Estimating Reconstruction Error due to Jitter of Gaussian Markov Processes

Estimating Reconstruction Error due to Jitter of Gaussian Markov Processes
Mudassir Javed*
and Dawood Shah
Department of Electrical Engineering
College of Electrical and Mechanical Engineering,
National University of Sciences and Technology, Islamabad, Pakistan.
Email: mudassir.javed16@ee.ceme.edu.pk, dawoodshahee@gmail.com.
*
Corresponding Author
Abstract—This paper presents estimation of reconstruction error due to jitter of Gaussian Markov Processes. Two samples are considered
for the analysis in two different situations. In one situation, the first sample does not have jitter while the other one is effected by jitter.
In the second situation, both the samples are effected by jitter. The probability density functions of the jitter are given by Uniform
Distribution and Erlang Distribution. Statistical averaging is applied to conditional expectation of random variable of jitter. From that,
conditional variance is obtained which is defined as reconstruction error function and by knowing that, the reconstruction error of a
Gaussian Markov Process is determined.
Keywords—Jitter, Uniform Distribution, Erlang Distribution, Gaussian Markov Process, Probability Density Function.
I. INTRODUCTION
A Gaussian Markov process is the process which satisfies the
requirements of both Gaussian and Markov processes. These are
the stochastic processes that are desired to be recovered. While
variation in the precise location of sampling instants Tn is
referred to as Jitter. Thus the sampling instants become random
too.
The work on the reconstruction problem of a function was
first done in the year 1957 by A.V. Balakrishnan [1] in which he
tried to reconstruct a Wide Sense Stationary process. Later on in
the year 1962, the same author worked on the reconstruction of
random processes with jitter [2]. The same work was then
continued by Brown [3].
Later on, the work of Balakrishnan and Brown were
generalized by B. Liu and T. P. Stanley. Their work includes
calculation of error bounds introduced because of jitter in the
sampling stage.
Most of the researchers in their articles like W. M. Brown
[3], B. Liu et.al [4], M. Shinagawa et.al [5], Kurosawa et.al [6],
A. Tarczynski et.al [7], C. Feng et.al [8], B. Atakan [9], E. Van
der Onderaa and J. Renneboog [10], P. Marziliano and M.
Vatterli [11], N. Da Dalt, et.al [12] have claimed that the best
way to reconstruct from samples in the presence of jitter in
sampling instants is the way so defined in [2]. In all of the stated
research articles, the signal was defined to be real and stationary
in the wide sense and its power spectral density (PSD)
disappears outside the range of angular frequency (-ωb, ωb)
whereas ωb = 2πW.
There are some principal drawbacks of majority mentioned
publications:
1) The number of samples is equal to infinity,
2) The information about the probability density function
(pdf) of the sampled process is not used,
3) All samples have the same jitter distribution.
In this article Conditional Expectation is referred to as ideal
algorithm for the reconstruction of Gaussian processes. With
this approach being applied, new aspects of the problem are
investigated:
1) The number of samples are arbitrary.
2) The covariance function of the random process is taken
into use, without any concern of its power spectral density
being bandlimited or not.
3) The reconstruction error function is represented on the
whole time domain, thus a greater detail of the reconstruction
error of a process between any sampling intervals is
obtained.
II. PROPOSED METHODOLOGY
Let a non-stationary random process x(t) which is discretized in
time instants τ = (τ1, τ2, …, τN). Thus we obtain a multiple
samples of the process x(τ), where the number of samples N and
sampling instants τ are arbitrary. And so we get a new random
process having conditional probability density function f(x),
and central moments depending upon each sample value.
𝑓(𝑥(𝑡)|𝑋, 𝜏) = 𝑓(𝑥(𝑡)|𝑥(τ1), x(τ2), … , x(τ 𝑁)), (1)
In case of arbitrary pdfs, there exist a statistical approach
referred to as conditional expectation which promises minimum
error of estimate. Following this approach, the conditional
expectation also known as the conditional mean function is used
as reconstruction function while conditional variance function
is used as reconstruction error function. In case of Gaussian
2017 2nd
International Electrical Engineering Conference (IEEC 2017)
May 19 th
-20 th
, 2017 at IEP Centre, Karachi, Pakistan
1

