>Springer-Verlag 2003
A Bayesian-like estimator of the process capability
index C
pmkW. L. Pearn1 and G. H. Lin2*
1 Department of Industrial Engineering & Management, National Chiao Tung University, 2 Department of Communication Engineering, National Penghu, Institute of Technology, Penghu, Taiwan, ROC
Abstract. Pearn et al. (1992) proposed the capability index Cpmk, and
inves-tigated the statistical properties of its natural estimator ^CCpmkfor stable normal
processes with constant mean m. Chen and Hsu (1995) showed that under general conditions the asymptotic distribution of ^CCpmkis normal if m 0 m, and
is a linear combination of the normal and the folded-normal distributions if m¼ m, where m is the mid-point between the upper and the lower specifi-cation limits. In this paper, we consider a new estimator ~CCpmk for stable
pro-cesses under a di¤erent (more realistic) condition on process mean, namely, Pðm b mÞ ¼ p, 0 a p a 1. We obtain the exact distribution, the expected value, and the variance of ~CCpmk under normality assumption. We show that
for Pðm b mÞ ¼ 0, or 1, the new estimator ~CCpmk is the MLE of Cpmk, which
is asymptotically e‰cient. In addition, we show that under general condi-tions ~CCpmkis consistent and is asymptotically unbiased. We also show that the
asymptotic distribution of ~CCpmkis a mixture of two normal distributions.
Keywords and Phrases: process capability index; Bayesian-like estimator; consistent; mixture distribution
1. Introduction
Pearn et al. (1992) proposed the process capability index Cpmk, which
com-bines the merits of two earlier indices Cpk(Kane (1986)) and Cpm (Chan et al.
(1988)). The index Cpmk alerts the user if the process variance increases and/
or the process mean deviates from its target value, and is designed to monitor the normal and the near-normal processes. The index Cpmkis considered
argu-ably the most useful index to date for processes with two-sided specification * The research was partially supported by National Science Council of the Republic of China (NSC-89-2213-E-346-003).
limits (Boyles (1994), Wright (1995)). The index Cpmk, referred to as the
third-generation capability index, has been defined as the following: Cpmk¼ min USL m 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ; m LSL 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q 8 > < > : 9 > = > ;; ð1Þ
where USL and LSL are the upper and the lower specification limits, respec-tively, m is the process mean, s is the process standard deviation, and T is the target value. We note that Cpmkcan be rewritten as:
Cpmk¼ d jm mj 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ; ð2Þ
where m is the mid-point between the upper and the lower specification limits, and d is the half length of the specification interval ½LSL; USL. That is, m¼ ðUSL þ LSLÞ=2, and d ¼ ðUSL LSLÞ=2. For stable processes where the process mean m is assumed to be a constant (unknown), Pearn et al. (1992) considered the natural estimator of Cpmkwhich is defined as:
^ C Cpmk¼ d jX mj 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 n þ ðX TÞ 2 q ; ð3Þ
where X ¼ ðPi¼1n XiÞ=n and Sn ¼ fn1Pi¼1n ðXi X Þ2g1=2 are conventional
estimators of the process mean and the process standard deviation, m and s, respectively. If the process characteristic follows the normal distribution, Pearn et al. (1992) showed that for the case with T¼ m (symmetric tolerance) the distribution of the natural estimator ^CCpmk is a mixture of the chi-square
distribution and the non-central chi-square distribution, as expressed in the following: ^ C Cpmk@ dpffiffiffin s w 0 1ðlÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiw2 n1þ wn102 ðlÞ q ; ð4Þ where w2
n1 is the chi-square distribution with n 1 degrees of freedom, w10ðlÞ
is the central chi distribution with one degree of freedom and non-centrality parameter l, and w02
n1ðlÞ is the non-central chi-square distribution
with n 1 degrees of freedom and non-centrality parameter l, where l ¼ nðm TÞ2=s2. Chen and Hsu (1995) showed that the natural estimator ^CC
pmkis
asymptotically unbiased. Chen and Hsu (1995) also showed that under general conditions the natural estimator ^CCpmk converges to the normal distribution
Nð0; s2 pmkÞ, where s2pmk¼ s 2 9½s2þ ðm TÞ2þ 12ðm TÞs2 6m 3 18½s2þ ðm TÞ23=2 ( ) Cpmk þ 144ðm TÞ 2 s2 144ðm TÞm 3þ 36ðm4 s4Þ 144½s2þ ðm TÞ23=2 ( ) Cpmk2 ; ð5Þ m3, m4 are the third and fourth central moment of the process, respectively.
