MANAGEMENT SCIENCE

doi 10.1287/mnsc.1070.0741ec
pp. ec1ec16

informs

2007 INFORMS

e-companion
ONLY AVAILABLE IN ELECTRONIC FORM

Electronic CompanionPerformance Contracting in After-Sales Service

Supply Chains by Sang-Hyun Kim, Morris A. Cohen,
Serguei Netessine, Management Science, doi 10.1287/mnsc.1070.0741.

Technical Appendix to:

Performance Contracting in After-Sales Service Supply Chains
Sang-Hyun Kim, Morris A. Cohen, and Serguei Netessine
The Wharton School, University of Pennsylvania
Philadelphia, PA 19104
November 13, 2007

Abstract
This Technical Appendix accompanies the paper by Kim et al. 2007, Performance Contracting
in After-Sales Service Supply Chains. There are three appendices. In Appendix A, a numerical
example of two symmetric suppliers case is presented. Appendix B shows proofs of all analytical
results. Finally, Appendix C presents a discussion on using an alternative backorder measure (see
Section 3.4 of the main paper for the motivation).

Appendix A: Numerical Example of Two Symmetric Suppliers Case

In this example we assume that all parameter values are symmetric across the two suppliers except
for either fri g or fVar["i ]g, thereby isolating the trade-o between incentives and risk. Default values
are

2
i

b0 = 4. A normal distribution of the

= 10, ci = 1, ki = 0:2, r1 = 0:1, Var["1 ] = 10, and B

inventory on-order is chosen, in keeping with our continuous approximation for the underlying inventory
model. We vary supplier 2s risk aversion r2 and cost uncertainty Var["2 ] as well as the customers risk
aversion r0 in order to observe their eects on f

SB ; v SB g
i
i

SB
and faSB
i ; si g. We note that the rst-best

inventory positions are sF1 B = sF2 B = 8:725. Table 1 summarizes the results of varying r2 and r0 .
We observe only minimal changes in sSB
and sSB
as parameters change, with the greatest change
1
2
occurring when r0 is large. In contrast,

SB
2

changes widely and so does v2SB , but in the opposite

direction. As we observed in the single supplier case earlier, the complementarity between the cost
reduction incentive 1

and the performance incentive vi continues to exist in the multiple supplier

setting. fsSB
i g do not vary much because they are subject to an externality, namely, the overall
shkim@wharton.upenn.edu, cohen@wharton.upenn.edu, netessine@wharton.upenn.edu

r0 = 0:01
0:01
0:1

r2
FB
2
SB
1
SB
2
v1SB
v2SB
aSB
1
aSB
2
sSB
1
sSB
2

0.500
0.279
0.060
0.994
1.408
3.603
4.699
8.729
8.722
13.342

U0

0.909
0.279
0.279
0.993
0.993
3.603
3.603
8.725
8.725
13.923

r0 = 0:1
0:01
0:1

1
0.990
0.279
0.755
0.987
0.288
3.603
1.224
8.678
8.773
15.153

0.091
0.318
0.184
0.952
1.215
3.409
4.082
8.779
8.672
14.000

0.500
0.318
0.318
0.944
0.944
3.409
3.409
8.725
8.725
14.435

0:01

r0 = 1
0:1

0.909
0.318
0.726
0.934
0.317
3.409
1.371
8.650
8.802
15.676

0.010
0.430
0.436
0.823
0.826
2.849
2.820
8.905
8.559
19.104

0.091
0.430
0.430
0.801
0.801
2.852
2.852
8.725
8.725
19.153

1
0.500
0.429
0.539
0.758
0.490
2.856
2.303
8.330
9.151
20.141

Table 1: Eects of changing r2 .

Var["2 ]
SB
1
SB
2
SB
v1
v2SB
aSB
1
aSB
2
sSB
1
sSB
2

U0

(
1
0.279
0.172
0.993
1.127
3.603
4.138
8.724
8.727
13.653

r0 = 0:01
F B = 0:909)
2
10

100

0.279
0.279
0.993
0.993
3.603
3.603
8.725
8.725
13.923

0.279
0.650
0.992
0.507
3.603
1.752
8.718
8.733
15.139

r0 = 0:1
( F2 B = 0:5)
1
10
100
0.318
0.268
0.944
1.007
3.409
3.658
8.725
8.726
14.172

0.318
0.318
0.944
0.944
3.409
3.409
8.725
8.725
14.435

0.318
0.442
0.944
0.785
3.409
2.790
8.723
8.728
16.780

(
1

r0 = 1
= 0:091)
10
100

FB
2

0.429
0.654
0.794
0.505
2.853
1.731
8.670
8.782
17.794

0.430
0.430
0.801
0.801
2.852
2.852
8.725
8.725
19.153

0.430
0.156
0.800
1.149
2.852
4.221
8.715
8.735
24.341

Table 2: Eects of changing Var["2 ] when r1 = r2 = 0:1.

