMS) to measure this noise. It's like taking a photo of a wobbly hand and measuring the blur.
• The Flaw: This old ruler is easily tricked. If the hand moves slowly in a big circle (a trend) or if there's a sudden jerk (a jump), the old ruler thinks the noise is huge, even if the hand is actually quite steady in between.

2. The Solution: The "Noise Detective" (AVAR)

Enter the Allan Variance (AVAR).

• How it works: Instead of looking at the whole picture at once, AVAR looks at how much the data changes from one moment to the next.
• The Superpower: It is very good at ignoring slow, boring trends (like the Earth's slow wobble) and focusing only on the random jitter. It's like a noise-canceling headphone that filters out the hum of the engine so you can hear the music.
• The Bonus: It can tell you what kind of noise you have. Is it "White Noise" (like static on a TV)? "Flicker Noise" (like a flickering lightbulb)? Or "Random Walk" (like a drunk person stumbling)? Knowing this helps scientists fix their data better.

3. The New Upgrades: Handling Messy Data

The original AVAR was designed for perfect, clock-like data where every measurement is equally reliable and happens at the exact same time. But real-world science is messy. The author, Zinovy Malkin, explains that real data has two big problems:

A. The "Unequal Trust" Problem (Weighted AVAR)

Imagine you are asking a group of people for the time.

• Person A has a fancy atomic clock.
• Person B has a broken wristwatch.
• Person C has a phone with a bad signal.

If you just average their answers, the broken watch ruins the result.

• The Fix: The paper introduces WAVAR (Weighted AVAR). This tool says, "I trust the atomic clock more, so I'll listen to it louder. I'll ignore the broken watch." It handles data where some measurements are "noisier" than others.

B. The "3D Puzzle" Problem (Multi-dimensional AVAR)

Sometimes, you aren't measuring just one thing (like time). You are measuring a 3D object moving in space (Up/Down, Left/Right, Forward/Backward).

• Imagine trying to measure how much a spinning top wobbles. You can't just look at the "Up" movement; you have to look at the whole 3D wobble.
• The Fix: The paper introduces MAVAR and WMAVAR. These tools treat the data as a single 3D (or 2D) arrow instead of three separate lines. This gives a much more accurate picture of how the object is actually moving.

4. Real-World Examples from the Paper

The author shows how these tools saved the day in several scenarios:

• The "Bad Data" Filter: In one example, a GPS station had a few measurements that were way off (outliers). The old method thought the whole station was shaking violently. The new Weighted AVAR realized, "Wait, those bad measurements had huge error bars attached to them," and ignored them, revealing the station was actually quite stable.
• The "Jump" Problem: Sometimes data has sudden "jumps" (like a sensor glitch). The paper admits that even the new tools can get confused by sudden jumps, but they are much better than the old methods.
• Star Maps: Scientists used these tools to check the "International Celestial Reference Frame" (basically, the universe's map). They found that a newer map (ICRF2) was much less "noisy" and more stable than the old one (ICRF1), proving the new map was better.
• Earth's Spin: They analyzed how the Earth spins. They found that the "noise" in the Earth's rotation data is mostly "Flicker Noise" (unpredictable jitter), which helps them understand how the Earth's core and atmosphere interact.

5. The Bottom Line

The paper concludes that while no tool is perfect, the Allan Variance and its new "supercharged" versions (Weighted and Multi-dimensional) are essential for modern science.

• Old Way: "The data is messy, so the error is huge."
• New Way: "The data has some messy parts, but if we weigh them correctly and look at the 3D movement, we can see the true signal hidden underneath."

It's like upgrading from a blurry, black-and-white photo to a high-definition, 3D video that can automatically filter out the static, allowing scientists to see the universe and our planet with incredible clarity.

Technical Summary

1. Problem Statement

The assessment of noise characteristics in physical measurement time series is critical for determining the quality and statistical reliability of data in astrometry and geodesy. While the Allan Variance (AVAR) has been a standard tool in metrology for assessing frequency standards, its direct application to astro-geodetic data faces specific challenges that the classical definition does not address:

• Unequal Uncertainties: Unlike clock comparisons where measurements often have uniform precision, geodetic and astrometric observations (e.g., VLBI, GPS, SLR) typically consist of data points with varying levels of uncertainty. Classical AVAR assumes uniform precision, leading to biased results if proper weighting is not applied.
• Multi-dimensional Data: Many astro-geodetic quantities are inherently vector-based (e.g., 3D station coordinates $X, Y, Z$ or 2D celestial object positions in Right Ascension and Declination). Classical AVAR is designed for scalar time series and cannot natively process multi-dimensional vectors without treating components independently, which ignores correlations between dimensions.
• Non-Stationarity and Outliers: Astro-geodetic time series often contain jumps, outliers, and non-stationary trends. Classical statistical measures like Root Mean Square (RMS) are heavily dependent on the removal of low-frequency systematic components (trends), whereas AVAR is theoretically more robust but requires specific handling for unevenly weighted data.

2. Methodology

The paper reviews the application of AVAR and introduces/modifies its mathematical framework to suit astro-geodetic data.

