SAND2026-25291 SANDIA REPORT SAND2026-25291 Printed August 2026 Allan Variance is Bootstrap Aggregation for Spectral Estimation Michael R. Walker II Prepared by Sandia National Laboratories Albuquerque, New Mexico 87185 Livermore, California 94550 CONTENTS 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2. Noise Color . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3. From Clocks to Inertial Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.1. Modeling Clock Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.2. Extension to Inertial Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.3. Labeling Noise Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 4. Allan variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4.1. 𝑀-sample variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4.2. Continuous-time signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4.3. Discrete-time signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 5. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 7 8 1. INTRODUCTION Allan variance is an important tool for analyzing non-stationary noise sequences. The initial motivation dates back to the 1960s and focused on quantifying performance of atomic clocks [1]. Allan variance has has since become the de-facto method for characterizing noise in clocks [2], accelerometers[3] and gyros[4]. Piecing together a generalized theoretical framework for Allan variance is challenging for a few reason. First, so-called Allan variance was not presented in Allan’s seminal publication [1]. Confounding the issue, application-specific jargon is typically used when interpreting results (c.f. Table 3-1)[2, 3, 4]. The distinction between continuous-time and discrete-time signals is often blurred. Perhaps more concerning is the finding that high-resolution acquisition methods can distort the Allan variance[5]. In the broad scope of signal processing applications, periodograms are arguably more popular than Allan variance. Roughly 60 years since inception, it is reasonable to ask whether Allan variance is still relevant today [6]. Indeed we find Allan variance analysis useful, particularly for estimating low-frequency spectral content in long sequences. Additionally, reference methods for parameterizing noise simulators and sensor fusion models begin with Allan variance analysis[7]. 9 10 π‘₯ W (𝑑) 𝑑/𝑑𝑑 π‘₯ V (𝑑) (a) differentiation ∫ π‘₯ W (𝑑) π‘₯R (𝑑) (b) integration Figure 2-1 Generation of violet and red noise from white noise. 2. NOISE COLOR Color is used in many fields to distinguish noise sources according to trends in their power spectral densities [8]. Here we briefly define several noise colors, how they are related, and notionally how to generate them from white noise. For stochastic signal, π‘₯ π‘˜ (𝑑), with power spectral density 𝑋 π‘˜ ( 𝑓 ), the color of π‘₯ π‘˜ (𝑑) refers to a power-law relationship |𝑋 π‘˜ ( 𝑓 )| 2 β‰ˆ π›Όπ‘˜ 𝑓 π‘˜ . (1) for some constant scale factor 𝛼 π‘˜ . Specific color names are typically associated with specific1 integers π‘˜. For example, white noise describes π‘˜ = 0, or a flat power spectra. Noise spectra that increases with 𝑓 quadratically (π‘˜ = 2) is referred to as violet noise. Noise spectra that decrease with 𝑓 quadratically (π‘˜ = βˆ’2) are referred to as red noise. The color naming convention intuitively relates the spectral content of the colors themselves. Red and violet are on opposite ends of the visible light spectrum. Spectral content of red noise is concentrated at low frequencies, while spectral content of violet noise is concentrated at high frequencies. White light comprises all light colors with