backorder constraint. Although fsSB
i g can be used as instruments for protection from performance
risk, their ranges are limited by the constraint. Hence, risk allocation (including allocation of the
performance risk) is primarily achieved through varying f i g.
From the table we also conrm that

SB
2

>

FB
2

for a relatively large ratio r0 =r2 , while the opposite

is true for a small ratio r0 =r2 , just as predicted by Proposition 5 for a single supplier. Furthermore, we
notice that

SB
2

increases monotonically in r2 for small r0 (= 0.01), but we do not observe the same

monotonicity when r0 is large. Specically, when r0 = 1,

SB
2

initially decreases from 0.436 to 0.430 but

then increases to 0.539. The explanation is as follows. When r0 is small, increasing

tends to reduce

both the cost and performance premiums, as can be inferred from inspection of (12) and (13). However,
when r0 is large, tension exists between the two risk premium terms. As before, increasing

reduces

the performance risk for both the customer and the suppliers. But it now exposes the customer to the
risk of greater cost (since the rst term in the cost premium r0

2
i

+ ri (1

2
i)

Var["i ]=2 dominates).

These two opposing forces break down the monotonicity.

Next, Table 2 illustrates the eect of varying Var["2 ]. Once again we observe that sSB
is not very
2
sensitive to changes in Var["2 ]; but that

SB
2

is. From the table, we see that

2

SB
2

moves toward

FB
2

b0
B

(sSB
2

SB
1
SB
2
SB
v1
v2SB
aSB
1
aSB
2
sSB
1
sSB
2
FB
sSB
1 )=s1

IIR1
IIR2
NCR1
NCR2
CRP1
CRP2
PRP1
PRP2

0.421
0.592
1.678
0.792
2.896
2.039
11.852
12.205
2.94%
11.852
12.205
2.057
1.623
0.610
1.708
1.252
0.574
24.521

0.404
0.604
1.147
0.557
2.980
1.981
10.384
10.752
3.48%
10.384
10.752
2.092
1.589
0.586
1.696
1.160
0.587
21.484

0.398
0.614
0.935
0.463
3.011
1.929
9.355
9.751
4.15%
9.355
9.751
2.104
1.557
0.577
1.687
1.116
0.600
19.425

0.396
0.624
0.820
0.412
3.021
1.879
8.511
8.948
5.00%
8.511
8.948
2.108
1.526
0.574
1.680
1.090
0.611
17.781

0.396
0.634
0.748
0.379
3.021
1.830
7.765
8.259
6.17%
7.765
8.259
2.108
1.495
0.574
1.675
1.074
0.621
16.364

0.397
0.644
0.699
0.357
3.015
1.782
7.078
7.644
7.69%
7.078
7.644
2.106
1.465
0.576
1.671
1.062
0.628
15.088

0.399
0.653
0.666
0.341
3.006
1.737
6.425
7.083
9.75%
6.425
7.083
2.102
1.435
0.578
1.668
1.055
0.632
13.905

U0
^0 , with r0 = 0:5, r1 = 0:1 and r2 = 1. See the caption below Table 3 of
Table 3: Eects of changing B
the main paper for interpretation of acronyms.
as Var["2 ] increases (regardless of the value of r0 ), conrming the prediction from Proposition 5 for the
single supplier case. In addition, numbers in the table indicate that @
whereas @

SB =@r
0
2

SB =@r
0
2

> 0 for small Var["2 ]

< 0 for large Var["2 ] (this can be proven analytically in the single supplier case,

but we omit the derivation). In other words, when cost uncertainty is relatively low, a more risk-averse
customer moves toward a C+ contract and takes up a larger portion of the cost risk, a nonintuitive
result. What actually happens is that the performance premium is more important in this situation,
outweighing the concern for the cost risk. Clearly, in the presence of performance risk, the intuition
regarding cost sharing is not always straightforward.
Finally, Table 3 shows how the optimal contract parameters and the suppliersactions vary as the
overall backorder constraint changes. In this example, suppliers 1 and 2 are asymmetric only in their
b0
attitude toward risk: r1 = 0:1 and r2 = 1; while r0 = 0:5. As expected, v1SB and v2SB decrease as B

increases, since a less stringent backorder constraint allows for smaller inventories and hence reduces the
need for performance incentives. Changes in

SB
1

and

SB
2

are relatively small. We see that distortion

SB
in fsSB
i g becomes larger as we relax the constraint (measured by the quantity (s2

FB
sSB
1 )=s1 ; it

b0 = 1 to 9.75% at B
b0 = 7). Intuitively, this happens because the less stringent
grows from 2.94% at B
backorder constraint results in a larger range in which inventories can be adjusted without violating the
constraint. However, the magnitude of the distortion is still small, conrming our previous observation
that the presence of the backorder constraint limits the customers contract parameter choices ( and

v) such that they induce the base stock levels close to the rst-best values fsFi B g.