Core Concepts

• Classical AVAR: Defined as $AVAR = \frac{1}{2(n-1)} \sum (y_i - y_{i+1})^2$. It is used to estimate noise levels and spectral characteristics (identifying white noise, flicker noise, random walk, etc.) via the slope of the log-log plot of AVAR vs. sampling interval $\tau$.
• Weighted AVAR (WAVAR): To handle measurements with different uncertainties ($s_i$), the paper adopts a weighted estimator where weights $p_i$ are inversely proportional to the sum of the variances of adjacent points: $p_i = (s_i^2 + s_{i+1}^2)^{-1}$.
• Multi-dimensional AVAR (MAVAR): Extends AVAR to $k$-dimensional vectors by calculating the Euclidean distance between adjacent vector points.
• Weighted Multi-dimensional AVAR (WMAVAR): The most general estimator proposed. It combines weighting and multi-dimensional analysis.
  • It defines the difference between adjacent measurements as the Euclidean norm $d_i = |y_i - y_{i+1}|$.
  • It utilizes a weight formula derived from error propagation laws. The paper notes a simplified formula (Eq. 4) is often used in practice to avoid singularities when $d_i = 0$, showing negligible practical difference from the rigorous derivation.
• Dynamic AVAR (DAVAR): Introduced for non-stationary time series where noise characteristics change over time, computed over a sliding window.

Spectral Analysis

The paper details how AVAR is used to identify noise types based on the spectral index $\alpha$ in the power law $S_y(f) = S_0 f^\alpha$. The relationship between the log-log slope $\mu$ of the AVAR and $\alpha$ is given by $\alpha = -(\mu + 1)$.

• $\mu = -1 \rightarrow$ White noise ($\alpha = 0$)
• $\mu = 0 \rightarrow$ Flicker noise ($\alpha = -1$)
• $\mu = 1 \rightarrow$ Random walk ($\alpha = -2$)

3. Key Contributions

1. Formalization of WMAVAR: The paper consolidates the definitions of WAVAR, MAVAR, and WMAVAR, presenting WMAVAR as a universal estimator that encompasses the others as special cases. It provides the mathematical framework for applying these to $k$-dimensional vectors with heterogeneous uncertainties.
2. Demonstration of Robustness: Through case studies (e.g., station height and Celestial Pole Offsets), the paper demonstrates that WADEV (Allan Deviation derived from WAVAR) is significantly more robust to outliers than classical ADEV, provided the outliers have larger associated uncertainties.
3. Independence from Low-Frequency Trends: The paper highlights that AVAR/WADEV is practically independent of long-term systematic trends and low-frequency harmonics (e.g., Free Core Nutation signals), whereas Weighted RMS (WRMS) is heavily dependent on the accuracy of the model used to remove these trends.
4. Handling of Data Gaps: The paper discusses methods for handling unevenly spaced data, such as forming "normal points" (averaging over intervals) or using specific techniques to fill gaps, though it notes that averaging may lose high-frequency noise information.

4. Results and Applications

The paper reviews extensive applications of AVAR in three main domains:

• Celestial Reference Frame (CRF):

  • AVAR (specifically WMADEV) was used to compare radio source catalogs (e.g., ICRF1 vs. ICRF2 and Pulkovo catalogs).
  • Result: WMADEV proved to be a sensitive, independent metric for catalog quality. Comparisons showed that newer catalogs (ICRF2, Pulkovo RSC) had lower noise levels in Celestial Pole Offset (CPO) series than older ones, confirming their higher accuracy.
  • AVAR analysis revealed that ~60% of source position series exhibit white noise, while ~40% show flicker noise, aiding in the selection of stable "core sources" for reference frames.

• Geodesy (Station Positions):

  • AVAR was applied to station position time series from VLBI, GPS, SLR, and DORIS.
  • Result: GPS station motions were found to be dominated by flicker noise, whereas VLBI, SLR, and DORIS series were dominated by white noise.
  • The method successfully distinguished between genuine geophysical signals and systematic errors in baseline length monitoring.
  • It was shown that AVAR provides a more realistic noise estimate for station heights even in the presence of undetected outliers compared to classical methods.

• Earth Rotation and Geodynamics:

  • AVAR is a standard step in the International Earth Rotation and Reference Systems Service (IERS) for combining Earth Orientation Parameters (EOP).
  • Result: It characterizes internal random deviations and helps assign proper weights to different EOP series during combination.
  • Analysis of CPO series confirmed that WRMS is highly model-dependent, while WADEV is not, making WADEV superior for evaluating the stability of Earth rotation models.

5. Significance

• Superior Noise Characterization: The paper establishes that AVAR and its modifications (WAVAR, WMAVAR) are superior to classical variance (STD, WRMS) for astro-geodetic data because they are less sensitive to low-frequency trends and outliers.
• Standardization for Vector Data: By formalizing WMAVAR, the paper provides a rigorous statistical tool for analyzing 2D and 3D geodetic vectors, which is essential for modern space geodesy where coordinate components are interdependent.
• Improved Data Processing: The ability to accurately identify noise types (white vs. flicker) allows for better construction of covariance matrices, more realistic uncertainty estimation for station velocities, and improved regularization of Earth Orientation Parameter series.
• Limitations and Future Work: The author notes that AVAR still faces challenges with non-stationary time series (requiring DAVAR) and potential correlations between measurements. Additionally, the handling of unevenly spaced data with gaps remains an area requiring further refinement.

In conclusion, the paper argues that while AVAR originated in metrology, its modified forms (WAVAR, WMAVAR) are now indispensable tools for the rigorous analysis of modern astrometric and geodetic time series, offering robustness and spectral insight that classical statistical methods cannot provide.