Noise Induces Biased Estimation of the Correction Gain

Jooeun Ahn; Zhaoran Zhang; Dagmar Sternad

doi:10.1371/journal.pone.0158466

Abstract

The detection of an error in the motor output and the correction in the next movement are critical components of any form of motor learning. Accordingly, a variety of iterative learning models have assumed that a fraction of the error is adjusted in the next trial. This critical fraction, the correction gain, learning rate, or feedback gain, has been frequently estimated via least-square regression of the obtained data set. Such data contain not only the inevitable noise from motor execution, but also noise from measurement. It is generally assumed that this noise averages out with large data sets and does not affect the parameter estimation. This study demonstrates that this is not the case and that in the presence of noise the conventional estimate of the correction gain has a significant bias, even with the simplest model. Furthermore, this bias does not decrease with increasing length of the data set. This study reveals this limitation of current system identification methods and proposes a new method that overcomes this limitation. We derive an analytical form of the bias from a simple regression method (Yule-Walker) and develop an improved identification method. This bias is discussed as one of other examples for how the dynamics of noise can introduce significant distortions in data analysis.

Citation: Ahn J, Zhang Z, Sternad D (2016) Noise Induces Biased Estimation of the Correction Gain. PLoS ONE 11(7): e0158466. https://doi.org/10.1371/journal.pone.0158466

Editor: Ramesh Balasubramaniam, University of California, Merced, UNITED STATES

Received: March 20, 2015; Accepted: June 16, 2016; Published: July 27, 2016

Copyright: © 2016 Ahn et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All the results are obtained from numerical analysis and analytical approach that are fully explained in the manuscript and therefore can be reproduced by anyone. In that sense the data are fully available.

Funding: This work was supported by the National Institutes of Health (NIH) R01-HD045639, NIH R01-HD087089, and the National Science Foundation NSF-EAGER 1548514 awarded to DS; and the Natural Science and Engineering Research Council of Canada (NSERC) RGPIN 2015-04395 awarded to JA. DS was also supported by a visiting scientist appointment at the Max-Planck Institute for Intelligent Systems in Tübingen, Germany. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding organizations.

Competing interests: The authors have declared that no competing interests exist.

Introduction

Error detection and correction is one of the most fundamental elements in motor learning, in fact in almost every type of learning. In engineering, the notion that the error in the current execution is used to improve the next execution is the essence of any controller design for systems that perform tasks repeatedly [1–3]. In human learning, it is universally agreed upon that the perception and correction of errors plays a significant role in the process of motor learning [4, 5], although the neural and cognitive mechanisms underlying error-based learning are not yet fully understood [6, 7]. Feedback of error, and knowledge of results or performance have long been shown to facilitate and accelerate acquisition of new motor skills [8]. Based on this understanding, motor learning can be enhanced by magnifying the perceived error [9–12]. Errors are not only instrumental during early practice, but smaller corrections also continue into later stages of skilled performance. Prior studies highlighted the critical role of error corrections during the repeated execution of goal-directed movements near the goal [13–15]. In fact, expert performance at a highly skilled level implies that any small error is readily compensated for.

While ubiquitous, it remains an open question to what degree each error is corrected for. Should a learner completely compensate for each error in the subsequent attempt, or should he/she be more conservative and only correct a smaller portion of the error? If random noise is part of the observed fluctuations, then complete error cancellation may lead to an overshoot and ultimately unstable performance. As errors become smaller with practice, should this proportion or “gain” of the correction change? With these questions, the sensitivity of error correction, the gain or learning rate, has been a key variable in several areas of research.

With the goal to shed light on error-based learning, various mathematical models have been used to quantify detection and correction of error. For example, Galea et al. used a simple deterministic model to reveal differences in the learning rate under positive and negative reward in a motor adaptation task [16]. Smith et al. introduced a learning model with error feedback on two different time scales that reproduced several features of human motor adaptation, including savings [17]. However, neither of the two models considered noise, which is inevitable and, arguably, even beneficial to motor learning [18–21]. Heeding to the critical role of random fluctuations, Scheidt et al. proposed an exponential model for learning of goal-directed reaching under random perturbations [22]. Baddeley et al. compared a series of simple models that included noise sources to find the best fit for human visuo-motor performance; results suggested that exponential weighting of recent errors may best account for the high efficacy of human performance [23]. Diedrichsen et al. also used a stochastic state-space model of trial-by-trial adaptation and suggested a dissociation between errors that guide behavioral goals and those that guide neural adaptation of internal models [24]. Several other studies used stochastic iterative models to address how error corrections are associated with a structural change of variability using skills ranging from simple reaching [13, 25] to throwing a ball [14].

The most critical parameter in all these learning models is the correction gain, learning rate or feedback gain, which determines how much of the error in the current trial is corrected in the following trial. The simplest form of an iterative learning model that includes a correction gain and noise can be described by a set of three linear equations [13]: (1) (2) (3)

Eq 1 defines the motor output, x at the i^th trial as a sum of the planned execution m and noise N; the latter is a random vector drawn from a zero mean distribution with a known covariance matrix. Most learning models assume that this noise comes from a Gaussian distribution. Eq 2 defines the error as the difference between the output and the target. Eq 3 models the learning process as an update of the planned execution based on the error information. The constant B is the correction gain, which defines the fraction of the error that is adjusted in the next trial.

Conventionally, estimation of this critical parameter B in a given time series of data has relied on a convenient and common mathematical tool—the method of least-squares regression. Different parameter optimization algorithms include the Gauss-Newton algorithm, gradient descent, and the Levenberg-Marquardt algorithm [26–29]. These model-based methods have been applied regardless of possible noise in the model or in the measured data. Implicitly or explicitly, these approaches relied on the assumption that the effect of noise averages out. If noise is explicitly taken into account using stochastic models, a widely-used method is the Expectation-Maximization (EM) algorithm [30]. This algorithm estimates the model parameters including the noise via maximum likelihood estimation [31, 32]. However, this method is confined to models with independent noise sources and cannot be applied to auto-regressive processes [33].

The present study closely examined the viability of least-square methods for estimating model parameters, scrutinizing the common assumption that noise averages out. As we will demonstrate, the subtle dynamics of noise significantly influence the estimates of the correction gain and produce a bias. This bias may be relatively small when the difference between initial and final motor output is by orders of magnitude larger than the noise variance, as is the case at initial stages of learning. However, when the learning process approaches steady-state, or the error is comparable with the variability of the motor output, the bias in the estimated correction gain can become significant.

We devise a new method that improves the accuracy of identification of the correction gain, for the case when motor learning is close to steady state and the error is of similar magnitude as the overall variability. Analytical approaches and numerical simulations show that the quantified bias and the adjustment by the new method are robust and are insensitive to the magnitude and the distribution of the noise.

Methods

A Simple Iterative Model

To exemplify the problem of least-square estimation of the correction gain and demonstrate the bias of the conventional estimate, we examined a simple model that represents the basic structure of a large number of learning models [13, 14, 23–25]. For the simplest one-dimensional case, the vector variables in Eqs 1 to 3 become scalars. Defining the origin of the coordinate as the desired motor output, or the target state (x_target = 0), (4) (5) (6) where N_i is a random variable from a distribution with zero mean and a standard deviation of σ_N, and B is assumed to satisfy 0 < B < 1. From Eqs 4, 5 and 6,

Therefore, the learning process can be described by a simple linear equation: (7)

In this auto-regressive form, least-square regression can be used to estimate B from a time series of measured data.

