You are on page 1of 12

Statistics and Computing 15: 341352, 2005

C 2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.




Accurate ARL computation for EWMA-S 2


control charts
SVEN KNOTH
Advanced Mask Technology Center, Postfach 110161, D-01330 Dresden

Received September 2004 and accepted June 2005

Originally, the exponentially weighted moving average (EWMA) control chart was developed for
detecting changes in the process mean. The average run length (ARL) became the most popular
performance measure for schemes with this objective. When monitoring the mean of independent
and normally distributed observations the ARL can be determined with high precision. Nowadays,
EWMA control charts are also used for monitoring the variance. Charts based on the sample variance
S 2 are an appropriate choice. The usage of ARL evaluation techniques known from mean monitoring
charts, however, is difficult. The most accurate methodsolving a Fredholm integral equation with
the Nystrom methodfails due to an improper kernel in the case of chi-squared distributions. Here,
we exploit the collocation method and the product Nystrom method. These methods are compared
to Markov chain based approaches. We see that collocation leads to higher accuracy than currently
established methods.
Keywords: control charts, monitoring variance, Fredholm integral equations, collocation, product
Nystrom method

1. Introduction
Statistical Process Control (SPC) refers to statistical methods
that are used for monitoring certain process parameters. Usually, it is assumed that the observations are independent and
normally distributed. In this case, mean and variance are monitored. In SPC literature one can find many papers dealing with
detecting changes in the mean level of normal data. The considered procedures were introduced about fifty years ago or earlier,
the most popular being the original Shewhart chart (Shewhart
1931), the SPRT (sequential probability ratio test) connected
CUSUM (cumulative sum) scheme (Page 1954), the ShiryaevRoberts scheme (Girshick and Rubin 1952) as a quasi-Bayesian
scheme, and the EWMA control chart (Roberts 1959). The more
recent examples were developed to detect single changes as
quickly as possible for a given bound of false alarms. Most
frequently, the so-called average run length (ARL) is taken as
a performance measure. The ARL is the expected number of
observations (or subgroups) until a signal is raised, assuming
a constant parameter situation. Naturally, the ARL should be
large for an undisturbed process, but small for shifted mean,
scale or further changed parameters. In the literature further
measures are assessed, which could be more relevant than the
ARL. In any case, in dealing with these more complex quantiC 2005 Springer Science + Business Media, Inc.
0960-3174 

ties one observes the same difficulties as described here for the
ARL.
The current paper investigates exponentially weighted moving average (EWMA) control charts for detecting scale changes,
that is, shifts in the variance parameter of normally distributed
data. Recall that the EWMA mean control chart experienced,
after introduction by Roberts (1959), a kind of rebirth with
Hunter (1986), Waldmann (1986), Crowder (1987), and eventually with Lucas and Saccucci (1990). The first papers dealing with EWMA control charts for monitoring the variance
came from Wortham and Ringer (1971) and, a decade later,
Sweet (1986), where only rough recommendations for the control chart design parameters were given, without evaluation measures like the ARL. Thereafter, Domangue and Patch (1991),
Crowder and Hamilton (1992), MacGregor and Harris (1993),
Chang and Gan (1994), Tewes (1995), and Mittag et al. (1998)
investigate EWMA control charts based on the sample variance, standard deviation and the natural log of the sample variance. Additionally, in papers by Srivastava (1994), Gan (1995),
and Reynolds Jr. and Stoumbos (2001) the joint monitoring of
mean and variance with EWMA schemes was considered. In
Lowry et al. (1995), and later in Acosta-Meja et al. (1999),
and Acosta-Meja and Pignatiello Jr. (2000), extensive comparison studies for a large number of control charts for monitoring

342

Knoth

the variance were performed, though other papers dealing with


variance monitoring by applying CUSUM schemes are available. Here we consider, however, the numerical analysis of the
ARL of EWMA control charts, which is more subtle than that
of CUSUM schemes.
It is interesting that EWMA smoothing of ln S 2 (the natural log) reached great popularity. There are different reasons
for this phenomenon. Mainly, it is the transition from a scalechange model to a level-change model that motivates the lntransformation. Now, changes in the scale do not affect the variance of the new chart statistic. Furthermore, ln S 2 is nearly
normally distributed and so one can transfer the known results from the EWMA mean control chart. Finally, the distribution of ln S 2 is more symmetric than that of S 2 , so that twosided control charts are simpler to design. There is also a further reason, which is rather a technical one. While for ln S 2
the computational methods from the mean case can be used
without any restriction, the original S 2 statistic causes trouble. This paper provides a feasible solution of that issue. Thus,
we pick EWMA control charts based just on the sample variance S 2 in order to obtain an appropriate variance monitoring
scheme.
Let {X i j } be a sequence of subgroups of independent and
normally distributed data. Each subgroup i consists of n observations X i1 , . . . , X in . The subgroup size n can be equal to 1
assuming that the mean = 0 is known or is simultaneously
monitored. For n > 1 the mean can be unknown and changing
between samples. We consider the following change point model
for the variance 2 :
 2
i <m
,
2 = 02
.
2
1 = 0 , i m
The parameters 02 and 12 denote the in- and out-of-control
values of the monitored variance. Then, based on the sequence
of sequentially observed data we want to detect the unknown
integer value m, which is called the change point. Similar to
the EWMA control chart for monitoring the mean we construct our chart statistic by exponentially weighted smoothing of an appropriate estimator. For n > 1 and unknown or
changing
Si2 =

n
1 
(X i j X i )2 ,
n 1 j=1

n
1
X i =
Xi j ,
n j=1

while for n 1 and known = 0


n
1
Si2 =
(X i j 0 )2
n j=1

is exploited. Here we consider only control charts, which exclusively monitor the variance. The accurate ARL calculation of
EWMA- X -S 2 control charts will be published in a later paper.
The results of the current paper can be applied for using S2 as an
omnibus chart statistic in order to monitor mean and variance at
the same time, as in Domangue and Patch (1991). Gan (1995),
however, demonstrated that these omnibus charts are evidently

dominated by the combined control chart schemes like, e. g., by


EWMA- X -S 2 control charts.
By writing Vi as a dummy for the related variance estimator,
the EWMA sequence with smoothing constant (0, 1] is
given by:
Z 0 = z 0 = 02 ,

(1)

Z i = (1 ) Z i1 + Vi ,

i 1.

(2)

We are mostly interested in detecting variance increases. Hence,


a signal after processing subgroup i is given if for the first time



2 2
2
Z i > 0 + c
.
2 k 0
In Section 4 we consider also two-sided EWMA-S 2 charts. The
so-called critical value c and the EWMA parameter are determined to ensure a certain chart performance. The normalization
factor in the above stopping rule is equal to the asymptotic standard deviation of Z i . Recall that (using the independence)
Z i = (1 ) 02 +

i


(1 )i j V j ,

j=1

Var (Z i ) =

(1 (1 )2i )Var (Vi ).


