Approximation in Stochastic Integer Programming

Approximation in Stochastic Integer Programming
∗
Leen Stougie Maarten H. van der Vlerk
Eindhoven Technical University University of Groningen
and CWI Amsterdam The Netherlands
The Netherlands m.h.van.der.vlerk@eco.rug.nl
leen@win.tue.nl

May 7, 2003

Approximation algorithms are the prevalent solution methods in the field of
stochastic programming. Problems in this field are very hard to solve. Indeed, most
of the research in this field has concentrated on designing solution methods that
approximate the optimal solutions. However, efficiency in the complexity theoretical
sense is usually not taken into account. Quality statements mostly remain restricted
to convergence to an optimal solution without accompanying implications on the
running time of the algorithms for attaining more and more accurate solutions.
However, over the last twenty years also some studies on performance analysis
of approximation algorithms for stochastic programming have appeared. In this
direction we find both probabilistic analysis and worst-case analysis. There have
been studies on performance ratios and on absolute divergence from optimality.
Only recently the complexity of stochastic programming problems has been ad-
dressed, indeed confirming that these problems are harder than most combinatorial
optimization problems.
Approximation in the traditional stochastic programming sense will not be dis-
cussed in this chapter. The reader interested in this issue is referred to surveys on
stochastic programming, like the Handbook on Stochastic Programming [31] or the
text books [2, 16, 29]. We concentrate on the studies of approximation algorithms
which are more similar in nature to those for combinatorial optimization.

1 Introduction
Stochastic programming models arise as reformulations or extensions of optimiza-
tion problems with random parameters. To set the stage for our review of approxi-
mation in stochastic (integer) programming, we first introduce the models and give
an overview of relevant mathematical properties.
Consider the optimization problem
min cx
x
s.t. Ax = b
Tx = h
x ∈ X,
where X ⊂ Rn specifies nonnegativity of and possibly integrality constraints on
the decision variables x. In addition to the m1 deterministic constraints Ax = b,
there is a set of m constraints T x = h, whose parameters T and h depend on
∗ The research of this author has been made possible by a fellowship of the Royal Netherlands

Academy of Arts and Sciences.

1

information which becomes available only after a decision x is made. The stochastic
programming approach to such problems is to assume that this uncertainty can
be modeled by random variables with known probability distribution, and then
to reformulate the model to obtain a meaningful and well-defined optimization
problem. In this chapter we will use bold face characters for random variables, and
plain face to indicate their realizations.

1.1 Stochastic programming models
The first important class of stochastic programming models, known as recourse
models, is obtained by allowing additional or recourse decisions after observing the
realizations of the random variables (T , h). Thus, recourse models are dynamic:
time is modeled discretely by means of stages, corresponding to the available infor-
mation. If all uncertainty is dissolved at the same moment, this is captured by a
recourse model with two stages: ‘present’ and ‘future’. Given a first-stage decision
x, for every possible realization q, T, h of q, T , h, infeasibilities h − T x are compen-
sated at minimal costs by choosing second-stage decisions as an optimal solution of
the second-stage problem
min qy
y
s.t. W y = h − T x,
y ∈ Y,

where q is the (random) recourse unit cost vector, the recourse matrix W specifies
the available technology, and the set Y ⊂ Rn2 is defined analogously to X. We will
+
use ξ = (q, T , h) to denote the random object representing all randomness in the
problem. The value function of this second-stage problem, specifying the minimal
recourse costs as a function of the first-stage decision x and a realization of ξ, will be
denoted by v(x, ξ); its expectation Q(x) := Eξ [v(x, ξ)] gives the expected recourse
costs associated with a first-stage decision x. Thus, the two-stage recourse model is
min cx + Q(x)
x
s.t. Ax = b (1)
x ∈ X,
where the objective function cx+Q(x) specifies the total expected costs of a decision
x.