Testing the Validity of the Conventional Least-Square Regression

To test whether the least-square estimation extracts the real parameters with sufficient accuracy, the first step was to generate a set of time series {x_k} using the simple model of Eq 7 with a known value of B and random noise samples. The lengths of the time series n varied from 10 to 800. To extract the correction gain and the variance of the noise, Eq 7 was analyzed using the standard regression approach. We opted to use the Levenberg-Marquardt algorithm [28] among many other alternative least-square algorithms because this algorithm is an effective numerical method for regression. It has therefore been a common choice for least-squares problems in commercialized numerical computing environments including Matlab (Mathworks Inc., Natick, MA). The Levenberg-Marquardt algorithm has also been widely applied in motor learning studies [34, 35] and for the training of neural networks [36–38].

We generated time series with three different B values of 0.25, 0.50, and 0.75. As little is known about the origin and magnitude of the noise in actual data, we also investigated the effect of the noise distribution with three types of distributions—normal, uniform, and asymmetric lognormal. In addition, we simulated different magnitudes of the noise level σ_N. The probability density functions of the added noise are demonstrated in Fig 1.

Download:

Fig 1. Probability density functions of the applied noise.

Three types of noise distributions (normal, uniform and lognormal) with three different standard deviations were considered. For each type of distribution, the expectation value of the noise is zero, whereas the standard deviation σ_N varies from 0.5 to 2, resulting in different noise levels.

https://doi.org/10.1371/journal.pone.0158466.g001

The simulation and subsequent parameter estimation were repeated 1000 times for each parameter set and time series length n. The difference between the true B and the mean of the 1000 estimated values was evaluated for each noise distribution, noise level, and n. Numerical simulations and analyses were implemented in Matlab (Mathworks Inc., Natick, MA).

Quantification of the Bias and Development of Improved Estimation: Adjusted Yule-Walker (AYW) Method

Following the demonstration of a significant bias as reported in the results, we proceeded to show that this bias can be removed if an analytical expression of the bias is obtained. To attain a closed form of the bias, we considered one of the simplest estimators, the Yule-Walker algorithm [39, 40]. Note that the Yule-Walker estimation method does not reduce the bias by itself. However, it lends itself to obtain an analytical expression of the bias due to its simple structure. The same analytical derivation might be done for the commonly used Levenberg-Marquardt algorithm, but would be more challenging and less transparent. Using the derived correction term, this bias can then be eliminated, which significantly improves the accuracy of the parameter estimation.

For any autoregressive (AR) process, the general Yule-Walker method calculates the AR parameters. The algebraic expression of the simplest AR process of order one is (8) where N_i is a random variable. The Yule-Walker equation estimates the parameter A as (9) which is identical to the least-square linear regression. Note that Eq 8, which is a first-order AR process, has a similar structure as the simple learning model in Eq 7. The assumption that the noise in each iteration is an independent random sample allowed to approximate the expected bias of the Yule-Walker estimation in a simple closed form. Using the closed form of the bias, we could then quantify the correction gain more accurately. We called the procedure the Adjusted Yule-Walker method (AYW).

We verified the reliability of this AYW method in the same way as we tested the reliability of the conventional least-square method: with a fixed value of the correction gain B, we constructed a set of time series {x_k} with added noise; the length of the time series n varied from 10 to 800. As above, the simulations and estimation were performed with three different noise distributions and three different noise levels (Fig 1). The simulation was repeated 1000 times for each n and each parameter set. The correction gain B was estimated by the AYW method, and the difference between the true B value and the mean of the 1000 estimated values was evaluated for each noise distribution, noise level, and n. The entire procedure is overviewed in Fig 2.

Download:

Fig 2. Illustration of the analysis method and a representative data set.

https://doi.org/10.1371/journal.pone.0158466.g002

Results

Demonstration of the Bias from Least-Square Estimation

The conventional Levenberg-Marquardt (LM) algorithm based on the least-square method yielded a substantial bias. Fig 3 shows the B values for different time series length n estimated with the LM-algorithm (red) and the AYW method (blue) against the real B value (green). The estimated correction gains had higher values than the actual correction gains B. Table 1 summarizes and compares the magnitude of the biases when the correction gain was estimated by the two methods. To facilitate comparison, a normalized error was calculated as , where is the mean of the 1000 estimates, and B is the actual correction gain. Depending on the true value of B, the bias from the LM estimates was at 16 ~ 170% of B, even when the length of the time series n reached 800. The biases in the AYW method became less than 5% as n reached 800. Further, in Fig 3, the panels A to C show that the significant bias did not change much whether the noise came from a normal, uniform, or asymmetric lognormal distribution. Panel D in Fig 3 also shows that the bias was largely unaffected by the level of the noise.

Download:

Fig 3. Actual and estimated correction gain B.

Time series with actual correction gains B of 0.25, 0.5 and 0.75; different trial length n (from 10 to 800); and three distributions of noise (normal, uniform, and lognormal) were generated using Eq 7. The correction gain B was estimated for each time series using both the Levenberg-Marquardt least-square algorithm (red) and the Adjusted Yule-Walker (AYW) method (blue), and then compared with the actual correction gain B (green). The simulations were repeated 1000 times for each algorithm and each n. The estimation by the Levenberg-Marquardt algorithm yielded a substantial bias, whereas the AYW method significantly reduced the bias for large enough n. These results remained unchanged for the noise sampled from a Gaussian (A), a uniform (B), or an asymmetric lognomal (C) distribution. In addition, the bias from the Levenberg-Marquardt algorithm and the improvement by the AYW method were not affected by the magnitude of the noise level σ_N (D).

https://doi.org/10.1371/journal.pone.0158466.g003

Download:

Table 1. Normalized biases in the estimates of the correction gain B using the least-square methods and the Adjusted Yule-walker method.

https://doi.org/10.1371/journal.pone.0158466.t001

The biases were calculated for each type of noise distribution (normal, uniform, or lognormal) and each trial length in a normalized form of , where is the mean of the 1000 estimates, and B is the actual value of the correction gain. Note that the conventional least-square method with the Levenberg-Marquardt algorithm yielded biases of 16 ~ 170%, even with a large trial number. The biases with the AYW method became less than 5% as n reached 800.

Analytical Derivation of the Bias from the Yule-Walker Equation

We analytically derived the bias in the estimates of the correction gain using the Yule-Walker equation [39, 40], one of the simplest regression methods. The notation of the one-dimensional learning process of Eq 7 is simplified by substituting (1 –B) with C, (10) where N_i is a random variable from a distribution with zero mean and the standard deviation of σ_N, and C is between 0 and 1 because 0 < B < 1.

