QUALITY AND RELIABILITY ENGINEERING INTERNATIONAL

Qual. Reliab. Engng. Int. 2005; 21:781–794

Published online 13 May 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/qre.686

Research A New S 2-EWMA Control

Chart for Monitoring the
Process Variance
Philippe Castagliola∗,†
Institut Universitaire de Technologie de Nantes et IRCCyN, Rue Christian Pauc, La Chantrerie, BP 50609, 44306 Nantes,
France

This article demonstrates how a three-parameter logarithmic transformation

combined with an exponentially weighted moving average (EWMA) approach can
be used to monitor the sample variance of a process. The computation of the
parameters of the logarithmic transformation and the control limits are explained.
An easy-to-use table is provided and an illustrative example is given. The performance
of the logarithmic transformation is evaluated under the assumption of normality.
An optimal design strategy based on the ARL is presented and comparisons with
other procedures are performed. Copyright  c 2005 John Wiley & Sons, Ltd.

KEY WORDS: sample variance; EWMA; logarithmic transformation

INTRODUCTION

D
uring the last decade, the use of the exponentially weighted moving average (EWMA) statistic as a
process monitoring tool has become more and more popular in the statistical process control field.
The properties and design strategies of the EWMA control chart for the mean (introduced by Roberts1 )
have been thoroughly investigated by Robinson and Ho2 , Crowder3,4 , Lucas and Saccucci5 and Steiner6 . The use
of the EWMA as a tool for monitoring the process variability has received attention by Wortham and Ringer7,
Sweet8 , Ng and Case9 , Crowder and Hamilton10, Hamilton and Crowder11 and MacGregor and Harris12,
Gan13, Amin et al.14, Lu and Reynolds15 and Acosta-Mejia et al.16 . The goal of this paper is to present a new
EWMA-type control chart for monitoring the variance of the process. Our paper is, in a certain way, an extension
of Crowder and Hamilton’s10 initial approach because we use a logarithmic transformation. We chose to use
a two-sided version of the EWMA chart whereas Crowder and Hamilton used a one-sided version to detect
only increases in the variability. The two-sided version we suggest is constructed to detect both increases and
decreases in variability.

PROPOSED EWMA CHART FOR THE SAMPLE VARIANCE

Let Xk,1 , . . . , Xk,n be a sample of n independent normal, N(µ, σ0 ), random variables, where µ is the process
mean, σ0 is the nominal process standard deviation, and k is the subgroup number. In this paper we will

∗ Correspondence to: Philippe Castagliola, Institut Universitaire de Technologie de Nantes et IRCCyN, Rue Christian Pauc, La Chantrerie,
BP 50609, 44306 Nantes, France.
† E-mail: [email protected]

Received 8 April 2003

Copyright 
c 2005 John Wiley & Sons, Ltd. Revised 2 October 2003
782 P. CASTAGLIOLA

assume that σ0 is known, even if, in practice, this parameter has to be estimated from an in-control population.
The problem of the statistical design of an EWMA-type control chart with estimated parameters has been
recently tackled by Jones17 , but will not be covered in this paper. Let Sk2 be the sample variance of subgroup k,
i.e.

1
n
Sk2 = (Xk,j − X̄k )2
n − 1 j =1

where X̄k is the sample mean of subgroup k. In order to monitor the process variance, Crowder and Hamilton10,
following a recommendation by Box et al.18 , suggest the application of the classical EWMA approach on the
logarithm of the successive sample variances, Tk = ln Sk2 , i.e.

Zk = (1 − λ)Zk−1 + λTk (1)

where λ is a smoothing constant satisfying 0 < λ ≤ 1. The main motivation of this approach is that Tk = ln Sk2 ,
which has a log-Gamma distribution (see Johnson et al.19 ), tends to be more normally distributed than the
sample variance Sk2 . The main idea developed in this paper is that of applying the following three-parameter
transformation to Sk2

Tk = a + b ln(Sk2 + c)

with c > 0 (in order to avoid problems with the logarithmic transformation) and then using the EWMA statistic
in Equation (1). It can be noticed that this transformation belongs to a class of transformations originally
proposed by Johnson20. The main expectation of this approach is that if a, b and c are judiciously selected,
then this transformation may result in approximate normality to Tk better than the approach in Crowder and
Hamilton10. Trying to make Tk more normally distributed is related to making the distribution of Tk more
symmetrical. Concerning the value of Z0 , we will demonstrate in Appendix A that Z0 has a value close to 0
which depends only on n and not on σ0 . If the value of the mean E(Tk ) and the standard deviation σ (Tk ) of Tk
corresponding to parameters a, b and c are known, then the EWMA control limits for the transformed sample
variance will be set at
 
λ 1/2
LCL = E(Tk ) − K σ (Tk ) (2)
2−λ
 
λ 1/2
UCL = E(Tk ) + K σ (Tk ) (3)
2−λ

where K is a positive constant (for example, K = 3). These limits are evaluated using the limiting value for the
standard deviation of the EWMA statistic. It should be mentioned that the use of these fixed control limits is
easier, but makes the control chart less sensitive at the beginning. This approximation is widely used in industrial
applications and we will use it in the rest of the paper. Nevertheless, if the quality practitioner prefers to use the
exact value for the standard deviation of the EWMA statistic (see Montgomery21 and Ryan22 ), then the control
limits become
 1/2
λ{1 − (1 − λ)2k }
LCL = E(Tk ) − K σ (Tk )
2−λ
 1/2
λ{1 − (1 − λ)2k }
UCL = E(Tk ) + K σ (Tk )
2−λ