2. A Bayesian-like estimator
In real-world applications, the production may require multiple supplies with di¤erent quality characteristics on each single shipment of the raw materials, multiple manufacturing lines with inconsistent precision in machine settings and engineering e¤ort for each manufacturing line, or multiple workmanship shifts with unequal performance level on each shift. Therefore, the basic and common assumption that the process mean stay as a constant may not be satisfied in real situations. Consequently, using the natural estimator ^CCpmk to
measure the potential and performance of such a process is inappropriate as the resulting capability measure would not be accurate. For stable processes under those conditions, if the knowledge on the process mean, Pðm b mÞ ¼ p, 0 a p a 1, is available, then we can consider the following new estimator ~CCpmk.
In general, the probability Pðm b mÞ ¼ p, 0 a p a 1, can be obtained from historical information of a stable process.
~ C Cpmk¼ bn1½d ðX mÞIAðmÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q ; ð6Þ where bn1¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2=ðn 1Þ p
fG½ðn 1Þ=2=G½ðn 2Þ=2g is the correction factor, IAðÞ is the indicator function defined as IAðmÞ ¼ 1 if m A A, and IAðmÞ ¼ 1
if m B A, where A¼ fm j m b mg. We note that the new estimator ~CCpmkcan be
rewritten as the following:
~ C Cpmk¼ bn1½d ðX mÞIAðmÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q ¼ d ðX mÞIAðmÞ 3b1 n1Sn ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1þðX TÞ 2 S2 n s ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiCC~pk 1þðX TÞ 2 S2 n s ; ð7Þ where ~CCpk¼ bn1½d ðX mÞIAðmÞ=ð3SnÞ as defined by Pearn and Chen
(1996). If the process characteristic follows the normal distribution, Nðm; s2Þ,
then we can show the following Theorem.
Theorem 1. If the process characteristic follows the normal distribution, then
~ C Cpmk@ bn1 ffiffiffin p Cp Nðh; 1Þ 3 ffiffiffiffiffiffi w02 n
p , where Nðh; 1Þ is the normal distribution with mean h¼ 3pffiffiffinðCp CpkÞ, wn02 is the non-central chi-square distribution with n
degrees of freedom and non-centrality parameter l¼ nðm TÞ2=s2.
Proof: We note that 3b1
n1SnCC~pk¼ d ðX mÞIAðmÞ is distributed as the
normal distribution Nð3sCpk;s2=nÞ. Therefore, bn1½d ðX mÞIAðmÞ=
ð3sÞ ¼ bn1f½d=ð3sÞ ½ðX mÞIAðmÞ=ð3sÞ is distributed as bn1fCp
½Nðh; 1Þ=ð3pffiffiffinÞg, where Nðh; 1Þ is the normal distribution with mean h ¼ 3pffiffiffinðCp CpkÞ. We also note that ½nSn2þ nðX TÞ
2=s2 ¼Pn
i¼1ðXi TÞ2=s2
is distributed as wn02, the non-central chi-square distribution with n degrees of freedom and non-centrality parameter l¼ nðm TÞ2=s2. Therefore, ~CC
pmk is distributed as bn1f ffiffiffin p Cp ½Nðh; 1Þ=3g= ffiffiffiffiffiffi w02 n p .