equal intensity which is analogous to white noise comprising equal spectral power at all frequencies. Violet noise can be generated by differentiating white noise as illustrated in Figure 2-1a. The relationship between these noise colors can be easily verified through the differentiation property of the Fourier transform. Analytically violet noise approaches infinity with frequency, however all practical systems are arguably band-limited for some 𝑓 < ∞. Red noise can be generated by integrating a white noise process as illustrated in Figure 2-1b. Red noise is also referred to as Bownian noise, brown noise, or random walk. Red noise presents some analytic challenges, such as its singularity at 𝑋R (0). Red noise is not stationary as its variance increases with 𝑑. While non-stationarity does not preclude computation of the power spectral density, non-ergodicity complicates low-frequency spectral analysis. Additionally, blue and pink noise are associated with π‘˜ = 1 and π‘˜ = βˆ’1, respectively. In electronics, pink noise is often referred to as flicker. Generating blue and pink noise from white noise is relatively challenging. The typical approach involves constructing a filter to achieve the linear frequency trend over a sufficient bandwidth [9, 10]. 1 Some authors use colors to reference ranges of k (e.g. (1, 2)) 11 On a log-scale, power-law models are piecewise linear approximations of the power spectra. Most signals does not fit a single power-law relationship over the full signal bandwidth. Instead, the signal log-power spectrum can be modeled as piecewise-continuous function. Since the slope of the piecewise segments are defined, we need only determine the vertical offset (gains) of each noise component. 12 3. FROM CLOCKS TO INERTIAL SENSORS Many of the techniques used to characterize inertial sensors were initially motivated by the need to quantify performance of atomic clocks [1, 2]. For a clock, noise typically refers to instability in the frequency. In the following we derive a model for the noise signal of interest for clocks. We then extend the model to noise in the output of inertial signals. Finally we conclude with some notes on terminology. 3.1. Modeling Clock Errors We consider the sinusoidal clock model 𝑠(𝑑) = (𝑉0 + πœ– (𝑑)) sin (2πœ‹π‘£ 0 𝑑 + πœ™(𝑑)) . (2) Focusing on the phase of the clock, we define Ξ¦(𝑑) := 2πœ‹π‘£ 0 𝑑 + πœ™(𝑑), (3) and its instantaneous frequency 1 πœ• 𝑣(𝑑) := Ξ¦(𝑑) (4) 2πœ‹ πœ•π‘‘ 1 Β€ = 𝑣0 + πœ™(𝑑). (5) 2πœ‹ Instantaneous frequency is an elusive concept which we cannot measure directly. Instead, we consider what is often referred to as the normalized frequency deviation 𝑣(𝑑) βˆ’ 𝑣 0 𝑦(𝑑) := (6) 𝑣0 1 Β€ = πœ™(𝑑). (7) 2πœ‹π‘£ 0 Similarly, we define the instantaneous phase deviation 1 π‘₯(𝑑) := πœ™(𝑑). (8) 2πœ‹π‘£ 0 Clearly, (7) and (8) are related Β€ 𝑦(𝑑) = π‘₯(𝑑). (9) 3.2. Extension to Inertial Sensors Clock stability is important for many applications, but the associated characterization methodologies generalize well to other sensors. In the clock model, instantaneous frequency is not observed directly. Similarly, instantaneous acceleration cannot be digitized by an accelerometer. 