In the next section it is shown that whatever the value of n and σ0 , the value of E(Tk ) is so close to 0 that it can
be assumed to be 0. Consequently, in practice, the quality practitioner can use the following simplified control

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
A NEW S 2 -EWMA CONTROL CHART FOR MONITORING THE PROCESS VARIANCE 783

limits for the S 2 -EWMA control chart

 1/2
λ
LCL = −K σ (Tk )
2−λ
 1/2
λ
UCL = +K σ (Tk )
2−λ
The control limits given above correspond to a two-sided EWMA control chart, but one-sided EWMA control
limits can also be considered. The advantage of the approach proposed in this paper is that, even if the control
limits of the standard Shewhart S 2 chart are not symmetrical, the use of the suggested transformation makes the
S 2 -EWMA control limits symmetrical and a decrease or an increase of the process variance should be processed
in roughly the same way in both low and high sides of this chart. This is not a demonstration but just an intuitive
reasoning. This will be highlighted in the ‘Illustrative example’ section. The use of a FIR (fast initial response)
feature, proposed by Lucas and Crosier23 for the CUSUM control chart and modified by Lucas and Saccucci5
and Rhoads et al.24 for the EWMA control chart can also be considered for the S 2 -EWMA control chart by
setting the first value Z0 to Z0 = LCL/2 or Z0 = UCL/2 depending on whether the previous out-of-control
signal corresponded to an increase or a decrease in the nominal standard deviation. If the previous out-of-control
signal corresponds to an increasing variance, then Z0 = UCL/2 and if it corresponds to a decreasing variance,
then Z0 = LCL/2. When using the FIR feature, Lucas and Saccucci5 and Rhoads et al.24 showed considerable
reduction in the ARL to detect an off-target value. This conclusion can reasonably be drawn for the S 2 -EWMA
control chart too. In the following section, we present the main results concerning the computation of a, b, c,
E(Tk ), σ (Tk ) and Z0 .

HOW TO FIND a, b, c, E(Tk ), σ (Tk ) AND Z0

In Appendix A we prove that a, b and c can be defined as

b = B(n) (4)
c = C(n)σ02 (5)
a = A(n) − 2B(n) ln(σ0 ) (6)

where A(n), B(n), and C(n) are three functions depending only on the sample size n. The closed forms of
these functions are derived in Appendix A and their values can be found in Table I for n = 3, . . . , 15. Also in
Appendix A, we derive the distribution fTk (t|n) of Tk and show that this distribution depends only on n.
This property is important since it allows the calculation of the values of E(Tk ) and σ (Tk ) independently
of the value of σ0 . The computation of E(Tk ) and σ (Tk ) has been achieved by numerical quadrature, for
n = 3, . . . , 15, and the results are also shown in Table I. One can notice that values for E(Tk ) are very close to 0.
In fact they are so close to 0 that assuming E(Tk ) = 0 is a very good approximation. Finally, we demonstrate in
Appendix A that a reasonable value for Z0 is

Z0 = A(n) + B(n) ln{1 + C(n)}

This value depends only on n and not on σ0 . The values for Z0 are given in Table I. One can see that these values
are also close to 0 and can be replaced by 0 in practice with little practical effect.

AN ILLUSTRATIVE EXAMPLE

The goal of this example is to show how the S 2 -EWMA control chart behaves in the case of an increase
and a decrease in the nominal variance. The first 100 data points plotted in Figures 1(a) and (b) consist of

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
784 P. CASTAGLIOLA

Table I. Constants A(n), B(n), C(n), E(Tk ), σ (Tk ), and Z0 for n = 3, . . . , 15

n A(n) B(n) C(n) E(Tk ) σ (Tk ) Z0

3 −0.6627 1.8136 0.6777 0.024 72 0.9165 0.276
4 −0.7882 2.1089 0.6261 0.012 66 0.9502 0.237
5 −0.8969 2.3647 0.5979 0.007 48 0.9670 0.211
6 −0.9940 2.5941 0.5801 0.004 85 0.9765 0.193
7 −1.0827 2.8042 0.5678 0.003 35 0.9825 0.178
8 −1.1647 2.9992 0.5588 0.002 43 0.9864 0.167
9 −1.2413 3.1820 0.5519 0.001 82 0.9892 0.157
10 −1.3135 3.3548 0.5465 0.001 41 0.9912 0.149
11 −1.3820 3.5189 0.5421 0.001 12 0.9927 0.142
12 −1.4473 3.6757 0.5384 0.000 90 0.9938 0.136
13 −1.5097 3.8260 0.5354 0.000 74 0.9947 0.131
14 −1.5697 3.9705 0.5327 0.000 62 0.9955 0.126
15 −1.6275 4.1100 0.5305 0.000 52 0.9960 0.122