Example 1.1 Consider the following production planning problem. Using n pro-
duction resources, denoted by x ∈ Rn with corresponding unit cost vector c, a
+
production plan needs to be made such that the uncertain future demand for m
products, denoted by h ∈ Rm , is satisfied at minimal costs. The available produc-
tion technology suffers from failures: deploying resources x yields uncertain amounts
of products T i x, i = 1, . . . , m. Restrictions on the use of x are captured by the
constraints Ax = b.
We assume that the uncertainty about future demand and the production tech-
nology can be modelled by the random vector (T , h), whose joint distribution is
known, for example based on historical data.
A possible two-stage recourse model for this problem is based on the following
extension of the model. For each of the individual products, if the demand hi turns
out to be larger than the production Ti x, the demand surplus hi − Ti x is bought
1
from a competitor at unit costs qi . On the other hand, a demand shortage gives
2
rise to storage costs of qi per unit. The corresponding second-stage problem and

2

its value function are
v(x, ξ) = min q 1 y 1 + q 2 y 2
y
s.t. y 1 − y 2 = h − T x, ξ ∈ Ξ,
y = (y 1 , y 2 ) ∈ R2m .
+

Defining Q as the expectation of this value function, we obtain a two-stage recourse
model that fits the general form (1).
This particular model type with recourse matrix W = (Im , −Im ), where Im
is the m-dimensional identity matrix, is known as a simple recourse model. The
integer recourse version of this model, for example corresponding to the case that
only batches of fixed size can be bought, will be discussed in Section 3.

So far, we have introduced the recourse concept as a modelling tool to handle
random constraints, by means of specifying recourse actions with corresponding re-
course costs. There is however another class of problems for which the (two-stage)
recourse model is a natural approach, namely hierarchical planning models (HPM).
Such problems involve decisions at two distinct levels: strategic decisions which have
a long-term impact, and operational decisions which are depending on the strategic
decisions. For example, in the hierarchical scheduling problem discussed in Section
4, the strategic decision is the number of machines to be installed, and the opera-
tional decisions involve the day-to-day scheduling of jobs on these machines. At the
time that the strategic decision needs to be made, only probabilistic information on
the operational level problems (e.g. the number of jobs to be scheduled) is available.
Hierarchical planning models fit the structure of two-stage recourse models, with
strategic and operational decisions corresponding to first-stage and second-stage
variables, respectively. Moreover, since strategic decisions are typically fixed for a
relatively long period of time, it is natural to use the expectation of the operational
costs as a measure of future costs.
Unlike conventional linear recourse models (1), HPM are not necessarily formu-
lated as (mixed-integer) LP problems, see our example in Section 4. Nevertheless,
despite these differences in interpretation and formulation, we use the generic name
(two-stage) recourse model to refer to both model types, in line with the stochastic
programming literature.
In many applications new information becomes available at several distinct mo-
ments, say t = 1, . . . , H, where H is the planning horizon. That is, we assume that
realizations of random vectors ξ t = (q t , T t , ht ) become known at time t. This can
be modelled explicitly in a multistage recourse structure: for each such moment
t = 1, . . . , H, a time stage with corresponding recourse decisions is defined. In
compact notation, the multistage recourse model is
min
0
cx0 + Q1 (x0 )
x
s.t. Ax0 = b
x0 ∈ X,
where the functions Qt , t = 1, . . . , H, representing expected recourse costs, are
recursively defined as
Qt (xt−1 ) := Eξ t v t (xt−1 , ξ t ) ξ 1 , . . . , ξ t−1 ,

where the expectation is with respect to the conditional distribution of ξ t given
ξ 1 , . . . , ξ t−1 ,
v t (xt−1 , ξ t ) := min q t xt + Qt+1 (xt )
t
x
s.t. W t xt = ht − T t xt−1
xt ∈ X t ,

3

and QH+1 ≡ 0 (or some other suitable choice). In this chapter we concentrate on
two-stage problems only.
The second main class of stochastic programming problems consists of probabilis-
tic or chance-constrained problems, which model random constraints1 by requiring
that they should be satisfied with some prescribed reliability α ∈ [0, 1]; typically,
α ∈ (.5, 1). Thus, the random constraints T x ≥ h are replaced by the joint chance
constraint
Pr{T x ≥ h} ≥ α,
or by m individual chance constraints
Pr{T i x ≥ hi } ≥ αi , i = 1, . . . , m.
Since we will not consider chance-constrained models in our discussion of ap-
proximation results, we do not present them in more detail here.