2

The normalized variance estimator k Vi / 2 (k = n 1 for S 2


and k = n for S2 ) is chi-squared distributed with k degrees of
freedom. If we apply S2 and the actual mean differs from 0
used in S2 , then the normalized statistic k Vi / 2 follows a noncentral chi-squared distribution. To be precise, the normalized
version n S2 / 2 follows a noncentral chi-squared distribution
2
n,q
with n degrees of freedom and the noncentrality parameter
q = n (0 )2 / 2 . The probability density and the cumulative
2
distribution function of n,q
are numerically more demanding
2
than for n . Finally, by using the non-central instead of the central chi-squared distribution we can also assess the ARL performance of the omnibus charts given in Domangue and Patch
(1991). Later, we consider only configurations, which lead to
the central chi-squared distribution. The problems demonstrated
and the solutions applied afterwards are valid for the non-central
case as well. For the central chi-squared statistic we conclude
that
Var (Vi ) =

24
,
k

Var (Z i )

24
2 k

for i .

In the above stopping level we utilize the undisturbed process


variance 2 = 02 . More formally, the EWMA-S 2 control chart
is characterized by the stopping time





2 2
2
L = inf i IN : Z i > 0 + c
.
2 k 0

(3)

Accurate ARL computation for EWMA-S 2 control charts


Now, we are ready to define more rigorously the ARL by denoting E m () as the expectation for a fixed change point m. Thus,

E (L), in-control case 2 = 02
ARL =
.
E 1 (L), out-of-control case 2 = 12 = 02
As previously mentioned, we evaluate these expectations only
for the two special casesthe ARL scenario, where 2 is constant for the whole sequence of data.
In the next section ARL computation techniques of mean
EWMA charts are revisited and drawbacks in applying the
most precise numerical methods to variance EWMA charts are
demonstrated. Then in the following section the collocation
method and the product Nystrom method are presented, which
allow more accurate results. Finally, the new methods are compared with established ones by re-analyzing results of previous
papers dealing with EWMA-S 2 charts, with a comparative study
of EWMA-S 2 and EWMA-ln S 2 in terms of the ARL.

2. State-of-the-art of ARL computation


for EWMA control charts
Naturally, there are various ways of reviewing techniques for
computation EWMA ARLs (or CUSUM and similar control
charts). Here, we want to start at the well-known integral equation, which characterizes the ARL as function of the starting
value z 0 in (1). The first to consider integral equations in the
context of evaluating the performance of control charts was Page
(1954), who introduced the CUSUM chart. Crowder (1987) provided the related EWMA ARL integral equation. If L(z) is the
ARL of an EWMA scheme (1), (2), and (3) with z 0 = z, then
the following holds:


 cu
1
x (1 ) z
L(z) = 1 +
L(x) f
d x, z [cl , cu ].

cl
(4)

Thus, for the stopping


 (3) the integral limits are cl = 0
 time

2 2
. Note that for the ease of preand cu = 02 + c 2
k 0
sentation we distinguish between the critical values cl and cu ,
and the critical limits cl and cu as the resulting threshold values
for the considered chart statistic Z i . Some authors employ the
final limits, others the normalizing approach. The function f ()
stands for the density of Vi . Thus, f () is the probability density function (pdf) of a chi-squared distribution. In the original
case consideredby Crowder (1987) the corresponding values are

cl = cu = c 2
0 and f () the pdf of a normal distribution. Because this is the most popular and best investigated approach, our review will be performed for normal mean EWMA
control charts. Crowder (1987) solves the integral equation (4)
by applying the Nystrom method and Gaussian quadrature.
The main idea of the Nystrom method is simple and leads to
excellent results for the considered situation. We replace the integral by an appropriate quadrature (Crowder takes a GaussianLegendre one), consider the equation only at the related nodes

343
of the quadrature and solve the resulting linear equation system. More formally, from the quadrature rule we obtain nodes
z 1 , z 2 , . . . , z N and weights w1 , w2 , . . . , w N for given N IN.
The resulting linear equation system looks like:


N

z j (1 )z i
1
L(z i ) = 1 +
wj f
L(z j ),

j=1
i = 1, 2, . . . , N .

(5)

After solving the linear equation system we can evaluate L(z)


for arbitrary z [cl , cu ] by replacing z i by z in (5).
For details we refer to the literature concerned with the numerical solution of Fredholm integral equations of the second kind,
like Atkinson (1976) or Hackbusch (1995). Besides, instead of
Gaussian quadrature rules we could plug in the Simpson rule
which is based on piecewise interpolation by quadratic splines
(and integrates exactly up to degree 3) and provides sufficient
accuracy.
It is surprising that the most popular ARL computation
method, the Markov chain approximation originally introduced
for CUSUM by Brook and Evans (1972), is nothing else than
the Nystrom method equipped with the midpoint rule, which is a
very simple quadrature rule. A more comprehensive study can be
found in Champ and Rigdon (1991). They conclude that the more
sophisticated methods like Gaussian or Simpson quadrature
based approaches should be preferred. It is interesting that Page
employed the midpoint rule for solving the ARL integral equation in 1963. In order to demonstrate this connection we apply
the midpoint rule to equation (4). Assuming the decomposition
[cl , cu ] = [cl , cl + w) [cl + w, cl + 2w) . . . [cl + N w, cu ]
and using the midpoints z i = cl + iw w/2 , i = 1, 2, . . . , N
we can approximate equation (4) with:


 cl + jw
N

1
x (1 ) z i
L(z i ) = 1 +
L(z j )
f
dx

cl +( j1)w
j=1
= 1+

N


qi j L(z j ),

j=1


cl + jw (1 )z i
qi j = F



c + ( j 1)w (1 )z i
F l
.