20 identical subgroups of n = 5 observations randomly generated from a N(20, 0.1) distribution (corresponding
to an in-control process), while the last 50 data points of Figure 1(a) consist of 10 subgroups of n = 5
observations randomly generated from a N(20, 0.2) distribution (the nominal process standard deviation σ0
has increased by a factor of 2), and the last 50 data points of Figure 1(b) consist of 10 subgroups of n = 5
observations randomly generated from a N(20, 0.05) distribution (the nominal process standard deviation σ0
has decreased by a factor of 2). The corresponding 30 sample variances are plotted in Figure 1(c), for the
two cases (increasing and decreasing). At this step, the asymmetry between increasing and decreasing sample
variances is particularly noticeable. If n = 5 and σ0 = 0.1, then b = 2.3647, c = 0.5979 × 0.12 = 0.005 979 and
a = −0.8969 − 2 × 2.3647 × ln(0.1) = 9.9929. The 30 transformed sample variances are plotted in Figure 1(d)
and the S 2 -EWMA sequence along with the S 2 -EWMA control limits LCL = 0.0075 − 3√0.1/1.9 × 0.967 =
√
−0.658 and UCL = 0.0075 + 3 0.1/1.9 × 0.967 = 0.673 (λ = 0.1 and K = 3) are plotted in Figure 1(e).
The S -EWMA control chart clearly detects an out-of-control signal at the 25th subgroup (in the increasing
2

case) and at the 26th subgroup (in the decreasing case), pointing out that an increase/decrease in the variance
occurred. As previously noted, the S 2 -EWMA control chart seems to detect the increasing and decreasing of the
nominal variance in a very symmetrical way. At this point, this symmetric performance is only postulated and
an ARL comparison should be conducted to demonstrate it.

PERFORMANCE IN TERMS OF ARL

Let σ1 be the new process standard deviation and let τ = σ1 /σ0 be the variable reflecting the shift in the process
variability. For the quality practitioner, it could be interesting to use different (λ, K) combinations which give
the same in-control ARL = ARL0 (i.e. the ARL when the process is functioning at the nominal variability
σ = σ0 or equivalently τ = 1) and then find, for a specified value of the shift τ , the combination (λ∗ , K ∗ ) which
yields the smallest possible out-of-control ARL, i.e.

ARL(1, λ∗ , K ∗ , n, t) = ARL0 (7)

∗ ∗
ARL(τ, λ , K , n, t) = min ARL(τ, λ, K, n, t) (8)
λ,K

The computation of ARL(τ, λ, K, n, t) is presented in Appendix B. The obtained combination (λ∗ , K ∗ ) is then
‘optimally’ designed for detecting a particular shift τ . In this paper, ARL0 = 370 (corresponding to the classical
3σ limits for a Shewhart control chart) and ARL0 = 200 (cf. Crowder and Hamilton10) have been chosen for
the in-control ARL. In order to compute the optimal combinations (λ∗ , K ∗ ), Equations (7) and (8) are solved
using a nonlinear optimization algorithm with 0.05 ≤ λ as a supplementary constraint. The choice of 0.05 as

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
A NEW S 2 -EWMA CONTROL CHART FOR MONITORING THE PROCESS VARIANCE 785

(a) (b)
20.6 20.6

20.4 20.4

20.2 20.2
Observations

Observations
20 20

19.8 19.8

19.6 19.6

19.4 19.4

19.2 19.2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Subgroups Subgroups

(c) (d)
0.12 increasing case 5 increasing case
decreasing case decreasing case

4
0.1

3
T=a+b*log(S2+c)

0.08
2
S2

0.06 1

0
0.04

-1

0.02
-2

0 -3
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Subgroups Subgroups

(e)
2 increasing case
decreasing case

1.5

1
S2-EWMA

UCL
0.5

-0.5
LCL
-1
0 5 10 15 20 25 30
Subgroups

Figure 1. (a) Data with an increasing variance; (b) data with a decreasing variance; (c) sample variances; (d) transformed
sample variances; (e) S 2 -EWMA control chart, λ = 0.1

a lower bound for λ is due to numerical difficulty in evaluating ARLs for small values of λ. This problem is
also mentioned by Crowder and Hamilton10. The (λ, K) combinations such that ARL0 = 370 and ARL0 = 200
are plotted in Figure 2. In both cases, the trajectory of the curve corresponding to n = 3 looks very different
from the other curves. This can be explained by the fact that the three-parameter log-normal transformation is
not as ‘smooth’ as the simple logarithm transformation. In Figures 3 and 4, ARL functions are plotted versus λ
(and the corresponding value of K) for different values of n and τ . Three interesting conclusions can be drawn
from these plots.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
786 P. CASTAGLIOLA

(a)
3
n=3

2.9

n=5
2.8 n=9
n=7

2.7
K

2.6

2.5

2.4

2.3

2.2
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
O
(b)
3

2.9

2.8 n=3

2.7
n=5
n=9
K

2.6 n=7

2.5

2.4

2.3

2.2
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
O
Figure 2. (λ, K) such that (a) ARL0 = 370 and (b) ARL0 = 200