processes, the covariance function 𝐾(τ𝑖, τ𝑗) is widely used. The
conditional expectation and variance of a non-stationary
Gaussian process are given below:
𝑚̃(𝑡) = 𝑚(𝑡) + ∑ ∑ 𝐾𝑥
𝑁
𝑗=1 (𝑡, τ𝑖)𝑎𝑖𝑗[𝑥(τ𝑖) − 𝑚(τ𝑗)]𝑁
𝑖=1 (2)
𝜎̃2(𝑡) = 𝜎(𝑡) − ∑ ∑ 𝐾𝑥(𝑡, τ𝑖). 𝑎𝑖𝑗 𝐾𝑥(τ𝑗, 𝑡)𝑁
𝑗=1
𝑁
𝑖=1 (3)
Whereas the term 𝑎𝑖𝑗 represents each term of the inverse
covariance matrix:
𝐴 = [
𝐾𝑥(τ1, τ1) ⋯ 𝐾𝑥(τ1, τ 𝑁)
⋮ ⋱ ⋮
𝐾𝑥(τ 𝑁, τ1) ⋯ 𝐾𝑥(τ 𝑁, τ 𝑁)
]
−1
(4)
The expressions (2) and (3) depend on the current number of
sample j, on the total number of samples N, on the set of
arbitrary sampling instants 𝜏𝑖, on the covariance moment
among process sections at the instants 𝜏𝑖 and 𝜏𝑗 and on the
covariance moment 𝐾𝑥(𝑡, τ𝑖) among the current sections of the
time t.
Conditional expectation is used directly in case of
known or definite sampling instants. But in case of jitter, when
the sampling instants are not known, but random then statistical
average with respect to the pdf of jitter is applied to the above
stated rule in order to estimate the conditional functions.
With the help of this rule, and having the number of
samples N finite, the reconstruction of a random process inside
the sampling region and outside the sampling region can be
described.
III. JITTER
In this paper the sampling instants i.e.𝜏̃ are random because
of presence of jitter. Also the jitter present in each sample isn’t
necessary to be the same as it can variate from sample to sample.
Thus the pdf of jitter in each sample differ from each other. The
pdfs of the jitter considered in this article are uniform or
rectangular distribution as shown in figure 1,
𝑓(𝜏̃) = {
0 , 𝜏̃ < 𝑎
1
𝑏−𝑎
, 𝑎 ≤ 𝜏̃ ≤ 𝑏
0, 𝜏̃ ≥ 𝑏
, (5)
Figure 1. pdf of Uniform Distribution
and the Erlang distribution [13]. Erlang Distribution was
developed by A. K. Erlang to examine the number of telephone
calls which might be made at the same time to the operators of
the switching stations. This work on telephone traffic
engineering has been expanded to consider waiting times in
queueing systems in general. The distribution is now used in the
fields of stochastic processes and of biomathematics. It can be
used to model the time to complete n operations in series, where
each operation requires an exponential period of time to
complete. The pdf of Erlang distribution is shown in figure 2 and
its mathematical form is given below:
𝑓(𝜏̃) = 𝜆 . 𝑒−𝜆𝜏̃
.
(𝜆𝜏̃) 𝑘−1
(𝑘−1)!
, 𝜆, 𝜏̃ ≥ 0 (6)
Erlang distribution has two parameters as seen in the pdf, k
and 𝜆. The k parameter is referred to as the shape parameter
while the 𝜆 is referred to as the rate parameter. When the value
of shape parameter k equals 1, it takes the form of exponential
distribution.
An alternative, but equivalent, parametrization uses the scale
parameter µ which is the reciprocal of the rate parameter (i.e. µ
=1/ 𝜆).
Figure 2. pdf of Erlang Distribution
Two different cases are considered. 1) When the first sample
is definite while the other sample has jitter i.e. the position of the
second sample is random 𝜏̃. 2) When both the samples have jitter
i.e. both the sampling instants are random variables.
For the first case, the average reconstruction function which
is the average conditional expectation depends upon the pdf of
the jitter present in the second sample 𝑓(𝜏̃), so we get:
〈𝜎̃2(𝑡)〉 = ∬ 𝜎̃2(𝑡). 𝑓(𝜏̃2)𝑑𝜏̃2 (7)
For the second case, the average conditional expectation
depends upon the pdfs of jitter of both samples i.e. 𝑓(𝜏̃1), 𝑓(𝜏̃2).
Thus the average reconstruction error is calculated as follows:
〈𝜎̃2(𝑡)〉 = ∬ 𝜎̃2(𝑡). 𝑓(𝜏̃1) 𝑓(𝜏̃2) 𝑑𝜏̃1 𝑑𝜏̃2 (8)
In this article, the random process used throughout is obtained at
the output of a low pass RC filter with white noise as the input.
2