To illustrate, (11)

To warrant using the Yule-Walker equation, or Eq 9, we regard N_i+1 –N_i in Eq 8 as another single noise term, such that the coefficient C can be estimated as (12)

From Eq 11, the expectation of x_n is a weighted sum of the expectation of N_i, each of which is zero. Therefore, (13)

By definition, N_p and N_q are independent when p ≠ q. Therefore, (14) because where f is the probability density function of the noise. Note that the validity of Eq 14 does not depend on the specific property of the distribution. Eq 14 remains valid, regardless of whether the probability density function f is symmetric like a normal and a uniform distribution, or asymmetric like a lognormal distribution. From Eq 11, it follows that

By Eq 14, terms with N_p N_q (p ≠ q) do not contribute to the expectation. In addition, by the definition of N_i, the expectation of is σ_N². Therefore,

The common ratio of C² is between 0 and 1. Therefore, as n goes to infinity,

Therefore, using Eq 13, (15)

Assuming large enough n, the denominator of Eq 12 approximately becomes the variance of x, or σ_x² multiplied by (n– 1), or (16)

The variance of x can also be simply obtained as a stable solution of or (17)

From Eq 11, x_i is a weighted sum of N₁, N₂, …, and N_i. Therefore, using Eq 14, the terms with N_i+1 N_i and x_i N_i+1 do not contribute to the expectation E[.] in Eq 17. In addition, as only N_i² in the x_i N_i contributes to the expectation, Eq 17 becomes (18)

By definition, the expectation of is σ_N². Therefore, the expectation of x², or σ_x² satisfies or which is same as Eq 16. Then, the expectation of C_YW becomes (19)

The numerator of Eq 19 becomes (20)

From Eq 11,

By Eq 14, only terms with N_i² contribute to the expectation E[.] in Eq 20, whereas terms with N_p N_q (p ≠ q) do not contribute to the expectation. Therefore, (21)

By definition of N_i, the expectation of is σ_N². Therefore, Eq 21 becomes (22)

Eq 22 can be re-written as (23)

The sum of the geometric sequence becomes

Therefore, Eq 23 becomes (24)

From Eqs 19 and 24, (25)

This expression shows that the bias is hardly reduced, even when n is large. Assuming large n, (26)

Note that the expectation of the estimate of C by the Yule-Walker method for large n is negative because 0 < C < 1. This means that the estimation is severely biased, if we use the Yule-Walker algorithm to measure the correction gain of the learning model. For a general autoregressive (AR) process, the Yule-Walker equation shows a finite bias [41, 42], but not to the same extent. The substantial bias in this case originated from the anti-correlated noise added to the linear relation between x_i and x_i+1. This contrasts with the added noise in the general AR process that has no correlation.

Equation for the Improved Estimation

From Eq 25, for large enough n, the expectation of C can be written as or (27)

From Eq 26, the term 1 + 2E[C_YW] is positive. Therefore, the quadratic equation (Eq 27) has one positive root and one negative root. We are interested in the positive root, and the coefficient C is estimated as (28)

Accordingly, the true correction gain B can be approximated as (29)

Simulation confirmed that Eq 29, the Adjusted Yule Walker or AYW method improves the estimate. Fig 3 and Table 1 show that the bias was significantly reduced by the AYW method to less than 5% as n increased. This improvement was neither affected by the type of the noise distribution, irrespective of whether the noise came from a normal, a uniform, or an asymmetric lognormal distribution. It was also unaffected by the magnitude of the noise level σ_N.

If the actual noise variance is of interest, the time series of N_i can also be obtained in the following manner. The motor output x_i is directly measured. After C is identified from Eq 28, the time series of (N_i+1 –N_i) can be computed from Eq 10. With an arbitrarily assumed value of N₁, N_i can be obtained. Then, N₁ and N_i can be recomputed so that N_i has zero mean.

Discussion

A variety of studies on error-based adaptation and learning in computational neuroscience have regarded the correction gains of the learning process as an essential piece of information. These correction gains, feedback gains, or learning rates have been typically identified by least-square curve fitting, frequently using the widely available Levenberg-Marquardt algorithm. These methods were based on the premise that the effect of noise averages out by the regression. Indeed, the least-square curve fitting provides a reasonably accurate measure of the correction gain when the performance changes are considerably higher than the amplitude of the noise processes. However, later in learning when performance approaches a plateau, we showed that the noise in the system induces a substantial bias in the estimate of the correction gain, when relying on conventional least-square regression. Using a simple autoregressive learning model, we then devised an improved method that quantifies and corrects the bias.

Necessity of Improved Estimation Method

Our simulation results demonstrate that the conventional methods based on regression always overestimate the correction gain. The noise induces uncertainty (i.e., non-zero variance in the estimation) as well as a clear bias in the mean or expected value. As shown in Fig 3 and Table 1, the bias can be substantial, even in the simplest case of a learning model. Furthermore, the bias was not reduced by increasing the trial number n in the estimates by the Levenberg-Marquardt least-square algorithm. The subsequent analytical derivation proved that the bias from a linear regression method, such as the Yule-Walker equation, converges to a non-zero value as n approaches infinity, as shown in Eq 26. The same behavior is the likely cause for the observed bias in the Levenberg-Marquardt estimates, evidenced in our simulations. This robustness of the bias against the data size highlights the serious limitation of the conventional estimation of the correction gain.

A possible alternative to the least-square regression methods is the Expectation Maximization (EM) algorithm that has been used to identify parameters of a stochastic model from the observed data [43, 44]. The EM algorithm can reduce the bias due to noise with long enough time series, but it cannot estimate model parameters that are assumed to be deterministic [33]. Consequently, for the simple iterative learning process as in Eq 7, which is a kind of Autoregressive Moving Average (ARMA) model, the parameter B cannot be estimated by the EM algorithm [33]. Hence, other methods need to be identified.

Insensitivity of Bias to Noise Magnitude and Distribution

The results show that the bias due to conventional estimates does not depend on the magnitude of the noise in the data. This robustness of the bias against the noise magnitude further underlines the importance of the bias reduction, suggesting that the bias needs to be adjusted, even if the noise magnitude is relatively low.

Another important observation is that the bias is insensitive to the noise distribution. Most of mathematical learning models have typically assumed white Gaussian noise, which is mathematically well-defined and convenient to analyze. However, several studies convincingly suggest that the noise distribution in biological systems is closer to an asymmetric lognormal distribution rather than a Gaussian distribution [45–48]. The insensitivity of bias to the noise distribution enables the AYW method to provide an improved estimation without relying on the common but vulnerable assumption of most learning models.

As the bias is induced by noise, it may appear necessary to identify the noise properties in the data set to assess and eliminate the bias. However, our analytical analysis shows that we do not need detailed knowledge of the noise properties to estimate the bias. Eq 25 demonstrates that the bias does not depend on the noise level σ_N. In addition, we did not assume any specific distribution of the noise to derive the closed form of the bias; we only assumed that the newly added noise in the current trial is independent of the noise in the previous trial. This implies that we can apply the same method for a variety of systems with different levels and distributions of noise. For example, we can use the same AYW method to estimate the correction gain of young and healthy subjects or patients who may have different types and degrees of noise.