Appendix B: Proofs
Proof of Proposition 1.
i.e., wi

(1

i )(ci si

We rst prove that at the equilibrium, all (IRi ) constraints are binding,

ai )

ki a2i =2

vi E [Bi j si ]

2
i ) Var["i ]=2

ri (1

all i. Suppose otherwise, i.e., that there exists j such that wj

kj a2j =2

2
j ) Var["j ]=2

rj (1

(1

ri vi2 Var[Bi j si ]=2 = 0 for

j )(cj sj

aj )

vj E [Bj j sj ]

rj vj2 Var[Bj j sj ]=2 > 0. By reducing wj by , the customers utility (5)

is increased by while the (AR) constraint is unaected. This result allows us to transform (AF B ) into

min

i ;vi ;ai ;si g

s.t:

Pn

i=1

Pn

ci si

i=1 E

ai + ki a2i =2 + r0

[Bi j si ]

2
i

+ ri (1

2
i)

Var["i ]=2 + (r0 + ri ) vi2 Var[Bi j si ]=2 ;

b0 :
B

Clearly, the objective function is minimized when vi = 0 for all i. With this observation, the Lagrangian
with the associated multiplier
L =
=

becomes

Pn

2
ci si ai + ki a2i =2 + r0 2i + ri (1
i ) Var["i ] =2 +
b0 + Pn ci si + E [Bi j si ] ai + ki a2 =2 + r0 2 + ri (1
B
i
i
i=1

i=1

Pn

i=1 E

2
i)

[Bi j si ]

Var["i ]=2 :

It is apparent that the minimization can be done separately for each supplier. Let Li
E [Bi j si ]

ai + ki a2i =2 + r0

2
i

+ ri (1

2
i)

b0
B

(B1)
ci si +

Var["i ]=2. As the objective is a decreasing function

of fsi g the optimal values are always at the corner and the (AR) constraint is binding, implying that
> 0. Note @ 2 Li =@a2i = ki > 0, @ 2 Li =@s2i = f (si ) > 0, and @ 2 Li =@
of cross partial terms @ 2 Li =@ai si = @ 2 Li =@si

= @ 2 Li =@

i ai

2
i

= r0 + r

0. In the absence

= 0, so the Hessian for supplier i is pos-

itive denite and hence the problem is convex, establishing the uniqueness of the equilibrium solution.
(6), (7), and (9) are obtained from the rst-order condition of supplier i. Clearly, the optimal si is a
function of , which is determined from the (AR) constraint, as in (8). The suppliers prot and the
customers expenditures follow immediately.
Proof of Proposition 2. Note the following identities are needed:
dVar[Bi j si ]=dsi =
d2 Var[Bi j si ]=ds2i =

2Fi (si )E[Bi j si ]

0;

2fi (si )E[Bi j si ] + 2Fi (si )[1

(B2)
Fi (si )]:

(B3)

Let us drop the subscript i for notational convenience. Dierentiating the suppliers expected utility

function (4) with respect to s and substituting (B2), we nd that

@U=@s =

(1

which is greater than zero for all s if

If

< 1 and v[1

F (0)]

(1

also that lims!1 @U=@s =

F (s)] + rv 2 F (s)E[B j s];

)c + v[1

= 1 and v > 0; because U increases without bound in this case.

)c, then @U=@s

(1

(B4)

0 at s = 0; so U is nondecreasing initially. Notice

)c < 0, so there exists at least one critical point on [0; 1). Setting

@U=@s = 0, we get E[B j s ] = ((1

)c

v[1

F (s )]) = rv 2 F (s ) . Substituting this result into the

second derivative (use (B3))

@ 2 U=@s2 =

vf (s) + rv 2 f (s)E[B j s]

rv 2 F (s)[1

F (s)]

(B5)

we obtain
@ 2 U=@s2

s=s

[v

(1

)c]f (s )=F (s )

where the inequality follows from the condition v

v[1

rv 2 F (s )[1

F (0)]

(1

F (s )] < 0;

)c. Since the second derivative

is negative at every critical point, s cannot be a minimizer. Combining this result with @U=@sjs=0 > 0
and lims!1 @U=@s < 0, we conclude that U has a unique maximizer. Optimal solutions follow from
the rst-order conditions.
Proof of Corollary 1. We drop the subscript i for notational convenience. After dierentiating the
rst-order condition (B4) implicitly with respect to r (optimal s is a function of r, i.e., s = s(r)) and
collecting the terms we obtain
@s
=
@r
vf (s )

v 2 F (s )E[B j s ]
rv 2 f (s )E[B j s ] + rv 2 F (s )[1

F (s )]

Notice that the denominator has the sign opposite of that in (B5). Hence @s =@r > 0. Similarly,
@a =@
@s
@
@s
@v

=
=
=

Proof of Proposition 3.