Thus a Gaussian Markov process well defined by the covariance
function given below is obtained:
𝐾(τ𝑖, τ𝑗) = 𝜎2
. 𝑒−𝛼|τ 𝑖,τ 𝑗|
(9)
Whereas α=1/RC. In the said filter, the value of α is assumed to
be 1, so that the covariance time has a value 1.
Figure 3. Reconstruction Error Function with both samples
having no Jitter.
IV. EXAMPLES
The conditional expectation also referred to as conditional mean
defines the conditional variance which is defined as the
reconstruction error function.
Here, in all the examples considered, the two sampling
instants taken are at 0 and 0.5. While the covariance time used
is 1.
The reconstruction error function without any jitter as
defined by (3) is represented in the figure 3. From the figure it is
visible that the function is defined on the whole time domain.
Also it can be seen that the reconstruction error at sampling
points have value 0 as at these points the value of the process is
exactly known. While exactly in the center of both the sampling
intervals, the value of the reconstruction error is maximum
which is obvious as there is greater uncertainty about the exact
value of the process at this instant.
On the basis of (7), (8) a set of examples are presented to
clearly show the effect of jitter during reconstruction. These
expressions describe the average reconstruction function with
random sampling intervals.
The multiple scenarios we have considered are as follows:
 One sample with random sampling intervals.
o The jitter of the sample described by uniform
pdf.
o The jitter of the sample described by Erlang
Distribution function.
 Both the samples with random sampling intervals with
jitter defined by Uniform pdf.
o Both samples having same jitter.
o Both samples with different jitter.
A. When the second sample has random sampling instant i.e
it has jitter which is defined by Uniform Probability
Distribution
In this case, the process is carried out with the help of average
reconstruction error function (7). The second sampling instant
is effected by jitter represented by ε2, thus the sampling instant
becomes:
𝜏̃2 = 𝜏2 + 𝜀2 (10)
The random variable 𝜀2 in this case is defined by Uniform
Probability Density Function and is limited between the interval
of 0.05seconds on either side of the sampling point. So 𝜏̃2
becomes as shown in (11):
[𝜏̃2 − 0.05, 𝜏̃2 + 0.05] = [0.45,0.55] (11)
The reconstruction error function is shown in figure 4. In the
figure it can be seen that the reconstruction estimate is erroneous
only in the region where there is effect of jitter i.e. between 0.45
and 0.55. As compared with the original samples without jitter,
the jitter in here causes 5% error in the sample.
B. When the jitter of the second sample is defined using Erlag
Distribution
In this scenario, the jitter in the second sample is described
using Erlang Distribution.
For instance four different distributions are used
having the following values.
Figure 4. The reconstruction error function when one sample
is without jitter and the other sample have the presence of
jitter defined by Uniform pdf.
3

𝑘 = 1, 𝜆 = 85
𝑘 = 2, 𝜆 = 85
𝑘 = 2, 𝜆 = 65
𝑘 = 2, 𝜆 = 45
The pdfs of the jitter is given in figure 5. The reconstruction
error function is shown in figure 06. Where it is clearly visible
that the maximum error between all the four different samples
is different in both values and position.
It is because of the difference between average
distances of the samples. Also, with the increase in this
difference, the maximum error which lies exactly in middle of
the two sampling points where the uncertainty is maximum
increases. Also, from the figure, it is clear that the error is
effecting the value at the sampling instant too with the variance
of the jitter.
C. When both the samples have same jitter defined by
Uniform Distribution Function
In this case, the jitter is present in both the samples, defined by
Uniform Distribution. The jitter is exactly the same. In order to
clarify the case in a more better form, three different jitter
widths are taken with the following limits.
o [-0.025, 0.025] i.e. total width = 0.05.
o [-0.05, 0.05] i.e. total width = 0.1.
o [-0.1, 0.1] i.e. total width = 0.2.
These widths are represented using blue, green and red dotted
line in figure 7.
Figure 7 gives reconstruction error function for each of the jitter
width in both samples. From this example, it is clear that the
error increases with the increase in jitter width as it makes the
exact value of sample at the sampling instant more uncertain.
Figure 5. Four different Jitters represented by Erlang
distribution of different parameters
Figure 6. Reconstruction Error Function when one sample has
jitter defined by Erlang distributions of different parameters as
shown in figure 5.
D. When both the samples have jitter of different widths
defined by Unifrom Distribution Function
And lastly the case when both the samples have different jitter
widths, well described by Uniform Probability Density
Function is studied. The jitter width of the first sample is [-
0.025, 0.025] with total width of 0.5 while the width of the jitter
of the other sample is [-01, 0.1] of total width 0.2.
The reconstruction error function is given in figure 8.
Where it can be seen that error in the sample with smaller jitter
width is small as compared to the other one of larger jitter
width.
From this example, it can be concluded that the
characterization of jitter of samples with different distributions
is also possible.
4