13 We first define velocity as the integral of acceleration ∫ 𝑑 𝑣(𝑑) = π‘Ž(𝜏)π‘‘πœ + 𝑣 0 . (10) 0 To be clear, here we reappropriate 𝑣(𝑑) and 𝑣 0 to indicate the velocity at time 𝑑 and 𝑑 = 0, respectively. Accelerometers report the change in velocity Δ𝑣(𝑑) := 𝑣(𝑑 + 𝜏) βˆ’ 𝑣(𝑑) (11) ∫ 𝑑+𝜏 = π‘Ž(𝑠)𝑑𝑠 (12) 𝑑 which integrate acceleration over the sampling period 𝜏 (typically on the order of ms). To put accelerometer equations on the same footing as the clock equations, we model the acceleration Β€ π‘Ž(𝑑) = π‘Ž 0 + πœ™(𝑑), (13) drawing an analogy with the clock frequency (5). Here π‘Ž 0 represents the nominal inertial acceleration observed by the sensor and πœ™(𝑑) represents random fluctuations in velocity. Plugging (13) into (10), we have 𝑣(𝑑) = π‘Ž 0 𝑑 + πœ™(𝑑) βˆ’ πœ™(0) + 𝑣 0 . (14) Contrasting (14) and (3), velocity is analogous to clock phase. Here, πœ™(0) and 𝑣 0 are initial conditions which were assumed 0 in the clock model (phase 0 at time 0). For an accelerometer we consider the normalized acceleration deviation π‘Ž(𝑑) βˆ’ π‘Ž 0 𝑦(𝑑) := (15) π‘Ž0 1 Β€ = πœ™(𝑑) (16) π‘Ž0 which is analogous to the normalized frequency deviation (7). Similarly, we define the instantaneous velocity deviation 1 π‘₯(𝑑) := πœ™(𝑑) (17) π‘Ž0 such that (9) holds. We define an approximation to the normalized acceleration deviation   1 Δ𝑣(𝑑) 𝑦¯ (𝑑) := βˆ’ π‘Ž0 . (18) π‘Ž0 𝜏 Plugging (13) into (12), we have Δ𝑣(𝑑) = π‘Ž 0 𝜏 + πœ™(𝑑 + 𝜏) βˆ’ πœ™(𝑑). (19) Therefore,(18) approaches (16) in the limiting sense 1 Β€ lim 𝑦¯ (𝑑) = πœ™(𝑑). (20) πœβ†’0 π‘Ž0 14 Table 3-1 Noise terms for inertial sensors and clocks power ASD AVAR noise law 𝜎-slope 𝜎 2 -slope color accelerometers gyroscopes clocks 𝑓2 βˆ’3/2 βˆ’3 violet velocity white angle white phase white quantization quantization 𝑓 βˆ’1 βˆ’2 blue (velocity flicker) (angle flicker) phase flicker velocity RW angle RW phase RW 𝑓0 βˆ’1/2 βˆ’1 white (accel. white) (rate white) (freq. white) pink bias instability bias instability 𝑓 βˆ’1 0 0 (flicker) (accel. flicker) (rate flicker) freq. flicker red 𝑓 βˆ’2 +1/2 +1 (RW) accel. RW rate RW freq. RW 𝑓 βˆ’3 +1 +2 trend rate ramp freq. ramp 3.3. Labeling Noise Terms For clocks and inertial sensors, power-law noise models are often parameterized using Allan variance. Unfortunately there are many different labels used to reference the same noise terms as shown in Table 3-1. For practical examples, see Fig. 3 [7] and [4]. Here we abbreviate Allan standard deviation (ASD), Allan variance (AVAR), and random walk (RW). Additionally, Allan variance is sometimes used as a generalized term and the distinction between standard deviation and variance must be parsed in-context. For example, Allan variance may be illustrated using plots of 𝜎 and not specifically 𝜎 2 [4]. Highlighting noise terms associated with an accelerometer, consider the accelerometer output is acceleration. White noise on the output will effect white acceleration noise. If instead, the output was first modeled as a velocity signal, integration would be required to convert the velocity signal to an acceleration signal. Alternatively, consider a velocity signal comprising both white and red noise. Differentiating the velocity signal increases the power of the noise terms which produces violet and white noise in the resulting acceleration signal. For this reason, violet noise for accelerometers are often referenced as white velocity noise. Similarly, white acceleration noise can be considered velocity random-walk which could be obtained by differentiating a velocity signal with red noise. 15 16 4. ALLAN VARIANCE Allan variance is typically motivated by continuous-time signals yet applied to digitized signals. Here we derive an expression for Allan variance of continuous-time signals and use the result to motivate expressions for discrete-time signals. Additionally we distinguish the so-called overlapping Allan variance from the non-overlapping Allan variance. However, we first establish notation by describing the generalized M-sample variance. 4.1. 𝑀 -sample variance Consider an indexed set of scalar values {π‘₯1 , π‘₯2 , . . .