Considerations and Limitations

While the noise distribution is of no concern, our analysis assumes that the noise samples added to each trial have no correlation. While the assumption of uncorrelated noise in human movement has been generally accepted in learning models and is supported by some experimental observations [45], other studies have also shown significant correlations, positive and negative, in the variability of the motor output. For example, time series of stride intervals in human walking and heart beats exhibit long-range correlations, suggesting that the variability in those motor behaviors cannot be modeled as white noise [49, 50]. Hence, caution is necessary when applying the AYW method to a system with correlated or anti-correlated noise.

However, we need to distinguish the variability of the observed motor output from the assumed noise that causes the variability. In Eq 7, the variability of x may exhibit correlation, even if N has no correlation. Actually, a recent study showed that the long-range correlations in stride intervals of human walking and other rhythmic motor behavior may be explained by uncorrelated noise filtered by stable rhythmic dynamics of the sensorimotor system [51]. The assumption that the original unfiltered source of variability is white noise is also common and highly effective in control and signal processing theory as well as in computational biology. It is also the basis of the widely-applicable Kalman filter [52]. Hence, the observed correlated variability in the motor output does not invalidate the AYW method, which assumes uncorrelated noise.

The AYW method significantly reduces the bias when the length of the time series is sufficiently long (n ≥ 400). However, the accuracy of the AYW method is relatively low when the length of the time series is short (n ≤ 50), although the bias is always smaller than the bias in the least-square method (see Table 1). This inaccuracy with insufficient data is due to the fact that the bias reduction by the AYW method depends on Eq 16, which assumes large n. Future work may resolve this inaccuracy when only insufficient data are available.

While the AYW method provides a less biased estimation, it shows higher variability than the least-square method, particularly when the size of the data is small. This high variability may limit the efficacy of the AYW method when a distinction between correction gains of two groups is required and only a small data set is available. This trade-off between bias and variance has already been identified in other estimation methods and discussed in prior studies [53, 54]. Development of a new method that optimizes the efficacy considering the trade-off between bias and variance is deferred to future work.

Like many other studies, our approach assumed that the correction gain or the learning rate is constant. This assumption, though widely used, may not be strictly accurate; the learning rate is approximately constant for small errors, but may change when the errors become larger [55, 56]. This study addresses the bias of the correction gain estimation when the learning process approaches skilled levels or steady state, where small errors are prevalent. Therefore, the assumption of a constant correction gain is sufficiently valid within the scope of this study.

For initial study, our analysis used a highly simplified one-dimensional model of motor learning, and therefore, generalization to other models and multi-dimensional cases will require further considerations. For example, for rapid pointing movements, van Beers showed that this simple model was insufficient to account for the observed structure of variability [13]. However, our chosen model contains the essential components of motor learning: feedback through error information and unplanned variability due to noise. This basic simple learning model was used to clearly demonstrate the influence of noise on estimation.

A Representative Problem for the Influence of Noise on Data Analysis

The observed bias is an example for how the dynamics of noise can introduce significant distortions in the analysis. It is widely assumed that noise is “neutral” if it has zero mean, as the effect of noise averages out with a large amount of data. However, the accumulated effect of noise may not always remain neutral, if the noise is processed by the analysis, such as in the conventional least-square curve fitting. Another non-intuitive effect of noise on estimation was highlighted in a recent study on Floquet multipliers, a method for assessing orbital stability of rhythmic movements. Noise induces a bias in the estimation of orbital stability [57]. Another example for the surprising effect of noise is the two-thirds power law that is widely observed in human movements [58–61]. It has been shown that this subtle relation between velocity and curvature can be generated by Gaussian noise alone [62]. Hence, experimental assessment and interpretation of the power law as a fundamental principle of movement generation needs caution.

Noise is ubiquitous and deserves more attention as its dynamics may influence the estimation of the “deterministic” elements of the system. Our study is only one example to show that the common belief that the accumulated effect of noise is neutral needs revision. More work is needed to examine more complex learning models and the estimation of noise processes from the data.

Author Contributions

Conceived and designed the experiments: JA ZZ DS. Performed the experiments: JA ZZ. Analyzed the data: JA ZZ. Wrote the paper: JA ZZ DS. Derived the closed form of the bias of the correction gain: JA. Developed Adjusted Yule-Walker (AYW) method: JA.