Here, F() denotes the cumulative distribution function of the


smoothed statistic. By comparing it with the Markov chain approach like in Lucas and Saccucci (1990) we observe that both
attempts coincide. Hence, with respect to results of the numerical
treatment of Fredholm integral equations, we might prefer the
more precise methods based on Gaussian or Simpson quadrature.
The attempt of Waldmann (1986), which relies on appropriate
quadrature rules as well, provides the corresponding results for
higher moments of L, quantiles, and, of course, the whole survival function P(L > n) with the same precision as the Nystrom
solution of the above integral equation if the same quadrature

344
rule is taken. Eventually, higher moments of the stopping time
L were also computed by Crowder (1987)
With the following example the precision of the competing
concepts in case of monitoring the mean will be demonstrated.
Let the EWMA smoothing constant = 0.1, which is a common
choice for fast detection of small mean shifts. The critical value
c = 2.814 310 is chosen in order to get an in-control ARL
(E (L)) of 500. The methods under considerations are
Nystrom method with Gauss-Legendre quadrature, N nodes,
Nystrom method with Simpson rule, N nodes,
Markov chain approach called Brook/Evans method (Brook
and Evans 1972), N states (nodes),
Interpolation A R L(N ) = a + b/N + c/N 2 for the Brook/
Evans method at matrix dimensions (node counts) N 20,
N 10, and N similar to Brook and Evans (1972) who chose
5, 10, and 15, and using A R L() = a as final approximation,
OLS fit A R L(N ) = a + b/N + c/N 2 for the Brook/Evans
method at matrix dimensions N 32, N 24, N 16, N 8
and N like in Lucas and Saccucci (1990), who employed 51,
59, . . . 83, and again A R L() = a,
Collocation like in Gianino et al. (1990)not a frequently
applied approach, but the final choice in this paper for the
EWMA-S 2 .
There are, of course, more methods. Here, the most popular
ones plus collocation are taken. Moreover, the most precise one
is amongst them. In Fig. 1 the results of the above methods
for odd N (Simpson rule demands odd node values) up to 301
are plotted. Tracking this figure we conclude that the most precise method is the Gauss-Legendre Nystrom method followed
by the collocation method and Simpson rule based Nystrom.
The regression methods provide considerably good precision,
while the pure Markov chain approach gives the weakest preci-

Fig. 1. Precision of different ARL computation methods for the incontrol ARL of a two-sided EWMA chart monitoring the mean of normally distributed data, the true value is 500 (c = 2.814 310) and the
smoothing value = 0.1

Knoth
sion. However, all methods are sufficient accurate. Remark that
Lucas and Saccucci (1990) do not use the simpler regression
interpolation of Brook and Evans (1972) despite its better performance. Finally, for smaller or larger we can observe the
same situation. Hence, we learn from this figure that we have to
prefer the Gauss-Legendre Nystrom method.
Unfortunately, that method fails for non-smooth kernels,
which appear in the case of EWMA control charts based on
chi-squared distributed statistics like S 2 . The most unpleasant
case occurs for one degree of freedom. Now, the kernel is unbounded at zero, that is, for x = (1 ) z in (4). Reconsidering
the results of Domangue and Patch (1991) provides Fig. 2. Recall that the sample size is n = 1 and the statistic S2 = (X 0 )2
is used. For Fig. 2 the smoothing value is set to = 0.025 and
the in-control ARL value is 250. The related critical value results in c = 1.661 865 (in Domangue and Patch (1991) the value
1.663 is given). Applying the above methods we have to realize
that Gaussian-Legendre and Simpson Nystrom method completely break down. Additionally, the regression based attempts
result in heavily oscillating values, which are rather useless.
The mere Markov chain approach only provides useful results
which are plotted for all matrix dimensions between 5 and 301.
The accuracy, however, is weak and the non-monotone behavior
does not allow a really precise approximation. Nevertheless, in
Domangue and Patch (1991) the Markov chain approach is employed. In the appendix of their paper the authors state that they
use a dimension N 200. From Fig. 2 we see the accuracy we
might expect. The methods developed in the current paper are
added in order to get an idea what could be attained with more
appropriate attempts.
Reynolds Jr. and Stoumbos (2001) experience the same problem for the same EWMA control chart as above, with the exception that they use a reflecting border at 02 (for details see

Fig. 2. Precision of the Markov chain method, product Nystrom and


collocation for the in-control ARL of a one-sided EWMA chart monitoring the variance of normally distributed data, the true value is 250,
the smoothing value = 0.025, sample size n = 1, statistic S2 , and
the critical value is c = 1.661 865

Accurate ARL computation for EWMA-S 2 control charts


Section 4). They utilize a matrix dimension of 400. This example is reconsidered in Section 4.
For one more degree of freedom we obtain the exponential
distribution, which leads to a bounded, but discontinuous (at
zero again) kernel. Now, there are some exact results by Gan
and Chang (2000). Unfortunately, for our one-sided EWMA
scheme without reflecting border it can be used only as a further
approximation (with considerable precision).
For larger numbers of degrees of freedom the kernel is
bounded and continuous, but still vanishes for negative arguments of f (). The latter remains a problem for the Nystrom
method based on the Gauss-Legendre quadrature rule.
It is difficult to attain high accuracy if one tries to apply
Gauss-Legendre quadratures to integrands which are constant
over a certain part of the integration interval. This happens if
we employ the Nystrom method. The main idea of the Nystrom
method is that we replace the integral by a quadrature with nodes
that do not depend on the argument z on the left hand side of
(4). Thus, we apply the quadrature for integrands of that subtle
type for nearly all z. The Simpson rule based approach behaves
slightly better. For the regression attempts the same problem
occurs as above. However, the Simpson rule behaves similarly
to the Markov chain method, especially concerning the monotonicity (see Fig. 3). Generally speaking, the accuracy problem
lessens with increasing number of degrees of freedom. If one
is interested in accurate ARL values for degree numbers larger
than 10, then one is free to use any of the presented methods. In
practice, however, the subgroup sizes used in SPC are usually
smaller than 10.
Let us now recapitulate how this problem is treated in literature. First, the pioneer of applying the integral equation method
for evaluating EWMA ARLs, Crowder, bypasses the problem
by using ln S 2 in Crowder and Hamilton (1992). The new statistic ln S 2 leads to a proper integral kernel. Moreover, Crowder
and Hamilton (1992) use, instead of (2), an EWMA sequence
with a reflecting border at 02 (more details in Section 4) and

Fig. 3. The true ARL value is 250, the smoothing value = 0.18,
sample size n = 5, statistic S 2 , and critical value c = 2.909 223

345
evaluate EWMA performance for sample size 5 (= 4 degrees
of freedom) and for 0.05. They are then able to use the
Gauss-Legendre quadrature again and get high accuracy. And
so do Chang and Gan (1994), Srivastava (1994), Gan (1995),
Acosta-Meja et al. (1999), and others. If EWMA smoothing
of S 2 is considered, then mostly the Markov chain approach
is employed like in Domangue and Patch (1991), Mittag et al.
(1998), Acosta-Meja and Pignatiello Jr. (2000), and Reynolds
Jr. and Stoumbos (2001). Finally, MacGregor and Harris (1993)
use Monte Carlo simulation.