• As expected, when τ corresponds to a ‘large’ increase in the process variance, say τ > 1.5, the value
of λ has very little effect on the ARL. For example, in the plot corresponding to ARL0 = 370, n = 5, in
Figure 3, the ARL corresponding to τ = 1.5 is 4.8 for λ = 0.05 and 5.6 for λ = 0.5. Consequently, if τ
is large, any value of λ in (0.05, 0.5) can be used as the optimal λ∗ value. In this case, we suggest using
λ∗ = 0.05 as the optimal value. The reader must be warned by the fact that the problem with the use of a
small value for λ like 0.05 is that the EWMA chart will be slow to react to a sudden large shift.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
 (a) (b)
140 W=0.7 W=0.5 W=0.9 W=0.7 W=0.5
140
130 W=1.1
130

Copyright 
120 120
110 110
W=1.1
100 100
90 90

80 80

70 70

ARL
ARL
60 60

50 50

c 2005 John Wiley & Sons, Ltd.

40 40
W=1.3
30 30

20 20 W=1.3
W=1.5
10 10 W=1.5
W=2.0 W=2.0
0 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
O O
(c) (d)
140 W=0.9 W=0.7 W=0.5 W=0.9 W=0.7
140
130 130
120 120
110 110
100 100
90 W=1.1 90
A NEW S 2 -EWMA CONTROL CHART FOR MONITORING THE PROCESS VARIANCE

W=1.1
80 80
70 70

ARL
ARL
60 60
50 50
40 40 W=0.5
30 30
20 20
W=1.3
10 10 W=1.3
W=1.5
W=2.0 W=1.5
0 0 W=2.0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
O O

Figure 3. ARL as a function of λ for different values of n and τ (ARL0 = 370): (a) n = 3; (b) n = 5; (c) n = 7; (d) n = 9

Qual. Reliab. Engng. Int. 2005; 21:781–794

787
 788
(a) (b)
W=0.9 W=0.7 W=0.5 140 W=0.9 W=0.7 W=0.5
140

130 130

Copyright 
120 120

110 110

100 100

90 90

80 W=1.1 80

70 70
W=1.1

ARL

ARL
60 60

50 50

c 2005 John Wiley & Sons, Ltd.

40 40

30 30
W=1.3 20
20
W=1.3
10 W=1.5 10
W=1.5
W=2.0 W=2.0
0 0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
O O
(c) (d)
W=0.9 W=0.7 W=0.5 140 W=0.9
140

130 130

120 120

110 110
W=0.7
100 100

90 90

80 80

70 70

ARL

ARL
60 60
W=1.1 W=1.1
50 50

40 40

30 30

20 20
W=0.5
10 W=1.3 10 W=1.3
W=1.5 W=1.5
0 W=2.0 0 W=2.0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
O O

Figure 4. ARL as a function of λ for different values of n and τ (ARL0 = 200): (a) n = 3; (b) n = 5; (c) n = 7; (d) n = 9

Qual. Reliab. Engng. Int. 2005; 21:781–794

P. CASTAGLIOLA
A NEW S 2 -EWMA CONTROL CHART FOR MONITORING THE PROCESS VARIANCE 789

Table II. Optimal combinations (λ∗ , K ∗ ) and smallest possible out-of-control ARL∗ for τ ∈ (0.5, 2),
n = {3, 5, 7, 9} and ARL0 = 370

n=3 n=5 n=7 n=9

τ λ∗ K∗ ARL∗ λ∗ K∗ ARL∗ λ∗ K∗ ARL∗ λ∗ K∗ ARL∗
0.50 0.14 2.755 12.3 0.24 2.819 6.2 0.34 2.850 4.2 0.44 2.866 3.2
0.60 0.10 2.689 17.7 0.18 2.790 9.0 0.25 2.836 6.1 0.31 2.863 4.7
0.70 0.08 2.642 28.4 0.12 2.724 14.6 0.17 2.791 10.0 0.20 2.822 7.7
0.80 0.05 2.546 54.2 0.08 2.634 28.1 0.10 2.689 19.5 0.12 2.732 15.1
0.90 0.05 2.546 155.2 0.05 2.514 80.8 0.05 2.505 56.7 0.06 2.555 44.5
0.95 0.05 2.546 323.5 0.05 2.514 202.1 0.05 2.505 150.1 0.05 2.501 120.5

1.05 0.05 2.546 153.5 0.05 2.514 121.2 0.05 2.505 99.3 0.05 2.501 84.1
1.10 0.05 2.546 65.1 0.05 2.514 44.9 0.05 2.505 35.0 0.05 2.501 29.1
1.20 0.05 2.546 21.8 0.05 2.514 15.3 0.05 2.505 12.5 0.05 2.501 10.8
1.30 0.05 2.546 11.7 0.05 2.514 8.8 0.05 2.505 7.4 0.05 2.501 6.6
1.40 0.05 2.546 7.9 0.05 2.514 6.2 0.05 2.505 5.4 0.05 2.501 4.8
1.50 0.05 2.546 5.9 0.05 2.514 4.8 0.05 2.505 4.3 0.05 2.501 3.9
1.60 0.05 2.546 4.8 0.05 2.514 4.0 0.05 2.505 3.6 0.05 2.501 3.2
1.70 0.05 2.546 4.1 0.05 2.514 3.5 0.05 2.505 3.1 0.05 2.501 2.9
1.80 0.05 2.546 3.6 0.05 2.514 3.1 0.05 2.505 2.8 0.05 2.501 2.6
1.90 0.05 2.546 3.2 0.05 2.514 2.8 0.05 2.505 2.5 0.05 2.501 2.4
2.00 0.05 2.546 2.9 0.05 2.514 2.5 0.05 2.505 2.3 0.05 2.501 2.2