1=k < 0;
vf (s )
vf (s )

@a =@v = 0;
c
2
rv f (s )E[B j s ] + rv 2 F (s )[1
[1 F (s )] + 2rvF (s )E[B j s ]
rv 2 f (s )E[B j s ] + rv 2 F (s )[1

F (s )]
F (s )]

> 0;
> 0:

With r0 = r1 = ::: = rn = 0, the customers Lagrangian for supplier i

becomes Li (ai ; si ; ) = ci si + E [Bi j si ]

ai + ki a2i =2, and solutions are given by (6), (7), and (8).
5

From the suppliers utility Ui (ai ; si ; wi ; vi ;

clear that setting

= 0 and vi =

i)

= wi

(1

i )(ci si

ai )

ki a2i =2 it is

vi E [Bi j si ]

yields the same Lagrangian (with the reverse sign) as Li plus a

constant, reproducing the rst-best solutions.

The suppliers prot and the customers expenditure

follow immediately.
Proof of Lemma 1. Dene

F (s)]2 and note that (13) can be rewritten as

4cF (s)E [B j s] =[1

v( ) =

1
2c
1 F (s) r

1+

1 + r (1

) ;

from which we obtain

v0( ) =

1
c
p
F (s) 1 + r (1

1
21

v 00 ( ) =

c
r
F (s) [1 + r (1

)]3=2

Dierentiating the Lagrangian (12) and substituting @(v 2 )=@a = 2v(a)v 0 (a) and @ 2 (v 2 )=@a2 = 2(v 0 (a))2 +
2v(a)v 00 (a) in it, we nd that
@L
= + [(r0 + r)
@
k

r]Var["]

2(r0 + r)c2 1
[1 F (s)]2 r

p
1 + r (1

@2L
1
(r0 + r)c2
1
=
+
(r
+
r)Var["]
+
0
@ 2
k
[1 F (s)]2 [1 + r (1

)]3=2

Var[B j s];

Var[B j s] > 0:

Proof of Proposition 5. We use the results in the proof of Lemma 1. Dene

`eF B ( )

[(r0 + r)

`eAO ( )

[(r0 + r)

`eSO ( )

`eSB ( )
F B,

r]Var["];

+ [(r0 + r)

r]Var["];

@L
( ) = + [(r0 + r)
@
k
SO ,

AO ,

and

SB

FB

r]Var["]

(1 + r (1

2(r0 + r)c2 1
[1 F (s)]2 r

))
1

1=2

Var[B j s];

(1 + r (1

))

1=2

Var[B j s]:

are the solutions to `eF B ( ) = 0, `eSO ( ) = 0, `eAO ( ) = 0, and `eSB ( ) = 0,

respectively. Observe `eAO ( )

AO .

2(r0 + r)c2 1
[1 F (s)]2 r

r]Var["]

In contrast, both `eF B ( )

`eSB ( )

= r=(r0 + r) in `eSB to obtain

1 r
`eSB ( F B ) =
k r0 + r

`eSO ( ) for any . Since `e0j ( ) > 0 for all j,

`eSB ( ) and `eF B ( )

2(r0 + r)c2 1
[1 F (s)]2 r
6

"

SO

SB

`eSB ( ) are possible. To see this, substitute

r0 r
1+
r0 + r

1=2

Var[B j s]:

r=r0 and rewrite `eSB (

Let

`SB ( )

`eSB (

FB

F B)

as a function of :
2c2
1
[1 F (s)]2 r

1
)=
k1+

1+

"

1=2

r
1+
1+

Var[B j s]:

Dierentiating, we see that

1
2c2 Var[B j s]
`0SB ( ) =
+
k(1 + )2 [1 F (s)]2 r

"

unique

Hence, `SB ( ) is increasing. Notice lim

r
1+
1+

!0 `SB (

)=

1=2

1; lim

!1 `SB (

SB

<

F B;

Dierentiating `eSB (

since `eSB ( ) is also increasing. Likewise,

SB )

d SB
=
d(V ar["])

The numerator is negative if

> 0:

) = 1=k: Therefore there is a

such that `SB ( y ) = 0. Since `SB ( ) is increasing, `SB ( ) = `eSB (

implying that

3=2

r
1+
1+

(1 + 1= ) r
+
2(1 + )2

SB

>

F B)
FB

> 0 for all

for all

<

>
y

= 0 with respect to Var["] and collecting terms, we see that

r
1
k

(r0 +r)c2 1
1
[1 F (s)]2 r [1+r (1

+ (r0 + r)Var["] +
SB

>

F B,

dv SB =d(Var["]) is the reverse of that of d

SB

(r0 + r)

SB

zero if

SB =d(Var["])

F B,

:
SB )]3=2

Var[B j s]

and positive if

SB

<

F B.

The sign of

via (13).

Appendix C: Impact of Using Alternative Backorder Measure

C.1

Introduction

In this Appendix we show that all results and the insights derived in the paper remain valid when an
alternative backorder measure is used, namely, the time average of stationary backorders. Assume that
the repair processes have run for a su cient amount of time such that the steady state is reached at
time 0. Let

be the horizon at which the total backorders are counted (averaged) and let B(t) be

the stationary backorder random variable measure at time t. For notational convenience we drop the
subscript i denoting the suppliers. Dene
e )
B(

B(t)dt:

e ) j s] = E[B(t) j s]. The variance is

Because of stationarity E[B(
e ) j s] = 2
Var[B(

!(u j s) 1

du

where
!(u j s)

Cov[B(t); B(t + u) j s]

is the autocovariance of B(t) (see Papoulis and Pillai [6], p. 525) and where u is the time lag. Note
that !(u j s) is a function of u only because of the stationarity assumption. The ergodicity of B(t)
e ) j s] goes to zero as
means that Var[B(

approaches innity. In our problem context, however,

may

be comparable to 1= and in such instances the variance is not negligible.