3. Numerical solution of the ARL integral


equation for one-sided (upper)
EWMA-S2 charts
A promising suggestion of the literature concerned with the numerical solution of Fredholm integral equations for dealing with
improper kernels is use of the product Nystrom method. To use
this method we split the improper kernel (see the related chapters in Atkinson (1976) or Hackbusch (1995)) into a smooth part
and a non-smooth part. The latter is now used as a quadrature
weighting function. That means the quadrature weights have to
be recomputed. Unfortunately, we have to do this for each node
because the non-smooth part depends on the considered node z i
in (5). To be more precise, look at our kernel function in (4) for
the statistic Vi with k degrees of freedom:


x (1 )z
f


 k2 1 x(1)z 

exp k 2 2
x(1)z

, x > (1 )z
2  k2
=
(6)
(k/2) 2 k

0,
x (1 )z
where () denotes the complete Gamma function. One problem
we have do deal with is the vanishing part, i. e. for cl x
(1 ) z the kernel function is constantly zero. Additionally,
for k = 1 the power term in the above enumerator reduces
to ( x(1)z
)1/2 , which causes further trouble for x close to

(1 )z. So, we split off




x (1 ) z
f non-smooth


k
x (1 )z 2 1
, x > (1 )z
=

0
, x (1 )z
and use the remaining terms as the smooth part. For k > 1 it is
not necessary to split off the power term. It turns out, however,
that this splitting leads to better results than merely splitting off
the constant zero part.
Unfortunately, the calculation of the new weights wi j is quite
cumbersome, such that, e. g., in Press et al. (2002) instead of
the Gaussian quadrature rule, a generalization of Simpsons
3/8 rule is preferredwhich is again based on a uniform grid.

346

Knoth

Their C code is used here for the practical solution of our problem. Roughly speaking, the precision is similar to the precision of the standard Simpson rule applied to proper kernels.
For all the details see the Section Integral Equations with
Singular Kernels in Press et al. (2002). Because the product Nystrom method is inferior to the following method, we
leave it at that and proceed with more details for the collocation
method.
Usually, one can find the collocation method in textbooks on
numerical solution of integral or differential equations. We have
to fix a basis, that is a set of N independent interpolating functions, and approximate the solution of the integral equation (4) by
a linear combination of these functions. There are two EWMAARL papers on the market (as far as the author of the current
paper knows) which exploit collocation. The first one is Gianino
et al. (1990) who compute the ARL for mean EWMA charts.
The authors use the simple basis 1, z, z 2 , . . . , z N 1 , utilize
L(z)

N


c j z j1 ,

j=1

and determine the unknown constants c j by solving the resulting linear equation system. The latter system is obtained by
evaluating (4) at given nodes z i and plugging in the above approximation on both sides of (4). Gianino et al. (1990) use a
uniform mesh and compute the integrals by using the Simpson
rule. Usually, the collocation method is very fast if the integrals
can be determined exactly when applied to the basis functions.
For the considered case they have to be computed numerically.
The necessary quadrature node number, however, is small. In
Gianino et al. (1990) this number is not mentioned. Remark that
the smaller the smoothing constant the larger this number has
to be. Eventually, the authors were able to obtain the same accuracy as Lucas and Saccucci (1990) (see Fig. 1 also), with only
six basis functions. The collocation based results in Fig. 1 are
obtained by applying Gianino et al. (1990), that is, monomials
and a uniform mesh.
The second and very recent collocation paper is Calzada and
Scariano (2003), who solve the ARL integral equation for a
couple of distributions including the Gamma distribution, to
which chi-squared belongs. They employ, however, collocation
in a very simple way. The basis functions are defined piecewise on a uniform mesh. Each function vanishes outside the
supporting interval, while it is equal to one in this interval. In
this way we obtain a basis of step functions. Then, the authors
illustrate that this collocation leads in fact to the well-known
Markov chain approach. In other words, the midpoint rule again
is utilized. The Markov chain approach is Nystrom and collocation simultaneously. Thus, the Calzada and Scariano (2003)
collocation provides the same accuracy as the Markov chain
approach.
By tuning (see the next paragraphs) the collocation, we obtain an accuracy-efforts relation which is between the GaussianLegendre and Simpson Nystrom methods for the ARL of mean
EWMA control charts.

In what follows, we want to use


instead of the uniform grid the roots of the Chebyshev polynomial of order N , because they provide a better performance
than other grids if the number of basis functions is increased,
instead of the monomials 1, z, z 2 , . . . , z N 1 the Chebyshev
polynomials T j up to order N 1, i. e.
T j (z) = cos( j arccos(z)),
j = 0, 1, . . . , N 1, z [1, 1],
which ensure more stable numerical quadratures than monomials or Lagrange polynomials.
This is done to be able to calculate accurate ARL values for very
small like in Knoth (2004), where the smallest = 0.000 042.
This very small value demands large N .
We start with the EWMA-S 2 chart
by (2) and (3).

 defined

2 2
Recall that cl = 0 and cu = 02 +c 2
. The collocation
k 0
nodes are given by the roots of TN (z) shifted in the interval
[0, cu ]:



c
(2 i 1)
z i = u 1 + cos
, i = 1, 2, . . . , N .
2
2N

We consider for all i = 1, 2, . . . , N Chebyshev polynomials in


[0, cu ], i. e. T j (z) = T j1 ((2 z cu )/cu ) and replace the lower
limit 0 by (1 ) z i (function f () vanishes for x (1 ) z i )
 cu
N
N


1
c j T j (z i ) = 1 +
cj
T j (x)

(1) z i
j=1
j=1


x (1 )z i
f
d x,

where the integrals on the right hand side are determined by


Gauss-Legendre quadrature. Contrary to the quadrature for the
Nystrom method with its fixed node grid, we choose the quadrature grid for any new interval [(1 ) z i , cu ]. Thus, we need
only 20 or less quadrature nodes in order to obtain sufficient
precision. Additionally, substituting y 2 = x (1 ) z i within
these integrals ensures this for odd degrees of freedom also.
Figure 2 illustrates how the new methods work for the nasty
case k = 1. Remember, Domangue and Patch (1991) use a matrix dimension N 200. Based on this ARL approximation they
determine a critical value of c = 1.663 which is in fact 1.662.
The difference is small, but the collocation method provides for
N = 28 already the final in-control ARL value 249.9997, while
the product Nystrom method needs about 2 500 nodes for the
same accuracy. The Markov chain approach does not accomplish such an accuracy with state numbers up to 5 000. Thus,
the product Nystrom method behaves better than the Markov
chain approximation, but does not reach the collocation. Table 1
presents some numbers and CPU times for various matrix dimensions N . On the one hand, the new methods need more time
for fixed N . On the other hand, collocation needs at most 3 hundredths of a second to ensure an accuracy which the others do not