Table III. Optimal combinations (λ∗ , K ∗ ) and smallest possible out-of-control ARL∗ for τ ∈ (0.5, 2),
n = {3, 5, 7, 9} and ARL0 = 200

n=3 n=5 n=7 n=9

τ λ∗ K∗ ARL∗ λ∗ K∗ ARL∗ λ∗ K∗ ARL∗ λ∗ K∗ ARL∗
0.50 0.16 2.551 10.6 0.28 2.629 5.4 0.39 2.664 3.7 0.47 2.683 2.9
0.60 0.12 2.496 15.2 0.20 2.592 7.9 0.27 2.642 5.4 0.34 2.674 4.2
0.70 0.09 2.435 23.9 0.14 2.529 12.6 0.19 2.594 8.7 0.23 2.632 6.8
0.80 0.07 2.381 44.4 0.09 2.425 23.7 0.12 2.498 16.6 0.14 2.537 13.0
0.90 0.05 2.325 108.5 0.05 2.269 63.3 0.06 2.305 45.8 0.07 2.347 36.4
0.95 0.05 2.325 189.4 0.05 2.269 133.4 0.05 2.251 105.1 0.05 2.241 87.6

1.05 0.05 2.325 95.0 0.05 2.269 78.7 0.05 2.251 66.9 0.05 2.241 58.3
1.10 0.05 2.325 43.9 0.05 2.269 32.4 0.05 2.251 26.2 0.05 2.241 22.4
1.20 0.05 2.325 15.9 0.05 2.269 12.0 0.05 2.251 10.0 0.05 2.241 8.9
1.30 0.05 2.325 8.9 0.05 2.269 7.1 0.05 2.251 6.1 0.05 2.241 5.5
1.40 0.05 2.325 6.1 0.05 2.269 5.1 0.05 2.251 4.5 0.05 2.241 4.1
1.50 0.05 2.325 4.7 0.05 2.269 4.0 0.05 2.251 3.6 0.05 2.241 3.3
1.60 0.05 2.325 3.8 0.05 2.269 3.3 0.05 2.251 3.0 0.05 2.241 2.8
1.70 0.05 2.325 3.3 0.05 2.269 2.9 0.05 2.251 2.7 0.05 2.241 2.5
1.80 0.05 2.325 2.9 0.05 2.269 2.6 0.05 2.251 2.4 0.05 2.241 2.2
1.90 0.05 2.325 2.6 0.05 2.269 2.3 0.05 2.251 2.2 0.05 2.241 2.0
2.00 0.05 2.325 2.4 0.05 2.269 2.1 0.05 2.251 2.0 0.05 2.241 1.9

• When τ corresponds to a ‘small’ increase in the process variance, say 1 < τ < 1.5, the minimal ARL
is always reached for λ = 0.05. In that case, taking λ∗ = 0.05 seems to be the best solution in terms of
minimizing the ARL. In fact, because we restricted our search in (0.05, 1), the true optimal value could
be smaller.
• When τ corresponds to a decrease in the process variance (τ < 1), the ARL curves look very differently
and are ‘J’ shaped. In that case, there is an optimal value for λ in (0.05, 1) which minimizes the ARL.
One can also notice in this case that a bad choice of λ could have a dramatic impact on the out-of-control
ARL.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
790 P. CASTAGLIOLA

Table IV. Comparison in terms of ARL between different control

charts for n = 5 and ARL0 = 200

τ R CUSUM EWMA(1) EWMA(2) S 2 -EWMA

1.0 200 200 200 200 200
1.1 69 46 43 49 32.4
1.2 31 19 18 20 12.0
1.3 17 10 11 11 7.1
1.4 10 7.1 7.6 7.1 5.1
1.5 7.0 5.3 6.0 5.2 4.0
2.0 2.4 2.5 3.2 2.4 2.1