The r-th moment (about zero) of ~CCpmk, therefore, can be obtained as: Eð ~CCpmkr Þr¼ E 8 > > > < > > > : bn1 ffiffiffin p Cp Nðh; 1Þ 3 ffiffiffiffiffiffi w02 n p 9 > > > = > > > ; r ¼X r i¼0 bn1r r i E Nðh; 1Þ 3pffiffiffinCp i ffiffiffi n p Cp ffiffiffiffiffiffi w02 n p " #r ( ) ; ð8Þ
By setting r¼ 1, and r ¼ 2, we may obtain the first two moments and the variance as: Eð ~CCpmkÞ ¼ X1 i¼0 bn1 1 i E Nðh; 1Þ 3pffiffiffinCp i ffiffiffi n p Cp ffiffiffiffiffiffi w02 n p " # ( ) ; ð9Þ Eð ~CCpmk2 Þ ¼X 2 i¼0 b2n1 2 i E Nðh; 1Þ 3pffiffiffinCp i ffiffiffi n p Cp ffiffiffiffiffiffi w02 n p " #2 8 < : 9 = ;; ð10Þ
Varð ~CCpmkÞ ¼ Eð ~CCpmk2 Þ ½Eð ~CCpmkÞ2: ð11Þ
We note that for the case with pðm b mÞ ¼ 1, ~CCpmk< ^CCpmk for X b m
and ~CCpmk> ^CCpmk for X < m d½ð1 bn1Þ=ð1 þ bn1Þ. If the process
distri-bution is normal, then the probability PðX b mÞ ¼ Ffpffiffiffin½ðm mÞ=sg con-verges to 1. Thus, for large values of n, we expect to have ~CCpmk< ^CCpmk. On
the other hand, if Pðm b mÞ ¼ 0, then we have ~CCpmk< ^CCpmk for X a m and
~ C
Cpmk> ^CCpmkfor X > mþ d½ð1 bn1Þ=ð1 þ bn1Þ. If the process distribution
is normal, then the probability PðX a mÞ ¼ Ffpffiffiffin½ðm mÞ=sg converges to 1. Thus, for large values of n, we also expect to have ~CCpmk< ^CCpmk. Explicit
forms of the expected value and the variance of ~CCpmk are analytically
intrac-table. But, for the cases with Pðm b mÞ ¼ 1 or 0, the probability density function may be obtained (the proof is omitted for the simplicity of the pre-sentation).
3. Asymptotic distribution of ~CCpmk
In the following, we show that if the knowledge on the process mean, the probabilities Pðm b mÞ ¼ p, and Pðm < mÞ ¼ 1 p, with 0 a p a 1 is given, then the asymptotic distribution of the proposed new estimator ~CCpmk is a
mixture of two normal distributions. We first present some Lemmas. The proofs for these Lemmas can be found in the reference Serfling (1980). A direct consequence of our result is that for the cases with either Pðm b mÞ ¼ 1, or Pðm b mÞ ¼ 0, the asymptotic distribution will then be an ordinary normal distribution.
Lemma 1: If m4¼ EðX mÞ4 exists, then pffiffiffinðX m; S2
n s2Þ converges to Nðð0; 0Þ; SÞ in distribution, where S ¼ s 4 m 3 m3 m4 s4 .
Lemma 2: If gðx; yÞ is a real-valued di¤erentiable function, then ffiffiffi n p ½gðX ; S2 nÞ gðm; s2Þ converges to Nð0; DSD0Þ in distribution, if D ¼ qg qx ð m; s2Þ ;qg qy ð m; s2Þ 0ð0; 0Þ.
Lemma 3: If the random vector ðw1n; w2n; . . . ; wknÞ converges to the random
vectorðw1; w2; . . . ; wkÞ in distribution, and the random vector ðv1n; v2n; . . . ; vknÞ
converges to the random vector ðv1; v2; . . . ; vkÞ in probability, then the
ran-dom vectorðv1nw1n; v2nw2n; . . . ; vknwknÞ converges to the random vector ðv1w1;
v2w2; . . . ; vkwkÞ in distribution.
Lemma 4: If the random vector ðw1n; w2n; . . . ; wknÞ converges to the random
vector ðw1; w2; . . . ; wkÞ in distribution, and the function g is continuous with
probability one, then gðw1n; w2n; . . . ; wknÞ converges to gðw1; w2; . . . ; wkÞ in
dis-tribution.
Lemma 5: If the random vectorðv1n; v2n; . . . ; vknÞ converges to the random
vec-torðv1; v2; . . . ; vkÞ in probability, and the function g is continuous with
proba-bility one, then gðv1n; v2n; . . . ; vknÞ converges to gðv1; v2; . . . ; vkÞ in probability.
Lemma 6: Ifm4¼ EðX mÞ4exists, thenpffiffiffinðX m; X m; S2
n s2Þ converges to Nðð0; 0; 0Þ; SÞ in distribution, where S¼ s2 s2 m 3 s2 s2 m 3 m3 m3 m4 s4 2 6 4 3 7 5. Proof: See Chen and Hsu (1995).
Theorem 2: The estimator ~CCpmkis consistent.
Proof: We first note thatðX ; S2
nÞ converges to ðm; s2Þ in probability and bn1
converges to 1 as n! y. Since ~CCpmkis a continuous function ofðX ; Sn2Þ, then
it follows directly from Lemma 5 that ~CCpmk converges to Cpmk in probability.
Hence, ~CCpmkmust be consistent.