Accurate ARL computation for EWMA-S 2 control charts

347

Table 1. Resulting ARL approximations (upper entry) and CPU times


(lower entry, measured in hundredth seconds on an Athlon XP 1.4
GHz) for selected matrix dimensions N and 3 different computational
methods; true ARL value is 249.9997
Matrix dimension N
Method

25

51

101

201

Markov chain

103.7077
<1
<0
<1
249.9999
1

307.4809
1
376.0594
1
249.9997
3

254.6729
3
256.6795
2
249.9997
11

250.3782
13
250.4456
9
249.9997
40

Factor Nystrom
Collocation

301
249.3206
39
250.0908
27
249.9997
101

attain at all. Similar relations between the competing methods


are valid for all the calculations done in this paper.
Finally, a Monte Carlo estimate of the in-control ARL with
109 repetitions with 249.9892 (standard error 0.0080) supports
the above result. It is, however, difficult to use Monte Carlo
studies for verifying those accuracies.
For k larger than 1, the Markov chain and the Simpson rule
based Nystrom method perform better (see Fig. 3). By using
an example from Mittag et al. (1998) we consider an EWMA
control chart based on S 2 with sample size n = 5, smoothing
parameter = 0.18, and a critical value c = 2.909 223. In Fig. 3
we see how the new methods work. Mittag et al. (1998) do not
mention the matrix dimension they used, but in the related PhD
dissertation of Tewes (1995) the value is set to N = 126. The
higher accuracy of the collocation does not change the critical
value given in Mittag et al. (1998) as c = 2.91, which is no
surprise for the chosen accuracy of c. Collocation is again the
definite winner, while the three remaining competitors are close
to each other. The product Nystrom method at least shows a
nearly monotone behavior and provides for larger N the best
accuracy among the three weaker methods. Again, at small N
(N = 14) collocation reached the final value 249.9998, while
product Nystrom needs about 1 300 nodes and the Markov chain
method fails again for dimensions up to 5 000. Monte Carlo (109
rep.) results now in 249.9961 with standard error 0.0078.
We now check collocation and product Nystrom methods in
the case of k = 2, i. e. for exponentially distributed data. It
is known that for 2 Vi / 2 22 the quantity Vi is exponentially distributed with mean 2 . Gan (1998) and Gan and Chang
(2000) published exact results for a different chart setup. The
more recent paper, however, allows derivation of very precise
approximations. Originally, a two-sided EWMA chart is evaluated in Gan and Chang (2000). If we let their lower limit cl tend
to 0 (see next section), we can use the related ARL as approximation of the one-sided EWMA chart defined in (2) and (3).
Remark that in applying the method of Gan and Chang (2000) we
have to solve a linear equation system as well, whose dimension
increases (unbounded) for decreasing cl . Hence, in Fig. 4 the parameter N denotes, in case of this approximation sequence, the
dimension of the previously mentioned linear equation system.
For easy application of the results of Gan and Chang (2000) and

Fig. 4. The true value is 502.3431, the smoothing value = 0.18,


sample size n = 2, statistic 2 S2 E x p(1/ 2 ), and the critical limit is
cu = 4.273

later of Gan (1998) we set 02 = 2 and z 0 = 2 for the exponential


case (k = 2).
At first sight collocation has to share the first place with
the approximation sequence based on the results of Gan and
Chang (2000). Collocation needs 13 nodes to get the final value
502.3431. The approximation sequence based on Gan and Chang
(2000) quickly reaches the correct level. For obtaining a stable
result with the same accuracy, however, the dimension of the
linear equation system has to exceed 200 (for N = 202 we obtain 502.3478, for N = 220 we get the final value). We have
to keep in mind that the collocation involves quadratures while
the Gan and Chang (2000) solution needs simpler routines for
filling the system matrix. Nonetheless, collocation wins again.
Product Nystrom needs 1 400 nodes to achieve the same accuracy, while the Markov chain approach retains weaker precision
for dimensions up to 5 000. We complete this case by giving the
related Monte-Carlo results: The ARL estimate based on 1010
repetitions is equal to 502.3470 with standard error 0.0050.
What happens if we adapt the collocation and the product
Nystrom approach to deal with other EWMA setups? The next
section considers one-sided (upper) EWMA control charts with
reflecting (lower) barrier and two-sided EWMA control charts.
In this case, Gan (1998) and Gan and Chang (2000), respectively, provide exact results for the exponential distribution. For
other degrees of freedom we have to compare Markov chain
approximations and Monte Carlo studies again.

4. Computation of the ARL of one-sided


EWMA-S2 with reflecting barrier and of
two-sided EWMA-S2
The main difference between the EWMA setup treated in the
previous sections and the one-sided EWMA-S 2 with reflecting

348

Knoth

(lower) barrier is the following reflexion rule:




Z i = max 02 , (1 ) Z i1 + Vi .
Unlike mean EWMA charts (see, e. g., Yashchin (1989) among
others), the reflexion is very popular for variance EWMA control
charts. Here, we want to examine the EWMA-S 2 example used in
Reynolds Jr. and Stoumbos (2001), where again the nasty k = 1
case is considered. Before going into details, two statements
must be made. The reflexion case is numerically more complex
and evaluating the ARL for a starting value z 0 = 02 provides
a worst-case measure, while for the previous setup a kind of
average measure is computed.
First of all, the new integral equation for z [cl , cu ] with

cl = 02 is (see, for instance, Gan (1993))




cl (1 ) z
L(z) = 1 + F
L(cl )



 cu
1
x (1 ) z
+
L(x) f
d x,
(7)

cl
where F() is the cdf of the variance estimator Vi , i. e. of a chisquared distribution with k degrees of freedom.
The adaption of the product Nystrom method is straightforward. In the linear equation system for the unknown L(z i )
in each row in the first column (corresponding to L(cl )) only
c (1) z
F( l i ) is added.
The Markov chain approach can be improved if we copy
the original Brook and Evans (1972) approach for CUSUM instead of Lucas and Saccucci (1990) for EWMA and decompose
[cl , cu ] into N subintervals
[cl , cl + w/2] [cl + w/2, cl + 3/2 w]
[cl + (2 N 3)w/2, cu ].
The first subinterval with half width represents the reflecting
state.
Finally, we want to use an appropriate collocation scheme.
Contrary to the last section, it is now better to use basis functions,
which are defined piecewise. The reason behind this approach is
that the solution L(z) is no longer smooth over the whole interval [cl , cu ]. Hence, it seems to be appropriate to use piecewise
smooth basis functions. Here a setup is chosen which guarantees an approximation of L(z) that is continuous over [cl , cu ]. In
numerical analysis it is not unusual to take piecewise interpolating functions. Here, we take a very special piecewise structure,
which looks similar to the solution techniques in Gan (1998) and
Gan and Chang (2000), where exact results for the exponential
distribution are given. First, [cl , cu ] is decomposed into
[cl , cl /(1 )] [cl /(1 ), cl /(1 )2 ]
[cl /(1 ) M1 , cu ]
ln(c )ln(c )

u
l
with M = ln(1)
. Remark that the EWMA statistic Z i
based on the non-negative Vi can move at most one subinterval
downwards. The most obvious impact of that behavior consists

Fig. 5. In-control ARL function L(z) of two-sided EWMA-S 2 control


chart with smoothing value = 0.05, sample size n = 1, statistic S2 ,
and the critical limits are cl = 0.5329 and cu = 1.7161; see also Fig. 8

in the fact that the lower integral limit could be set to



cl ,
for z [cl , cl /(1 )]
,
(1 )z, for z > cl /(1 )

(8)

because for x in [cl , (1)z] the density function f () vanishes.