Table V. Comparison of the ARL between the S and the S 2 -EWMA control chart

S chart S 2 -EWMA chart

τ n=3 n=5 n=7 n=9 n=3 n=5 n=7 n=9
0.50 100.4 28.8 11.1 5.5 10.6 5.4 3.7 2.9
0.60 144.3 56.4 26.3 14.3 15.2 7.9 5.4 4.2
0.70 196.1 100.9 57.6 36.0 23.9 12.6 8.7 6.8
0.80 250.7 167.3 116.7 85.4 44.4 23.7 16.6 13.0
0.90 270.4 237.6 206.8 181.1 108.5 63.3 45.8 36.4
0.95 245.2 239.1 230.5 221.6 189.4 133.4 105.1 87.6
1.05 150.8 143.2 137.4 132.4 95.0 78.7 66.9 58.3
1.10 109.4 95.1 85.0 76.9 43.9 32.4 26.2 22.4
1.20 57.7 42.4 33.5 27.5 15.9 12.0 10.0 8.9
1.30 33.0 21.6 15.9 12.4 8.9 7.1 6.1 5.5
1.40 20.7 12.6 8.9 6.9 6.1 5.1 4.5 4.1
1.50 14.1 8.2 5.8 4.4 4.7 4.0 3.6 3.3
1.60 10.3 5.9 4.1 3.1 3.8 3.3 3.0 2.8
1.70 7.9 4.5 3.1 2.4 3.3 2.9 2.7 2.5
1.80 6.3 3.6 2.5 2.0 2.9 2.6 2.4 2.2
1.90 5.2 3.0 2.1 1.7 2.6 2.3 2.2 2.0
2.00 4.5 2.6 1.9 1.5 2.4 2.1 2.0 1.9

In Tables II and III we have summarized for τ ∈ (0.5, 2), for n = {3, 5, 7, 9} the optimal combinations (λ∗ , K ∗ )
which yield the smallest possible out-of-control ARL∗ . Table II corresponds to an in-control ARL0 = 370, while
Table III corresponds to an in-control ARL0 = 200. For example, the best combination for τ = 1.5, n = 5 and
ARL0 = 370 is (λ∗ = 0.05, K ∗ = 2.514) and the minimal ARL∗ = 4.8. For τ = 0.7, n = 9 and ARL0 = 200
the optimal combination is (λ∗ = 0.23, K ∗ = 2.632) and the minimal ARL∗ = 6.8. Table IV, which includes the
results of Crowder and Hamilton10, except for the last two columns, is used to compare the performance of the
S 2 -EWMA control chart with other existing control charts in the case of an increase in the sample variance
(ARL0 = 200 and n = 5 are assumed). The first two columns correspond to the range chart and the CUSUM
chart based on ranges. The EWMA(1) and EWMA(2) are two Crowder EWMA control charts dedicated to the
optimal detection of a 20% and 40% increase in the sample variance (i.e. τ = 1.2 and 1.4). The last column
corresponds to the S 2 -EWMA control chart with optimal parameters (λ∗ = 0.05, K ∗ = 2.269). This chart is
supposed to efficiently detect 20% and 40% increases in the sample variance as well (see Table III). It is clear
that whatever the value of τ , the ARL obtained with the S 2 -EWMA control chart is always smaller than the one
of the other control chart. In that case, the choice of λ∗ = 0.05 seems to be confirmed as a very good choice in
order to have a small out-of-control ARL. In Table V we compare the ARL of the S control chart with the ARL of
the S 2 -EWMA control chart, with optimized (λ∗ , K ∗ ) for different values of τ and n (ARL0 = 200 is assumed).
As expected, the S 2 -EWMA control chart mostly gives, a smaller ARL than the S control chart. The difference
is particularly impressive in the decreasing case. For example, for τ = 0.5 and n = 3, the ARL = 100.4 for the

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
A NEW S 2 -EWMA CONTROL CHART FOR MONITORING THE PROCESS VARIANCE 791

S control chart, while for the S 2 -EWMA control chart the ARL = 10.6. When the value of τ becomes too high,
the S 2 -EWMA becomes less interesting compared with the classical Shewhart S control chart, and in some
cases, the ARL is larger (for τ = 2 and n = 9, the ARL = 1.5 for the S control chart while for the S 2 -EWMA
control chart, we have ARL = 1.9). As proved by Moustakides25, the CUSUM chart is optimal among all
control charts. So it is surprising that the S 2 -EWMA gives better results than the CUSUM approach. In fact this
is not surprising since we are comparing the EWMA method applied to transformed variances versus traditional
CUSUM control charts. There is no doubt that the CUSUM version of our method (with transformation of the
variances) will give better results than the EWMA version.

CONCLUSIONS
The S 2 -EWMA chart presented in this paper is a process sample variance monitoring the control chart based on
a three-parameter logarithmic transformation. The main advantages of this control chart are (1) the easy-to-use
scheme for computing the control limits and the parameters a, b, and c of the logarithmic transformation;
(2) the improved performance in terms of normality/symmetry of the transformed sample variances; and (3) the
possibility of detecting increases and decreases in the nominal standard deviation in a more similar way. As for
the traditional average X̄-EWMA control chart, it is possible to compute the ARL of the S 2 -EWMA control
chart by numerically solving a Fredholm-type integral. The optimal parameter design (λ∗ , K ∗ ) procedure shows
that for small shifts in the nominal standard deviation the value λ∗ = 0.05 is highly recommended, while for
larger shifts, any values of λ∗ work well and give roughly the same minimal ARL, close to the ARL of the
S control chart. A comparison with Crowder and Hamilton10 shows that the S 2 -EWMA control chart gives
smaller out-of-control ARL than other procedures.