We digress at this point and motivate why the distinction between the steady state variable B
e ) is important. At the heart of the discussion are the following two related
and the time average B(
questions:

(a) Which measure should be used in order to structure contracts?

(b) Which measure is used in practice in order to compute contractual payments?
In an ideal situation, the answers to these two questions should be the same. However, our inquiries into
this matter revealed that a discrepancy exists in practice. The answer to (a) is quite unambiguous
various guidelines on the weapon system support and maintenance make it clear that it is the availability
observed at a random point in time that practitioners are concerned with (e.g., make sure that a eet
of systems is available with 95% chance if a war breaks out tomorrow). The DoD Guide for Achieving
Reliability, Availability, and Maintainability [2] denes availability as a measure of a degree to which
an item is in operable state and can be committed at the start of a mission when the mission is called
for at an unknown (random) point in time. In addition, the U.S. Air Force currently uses a condence
interval of availability, which is dened in terms of steady state random variable A = 1

B=N , in

assessing mission readiness and optimizing the stocking levels accordingly (Slay et al. [7]). Under this
denition B is the natural candidate for the performance metric that should be used to dene contract
parameters.
However, the implied method of measurement under this denition involves counting the number
of backorders once, i.e., at a random point in time. Understandably, the supplier is unlikely to agree
to this method of performance evaluation because of the perceived arbitrariness. Instead, the widely
accepted method of computing subsystem availability is counting the number of hours the subsystems
are ready for a mission (i.e., not grounded due to a backordered condition) and comparing that number
to the planned usage hours, forming a ratio between the two. This procedure is equivalent to counting
R
e=1
e )=N .
backorders over time and converting it to availability, i.e., A
( N ) = 1 B(
0 B(t)dt
Thus, we are faced with a situation where there exists a gap between what needs to be measured

e )) to compute contractual payments.

for managerial purpose (B) and what is actually measured (B(
8

Indeed, our conversations with practitioners indicate that this discrepancy is a source of controversy
that is generating ongoing debates in the aerospace and defense industries. Although this is a very
interesting observation that merits further in-depth study, we chose to employ B in the paper since (1)
it is consistent with the performance measure that practitioners intend to evaluate, and (2) it allows us
to develop a model that highlights the tradeo between risk and incentives, which is the focus of our
e ) as the performance measure
paper. For completeness, however, we show the implications of using B(
in the sequel.

e ) does not change any qualitative results, we note that Var[B j s] appears
To see that employing B(

in places where we discuss the impact of risk aversion with unobservable s. These sections are on

general problem formulation (Sections 4.2 and 4.3, including Proposition 2 and Corollary 1) and on
the analysis of the single supplier case (Section 5.3, including Lemma 1 and Proposition 5). All results
in these sections depend on two facts: (1) Var[B j s] decreases with s, and (2) the suppliers utility
function is unimodal (as shown in the proof of Proposition 2). Specically, we rely on the following
results (equations (15) and (16) appearing in the proof of Proposition 2):
dVar[B j s]=ds =
d2 Var[B j s]=ds2 =

2F (s)E[B j s]

0; and

2f (s)E[B j s] + 2F (s)[1

(C1)
F (s)]:

(C2)

When computing the optimal s we use (C1) that appears in the rst-order condition of the suppliers
problem, and derive the expression for v( ; s) (equation (13) in the paper). To show that the suppliers
utility U is unimodal in s under the condition specied in Proposition 2, we use (C2) and show that
the second derivative of the suppliers utility is always negative at a critical point, ensuring that this
point represents a global maximum.
The exact expressions in (C1) and (C2) are not essential. Rather, what matters are the signs of
dVar[B j s]=ds and of @ 2 U=@s2

s=s

. The reason is that all of our results use comparative statics after

identifying the optimal supplier choice s . Therefore, once we show that both signs continue to be
e ), we can simply substitute the terms on the right-hand
negative using the alternative measure B(
e ) j s]=ds and d2 Var[B(
e ) j s]=ds2
sides of (C1) and (C2) that appear anywhere in the paper by dVar[B(

without changing any qualitative insights.

To prove that dVar[B j s]=ds

0 for all s, one only needs to show that @!(u j s)=@s

0. This

result relies upon the multivariate stochastic dependence concept of association because !(u j s) is a
function of two identical but dependent random variables. We were able to show @!(u j s)=@s

0, and

the proof is given in the next section. Showing that U is unimodal turns out to be much more di cult.
However, we nd that U has to have an interior solution whenever the mild condition in Proposition 2
9

is met, and we also obtain the lower and upper bounds on s . Since the supplier will choose an interior
maximum, we are assured that @ 2 U=@s2

s=s

0. In fact, our numerical analysis indicates that U is

unimodal, although an analytical proof of this result is hard to obtain. We discuss this di culty in the
last section.