The non-varying lower limit for z < cl /(1 ) turns out to be
(i) the essential difference to the case analyzed in the previous
section, and (ii) the main driver for the non-smooth L(z) as
demonstrated in Fig. 5.
The quantity M depends heavily on cl , which is no problem
for the reflexion schemes. Below we will see that it becomes
more subtle for two-sided schemes. In Fig. 5 the ARL function
L(z) is given for a two-sided EWMA-S 2 chart with one degree of
freedom. Here we see that the number of subintervals M = 23
could be certainly decreased by combining some of them on
the right hand side. In this paper, however, we utilize the full
decomposition like in Gan (1998) and Gan and Chang (2000),
and in Knoth (1998) for the ARL calculation of CUSUM control
charts designed for Erlang distributed data.
Now, we take as grid points (nodes) for the collocation method
the arguments of the extrema of the Chebyshev polynomial
of order N 1 for each subinterval. More formally, the jth
node, j = 0, 1, . . . , N 1, at the ith subinterval [bi , bi+1 ],
i = 1, 2, . . . , M, is given by
z i j = bi +




( N 1 j)
bi+1 bi
.
1 + cos
2
N 1

Remark that z i, N 1 = z i+1,0 , i = 1, 2, . . . , M 1 and in this


way we obtain M( N 1)+1 nodes. These node points are known
as Chebyshev Gauss-Lobatto nodes and are used for Gaussian
quadrature where the interval borders are included as quadrature
nodes. Thus, our basis functions are constantly zero outside the
supporting subinterval and equal to the shifted Chebyshev polynomial on the supporting subinterval. Of course, more simple

Accurate ARL computation for EWMA-S 2 control charts

Fig. 6. The nominal ARL value is 370.4, the smoothing value = 0.1,
sample size n = 1, statistic S2 , reflexion at cl = 1, and the critical
value is c = 3.432

setups are possible like piecewise uniform grid and proper Lagrange polynomials. With this method, however, we obtain a very
useful set of N M interpolating functions. The continuity condition provides M 1 equations and the evaluation of the integral
equation at the nodes provide further M( N 1) + 1 equations
for the final linear equation system. The resulting dimension of
that system is N = N M. The reflexion term is treated similar
to the product Nystrom method and the collocation scheme for
the reflexion case is ready.
Next, we apply our adapted methods for k = 1. In Reynolds
Jr. and Stoumbos (2001) the smoothing constant is set to = 0.1
and c is taken to give an in-control ARL of 370.4. The authors
employ the Markov chain approach with dimension 400 and conclude that c = 3.432. In Fig. 6 the competing approaches are
presented. The number of subintervals used for the collocation
scheme is equal to M = 8. Again the collocation method dominates and reaches the final value of 369.5949 with dimension
56 = 8 7. The competing approaches perform better than in
the last section. The product Nystrom method needs now 377
nodes to obtain the same accuracy, while the Markov chain approach starting with dimension 2 000 provides 369.5947 (which
is not improved before dimension 5 000). The Monte Carlo estimate with 109 repetitions results in 369.5983 with standard
error 0.0115, which supports the previous values. Eventually,
we conclude that the actual critical value c for an in-control
ARL of 370 has to be c = 3.434 and not c = 3.432. In Table 2
the influence of the critical value c accuracy on the actual incontrol ARL value is illustrated for the above EWMA control
chart.
Table 2. Accuracy of the critical value c and the resulting in-control
ARL
c
E (L)

3.5
404.3920

3.44
373.5186

3.434
370.5717

3.4337
370.4250

3.43365
370.4006

3.433649
370.4001

349

Fig. 7. The exact ARL value is 342.7843, the smoothing value = 0.18,
sample size n = 2, statistic 2 S2 E x p(1/ 2 ), and the critical limit is
cu = 4.273 (reflexion at cl = 2)

As second example we compare the exact solutions given by


Gan (1998) with the above methods. We adapt the exponential example from the last section by reflecting at 02 = 2 and
use the same upper limit cu = 4.273 and smoothing parameter
= 0.18. By using the methods developed in Gan (1998) we
obtain 342.7843 as the exact in-control ARL value. In Fig. 7
the approximation results are presented. The collocation is now,
of course, only the second best scheme. At any rate, it needs
only matrix dimension 28 = 4 7 (M = 4) to reproduce the
exact value. Product Nystrom needs about 600 nodes, while the
Markov chain approach succeeds with dimension 2 500. It is
surprising that the product Nystrom performs worse for k = 2
than for k = 1 (especially if we take the larger and smaller
true ARL value for k = 2 into account, which should improve
the accuracy). Because of using an exact result based on Gan
(1998) we abstain from the Monte Carlo study.
Increasing the degrees of freedom k improves the behavior
of the collocation and the Markov chain approach, while the
product Nystrom deteriorates.
Finally, we want to consider the most unpleasant case (from
the numerical point of view)the two-sided EWMA-S 2 control
chart, which is only rarely discussed in literature. The numerical
treatment is similar to the above setup. For all methods the contribution of the reflecting barrier is removed and we return to an
integral equation (4). For the Markov chain method the decomposition of the previous section is applied, that is, all subintervals
are of equal size. For the collocation, however, we utilize again
the piecewise setup as described for the reflexion case.
By applying our three methods to the most subtle case, k = 1,
we use a chart setup given in MacGregor and Harris (1993). The
authors analyze their charts by using approximating chi-squared
distributions for the chart statistic Z i . They call their chart exponentially weighted mean square (EWMS) which corresponds
to our two-sided EWMA-S 2 with k = 1. For a nominal incontrol ARL of 500 and smoothing value = 0.05 they obtain
cl = 0.532 9 and cu = 1.716 1. Their final approximation of