References

1. Arimoto S, Kawamura S, Miyazaki F. Bettering operation of Robots by learning. Journal of Robotic Systems. 1984;1(2):123–40.
- View Article
- Google Scholar
2. Bristow DA, Tharayil M, Alleyne AG. A survey of iterative learning control. Control Systems, IEEE. 2006;26(3):96–114.
- View Article
- Google Scholar
3. Xu J-X, Tan Y. Linear and nonlinear iterative learning control: Springer; 2003.
4. Adams JA. A Closed-Loop Theory of Motor Learning. Journal of Motor Behavior. 1971;3(2):111–50. pmid:15155169
- View Article
- PubMed/NCBI
- Google Scholar
5. Schmidt RA. A schema theory of discrete motor skill learning. Psychological review. 1975;82(4):225.
- View Article
- Google Scholar
6. Richardson MJ, Riley MA, Shockley K, Seidler R, Kwak Y, Fling B, et al. Neurocognitive Mechanisms of Error-Based Motor Learning. Progress in Motor Control. Advances in Experimental Medicine and Biology. 782: Springer New York. p. 39–60.
7. Requin J, Stelmach G, Schmidt R. Frequent Augmented Feedback Can Degrade Learning: Evidence and Interpretations. Tutorials in Motor Neuroscience. NATO ASI Series. 62: Springer Netherlands; 1991. p. 59–75.
8. Salmoni AW, Schmidt RA, Walter CB. Knowledge of results and motor learning: a review and critical reappraisal. Psychological Bulletin. 1984;95(3):355. pmid:6399752
- View Article
- PubMed/NCBI
- Google Scholar
9. Wei Y, Bajaj P, Scheidt R, Patton J, editors. Visual error augmentation for enhancing motor learning and rehabilitative relearning. Rehabilitation Robotics, 2005 ICORR 2005 9th International Conference on; 2005 28 June-1 July 2005.
10. Wei Y, Patton J, Bajaj P, Scheidt R, editors. A Real-Time Haptic/Graphic Demonstration of how Error Augmentation can Enhance Learning. Robotics and Automation, 2005 ICRA 2005 Proceedings of the 2005 IEEE International Conference on; 2005 18–22 April 2005.
11. Patton J, Stoykov M, Kovic M, Mussa-Ivaldi F. Evaluation of robotic training forces that either enhance or reduce error in chronic hemiparetic stroke survivors. Experimental Brain Research. 2006;168(3):368–83. pmid:16249912
- View Article
- PubMed/NCBI
- Google Scholar
12. Sharp I, Huang F, Patton J. Visual error augmentation enhances learning in three dimensions. Journal of NeuroEngineering and Rehabilitation. 8(1):52.
- View Article
- Google Scholar
13. van Beers RJ. Motor Learning Is Optimally Tuned to the Properties of Motor Noise. Neuron. 2009;63(3):406–17. pmid:19679079
- View Article
- PubMed/NCBI
- Google Scholar
14. Abe MO, Sternad D. Directionality in Distribution and Temporal Structure of Variability in Skill Acquisition. Frontiers in Human Neuroscience. 2013;7.
- View Article
- Google Scholar
15. van Beers RJ, Brenner E, Smeets JB. Random walk of motor planning in task-irrelevant dimensions. Journal of Neurophysiology. 2013;109(4):969–77. pmid:23175799
- View Article
- PubMed/NCBI
- Google Scholar
16. Galea JM, Mallia E, Rothwell J, Diedrichsen J. The dissociable effects of punishment and reward on motor learning. Nature Neuroscience. 2015;18(4):597–602. pmid:25706473
- View Article
- PubMed/NCBI
- Google Scholar
17. Smith MA, Ghazizadeh A, Shadmehr R. Interacting Adaptive Processes with Different Timescales Underlie Short-Term Motor Learning. PLoS Biology 2006;4(6):e179. pmid:16700627
- View Article
- PubMed/NCBI
- Google Scholar
18. van Beers RJ, Haggard P, Wolpert DM. The role of execution noise in movement variability. Journal of Neurophysiology. 2004;91(2):1050–63. pmid:14561687
- View Article
- PubMed/NCBI
- Google Scholar
19. Churchland MM, Afshar A, Shenoy KV. A central source of movement variability. Neuron. 2006;52(6):1085–96. pmid:17178410
- View Article
- PubMed/NCBI
- Google Scholar
20. Jones KE, Hamilton AFdC, Wolpert DM. Sources of signal-dependent noise during isometric force production. Journal of Neurophysiology. 2002;88(3):1533–44. pmid:12205173
- View Article
- PubMed/NCBI
- Google Scholar
21. Sternad D, Abe M, Hu X, Muller H. Neuromotor noise, sensitivity to error and signal-dependent noise in trial-to-trial learning. PLoS Computational Biology 2011;7:e1002159.
- View Article
- Google Scholar
22. Scheidt RA, Dingwell JB, Mussa-Ivaldi FA. Learning to Move Amid Uncertainty. Journal of Neurophysiology. 2001;86(2):971–85. pmid:11495965
- View Article
- PubMed/NCBI
- Google Scholar
23. Baddeley RJ, Ingram HA, Miall RC. System Identification Applied to a Visuomotor Task: Near-Optimal Human Performance in a Noisy Changing Task. The Journal of Neuroscience. 2003;23(7):3066–75. pmid:12684493
- View Article
- PubMed/NCBI
- Google Scholar
24. Diedrichsen J, Hashambhoy Y, Rane T, Shadmehr R. Neural Correlates of Reach Errors. The Journal of Neuroscience. 2005;25(43):9919–31. pmid:16251440
- View Article
- PubMed/NCBI
- Google Scholar
25. Burge J, Ernst MO, Banks MS. The statistical determinants of adaptation rate in human reaching. Journal of Vision. 2008;8(4).
- View Article
- Google Scholar
26. Lawson CL, Hanson RJ. Solving least squares problems: SIAM; 1974.
27. Hartley HO. The modified Gauss-Newton method for the fitting of non-linear regression functions by least squares. Technometrics. 1961;3(2):269–80.
- View Article
- Google Scholar
28. Marquardt DW. An algorithm for least-squares estimation of nonlinear parameters. Journal of the Society for Industrial & Applied Mathematics. 1963;11(2):431–41.
- View Article
- Google Scholar
29. Moré JJ. The Levenberg-Marquardt algorithm: implementation and theory. Numerical analysis: Springer; 1978. p. 105–16.
30. Dempster AP, Laird NM, Rubin DB. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society Series B (Methodological). 1977:1–38.
- View Article
- Google Scholar
31. Shumway RH, Stoffer DS. An approach to time series smoothing and forecasting using the EM algorithm. Journal of Time Series Analysis. 1982;3(4):253–64.
- View Article
- Google Scholar
32. Ghahramani Z, Hinton GE. The EM algorithm for mixtures of factor analyzers. Technical Report CRG-TR-96-1, University of Toronto, 1996.
33. Thiesson B, Chickering DM, Heckerman D, Meek C, editors. ARMA time-series modeling with graphical models. Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence; 2004: AUAI Press.
34. van Beers RJ. How does our motor system determine its learning rate? PLoS ONE. 7(11):e49373. pmid:23152899
- View Article
- PubMed/NCBI
- Google Scholar
35. Wei K, Dijkstra TM, Sternad D. Passive stability and active control in a rhythmic task. Journal of Neurophysiology. 2007;98(5):2633–46. pmid:17881482
- View Article
- PubMed/NCBI
- Google Scholar
36. Ampazis N, Perantonis SJ. Two highly efficient second-order algorithms for training feedforward networks. Neural Networks, IEEE Transactions on. 2002;13(5):1064–74.
- View Article
- Google Scholar
37. Asirvadam VS, McLoone S, Irwin G, editors. Parallel and separable recursive Levenberg-Marquardt training algorithm. Neural Networks for Signal Processing, 2002 Proceedings of the 2002 12th IEEE Workshop on; 2002: IEEE.
38. Chen T-c Han D-j, Au FT, Tham L, editors. Acceleration of Levenberg-Marquardt training of neural networks with variable decay rate. Neural Networks, 2003 Proceedings of the International Joint Conference on; 2003: IEEE.
39. Yule GU. On a method of investigating periodicities in disturbed series, with special reference to Wolfer's sunspot numbers. Philosophical Transactions of the Royal Society of London Series A, Containing Papers of a Mathematical or Physical Character. 1927;226(636–646):267–98.
- View Article
- Google Scholar
40. Walker G. On periodicity in series of related terms. Proceedings of the Royal Society of London Series A. 1931;131(818):518–32.
- View Article
- Google Scholar
41. De Hoon M, Van der Hagen T, Schoonewelle H, Van Dam H. Why Yule-Walker should not be used for autoregressive modelling. Annals of Nuclear Energy. 1996;23(15):1219–28.
- View Article
- Google Scholar
42. Broersen PM. Finite-Sample Bias Propagation in Autoregressive Estimation With the Yule-Walker Method. Instrumentation and Measurement, IEEE Transactions on. 2009;58(5):1354–60.
- View Article
- Google Scholar
43. Moon TK. The expectation-maximization algorithm. Signal Processing Magazine, IEEE. 1996;13(6):47–60.
- View Article
- Google Scholar
44. Cheng S, Sabes PN. Modeling sensorimotor learning with linear dynamical systems. Neural Computation. 2006;18(4):760–93. pmid:16494690
- View Article
- PubMed/NCBI
- Google Scholar
45. Stocker AA, Simoncelli EP. Noise characteristics and prior expectations in human visual speed perception. Nature Neuroscience 2006;9(4):578–85. pmid:16547513
- View Article
- PubMed/NCBI
- Google Scholar
46. Furusawa C, Suzuki T, Kashiwagi A, Yomo T, Kaneko K. Ubiquity of log-normal distributions in intra-cellular reaction dynamic. arXiv preprint q-bio/0503040. 2005.
47. Munsky B, Trinh B, Khammash M. Listening to the noise: random fluctuations reveal gene network parameters. Molecular Systems Biology. 2009;5(1).
- View Article
- Google Scholar
48. Kreutz C, Rodriguez MB, Maiwald T, Seidl M, Blum H, Mohr L, et al. An error model for protein quantification. Bioinformatics. 2007;23(20):2747–53. pmid:17768165
- View Article
- PubMed/NCBI
- Google Scholar
49. Hausdorff JM, Peng CK, Ladin Z, Wei JY, Goldberger AL. Is walking a random walk? Evidence for long-range correlations in stride interval of human gait. Journal of Applied Physiology. 1995;78(1):349–58. pmid:7713836
- View Article
- PubMed/NCBI
- Google Scholar
50. Skinner JE, Goldberger AL, Mayer-Kress G, Ideker RE. Chaos in the Heart: Implications for Clinical Cardiology. Nature Biotechnology 1990;8(11):1018–24.
- View Article
- Google Scholar
51. Ahn J, Hogan N. Long-Range Correlations in Stride Intervals May Emerge from Non-Chaotic Walking Dynamics. PLoS ONE. 2013;8(9):e73239. pmid:24086274
- View Article
- PubMed/NCBI
- Google Scholar
52. Kalman RE. A new approach to linear filtering and prediction problems. Journal of Fluids Engineering. 1960;82(1):35–45.
- View Article
- Google Scholar
53. Faber NM. A closer look at the bias—variance trade-off in multivariate calibration. Journal of Chemometrics. 1999;13(2):185–92.
- View Article
- Google Scholar
54. Green RL, Kalivas JH. Graphical diagnostics for regression model determinations with consideration of the bias/variance trade-off. Chemometrics and Intelligent Laboratory Systems. 2002;60:173–88.
- View Article
- Google Scholar
55. Wei K, Körding K. Relevance of error: what drives motor adaptation? Journal of Neurophysiology. 2009;101(2):655–64. pmid:19019979
- View Article
- PubMed/NCBI
- Google Scholar
56. Marko MK, Haith AM, Harran MD, Shadmehr R. Sensitivity to prediction error in reach adaptation. Journal of Neurophysiology. 2012;108(6):1752–63. pmid:22773782
- View Article
- PubMed/NCBI
- Google Scholar
57. Ahn J, Hogan N. Improved Assessment of Orbital Stability of Rhythmic Motion with Noise. PLoS ONE. 2015;10(3):e0119596. pmid:25798610
- View Article
- PubMed/NCBI
- Google Scholar
58. Viviani P, Flash T. Minimum-jerk, two-thirds power law, and isochrony: converging approaches to movement planning. Journal of Experimental Psychology: Human Perception and Performance. 1995;21(1):32. pmid:7707032
- View Article
- PubMed/NCBI
- Google Scholar
59. Sternad D, Schaal S. Segmentation of endpoint trajectories does not imply segmented control. Experimental Brain Research. 1999;124(1):118–36. pmid:9928796
- View Article
- PubMed/NCBI
- Google Scholar
60. Schaal S, Sternad D. Origins and violations of the 2/3 power law in rhythmic three-dimensional arm movements. Experimental Brain Research. 2001;136(1):60–72. pmid:11204414
- View Article
- PubMed/NCBI
- Google Scholar
61. Ivanenko YP, Grasso R, Macellari V, Lacquaniti F. Two-thirds power law in human locomotion: role of ground contact forces. Neuroreport. 2002;13(9):1171–4. pmid:12151763
- View Article
- PubMed/NCBI
- Google Scholar
62. Maoz U, Portugaly E, Flash T, Weiss Y, editors. Noise and the two-thirds power law. Advances in Neural Information Processing Systems; 2005.