}. We define the 𝑀-sample mean 𝑀 1 βˆ‘οΈ πœ‡Β― π‘₯ (𝑀) = π‘₯π‘š . (21) 𝑀 π‘š=1 We define the 𝑀-sample variance 𝑀 1 βˆ‘οΈ Β― π‘₯2 (𝑀) = 𝜎 (π‘₯ π‘š βˆ’ πœ‡Β― π‘₯ (𝑀)) 2 (22a) 𝑀 βˆ’ 1 π‘š=1 ( 𝑀 ! ) 1 βˆ‘οΈ = π‘₯ 2 βˆ’ 𝑀 πœ‡Β― π‘₯ (𝑀) 2 (22b) 𝑀 βˆ’ 1 π‘š=1 π‘š ! !2ο£Ό ο£² 𝑀 2 𝑀 ο£± βˆ‘οΈ 1   1 βˆ‘οΈ  ο£½  = π‘₯π‘š βˆ’ 𝑀 π‘₯π‘š . (22c) 𝑀 βˆ’1   π‘š=1 𝑀 π‘š=1   ο£³ ο£Ύ The factor of 𝑀 βˆ’ 1 effects an unbiased estimator. The sample variance 𝜎 ¯𝑀2 is trivial for 𝑀 < 2. In the limiting case, we have the 2-sample variance 1 Β― π‘₯2 (2) = (π‘₯ 2 βˆ’ π‘₯ 1 ) 2 . 𝜎 (23) 2 4.2. Continuous-time signals Consider a latent signal of interest 𝑧(𝑑) which is only observed through the improper integral ∫ 𝑑 π‘₯(𝑑) := 𝑧(𝑠)πœ•π‘ . (24) βˆ’βˆž A related quantity of interest is the time-average of 𝑧 over a finite duration 𝜏 ∫ 1 𝑑+𝜏 𝑦(𝑑; 𝜏) := 𝑧(𝑠)πœ•π‘  (25a) 𝜏 𝑑 π‘₯(𝑑 + 𝜏) βˆ’ π‘₯(𝑑) = . (25b) 𝜏 17 Allan variance is broadly described as the (time-average of the) two-sample variance. Allan variance is expressed as a function of the integration window 𝜏. Using (23) and (25b), we have three equivalent expressions 2 𝜎AVAR (𝜏) = 𝜎 Β― 𝑦2 (2) (26a) 1 = (𝑦(𝑑 + 𝜏; 𝜏) βˆ’ 𝑦(𝑑; 𝜏)) 2 𝑑 (26b) 2 1 = (π‘₯(𝑑 + 2𝜏) βˆ’ 2π‘₯(𝑑 + 𝜏) + π‘₯(𝑑)) 2 𝑑 (26c) 2𝜏 Here we use ⟨·⟩ 𝑑 to indicate the time-average operator acting on continuous-time signals 𝑦, π‘₯. 4.3. Discrete-time signals Applying Allan variance to discrete (digitized) sequences leaves some implementation details open to interpretation. For discrete sequences, we assume uniform sampling 𝑦 𝜏 [𝑛] := 𝑦(π‘›πœ; 𝜏), βˆ€ 𝑛 = 0, 1, . . . . (27) Given an 𝑁-length sequence {𝑦 𝜏 [𝑛] | 𝑛 = 0, 1, . . . , 𝑁 βˆ’ 1}, the so-called non-overlapping Allan variance is given by π‘βˆ’2 1 βˆ‘οΈ 2 𝜎NAVAR (𝜏) := (𝑦 𝜏 [𝑛 + 1] βˆ’ 𝑦 𝜏 [𝑛]) 2 . (28) 2(𝑁 βˆ’ 1) 𝑛=0 Using the generalized window function ο£±   1/𝑀, 0≀𝑛<𝑀 ο£²  𝑀 𝑀 [𝑛] := βˆ’1/𝑀, 𝑀 ≀ 𝑛 ≀ 2𝑀 βˆ’ 1 , (29)   0,  otherwise. ο£³ we restate (28) π‘βˆ’2 1 !2 2 1 βˆ‘οΈ βˆ‘οΈ 𝜎NAVAR (𝜏) = 𝑦 𝜏 [𝑛 + 1 βˆ’ π‘š]𝑀 1 [π‘š] . (30) 2(𝑁 βˆ’ 1) 𝑛=0 π‘š=0 The windowing function (29) is analogous to a numeric differential. Accordingly, (29) acts as a high-pass filter on 𝑦 𝜏0 . Therefore, (30) represents the time-averaged power in 𝑦 𝜏0 after high-pass filtering. Herein lies the significance of Allan variance. Many naturally-occurring signals exhibit power spectra that roll off at a reciprocal of frequency - with a singularity at 𝑓 = 0. Low-frequency spectral content implies resulting signals are not ergodic. Traditional spectral analysis will not converge as sequence length increases. Differentiating a signal in the time-domain is equivalent to multiplying by a frequency ramp in the Fourier-domain. Numeric differentiation removes the singularity, for 𝑓 βˆ’1 and 𝑓 βˆ’2 processes, and enables characterization of low (non-zero) frequency spectral content. 18 In the following, we distinguish non-overlapping Allan variance from the more common overlapping Allan variance. To clarify the distinction, we increase the sampling rate and reduce  averaging time of 𝑦. Using 𝜏0 = 𝜏/𝑀, we obtain the 𝑁 𝑀-length sequence the 𝑦 𝜏0 [𝑛] | 𝑛 = 0, 1, . . . , 𝑁 𝑀 βˆ’ 1 such that π‘€βˆ’1 ∫ (𝑛𝑀+π‘š)𝜏0 +𝜏0 1 βˆ‘οΈ 𝑦 𝜏 [𝑛] = 𝑧(𝑠)πœ•π‘  (31a) 𝜏 π‘š=0 (𝑛𝑀+π‘š)𝜏0 π‘€βˆ’1 1 βˆ‘οΈ = 𝜏0 𝑦 𝜏0 [𝑛𝑀 + π‘š] (31b) 𝜏 π‘š=0 π‘€βˆ’1 1 βˆ‘οΈ = 𝑦 𝜏 [𝑛𝑀 + π‘š] (31c) 𝑀 π‘š=0 0 Substituting (31c) into (28) we have π‘βˆ’2 !2 1 βˆ‘οΈ 1 π‘€βˆ’1 βˆ‘οΈ 2 𝜎NAVAR (π‘€πœ0 ) := 𝑦 𝜏 [𝑛𝑀 + π‘š + 𝑀] βˆ’ 𝑦 𝜏0 [𝑛𝑀 + π‘š] . (32) 2(𝑁 βˆ’ 1) 𝑛=0 𝑀 π‘š=0 0 The inner summation computes a single, 2-sample variance, with an averaging time of 𝜏 = π‘€πœ0 . The outer summation combines the 2-sample variances with a step size of 𝑀. Reducing