1.2 Mathematical properties
In this section, we review mathematical properties of recourse models. This provides
the background and motivation for the discussion of approximation results.
First we consider properties of continuous recourse models. Some of the results
will be used when we discuss the complexity of this problem class, and furthermore
they facilitate the subsequent discussion of properties of mixed-integer recourse
models. We state all properties here without proof. In the Notes at the end of the
chapter references to the proofs are given.

Remark 1.1 As before, all models are discussed here in their canonical form, i.e.,
all constraints are either equalities or nonnegativities. The models in subsequent
sections, which also contain inequalities and/or simple bounds, can be written in
canonical form using standard transformations.

1.2.1 Continuous recourse
Properties of (two-stage) recourse models follow from those of the recourse function
Q. In case all second-stage variables are continuous, properties of the value function
v are well-known from duality and perturbation theory for linear programming, and
are summarized here for easy reference.

Lemma 1.1 The function v, defined for x ∈ Rn and ξ = (q, T, h) ∈ Rn2 +m(n+1) ,
v(x, ξ) = inf qy : W y = h − T x, y ∈ Rn2
+

takes values in [−∞, ∞].
It is a convex polyhedral function of x for each ξ ∈ Rn2 +m(n+1) , and it is concave
polyhedral in q and convex polyhedral in (h, T ) for all x ∈ Rn .

If for some x the function v takes on the value +∞ with positive probability, this
means that x is extremely unattractive since it has infinitely high expected recourse
costs Q(x). From a modelling point of view this is not necessarily a problem, but
in practice it may be desirable to exclude this situation.
On the other hand, the situation that v(x, ξ) equals −∞ with positive probability
should be excluded altogether. Indeed, the value −∞ indicates that the model does
not adequately represent our intention, which is penalization of infeasibilities.
Finiteness of v is often guaranteed by assuming that the recourse is complete
and sufficiently expensive.
1 Barring uninteresting cases, chance constraints make sense only for inequality constraints.

4

Definition 1.1 The recourse is complete if v < +∞, i.e., if for all t ∈ Rm there
exists a y ∈ Y such that W y = t.

Assuming that Y = Rn2 , completeness is a property of the recourse matrix W only.
+
Such a matrix is called a complete recourse matrix.

Definition 1.2 The recourse is sufficiently expensive if v > −∞ with probability
1, i.e., if Pr{ξ ∈ Ξ : ∃λ ∈ Rm such that q ≥ λW } = 1.

For example, the recourse is sufficiently expensive if Pr{q ≥ 0} = 1.
From now on we assume that the recourse is complete and sufficiently expensive.
Then the recourse or expected value function Q(x) is finite if the distribution of ξ
satisfies the following condition:
For all i, j, k the random functions q j hi and q j T ik have finite expectations.
Sufficiency of this weak covariance condition follows from the representation of basic
feasible solutions in terms of the problem parameters.
The following properties of the recourse function Q are inherited from the
second-stage value function v.
Theorem 1.1 Consider the continuous recourse function Q, defined by
Q(x) = Eξ inf qy : W y = h − T x, y ∈ Rn2
+ , x ∈ Rn .

Assume that the recourse is complete and sufficiently expensive.
(a) The function Q is convex, finite, and (Lipschitz) continuous.
(b) If ξ follows a finite discrete distribution, then Q is a convex polyhedral function.
(c) The function Q is subdifferentiable, with subdifferential

∂Q(x) = ∂v(x, ξ) dF (ξ), x ∈ Rn ,
Ξ

where F is the cdf of the random vector ξ.
If ξ follows a continuous distribution, then Q is continuously differentiable.