[ref1] 1. Arimoto S, Kawamura S, Miyazaki F. Bettering operation of Robots by learning. Journal of Robotic Systems. 1984;1(2):123–40.
View Article
Google Scholar

[2] View Article

[3] Google Scholar

[ref2] 2. Bristow DA, Tharayil M, Alleyne AG. A survey of iterative learning control. Control Systems, IEEE. 2006;26(3):96–114.
View Article
Google Scholar

[5] View Article

[6] Google Scholar

[ref3] 3. Xu J-X, Tan Y. Linear and nonlinear iterative learning control: Springer; 2003.

[ref4] 4. Adams JA. A Closed-Loop Theory of Motor Learning. Journal of Motor Behavior. 1971;3(2):111–50. pmid:15155169
View Article
PubMed/NCBI
Google Scholar

[9] View Article

[10] PubMed/NCBI

[11] Google Scholar

[ref5] 5. Schmidt RA. A schema theory of discrete motor skill learning. Psychological review. 1975;82(4):225.
View Article
Google Scholar

[13] View Article

[14] Google Scholar

[ref6] 6. Richardson MJ, Riley MA, Shockley K, Seidler R, Kwak Y, Fling B, et al. Neurocognitive Mechanisms of Error-Based Motor Learning. Progress in Motor Control. Advances in Experimental Medicine and Biology. 782: Springer New York. p. 39–60.

[ref7] 7. Requin J, Stelmach G, Schmidt R. Frequent Augmented Feedback Can Degrade Learning: Evidence and Interpretations. Tutorials in Motor Neuroscience. NATO ASI Series. 62: Springer Netherlands; 1991. p. 59–75.

[ref8] 8. Salmoni AW, Schmidt RA, Walter CB. Knowledge of results and motor learning: a review and critical reappraisal. Psychological Bulletin. 1984;95(3):355. pmid:6399752
View Article
PubMed/NCBI
Google Scholar

[18] View Article

[19] PubMed/NCBI

[20] Google Scholar

[ref9] 9. Wei Y, Bajaj P, Scheidt R, Patton J, editors. Visual error augmentation for enhancing motor learning and rehabilitative relearning. Rehabilitation Robotics, 2005 ICORR 2005 9th International Conference on; 2005 28 June-1 July 2005.

[ref10] 10. Wei Y, Patton J, Bajaj P, Scheidt R, editors. A Real-Time Haptic/Graphic Demonstration of how Error Augmentation can Enhance Learning. Robotics and Automation, 2005 ICRA 2005 Proceedings of the 2005 IEEE International Conference on; 2005 18–22 April 2005.