350

Fig. 8. The reported ARL value is 497 (MC-simulation with 4 000 rep.),
the true ARL value is 512.73, the smoothing value = 0.05, sample
size n = 1, statistic S2 , and the critical limits are cl = 0.532 9 and
cu = 1.716 1

the actual ARL results in 497 (Monte Carlo with 4 000 rep.). In
Fig. 8 our approximation techniques are illustrated. Contrary to
all preceding figures we use a rough scale of the ARL axis and
node counts up to 501. Nonetheless, the shapes of the graphs are
not satisfying. Here, the number of subintervals for the piecewise collocation is large with M = 23 (see Fig. 5). Consequently,
the resulting matrix dimension for reasonable accuracy becomes
larger than for the previous constellations. We need at least dimension 299 = 2313 to achieve accuracy for the first place after
the decimal point. The related ARL value is 512.7. The other
methods oscillate too heavily for getting any useful approximation. Collocation with dimension 667 = 2329 provides 512.73,
and eventually with dimension above 2 000 the value 512.7326
is obtained. The indestructible Monte Carlo approach results in
512.7379 with standard error 0.0158 for 109 repetitions.
Eventually, we consider a larger k and return to the exponential distribution (k = 2) with published exact results in Gan
and Chang (2000). Note that there are some slight errors in the
numerical results (and not in the analytical ones) given in their
paper. The true in-control ARL value for a two-sided EWMA
scheme with = 0.18, cl = 0.811 and cu = 4.273 is not equal
to 246.66 as published, but 247.0118. The latter value is checked
by writing a program based on the formulas presented in Gan
and Chang (2000). A Monte Carlo study with 1010 repetitions
results in 247.0081 with standard error 0.0024. Our approximation methods behave considerably better than for k = 1, as we
can see in Fig. 9. Now, the collocation parameter M is equal to
9 and already for dimension 63 = 9 7 the above ARL value
is reproduced. Product Nystrom and Markov chain method do
not provide the same accuracy for dimensions up to 5 000. The
first one achieves a (absolute) precision of about 0.001 and the
second of 0.01.
Remark that there are further EWMA control charts using
S 2 like lower one-sided charts with or without reflecting upper barrier. At least for the reflexion scheme it is possible to

Knoth

Fig. 9. The exact ARL value is 247.0118, the smoothing value = 0.18,
sample size n = 2, statistic 2 S2 E x p(1/ 2 ), and the critical limits
are cl = 0.811 and cu = 4.273

employ the above methods. In case of the halfopen control chart an artificial reflexion has to be introduced. These
types of EWMA control charts will be analyzed in a further
paper.
In the final section it is demonstrated that there is reason to
apply S 2 instead of ln S 2 control charts.

5. Comparison of EWMA-S2 and EWMA-ln S2


In this section a short comparison study is performed, which investigates one-sided upper EWMA control charts with a lower
reflecting barrier. The established one, which is based on ln S 2 ,
is given in Crowder and Hamilton (1992). The ARL of this
EWMA-ln S 2 chart is evaluated with the standard Nystrom
method, which ensures high precision for the ln S 2 case. Crowder
and Hamilton (1992) provide nomograms and hints for setting
up an appropriate scheme.
It is known that small s are better for detection of small
changes, and large ones for detection of large changes. For
instance, Crowder and Hamilton (1992) recommend to use
for given in-control ARL of 200, and sample size n = 5 the
following values:
Out-of-control 12
Smoothing

1.22
0.05

1.32
0.16

1.42
0.32

The corresponding ARL values for different values of 2 are


given in Table 3. The results for ln S 2 are recomputed based
on Crowder and Hamilton (1992). In order to perform a reliable comparison we take a higher accuracy than in the original paper, so that instead of the critical values 1.06, 1.45, and
1.61 given in Crowder and Hamilton (1992) we utilize 1.055 21,
1.452 31, and 1.608 61 for = 0.05, 0.16, and 0.32, respectively. The results are based on the Gauss-Legendre Nystrom

Accurate ARL computation for EWMA-S 2 control charts

351

Table 3. Comparison of EWMA-lnS 2 (with different reflection barriers, 0 and E (lnS 2 ) 0.267) and EWMA-S 2 : ARL values for different 2 ,
bold values mark the optimal ones of all considered (Crowder and Hamilton, 1992) lnS 2 schemes
= 0.05

= 0.16

ln S 2

= 0.32

ln S 2

ln S 2

.267

S2

.267

S2

.267

S2

1
1.12
1.22
1.32
1.42
1.52
22

200.00
43.04
18.10
10.75
7.63
5.97
3.17

200.00
41.55
19.92
13.11
9.93
8.11
4.67

200.00
37.78
16.71
10.32
7.39
5.74
2.74

200.00
45.59
18.54
10.52
7.20
5.49
2.77

200.00
43.19
18.46
11.11
7.96
6.27
3.37

200.00
43.44
17.42
9.85
6.68
5.03
2.33

200.00
48.93
19.63
10.73
7.08
5.24
2.44

200.00
46.90
19.12
10.79
7.34
5.57
2.78

200.00
50.50
20.05
10.74
6.93
5.03
2.18

method with 50 nodes. The new ARL values differ only slightly
from the original ones. Crowder and Hamilton (1992) compute
probably their ARL values with more precise critical values
and published only the rounded ones. Recall Table 2 for the
influence of the critical value accuracy on the ARL results.
The bold values in Table 3 mark the best that their EWMA
schemes can achieve. Crowder and Hamilton (1992) noted that
1
1
2
E (ln S 2 ) ln 02 n1
3(n1)
2 + 15(n1)4 0.267 (for sub2
group size n = 5 and 0 = 1). According to Gan (1995), Chang
(1993) showed in his unpublished M. Sc. thesis that decreased reflecting barriers lead to slightly better ARL performance. Thus,
we tabulate as well ARL values for an EWMA-ln S 2 chart that
has a reflecting barrier at E (ln S 2 ). This seems to be reasonable, because the S 2 counterpart deploys E (S 2 ) as reflecting
barrier.
For the same values the related S 2 schemes are evaluated. In
all cases the collocation is employed with the following details
(see Section 4). Recall that the sample size was given as n = 5.
smoothing
M subintervals
matrix dimension
N = M N
critical value c

0.05
29
203

0.16
9
54

0.32
7
28

2.317 64

2.963 52

3.383 32

Larger values of N (corresponds to larger N ), i. e. larger matrix dimensions, do not change the critical values c for the given
accuracy. In Table 3 the out-of-control ARL values complete the
ln S 2 based results. All results of Table 3 get along with related
values of a Monte-Carlo study with 108 repetitions.
From Table 3 we learn, what we lose if we are using ln S 2
instead S 2 . For = 0.05 and = 0.16 the S 2 based chart is better
for all considered out-of-control values of 2 than the related
ln S 2 charts except for 2 = 1.1, where in case of = 0.16
the chart with reflection at 0.267 is better. Thus, it would be
better to use the S 2 based scheme because of better properties
(in terms of the ARL).