Figure 7. Reconstruction Error Function when both the
samples have jitter of same width at a time. This figure depicts
three different cases at a time.
V. CONCLUSION
From this paper, it can be concluded that the conditional
expectation or conditional mean rule permit us to depict the
sampling-reconstruction process for Gaussian Markov
processes where there is or not the presence of jitter in the
sampling instants. With this, it is conceivable to study that how
the reconstruction of the signal from its samples is influenced
on the whole time domain in presence of jitter. It was shown
that it is conceivable to bring the investigation under different
distinctive cases, for instance: when there is no jitter; when just
a single sample has jitter; when both samples have jitter with
the same or different pdf of different distribution functions.
Shortly, the reconstruction error increases with the increase in
jitter.
Figure 8. Reconstruction Error Function when both the
samples have jitter of different widths represented by Uniform
Probability Density Function.
REFERENCES
[1] A. V. Balakrishnan, “A note on the sampling principle for continous
signals,” IRE Trans. on Information Theory, vol. IT-3, pp. 143-146,
1957.
[2] A. V. Balakrishnan, “On the problem of time jitter in sampling”, IRE
Trans. on Information Theory, vol. IT-8, pp. 226-236, April 1962.
[3] W. M. Brown, “Sampling with random jitter”, J. SIAM, Vol.11, pp. 460-
473, June 1963.
[4] B. Liu, T. P. Stanley, “Error bound for jittered sampling”, IEEE
Transactions on Automatic Control, vol. 10. no. 4. October, 1965.
[5] M. Shinagawa, Y. Akazawa, T. Wakimoto, “Jitter analysis of highspeed
sampling systems,” IEEE J. Solid-State Circuits, vol. 25, no. 1, pp. 22
224, 1990.
[6] N. Kurosawa, H. Kobayashi, H. Kogure, T. Komuro, H. Sakayori,
Sampling clock jitter effects in digital-to-analog converters,”
Measurement, vol. 31, pp. 187-199, 2002.
[7] A. Tarczynski, N. Allay, “Spectral analysis of randomly sampled signals:
suppression of aliasing and sampler jitter,” IEEE Transations on Signal
Processing, vol. 52 , no. 12, pp. 3324-3334, 2004.
[8] C. Feng, Z. Jia-xiong, Y. Ji-hao, “Frequency spectrum analysis of
sampling clock jittersystem based on LabVIEW,” International
Conference on Computational Problem-Solving (ICCP), pp. 325-327,
2012.
[9] B. Atakan, “On exploiting sampling jitter in vehicular sensor networks,”
IEEE Transactions on Vehicular Technology, vol. 63, no. 1, pp. 403-407,
2014.
[10] E. Van der Onderaa and J. Renneboog, “Some formulas and applications
for nonuniform sampling of bandwidth-limited signals,” IEEE Trans.
Instrum. Meas., vol. 37, no. 3, pp. 353-357, 1998.
[11] P. Marziliano and M. Vatterli, “Reconstruction of irregularly sampled
discrete-time bandlimited signal with known sampling locations,” IEEE
Transactions on Signal Processing, vol. 48, no. 12, pp. 3462- 3471, 2000.
[12] N. Da Dalt, M. Harteneck, C. Sander, and A. Wiesbauer, “On the jitter
requirements of the sampling clock for analog-to-digital converters,”
IEEE Trans. Circuits Syst. 1, vol. 49, no. 9, pp. 1354- 1360, 2002.
[13] Xycoon, Statistics - Econometrics - Forecasting, Office for Research
Development and Education, http://www.xycoon.com/ erlang.htm
5

Estimating Reconstruction Error due to Jitter of Gaussian Markov Processes

More Related Content

What's hot (17)

Similar to Estimating Reconstruction Error due to Jitter of Gaussian Markov Processes (20)

Recently uploaded (20)

Estimating Reconstruction Error due to Jitter of Gaussian Markov Processes