C.2

Backorder variance decreases in s

e j s) and B(t
e + u j s) are identically distributed, and the autocovariance is a
Because of stationarity, B(t

function of the lag u only. That is,

!(u j s) = Cov[B(t); B(t + u) j s] = Cov[B(0); B(u) j s]

(see Heyman and Sobel [4], p. 365). Thus one can restrict attention to B(0) and B(u) (and hence, also
O(0) and O(u)), which are, in general, not independent. Whenever there is no ambiguity we omit the
argument and write B and O. We need the following multivariate dependence concept. Two random
variables X and Y are called (positively) associated if
Cov[ (X; Y ); (X; Y )] = E[ (X; Y ) (X; Y )]

E[ (X; Y )]E[ (X; Y )]

for all nondecreasing functions (in each argument) ; : R2 ! R (see, for example, Mller and Stoyan
[5]). Eick, Massey, and Whitt [3] demonstrate that O(0) and O(u) are associated for each u. Let 1S (x)
be the indicator function, which is one if x 2 S and zero otherwise. Since nondecreasing functions
of associated random variables are associated (Barlow and Proschan [1], p. 30), and the functions
(x

s)+ and 1(s;1) (x) are nondecreasing in x, B(0) = (O(0)

s)+ , B(u) = (O(u)

s)+ , 1(s;1) (O(0)),

1(s;1) (O(u)) are associated among themselves. In particular,

E[B(0)B(u) j s]

E[B j s]2 ;

(C3)

E[1(s;1) (O(0)) B(u) j s]

E[1(s;1) (O(0))] E[B(u) j s] = [1

E[1(s;1) (O(u)) B(0) j s]

[1

F (s)]E[B j s]; and

F (s)]E[B j s];

(C4)
(C5)

where we have used the fact that O(0) and O(u) are identically distributed.
As Eick, Massey, and Whitt [3] point out, O(0) and O(u) can be decomposed into three independent
random variables. Let A1 (u) be the number of arrivals in the repair facility up to time 0 that depart
by time u; let A2 (u) be the number of arrivals between times 0 and u that remain in the repair facility
at time u; and let A3 (u) be the number of arrivals up to time 0 that remain in the facility at time

10

u. Clearly, these three variables are independent, and O(0) = A1 (u) + A3 (u), O(u) = A2 (u) + A3 (u).
Thus we can write

!(u j s) = E[(A1 (u) + A3 (u)

s)+ (A2 (u) + A3 (u)

s)+ ]

E[(O

s)+ ]2 :

Since the second term is fully characterized by the cdf F , we are mainly concerned with the rst term.
Let

i(

j u) be the pdf of Ai (u). For notational convenience we use the shorthand notation

i(

) and

Ai when u is xed. The rst term is

E[(A1 + A3 s)+ (A2 + A3 s)+ ]
Z 1Z 1
Z 1
=
(x + z s)(y + z s) 1 (x) 2 (y) 3 (z)dxdydz
0
(s z)+ (s z)+
Z sZ 1 Z 1
=
(x + z s)(y + z s) 1 (x) 2 (y) 3 (z)dxdydz
0
s z s z
Z 1Z 1Z 1
+
(x + z s)(y + z s) 1 (x) 2 (y) 3 (z)dxdydz:
s

Dierentiating with respect to s using Leibnitzs rule, we obtain

d
E[(A1 + A3
ds Z Z

s)+ ]
Z sZ 1 Z 1
1
1
xy 1 (x) 2 (y)dxdy
(x + y + 2z 2s) 1 (x) 2 (y) 3 (z)dxdydz
3 (s)
0
0
0
s z s z
Z 1Z 1
Z 1Z 1Z 1
(s)
xy
(x)
(y)dxdy
(x + y + 2z 2s) 1 (x) 2 (y) 3 (z)dxdydz
3
1
2
0
0
s
0
0
Z 1Z 1
Z 1
(x + y + 2z 2s) 1 (x) 2 (y) 3 (z)dxdydz
0

(s z)+

s)+ (A2 + A3

(s z)+

E[1f(s;1);(s;1)g (O(0); O(u)) (O(0) + O(u)

2s)]:

After combining, we obtain

@!(u j s)
=
@s

E[1f(s;1);(s;1)g (O(0); O(u)) (O(0) + O(u)

2s)] + 2[1

F (s)]E[(O

s)+ ]:

(C6)

The association property is needed to determine the sign of this quantity. Observe, for the rst