6. Summary and conclusions


By means of collocation we are able to assess the ARL performance of S 2 based EWMA control charts for monitoring the
variance of normally distributed data. Collocation allows fast
and very precise computation of the ARL. With the exception
of two-sided charts based on individual observations (leads to
k = 1) the collocation results are very accurate. Even in problematic cases collocation provides suitable results.
It seems to be possible to transfer the results to S 2 based
CUSUM schemes, such that more accurate comparison studies
become possible.
The main problem which arises with chi-squared distributed
data, that support of the distribution does not resemble the whole
real axis, might be observed for other distributions as well.
Then the collocation or the product Nystrom method can also
be applied.
Eventually, the above techniques can be employed as well to
assess the steady-state performance of EWMA control charts
based on S 2 .

References
Acosta-Meja C.A. and Pignatiello Jr. J.J. 2000. Monitoring process
dispersion without subgrouping. Journal of Quality Technology
32(2): 89102.
Acosta-Meja C.A., Pignatiello Jr. J.J. and Rao B.V. 1999. A comparison
of control charting procedures for monitoring process dispersion.
IIE Transactions 31: 569579.
Atkinson K.E. 1976. A survey of numerical methods for the solution
of Fredholm integral equations of the second kind. Philadelphia:
Society for Industrial and Applied Mathematics SIAM. VII, 230 p.
Brook D. and Evans D.A. 1972. An approach to the probability distribution of CUSUM run length. Biometrika 59(3): 539549.
Calzada M.E. and Scariano S.M. 2003. Reconciling the integral equation and Markov chain approaches for computing EWMA average
run lengths. Commun. Stat. Simula. Comput. 32(2): 591604.

352
Champ C.W. and Rigdon S.E. 1991. A comparison of the Markov chain
and the integral equation approaches for evaluating the run length
distribution of quality control charts. Commun. Stat. Simula. Comput. 20(1): 191204.
Chang T.C. and Gan F.F. 1994. Optimal designs of one-sided EWMA
charts for monitoring a process variance. J. Stat. Comput. Simulation 49: 3348.
Crowder S.V. 1987. A simple method for studying run-length distributions of exponentially weighted moving average charts. Technometrics 29: 401407.
Crowder S.V. and Hamilton M.D. 1992. An EWMA for monitoring a
process standard deviation. Journal of Quality Technology 24(1):
1221.
Domangue R. and Patch S. 1991. Some omnibus exponentially
weighted moving average statistical process monitoring schemes.
Technometrics 33(3): 299313.
Gan F.F. 1993. Exponentially weighted moving average control charts
with reflecting boundaries. J. Stat. Comput. Simulation 46: 4567.
Gan F.F. 1995. Joint monitoring of process mean and variance using
exponentially weighted moving average control charts. Technometrics 37: 446453.
Gan F.F. 1998. Designs of one- and two-sided exponential EWMA
charts. Journal of Quality Technology 30(1): 5569.
Gan F.F. and Chang T.C. 2000. Computing average run lengths of exponential EWMA charts. Journal of Quality Technology 32: 183
187.
Gianino A.B., Champ C.W. and Rigdon S.E. 1990. Solving integral
equations by the collocation method. In: ASA Proceedings of the
Statistical Computing Section, pp. 101102.
Girshick M.A. and Rubin H. 1952. A Bayes approach to a quality control
model. Ann. Math. Stat. 23: 114125.
Hackbusch W. 1995. Integral Equations: Theory and Numerical Treatment. International Series of Numerical Mathematics, Birkhauser
Verlag, Basel, Boston, Berlin. vol. 120
Hunter J.S. 1986. The exponentially weighted moving average. Journal
of Quality Technology 18: 203210.
Knoth S. 1998. Exact average run lengths of CUSUM schemes for
erlang distributions. Sequential Analysis 17: 173184.
Knoth S. 2004. The art of evaluating monitoring schemeshow to
measure the performance of control charts? In: P. Grzegorzewski,
E. Mrowka, O. Hryniewicz, H.-J. Lenz, and P.-T. Wilrich (Eds.),

Knoth
Proceedings of the VIIIth International Workshop on Intelligent
Statistical Quality Control, pp. 117143.
Lowry C.A., Champ C.W. and Woodall W.H. 1995. The performance
of control charts for monitoring process variation. Commun. Stat.
Simula. Comput. 24(2): 409437.
Lucas J.M. and Saccucci M.S. 1990. Exponentially weighted moving
average control schemes: Properties and enhancements. Technometrics 32: 112.
MacGregor J.F. and Harris T.J. 1993. The exponentially weighted moving variance. Journal of Quality Technology 25: 106118.
Mittag H.-J. Stemann D., and Tewes B. 1998. EWMA-Karten zur

Uberwachung
der Streuung von Qualitatsmerkmalen. Allgemeines Statistisches Archiv 82: 327338.
Page E.S. 1954. Continuous inspection schemes. Biometrika 41: 100
115.
Press W.H., Flannery B.P., Teukolsky S.A., and Vetterling W.T. 2002.
Numerical recipes source code in C and C++ CD-ROM with
Windows or Macintosh single-screen license. The art of scientific
computing. 2nd ed. Cambridge: Cambridge University Press.
Reynolds Jr. M.R. and Stoumbos Z.G. 2001. Monitoring the process
mean and variance using individual observations and variable sampling intervals. Journal of Quality Technology 33(2): 181205.
Roberts S.W. 1959. Control-charts-tests based on geometric moving
averages. Technometrics 1: 239250.
Shewhart W.A. 1931. Economic Control of Quality of Manufactured
Product. Reinhold Company, Princeton, N.J.
Srivastava M.S. 1994. Comparison of CUSUM and EWMA procedures
for detecting a shift in the mean or an increase in the variance. J.
Appl. Stat. Sci. 1(4): 445468.
Sweet A.L. 1986. Control charts using coupled exponentially weighted
moving averages. IEEE Transactions on Information Theory 18:
2633.
Tewes B. 1995. Fertigungsuberwachung mit Qualitatsregelkarten vom
EWMA-Typ. Ph. D. dissertation, Fernuniversitat Hagen.
Waldmann K.-H. 1986. Bounds for the distribution of the run length
of geometric moving average charts. J. R. Stat. Soc., Ser. C, Appl.
Stat. 35: 151158.
Wortham A.W. and Ringer L.J. 1971. Control via exponential smoothing. The Logistics Review 7(32): 3340.
Yashchin E. 1989. Weighted cumulative sum technique. Technometrics
31(1): 321338.

You might also like