REFERENCES
1. Roberts SW. Control chart tests based on geometric moving averages. Technometrics 1959; 1(3):239–250.
2. Robinson PB, Ho TY. Average run lengths of geometric moving average charts by numerical methods. Technometrics
1978; 20(1):85–93.
3. Crowder SV. A simple method for studying run-length distributions of exponentially weighted moving average charts.
Technometrics 1987; 29(4):401–407.
4. Crowder SV. Design of exponentially weighted moving average schemes. Journal of Quality Technology 1989;
21(3):155–162.
5. Lucas JM, Saccucci MS. Exponentially weighted moving average control schemes: Properties and enhancements.
Technometrics 1990; 32(1):1–12.
6. Steiner SH. Exponentially weighted moving average control charts with time varying control limits and fast initial
response. Journal of Quality Technology 1999; 31:75–86.
7. Wortham AW, Ringer LJ. Control via exponential smoothing. The Logistics Review 1971; 7(32):33–40.
8. Sweet AL. Control charts using coupled exponentially weighted moving averages. IIE Transactions 1986; 18:26–33.
9. Ng CH, Case KE. Development and evaluation of control charts using exponentially weighted moving averages.
Journal of Quality Technology 1989; 21(4):242–250.
10. Crowder SV, Hamilton MD. An EWMA for monitoring a process standard deviation. Journal of Quality Technology
1992; 24(1):12–21.
11. Hamilton MD, Crowder SV. Average run lengths of EWMA control charts for monitoring a process standard deviation.
Journal of Quality Technology 1992; 24(1):44–50.
12. MacGregor JF, Harris TJ. The exponentially weighted moving variance. Journal of Quality Technology 1993;
25(2):106–118.
13. Gan FF. Joint monitoring of process mean and variance using exponentially weighted moving average control charts.
Technometrics 1995; 37:446–453.
14. Amin RW, Wolff H, Besenfelder W, Baxley R Jr. EWMA control charts for the smallest and largest observations.
Journal of Quality Technology 1999; 31:189–206.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
792 P. CASTAGLIOLA

15. Lu CW, Reynolds MR Jr. Control charts for monitoring the mean and variance of autocorrelated processes. Journal of
Quality Technology 1999; 31:259–274.
16. Acosta-Mejia CA, Pignatiello JJ Jr, Rao BV. A comparison of control charting procedures for monitoring process
dispersion. IIE Transactions 1999; 31:569–579.
17. Jones LA. The statistical design of EWMA control charts with estimated parameters. Journal of Quality Technology
2002; 34:277–288.
18. Box GE, Hunter WG, Hunter JS. Statistics for Experimenters. Wiley: New York, 1978.
19. Johnson NL, Kotz S, Balakrishnan N. Continuous Univariate Distributions. Wiley: New York, 1994.
20. Johnson NL. Systems of frequency curves generated by methods of translation. Biometrika 1949; 36:149.
21. Montgomery DC. Introduction to Statistical Quality Control. Wiley: New York, 2001.
22. Ryan TP. Statistical Methods for Quality Improvement. Wiley: New York, 2000.
23. Lucas JM, Crosier RB. Fast initial response for CUSUM quality-control schemes. Technometrics 1982; 4:199–206.
24. Rhoads TR, Montgomery DC, Mastrangelo CM. A fast initial response scheme for the exponentially weighted moving
average control chart. Quality Engineering 1996; 9(2):317–327.
25. Moustakides GV. Optimal stopping rules for detecting changes in distributions. Annals of Statistics 1986;
14(4):1379–1387.
26. Stuart A, Ord JK. Kendall’s Advanced Theory of Statistics, vol. 1. Edward Arnold: London, 1994.
27. Press WH, Teukolsky SA, Vetterling WT, Flannery BP. Numerical Recipes in C. Cambridge University Press:
Cambridge, 1988.

APPENDIX A
Let S 2 be the sample variance of n independent N(µ, 1) random variables. The distribution fS 2 (s) of S 2 is
defined for s ≥ 0 and it is given by the relation (see Johnson et al.19 )
  
n − 1 2
fS 2 (s) = fG s  ,
2 n−1

where fG (x|u, v) is the pdf of the gamma (u, v) distribution, i.e.


0 x≤0
fG (x|u, v) = x u−1 exp(−x/v)
 x>0
v u (u)

Consequently, the mean E(S 2 ), variance µ2 (S 2 ), and skewness γ1 (S 2 ) of S 2 are equal to

E(S 2 ) = 1
2
µ2 (S 2 ) =
n−1
8
γ1 (S 2 ) =
n−1
As a first step, three parameters A(n), B(n) and C(n), depending only on n, must be computed such that
T = A(n) + B(n) ln{S 2 + C(n)} is approximately a N(0, 1) random variable. Remembering that S 2 has a
gamma distribution, i.e. a unimodal skewed distribution, we suggest finding the three-parameter log-normal
distribution (another skewed distribution)
B(n)
fL {x|A(n), B(n), C(n)} = φ{A(n) + B(n) ln(x + C(n))}
x

which is the closest to the distribution of S 2 according to a criterion based on the first three moments E(S 2 ),
µ2 (S 2 ), and γ1 (S 2 ) of S 2 , where φ() is the pdf of the N(0, 1) distribution. If S 2 would have a log-normal