11

term, that
E[1f(s;1);(s;1)g (O(0); O(u)) (O(0) + O(u) 2s)]
Z 1Z 1
Z 1
(y + z s) 1 (x) 2 (y) 3 (z)dxdydz
=
0
(s z)+ (s z)+
Z 1Z 1
Z 1
(x + z s) 1 (x) 2 (y) 3 (z)dxdydz
+
0
(s z)+ (s z)+
Z 1
Z 1
E[1((s
E[1((s z)+ ;1) (A1 ) (A2 + z s)+ ] 3 (z)dz +
=

z)+ ;1) (A2 )

s)+ ]

(A1 + z

s)+ j A3 ] + E E[1(s;1) (A2 + A3 ) (A1 + A3

= E E[1(s;1) (A1 + A3 ) (A2 + A3

s)+ ] + E[1(s;1) (O(u)) (O(0)

= E[1(s;1) (O(0)) (O(u)

2[1

3 (z)dz

s)+ j A3 ]

s)+ ]

s)+ ]

F (s)]E[(O

(C7)

where the last inequality comes from (C4) and (C5). Therefore we have shown that
@!(u j s)
@s

as desired.

C.3

Unimodality of the suppliers utility function

Using a technique similar to the above, one can show that

d2
E[(A1 + A3
dsZ2 Z
Z
1

= 2

(s z)+

s)+ (A2 + A3
1

1 (s

1 (x) 2 (x) 3 (z)dxdydz

(s z)+

z)

s)+ ]

(y + z

s z

= 2 Pr(O(0) > s; O(u) > s) + 2

s)
Z

2 (y)dy +

2 (s

1 (s

z)E[(A1 + z

z)

(x + z

s)

s z

s)+ ]

1 (x)dx

3 (z)dz

3 (z)dz

which yields
@ 2 !(u j s)
@s2

= 2 Pr(O(0) > s; O(u) > s) + 2

2f (s)E[(O

s)+ ]

2[1

s
1 (s

z j u)E[(A1 (u) + z

s)+ ]

3 (z j u)dz

F (s)]2 :

(C8)

e ) j s]=ds2 and hence in @ 2 U=@s2

The sign of this expression is ambiguous. Substituting (C8) in d2 Var[B(
does not lead to an analytically tractable expression that would allow us to show that @ 2 U=@s2

s=s

0.

Despite the di culty, we can infer the shape of the utility function U in two nonoverlapping domains

12

that include s = 0 and s = 1. Recall that

U (s) = w

(1

)(cs

a)

ka2 =2

vE[B j s]

)2 Var["]=2

r(1

Dene

e ) j s]=2:
rv 2 Var[B(

U 1 (s)

(1

)(cs

a)

vE[B j s]

ka2 =2

r(1

)2 Var["]=2;

U 2 (s)

(1

)(cs

a)

vE[B j s]

ka2 =2

r(1

)2 Var["]=2

rv 2 Var[B j s]=2:

They dier from U (s) only by the last term, the performance premium. Specically, U 1 (s) does not have
e ) j s]g =
this term, whereas the corresponding term in U 2 (s) is the maximum value, since max fVar[B(
lim

e ) j s] =Var[B j s]. With these denitions, we can bound U (s) by U 1 (s) and U 2 (s) and

!1 Var[B(

also the rst derivative of U (s) by the rst derivatives of U 1 (s) and U 2 (s).
Lemma 1 For s 2 [0; 1),
U 1 (s)

U (s)

dU 1 (s)=ds

U 2 (s) ;

dU (s)=ds

(C9)
dU 2 (s)=ds :

Proof. (C9) is trivial. To see (C10), note that dVar[B(s)]=ds =

dU (s)=ds =

(1

)c + v[1

F (s)]

rv 2

(C10)

2F (s)E[B(s)]. Then

@!(u j s)
u
1
du
@s
F (s)]E[B j s]

(1
)c + v[1 F (s)] rv 2 [1
Z
rv 2
E[1f(s;1);(s;1)g (O(0); O(u)) (O(0) + O(u)
+

2s)] 1

du; (C11)

dU 1 (s)ds =
dU 2 (s)=ds =

(1

)c + v[1

F (s)];

(1

)c + v[1

F (s)] + rv 2 F (s)E[B j s]

where we have used @Var[B(s)]=@s =

dU 1 (s)=ds

dU (s)=ds, since @!(u j s)=@s

2F (s)E[B(s)] to obtain the last equation. Observe that

0. To compare dU (s)=ds and dU 2 (s)=ds, note that

E[1f(s;1);(s;1)g (O(0); O(u)) (O(0) + O(u)

= E[1f(s;1);(s;1)g (O(0); O(u)) ((O(0)
E[(O(0)

s)+ + (O(u)

s)+ + (O(u)

s)+ ] = 2E[B j s]:

13

2s)]
s)+ )]

Using (C11), we see that this implies that

dU (s)=ds
=
since

rv 2 [1

F (s)]E[B j s] + rv 2 E[B j s]

(1

)c + v[1

F (s)]

(1

)c + v[1

F (s)] + rv 2 F (s)E[B j s] = dU 2 (s)=ds

du = 1=2.