the step size of the outer summation (and increasing the upper bound), we obtain the so-called overlapping Allan variance[11] 𝑁 π‘€βˆ’2𝑀 π‘€βˆ’1 !2 2 1 βˆ‘οΈ 1 βˆ‘οΈ 𝜎OAVAR (π‘€πœ0 ) := 𝑦 𝜏 [𝑛 + π‘š + 𝑀] βˆ’ 𝑦 𝜏0 [𝑛 + π‘š] . (33) 2(𝑁 𝑀 βˆ’ 2𝑀 + 1) 𝑛=0 𝑀 π‘š=0 0 As an alternative statement of (33), we restate the inner summation as convolution with a 2𝑀-length windowing function (29) 𝑁 𝑀 2π‘€βˆ’1 !2 2 1 βˆ‘οΈ βˆ‘οΈ 𝜎OAVAR (π‘€πœ0 ) := 𝑦 𝜏 [𝑛 βˆ’ π‘š]𝑀 𝑀 [π‘š] . (34) 2(𝑁 𝑀 βˆ’ 2𝑀 + 1) 𝑛=2𝑀 π‘š=0 0 19 20 5. CONCLUSION Allan variance is a method of characterizing spectral content of time-domain signals. Computing two-sample variances mitigates effects of low-frequency content characteristic of non-ergodic processes. While producing a single two-sample variance measurement would be noisy, averaging multiple samples reduces uncertainty in the result. In the machine learning community, bootstrap aggregation (bagging) refers to averaging the result of weak learners trained on subsets of data[13]. If we consider a two-sample variance a weak-learner, Allan variance could be described as bagged spectral estimation. 21 22 REFERENCES [1] D. Allan, β€œStatistics of atomic frequency standards,” Proceedings of the IEEE, vol. 54, no. 2, pp. 221–230, 1966. [2] β€œIEEE standard definitions of physical quantities for fundamental frequency and time metrology-random instabilities,” IEEE Std 1139-1999, pp. 1–40, 1999. [3] β€œIEEE standard specification format guide and test procedure for linear single-axis, nongyroscopic accelerometers,” IEEE Std 1293-2018 (Revision of IEEE Std 1293-1998), pp. 1–271, 2019. [4] β€œIEEE standard for specifying and testing single-axis interferometric fiber optic gyros,” IEEE Std 952-2020 (Revision of IEEE Std 952-1997), pp. 1–93, 2021. [5] S. T. Dawkins, J. J. McFerran, and A. N. Luiten, β€œConsiderations on the measurement of the stability of oscillators with frequency counters,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 54, no. 5, pp. 918–925, 2007. [6] StackExchange contributors, β€œNoise - is allan variance still relevant,” 2022, [Online; accessed 26-October-2022]. [Online]. Available: https://dsp.stackexchange.com/questions/79121/is-allan-variance-still-relevant [7] J. A. Farrell, F. O. Silva, F. Rahman, and J. Wendel, β€œInertial Measurement Unit Error Modeling Tutorial: Inertial Navigation System State Estimation with Real-Time Sensor Calibration,” IEEE Control Systems Magazine, vol. 42, no. 6, pp. 40–66, Dec. 2022, conference Name: IEEE Control Systems Magazine. [8] β€œColors of noise,” Apr. 2023, page Version ID: 1149436666. [Online]. Available: https://en.wikipedia.org/w/index.php?title=Colors_of_noise&oldid=1149436666 [9] N. Kasdin, β€œDiscrete simulation of colored noise and stochastic processes and 1/f/sup /spl alpha// power law noise generation,” Proceedings of the IEEE, vol. 83, no. 5, pp. 802–827, 1995. [10] A. Makdissi, F. Vernotte, and E. D. Clercq, β€œStability variances: a filter approach,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, vol. 57, no. 5, pp. 1011–1028, 2010. [11] W. Riley and D. Howe, β€œHandbook of frequency stability analysis,” https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=50505, 2008. [12] β€œIEEE standard for specifying inertial measurement units (imus),” IEEE Std 1780-2022, pp. 1–61, 2022. [13] K. P. Murphy, Machine Learning: A Probabilistic Perspective, ser. Adaptive Computation and Machine Learning Series. Cambridge, MA: MIT Press, 2012. 23 24 DISTRIBUTION Emailβ€”Internal Name Org. Sandia Email Address Technical Library 1911 sanddocs@sandia.gov 25 Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.