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
A NEW S 2 -EWMA CONTROL CHART FOR MONITORING THE PROCESS VARIANCE 793

distribution, then T = A(n) + B(n) ln{S 2 + C(n)} would have exactly a N(0, 1) distribution. By looking for a
log-normal distribution which approximately fits the gamma distribution of S 2 , we expect that T will have an
approximate N(0, 1) distribution. What we suggest is finding parameters A(n), B(n), and C(n) such that the
log-normal distribution fL {x|A(n), B(n), C(n)} fits the first three moments E(S 2 ), µ2 (S 2 ), and γ1 (S 2 ) of S 2 .
This can be achieved (see Stuart and Ord26 (p. 243)) by first computing

w = [ (γ1 (S 2 )/2)2 + 1 + (γ1 (S 2 )/2)]1/3 − [ (γ1 (S 2 )/2)2 + 1 − (γ1 (S 2 )/2)]1/3

and then
1
B(n) =
ln(w2 + 1)
 
B(n) w2 (w2 + 1)
A(n) = ln
2 µ2 (S 2 )
µ2 (S 2 )
C(n) = − E(S 2 )
w
For the random variable Sk2 = σ02 S 2 (corresponding to the sample variance of n independent N(µ, σ0 ) random
variables), it is straightforward to see that
T = A(n) + B(n) ln{S 2 + C(n)}
= A(n) + B(n) ln{Sk2 /σ02 + C(n)}
= A(n) − 2B(n) ln(σ0 ) + B(n) ln{Sk2 + C(n)σ02 }

Consequently, if the parameters a, b and c are defined such that b = B(n), c = C(n)σ02 and a =
A(n) − 2B(n) ln(σ0 ), then T = a + b ln(Sk2 + c) = Tk . This ensures that Tk = a + b ln(Sk2 + c) will also be
approximately a N(0, 1) random variable. Because Tk = T , the distribution fTk (t|n) of Tk depends only on n.
This distribution is defined for t ≥ A(n) + B(n) ln{C(n)} and is equal to
     
1 t − A(n) t − A(n) n − 1 2
fTk (t|n) = exp fG exp − C(n) ,
B(n) B(n) B(n) 2 n−1
Finally, it seems logical to set the first value Z0 to Z0 = E{a + b ln(S 2 + c)}. Using the first-order expansion of
the expectance, we deduce Z0  a + b ln(σ02 + c). Then by replacing a, b and c by A(n) − 2B(n) ln(σ0 ), B(n)
and C(n)σ02 , we have

Z0  A(n) − 2B(n) ln(σ0 ) + B(n) ln{σ02 + C(n)σ02 }

After simplification
Z0  A(n) + B(n) ln{1 + C(n)}

APPENDIX B

Let Sk2 = σ12 S 2 be the sample variance of n independent N(µ, σ1 ) random variables (i.e. the sample variance
after a shift τ ) and let Tk = a + b ln(Sk2 + c). If a, b, c and Sk2 are respectively replaced by A(n) −
2B(n) ln(σ0 ), B(n), σ0 C(n) and σ12 S 2 , then

Tk = A(n) − 2B(n) ln(σ0 ) + B(n) ln{σ12 S 2 + σ02 C(n)}

= A(n) − 2B(n) ln(σ0 ) + B(n) ln{σ02 (τ 2 S 2 + C(n))}
= A(n) + B(n) ln{τ 2 S 2 + C(n)}

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794
794 P. CASTAGLIOLA

This result clearly shows that the distribution fTk (t|n, τ ) of Tk is equal to the distribution of the transformed
random variable τ 2 S 2 , i.e.
      
1 t − A(n) t − A(n)  n − 1 2τ 2
fTk (t|n, τ ) = exp fG exp − C(n) ,
B(n) B(n) B(n) 2 n−1

For this reason, the distribution fTk (t|n, τ ) of Tk depends only on n and τ . In order to compute the ARL for the
method described in this paper, the Crowder3 approach has been chosen. If ARL(τ, λ, K, n, t) is the ARL of
the sequence Z1 , . . . , Zn , . . . after a shift τ , given that the EWMA starts with Z0 = t, then it must satisfy
  
1 UCL z − (1 − λ)t 
ARL(τ, λ, K, n, t) = 1 + ARL(τ, λ, K, n, z)fTk  n, τ dz (B1)
λ LCL λ

where LCL and UCL are the fixed control limits defined by Equations (2) and (3). As pointed out by Crowder,
the equation above is a Fredholm integral equation of the second kind which can be approximated numerically
(some suitable methods are available, see Press et al.27, but are not described here).

Author’s biography
Philippe Castagliola graduated (PhD 1991) from the ‘Université de Technologie de Compiègne’ (France).
He worked at the ‘Ecole des Mines de Nantes’ (France) from 1991 to 2004 as an associate professor. He is
currently professor at the ‘Institut Universitaire de Technologie de Nantes’ (France). Since 2000, he has been
a member of the ‘Institut de Recherche en Communications et Cybernétique de Nantes’, UMR CNRS 6597
(France). His research activities include the development of new SPC techniques.

Copyright 
c 2005 John Wiley & Sons, Ltd. Qual. Reliab. Engng. Int. 2005; 21:781–794