As in Proposition 2 of the paper, we assume that condition v

v[1

F (0)]

(1

)c, which

ensures that U 1 is increasing initially, is met. Notice that

lim dU 1 (s)=ds =

s!0

(1

)c + v[1

lim dU (s)=ds = lim dU 2 (s)=ds =

s!0

lim dU (s)=ds =

s!1

s!0

F (0)]
(1

0;
)c + v[1

F (0)] + rv 2 F (0)

lim dU 1 (s)=ds = lim dU 2 (s)=ds =

s!1

s!1

(1

)c

0;

0:

Hence, U (s), U 1 (s), and U 2 (s) all start with positive slopes and approach the same negative slope as
s goes to innity. Thus, under the condition above, U (s) is guaranteed to have an interior maximum.
Let s , s1 , and s2 be the global maxima of U (s), U 1 (s), and U 2 (s), respectively. Dierentiation
proves that U 1 (s) is concave. We show that U 2 (s) is quasiconcave in the proof of Proposition 2. Hence,
they both have unique maxima. It is now also clear from (C10) that
s1

s2

since U (s) is increasing on [0; s1 ] and decreasing on [s2 ; 1).

Although we cannot show that U (s) is quasiconcave analytically, numerical experiments indicate
that this is indeed the case. Figure 1 shows a typical plot for U (s), which is representative of many
of the numerical experiments we attempted with a variety of problem parameters. In this particular
example,

= 0:1, L = 150, c = 1, r = 1,

= 0:5, and v = 1 when exponential service times are chosen.

We observe that U (s) approaches U 1 (s) asymptotically as

C.4

increases and unimodality is preserved.

Changes in contract terms

Although qualitative insights remain unchanged whether the steady state random variable B or the
e ) is used, the corresponding optimal contract parameters will assume dierent values.
time average B(

The following proposition summarizes how the the contract terms are modied in the single supplier
case.

14

-9

-10

-11

-12

-13

-14

-15

-16

-17

-18

10

30

25

20

15

Figure 1: U (s) for = 5; 10; 20; 30 from bottom to top. = 0:1, L = 150, c = 1, r = 1, = 0:5, and
v = 1 when exponential service time are chosen (note the scale in the y axis is negative because we
have chosen w = 0 and zero xed cost).

Proposition 1 Suppose there is a single supplier such that the conditions specied in Proposition 5
are satised. Let (

SB ; v SB )

be the optimal contract parameters when the steady state random variable

B is used as a backorder measure, and (e SB ( ); veSB ( )) be their counterparts when the time average
e ) is used. Then
B(
(i) e SB ( )

SB ,

(ii) veSB ( )

and

v SB .

Proof. Under the single supplier assumption, the optimal stocking level sSB is xed, as it is determined
R
SB .
e ) j s] = B
b0 . Let J( ) Var 1
from the binding backorder constraint E[B j s] = E[B(
0 B(t)dt j s
e ) j sSB ] and lim
With this denition, J( ) =Var[B(

we have J 0 ( )
for

!0 J(

e ) j sSB ]
) =Var[B j sSB ]. Since Var[B(

0. From the Proof of Proposition 5 (found in Appendix B), the optimality condition

is
k

where

Var[B j sSB ],

+ [(r0 + r)

r] Var["]

4cF (s)E[B j s]=[1

d
=
d

2(r0 + r)c2 1
[1 F (s)]2 r

1 + r (1

J( ) = 0;

F (s)]2 , which is independent of . Via implicit dierentiation, we get

2(r0 +r)c2 1
[1 F (s)]2 r
1
k

+ (r0 + r)Var["] +

1
1+r (1

J 0( )

(r0 +r)c2 1
J( )
[1 F (s)]2 r [1+r (1
)]3=2

Since the numerator is nonpositive and the denominator is positive, d =d

0, which in turn implies

dv=d
0 since sgn( ) = sgn(v) from Equation (13) of the paper. The results in (i) and (ii) follow
since (e SB ( ); veSB ( )) are associated with larger .
15

References
[1] Barlow, R.E. and F. Proschan. 1975. Statistical Theory of Reliability and Life Testing. Holt, Rinehart and Winston.
[2] Department of Defense. 2005. DoD guide for achieving reliability, availability, and maintainability.
http://www.dote.osd.mil/reports/RAMGuide.pdf.
[3] Eick, S. G., W. A. Massey, and W. Whitt. 1993. The physics of the Mt /G/1 queue. Operations
Research, 41(4), 731-742.
[4] Heyman, D. P. and M. J. Sobel. 2004. Stochastic Models in Operations Research, Volume 1. Dover,
Mineola, New York.
[5] Mller, A. and D. Stoyan. 2002. Comparison Methods for Stochastic Models and Risks. Wiley.
[6] Papoulis, A. and S. U. Pillai. 2002. Probability, Random Variables and Stochastic Processes, 4th
edition. McGraw-Hill, Boston.
[7] Slay F. M. et al. 1996. Optimizing spares support: the aircraft sustainability model. Logistics
Management Institute Report AF501MR1.

16