[ref11] 11. Patton J, Stoykov M, Kovic M, Mussa-Ivaldi F. Evaluation of robotic training forces that either enhance or reduce error in chronic hemiparetic stroke survivors. Experimental Brain Research. 2006;168(3):368–83. pmid:16249912
View Article
PubMed/NCBI
Google Scholar

[24] View Article

[25] PubMed/NCBI

[26] Google Scholar

[ref12] 12. Sharp I, Huang F, Patton J. Visual error augmentation enhances learning in three dimensions. Journal of NeuroEngineering and Rehabilitation. 8(1):52.
View Article
Google Scholar

[28] View Article

[29] Google Scholar

[ref13] 13. van Beers RJ. Motor Learning Is Optimally Tuned to the Properties of Motor Noise. Neuron. 2009;63(3):406–17. pmid:19679079
View Article
PubMed/NCBI
Google Scholar

[31] View Article

[32] PubMed/NCBI

[33] Google Scholar

[ref14] 14. Abe MO, Sternad D. Directionality in Distribution and Temporal Structure of Variability in Skill Acquisition. Frontiers in Human Neuroscience. 2013;7.
View Article
Google Scholar

[35] View Article

[36] Google Scholar

[ref15] 15. van Beers RJ, Brenner E, Smeets JB. Random walk of motor planning in task-irrelevant dimensions. Journal of Neurophysiology. 2013;109(4):969–77. pmid:23175799
View Article
PubMed/NCBI
Google Scholar

[38] View Article

[39] PubMed/NCBI

[40] Google Scholar

[ref16] 16. Galea JM, Mallia E, Rothwell J, Diedrichsen J. The dissociable effects of punishment and reward on motor learning. Nature Neuroscience. 2015;18(4):597–602. pmid:25706473
View Article
PubMed/NCBI
Google Scholar

[42] View Article

[43] PubMed/NCBI

[44] Google Scholar

[ref17] 17. Smith MA, Ghazizadeh A, Shadmehr R. Interacting Adaptive Processes with Different Timescales Underlie Short-Term Motor Learning. PLoS Biology 2006;4(6):e179. pmid:16700627
View Article
PubMed/NCBI
Google Scholar

[46] View Article

[47] PubMed/NCBI

[48] Google Scholar

[ref18] 18. van Beers RJ, Haggard P, Wolpert DM. The role of execution noise in movement variability. Journal of Neurophysiology. 2004;91(2):1050–63. pmid:14561687
View Article
PubMed/NCBI
Google Scholar

[50] View Article

[51] PubMed/NCBI

[52] Google Scholar

[ref19] 19. Churchland MM, Afshar A, Shenoy KV. A central source of movement variability. Neuron. 2006;52(6):1085–96. pmid:17178410
View Article
PubMed/NCBI
Google Scholar

[54] View Article

[55] PubMed/NCBI

[56] Google Scholar

[ref20] 20. Jones KE, Hamilton AFdC, Wolpert DM. Sources of signal-dependent noise during isometric force production. Journal of Neurophysiology. 2002;88(3):1533–44. pmid:12205173
View Article
PubMed/NCBI
Google Scholar

[58] View Article

[59] PubMed/NCBI

[60] Google Scholar

[ref21] 21. Sternad D, Abe M, Hu X, Muller H. Neuromotor noise, sensitivity to error and signal-dependent noise in trial-to-trial learning. PLoS Computational Biology 2011;7:e1002159.
View Article
Google Scholar

[62] View Article

[63] Google Scholar

[ref22] 22. Scheidt RA, Dingwell JB, Mussa-Ivaldi FA. Learning to Move Amid Uncertainty. Journal of Neurophysiology. 2001;86(2):971–85. pmid:11495965
View Article
PubMed/NCBI
Google Scholar

[65] View Article

[66] PubMed/NCBI

[67] Google Scholar

[ref23] 23. Baddeley RJ, Ingram HA, Miall RC. System Identification Applied to a Visuomotor Task: Near-Optimal Human Performance in a Noisy Changing Task. The Journal of Neuroscience. 2003;23(7):3066–75. pmid:12684493
View Article
PubMed/NCBI
Google Scholar

[69] View Article

[70] PubMed/NCBI

[71] Google Scholar

[ref24] 24. Diedrichsen J, Hashambhoy Y, Rane T, Shadmehr R. Neural Correlates of Reach Errors. The Journal of Neuroscience. 2005;25(43):9919–31. pmid:16251440
View Article
PubMed/NCBI
Google Scholar

[73] View Article

[74] PubMed/NCBI

[75] Google Scholar

[ref25] 25. Burge J, Ernst MO, Banks MS. The statistical determinants of adaptation rate in human reaching. Journal of Vision. 2008;8(4).
View Article
Google Scholar

[77] View Article

[78] Google Scholar

[ref26] 26. Lawson CL, Hanson RJ. Solving least squares problems: SIAM; 1974.

[ref27] 27. Hartley HO. The modified Gauss-Newton method for the fitting of non-linear regression functions by least squares. Technometrics. 1961;3(2):269–80.
View Article
Google Scholar

[81] View Article

[82] Google Scholar

[ref28] 28. Marquardt DW. An algorithm for least-squares estimation of nonlinear parameters. Journal of the Society for Industrial & Applied Mathematics. 1963;11(2):431–41.
View Article
Google Scholar

[84] View Article

[85] Google Scholar

[ref29] 29. Moré JJ. The Levenberg-Marquardt algorithm: implementation and theory. Numerical analysis: Springer; 1978. p. 105–16.

[ref30] 30. Dempster AP, Laird NM, Rubin DB. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society Series B (Methodological). 1977:1–38.
View Article
Google Scholar

[88] View Article

[89] Google Scholar

[ref31] 31. Shumway RH, Stoffer DS. An approach to time series smoothing and forecasting using the EM algorithm. Journal of Time Series Analysis. 1982;3(4):253–64.
View Article
Google Scholar

[91] View Article

[92] Google Scholar

[ref32] 32. Ghahramani Z, Hinton GE. The EM algorithm for mixtures of factor analyzers. Technical Report CRG-TR-96-1, University of Toronto, 1996.

[ref33] 33. Thiesson B, Chickering DM, Heckerman D, Meek C, editors. ARMA time-series modeling with graphical models. Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence; 2004: AUAI Press.

[ref34] 34. van Beers RJ. How does our motor system determine its learning rate? PLoS ONE. 7(11):e49373. pmid:23152899
View Article
PubMed/NCBI
Google Scholar

[96] View Article

[97] PubMed/NCBI

[98] Google Scholar

[ref35] 35. Wei K, Dijkstra TM, Sternad D. Passive stability and active control in a rhythmic task. Journal of Neurophysiology. 2007;98(5):2633–46. pmid:17881482
View Article
PubMed/NCBI
Google Scholar

[100] View Article

[101] PubMed/NCBI

[102] Google Scholar

[ref36] 36. Ampazis N, Perantonis SJ. Two highly efficient second-order algorithms for training feedforward networks. Neural Networks, IEEE Transactions on. 2002;13(5):1064–74.
View Article
Google Scholar

[104] View Article

[105] Google Scholar

[ref37] 37. Asirvadam VS, McLoone S, Irwin G, editors. Parallel and separable recursive Levenberg-Marquardt training algorithm. Neural Networks for Signal Processing, 2002 Proceedings of the 2002 12th IEEE Workshop on; 2002: IEEE.

[ref38] 38. Chen T-c Han D-j, Au FT, Tham L, editors. Acceleration of Levenberg-Marquardt training of neural networks with variable decay rate. Neural Networks, 2003 Proceedings of the International Joint Conference on; 2003: IEEE.