Theorem 3: Under general conditions, if the fourth central moment m4¼
EðX mÞ4exists, thenpffiffiffinð ~CCpmk CpmkÞ converges to p Nð0; spmk12 Þ þ ð1 pÞ
Nð0; s2 pmk2Þ in distribution, where spmk12 ¼D 2 1 9 1þ ðm TÞ2 s2 " #1 þD1 3 m3 s3 1þ ðm TÞ2 s2 " #3=2 Cpmk1 þ1 4 m4 s4 s4 1þ ðm TÞ2 s2 " #2 Cpmk12 spmk22 ¼D 2 2 9 1þ ðm TÞ2 s2 " #1 þD2 3 m3 s3 1þ ðm TÞ2 s2 " #3=2 Cpmk2 þ1 4 m4 s4 s4 1þ ðm TÞ2 s2 " #2 Cpmk22
D1¼ 9ðm TÞC2 pmk1 d ðm mÞ þ 1; Cpmk1¼ d ðm mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ; D2¼ 9ðm TÞC2 pmk2 dþ ðm mÞ 1; Cpmk2¼ dþ ðm mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q :
Proof: (CASE I) If m > m, we define the function g1ðx; yÞ ¼
d ðx mÞ 3
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi yþ ðx TÞ2
q ,
where x > m, y > 0. Since g1is di¤erentiable, then we have
qg1 qx ð m; s2Þ ¼ D1Cpmk1 d ðm mÞ; qg1 qy ð m; s2Þ ¼ 9 2 C3 pmk1 ½d ðm mÞ2; where D1¼ 9ðm TÞC2 pmk1 d ðm mÞ þ 1, and Cpmk1 ¼ d ðm mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q . If we define D1¼ qg1 qx ð m; s2Þ ;qg1 qy ð m; s2Þ ! , then D10ð0; 0Þ.
By Lemma 1 and Lemma 2, pffiffiffinðb1
n1CC~pmk CpmkÞ ¼ ffiffiffin
p
½g1ðX ; Sn2Þ
g1ðm; s2Þ converges to Nð0; spmk12 Þ in distribution, where
s2pmk1¼ D1SD10 ¼ D12 9 1þ ðm TÞ2 s2 " #1 þD1 3 m3 s3 1þ ðm TÞ2 s2 " #3=2 Cpmk1 þ1 4 m4 s4 s4 1þ ðm TÞ2 s2 " #2 Cpmk12
(CASE II) If m < m, we define the function g2ðx; yÞ ¼
dþ ðx mÞ 3
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi yþ ðx TÞ2
q ,
where x < m, y > 0. Since g1is di¤erentiable, then we have
qg2 qx ð m; s2Þ ¼ D2Cpmk2 dþ ðm mÞ; qg2 qy ð m; s2Þ ¼ 9 2 C3 pmk2 ½d þ ðm mÞ2; where D2¼ 9ðm TÞC2 pmk2 dþ ðm mÞ 1, and Cpmk2 ¼ dþ ðm mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q . If we define D2¼ qg2 qx ð m; s2Þ ;qg2 qy ð m; s2Þ ! , then D20ð0; 0Þ.
By Lemma 1 and Lemma 2, pffiffiffinðb1
n1CC~pmk CpmkÞ ¼ ffiffiffin
p
½g2ðX ; Sn2Þ
g2ðm; s2Þ converges to Nð0; spmk22 Þ in distribution, where
s2pmk2¼ D2SD20 ¼ D22 9 1þ ðm TÞ2 s2 " #1 þD2 3 m3 s3 1þ ðm TÞ2 s2 " #3=2 Cpmk2 þ1 4 m4 s4 s4 1þ ðm TÞ2 s2 " #2 Cpmk22
(CASE III) If m¼ m, then ffiffiffi n p ðb1 n1CC~pmk CpmkÞ ¼ ffiffiffi n p ðX mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q d 3 ½pffiffiffinðS2 n s2Þ þ ffiffiffi n p ðX2 m2Þ 2pffiffiffinðX mÞT ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q ½ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 n þ ðX TÞ 2 q : We define v1n¼ 1 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q , v2n¼ d 3½ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 n þ ðX TÞ 2 q ½ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 n þ ðX TÞ 2 q ; w1n¼ ffiffiffin p ðX mÞ; w2n¼ ffiffiffin p ðS2 n s2Þ þ ffiffiffi n p ðX2 m2Þ 2pffiffiffinðX mÞT: SinceðX ; S2
nÞ converges to ðm; s2Þ in probability, then ðv1n; v2nÞ converges to
ðv1; v2Þ in probability, where v1¼ Cpmk0 d , v2¼ 9C3 pmk0 2d2 ! , and Cpmk0¼ d 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2
q . Define the function Gðx; y; zÞ ¼ ðx; z þ xy 2xTÞ. Then by Lemma 4 and Lemma 6,ðw1n; w2nÞ ¼ ffiffiffin
p
½GðX ; X ; S2
nÞ Gðm; m; s2Þ
con-verges toðw1; w2Þ which is distributed as Nðð0; 0Þ; SGÞ, where
SG¼ 1 0 0 a m 1 s2 s2 m 3 s2 s2 m 3 m3 m3 b 2 6 4 3 7 5 1 a 0 m 0 1 2 4 3 5 ¼ s2 c c d ; with a¼ m 2T; b ¼ m4 s4; c¼ 2ðm TÞs2þ m3; d ¼ 4ðm TÞ2þ 4ðm TÞm3þ ðm4 s4Þ: Hence, by Lemma 3, 2 6 4 ffiffiffi n p ðX mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q ; 3 d ffiffiffi n p ðS2 n s2Þ þ ffiffiffi n p ðX2 m2Þ 2pffiffiffinðX mÞT ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q ½ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi S2 nþ ðX TÞ 2 q 3 7 5
converges toðv1w1; v2w2Þ ¼ Cpmk d w1; 9C3 pmk 2d2 w2 ! in distribution. Define Hðx; yÞ ¼ x þ y. Then, by Lemma 4,pffiffiffinðb1
n1CC~pmk CpmkÞ converges to Y ¼ Cpmk d w1 9 2 C3 pmk
d2 w2, which is a normal distribution with EðY Þ ¼ 0,
VarðY Þ ¼D 2 0 9 1þ ðm TÞ2 s2 " #1 þD0 3 m3 s3 1þ ðm TÞ2 s2 " #3=2 Cpmk0 þ1 4 m4 s4 s4 1þ ðm TÞ2 s2 " #2 Cpm02 where D0¼ 9ðm TÞC2 pmk0 d þ 1, Cpmk0¼ d 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q . Since Pfpffiffiffinðb1 n1CC~pmk CpmkÞ a rg ¼ Pfm b mgPf ffiffiffin p ðb1 n1CC~pmk CpmkÞ a rj m b mg þ Pfm < mgPfpffiffiffinðb1
n1CC~pmk CpmkÞ a r j m < mg for all real
number r, then it follows that pffiffiffinðb1
n1CC~pmk CpmkÞ converges to p Nð0; s2 pmk1Þ þ ð1 pÞ Nð0; spmk22 Þ in distribution. Since ffiffiffi n p ð ~CCpmk CpmkÞ ¼ ffiffiffi n p ðb1 n1CC~pmk CpmkÞ þ ffiffiffin p ð ~CCpmk b1n1CC~pmkÞ and bn1 converges to 1 as
n! y, thus by Slutsky’s theory, the theorem proved. Corollary 3.1: The estimator ~CCpmkis asymptotically unbiased.
Proof: From Theorem 3, pffiffiffinð ~CCpmk CpmkÞ converges to the following
p Nð0; s2
pmk1Þ þ ð1 pÞ Nð0; spmk22 Þ in distribution. Therefore,
Efpffiffiffinð ~CCpmk CpmkÞg converges to zero, and so ~CCpmk must be asymptotically
unbiased.
Corollary 3.2: If the process characteristic follows the normal distribution, thenffiffiffi n
p
ð ~CCpmk CpmkÞ converges to p Nð0; spmk12 0Þ þ ð1 pÞ Nð0; spmk22 0Þ, a mix-ture of two normal distributions, where
s2pmk10 ¼ D12 9 1þ ðm TÞ2 s2 " #1 þ1 2 1þ ðm TÞ2 s2 " #2 Cpmk12 ; s2pmk20 ¼ D22 9 1þ ðm TÞ2 s2 " #1 þ1 2 1þ ðm TÞ2 s2 " #2 Cpmk22 ; D1¼ 9ðm TÞC2 pmk1 d ðm mÞ þ 1; Cpmk1¼ d ðm mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q ; D2¼ 9ðm TÞC2 pmk2 dþ ðm mÞ 1; Cpmk2¼ dþ ðm mÞ 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi s2þ ðm TÞ2 q :
for the case with Pðm b mÞ ¼ 0, or 1, (i) ~CCpmkis the MLE of Cpmk, (ii) ~CCpmkis
asymptotically e‰cient.