[ref39] 39. Yule GU. On a method of investigating periodicities in disturbed series, with special reference to Wolfer's sunspot numbers. Philosophical Transactions of the Royal Society of London Series A, Containing Papers of a Mathematical or Physical Character. 1927;226(636–646):267–98.
View Article
Google Scholar

[109] View Article

[110] Google Scholar

[ref40] 40. Walker G. On periodicity in series of related terms. Proceedings of the Royal Society of London Series A. 1931;131(818):518–32.
View Article
Google Scholar

[112] View Article

[113] Google Scholar

[ref41] 41. De Hoon M, Van der Hagen T, Schoonewelle H, Van Dam H. Why Yule-Walker should not be used for autoregressive modelling. Annals of Nuclear Energy. 1996;23(15):1219–28.
View Article
Google Scholar

[115] View Article

[116] Google Scholar

[ref42] 42. Broersen PM. Finite-Sample Bias Propagation in Autoregressive Estimation With the Yule-Walker Method. Instrumentation and Measurement, IEEE Transactions on. 2009;58(5):1354–60.
View Article
Google Scholar

[118] View Article

[119] Google Scholar

[ref43] 43. Moon TK. The expectation-maximization algorithm. Signal Processing Magazine, IEEE. 1996;13(6):47–60.
View Article
Google Scholar

[121] View Article

[122] Google Scholar

[ref44] 44. Cheng S, Sabes PN. Modeling sensorimotor learning with linear dynamical systems. Neural Computation. 2006;18(4):760–93. pmid:16494690
View Article
PubMed/NCBI
Google Scholar

[124] View Article

[125] PubMed/NCBI

[126] Google Scholar

[ref45] 45. Stocker AA, Simoncelli EP. Noise characteristics and prior expectations in human visual speed perception. Nature Neuroscience 2006;9(4):578–85. pmid:16547513
View Article
PubMed/NCBI
Google Scholar

[128] View Article

[129] PubMed/NCBI

[130] Google Scholar

[ref46] 46. Furusawa C, Suzuki T, Kashiwagi A, Yomo T, Kaneko K. Ubiquity of log-normal distributions in intra-cellular reaction dynamic. arXiv preprint q-bio/0503040. 2005.

[ref47] 47. Munsky B, Trinh B, Khammash M. Listening to the noise: random fluctuations reveal gene network parameters. Molecular Systems Biology. 2009;5(1).
View Article
Google Scholar

[133] View Article

[134] Google Scholar

[ref48] 48. Kreutz C, Rodriguez MB, Maiwald T, Seidl M, Blum H, Mohr L, et al. An error model for protein quantification. Bioinformatics. 2007;23(20):2747–53. pmid:17768165
View Article
PubMed/NCBI
Google Scholar

[136] View Article

[137] PubMed/NCBI

[138] Google Scholar

[ref49] 49. Hausdorff JM, Peng CK, Ladin Z, Wei JY, Goldberger AL. Is walking a random walk? Evidence for long-range correlations in stride interval of human gait. Journal of Applied Physiology. 1995;78(1):349–58. pmid:7713836
View Article
PubMed/NCBI
Google Scholar

[140] View Article

[141] PubMed/NCBI

[142] Google Scholar

[ref50] 50. Skinner JE, Goldberger AL, Mayer-Kress G, Ideker RE. Chaos in the Heart: Implications for Clinical Cardiology. Nature Biotechnology 1990;8(11):1018–24.
View Article
Google Scholar

[144] View Article

[145] Google Scholar

[ref51] 51. Ahn J, Hogan N. Long-Range Correlations in Stride Intervals May Emerge from Non-Chaotic Walking Dynamics. PLoS ONE. 2013;8(9):e73239. pmid:24086274
View Article
PubMed/NCBI
Google Scholar

[147] View Article

[148] PubMed/NCBI

[149] Google Scholar

[ref52] 52. Kalman RE. A new approach to linear filtering and prediction problems. Journal of Fluids Engineering. 1960;82(1):35–45.
View Article
Google Scholar

[151] View Article

[152] Google Scholar

[ref53] 53. Faber NM. A closer look at the bias—variance trade-off in multivariate calibration. Journal of Chemometrics. 1999;13(2):185–92.
View Article
Google Scholar

[154] View Article

[155] Google Scholar

[ref54] 54. Green RL, Kalivas JH. Graphical diagnostics for regression model determinations with consideration of the bias/variance trade-off. Chemometrics and Intelligent Laboratory Systems. 2002;60:173–88.
View Article
Google Scholar

[157] View Article

[158] Google Scholar

[ref55] 55. Wei K, Körding K. Relevance of error: what drives motor adaptation? Journal of Neurophysiology. 2009;101(2):655–64. pmid:19019979
View Article
PubMed/NCBI
Google Scholar

[160] View Article

[161] PubMed/NCBI

[162] Google Scholar

[ref56] 56. Marko MK, Haith AM, Harran MD, Shadmehr R. Sensitivity to prediction error in reach adaptation. Journal of Neurophysiology. 2012;108(6):1752–63. pmid:22773782
View Article
PubMed/NCBI
Google Scholar

[164] View Article

[165] PubMed/NCBI

[166] Google Scholar

[ref57] 57. Ahn J, Hogan N. Improved Assessment of Orbital Stability of Rhythmic Motion with Noise. PLoS ONE. 2015;10(3):e0119596. pmid:25798610
View Article
PubMed/NCBI
Google Scholar

[168] View Article

[169] PubMed/NCBI

[170] Google Scholar

[ref58] 58. Viviani P, Flash T. Minimum-jerk, two-thirds power law, and isochrony: converging approaches to movement planning. Journal of Experimental Psychology: Human Perception and Performance. 1995;21(1):32. pmid:7707032
View Article
PubMed/NCBI
Google Scholar

[172] View Article

[173] PubMed/NCBI

[174] Google Scholar

[ref59] 59. Sternad D, Schaal S. Segmentation of endpoint trajectories does not imply segmented control. Experimental Brain Research. 1999;124(1):118–36. pmid:9928796
View Article
PubMed/NCBI
Google Scholar

[176] View Article

[177] PubMed/NCBI

[178] Google Scholar

[ref60] 60. Schaal S, Sternad D. Origins and violations of the 2/3 power law in rhythmic three-dimensional arm movements. Experimental Brain Research. 2001;136(1):60–72. pmid:11204414
View Article
PubMed/NCBI
Google Scholar

[180] View Article

[181] PubMed/NCBI

[182] Google Scholar

[ref61] 61. Ivanenko YP, Grasso R, Macellari V, Lacquaniti F. Two-thirds power law in human locomotion: role of ground contact forces. Neuroreport. 2002;13(9):1171–4. pmid:12151763
View Article
PubMed/NCBI
Google Scholar

[184] View Article

[185] PubMed/NCBI

[186] Google Scholar

[ref62] 62. Maoz U, Portugaly E, Flash T, Weiss Y, editors. Noise and the two-thirds power law. Advances in Neural Information Processing Systems; 2005.

Figures

Abstract

Introduction

Methods

A Simple Iterative Model

Testing the Validity of the Conventional Least-Square Regression

Quantification of the Bias and Development of Improved Estimation: Adjusted Yule-Walker (AYW) Method

Results

Demonstration of the Bias from Least-Square Estimation

Analytical Derivation of the Bias from the Yule-Walker Equation

Equation for the Improved Estimation

Discussion

Necessity of Improved Estimation Method

Insensitivity of Bias to Noise Magnitude and Distribution

Considerations and Limitations

A Representative Problem for the Influence of Noise on Data Analysis

Author Contributions

References