Proof: (i) For normal distributions, ðX ; S2
nÞ is the MLE of ðm; s2Þ. By the
invariance property, ~CCpmk is the MLE of Cpmk.
(ii) The Fisher information matrix can be calculated as: IðyÞ ¼ a 0 b0 c0 d0 ¼ s 2 0 0 ð2s2Þ2 ; where y¼ ðm; s2Þ; a0¼ E q qmln fðx; yÞ 2 ; b0¼ c0¼ E q qmln fðx; yÞ q qs2 ln fðx; yÞ ; and d0¼ E q qs2 ln fðx; yÞ 2 :
If Pðm b mÞ ¼ 1, then the information lower bound reduces to
q qmCpmk; q qs2Cpmk I1ðyÞ n q qmCpmk q qs2Cpmk 2 6 6 6 4 3 7 7 7 5 ¼D 2 1 9n 1þ ðm TÞ2 s2 ( )1 þC 2 pmk1 2 1þ ðm TÞ2 s2 ( )2 ¼s 2 pmk10 n :
On the other hand, if Pðm b mÞ ¼ 0, then the information lower bound reduces to q qmCpmk; q qs2Cpmk I1ðyÞ n q qmCpmk q qs2Cpmk 2 6 6 6 4 3 7 7 7 5 ¼D 2 2 9n 1þ ðm TÞ2 s2 ( )1 þC 2 pmk2 2 1þ ðm TÞ2 s2 ( )2 ¼s 2 pmk20 n :
Since the information lower bound is achieved (Corollary 3.2), then for the case with Pðm b mÞ ¼ 0, or 1, ~CCpmkis asymptotically e‰cient.
In practice, to evaluate the estimator ~CCpmk we need to determine the value
of the indicator which requires additionally the knowledge of Pðm b mÞ, or Pðm < mÞ. If historical information of the process shows Pðm b mÞ ¼ p, then we may determine the value IAðmÞ ¼ 1, or 1 using available random number
tables. For example, assume p¼ 0:375 is given, then IAðmÞ ¼ 1 if the
gen-erated 3-digit random number is no greater than 375, and IAðmÞ ¼ 1
4. Conclusions
Pearn et al. (1992) proposed the capability index Cpmk, which is designed to
monitor the normal and the near-normal processes. The index Cpmk is
con-sidered to be the most useful index to date for processes with two-sided spec-ification limits. Pearn et al. (1992) investigated the statistical properties of the natural estimator ^CCpmk for stable normal processes with constant mean m. In
this paper, we considered stable processes under a di¤erent condition (more realistic) where the process mean may not be a constant. For stable processes under such conditions with given knowledge of Pðm b mÞ ¼ p, 0 a p a 1, we investigated a new estimator ~CCpmkusing the given information.
We obtained the exact distribution of the new estimator, and derived its expected value and variance under normality assumption. For cases with Pðm b mÞ ¼ 0, or 1, we showed that the new estimator ~CCpmk is the MLE of
Cpmk. In addition, we showed that under general conditions ~CCpmkis consistent
and is asymptotically unbiased. We also showed that the asymptotic distribu-tion of ~CCpmk is a mixture of two normal distributions. The results obtained
in this paper allow us to perform a more accurate capability measure for processes under more realistic conditions in which using existing method (estimator) is inappropriate.
References
[1] Boyles RA (1994) Process capability with asymmetric tolerances. Communications in Statis-tics: Simulation and Computation 23(3):615–643
[2] Chan LK, Cheng SW, Spiring FA (1988) A new measure of process capability: Cpm. Journal
of Quality Technology 20(3):162–175
[3] Chen SM, Hsu NF (1995) The asymptotic distribution of the process capability index Cpmk.
Communications in Statistics: Theory and Methods 24(5):1279–1291
[4] Kane VE (1986) Process capability indices. Journal of Quality Technology 18(1):41–52 [5] Pearn WL, Kotz S, Johnson NL (1992) Distributional and inferential properties of process
capability indices. Journal of Quality Technology 24(4):216–233
[6] Pearn WL, Chen KS (1996) A Bayesian-like estimator of Cpk. Communications in Statistics:
Simulation & Computation 25(2):321–329
[7] Serfling RJ (1980) Approximation theory of mathematical statistics. John Wiley and Sons, New York
[8] Wright PA (1995) A process capability index sensitive to skewness. Communications in Sta-tistics: Simulation & Computation 52:195–203