Consider the special case that ξ follows a finite discrete distribution specified
by Pr{ξ = (q k , T k , hk )} = pk , k = 1, . . . , K. The finitely many possible realizations
(q k , T k , hk ) of the random parameters are also called scenarios. It is easy to see
that in this case the two-stage recourse model is equivalent to the large-scale linear
programming problem
K
max cx + pk q k y k
k=1
s.t. Ax =b (2)
k k k
T x + Wy = h , k = 1, . . . , K
x∈ Rn ,
+ y ∈
k
Rn2 .
+

Analogously, a mixed-integer recourse problem with finite discrete distribution
can be represented as a deterministic large-scale mixed-integer programming prob-
lem.

5

1.2.2 Mixed-integer recourse
Mixed-integer recourse models do not posses such nice mathematical properties;
in particular, convexity of the recourse function Q is not guaranteed. Indeed, the
underlying second-stage value function v is only lower semicontinuous (assuming
rationality of the recourse matrix W ), and discontinuous in general.
Also in this setting we are mostly interested in the case that v is finite. To have
v < +∞ we will assume complete recourse, see Definition 1.1. For example, this
¯ ¯
condition is satisfied if W is a complete recourse matrix, where W consists of the
columns of W corresponding to the continuous second-stage variables. On the other
hand, v > −∞ if the recourse is sufficiently expensive, see Definition 1.2, i.e., if the
dual of the LP relaxation of the second-stage problem is feasible with probability 1.

Theorem 1.2 Consider the mixed-integer recourse function Q, defined by
Q(x) = Eξ [inf {qy : W y = h − T x, y ∈ Y }] , x ∈ Rn ,

where Y := Zp × Rn2 −p . Assume that the recourse is complete and sufficiently
+ +
expensive, and that ξ = (h, T ) satisfies a weak covariance condition. Then
(a) The function Q is lower semicontinuous on Rn .
(b) Let D(x), x ∈ Rn , denote the set containing all ξ ∈ Ξ such that h − T x is a
discontinuity point of the mixed-integer value function v. Then Q is continuous
at x if Pr{ξ ∈ D(x)} = 0.
In particular, if ξ is continuously distributed, then Q is continuous on Rn .

1.3 Outline
As mentioned above, solving stochastic programming problems is very difficult in
general. Indeed, such problems are defined in terms of expectations of value func-
tions of linear (mixed-integer) programming problems or indicator functions (in the
case of chance constraints). This calls for the evaluation of multi-dimensional inte-
grals, which is computationally challenging already if the underlying random vector
ω has low dimension, and becomes a formidable task for problems of realistic size.
Even if the underlying distribution is discrete, the typically huge number of possible
realizations may render the frequent evaluation of function values impracticable. In
Section 2 the computational complexity of two-stage recourse models is addressed.
It is therefore not surprising that much of the stochastic programming literature
is devoted to approximation of some sorts. For example, a key issue for recourse
models is the construction of suitable discrete approximations of the distribution
of the underlying random vector. Such an approximation should have a relatively
small number of possible realizations, and at the same time result in a good approx-
imation of the recourse function, at least in a neighborhood of an optimal solution.
For chance-constrained problems such discrete approximations of the distribution
would destroy convexity of the problem. In this context, fast and accurate approx-
imation of high-dimensional (normal) distribution functions receives much research
attention.
We do not discuss these ‘typical’ stochastic programming approximation issues
here. They, as well as related subjects such as convergence and stability, are covered
in the Handbook on Stochastic Programming [31]. Instead, we consider approxi-
mations as they appear in a number of other ways in stochastic programming and
which are in spirit closer to approximation in combinatorial optimization.
Section 3 deals with convex approximations for integer recourse problems. Here
the problems themselves are approximated by perturbing the distribution functions

6

such as to achieve convex expected value functions. The strength of this approxi-
mation is that a bound on the absolute approximation error can be given, making
this an example of worst-case analysis of approximation algorithms.
Hierarchical planning problems, which are (integer) recourse problems, are dis-
cussed in Section 4. The key idea here is to replace hard second-stage problems by
easier ones, which asymptotically still give accurate results. Here the approach is
probabilistic analysis of approximation algorithms.
In Section 5 we will give one of the scarce examples of an approximation al-
gorithm for a stochastic programming problem for which a constant worst-case
performance ratio can be proved. The example also shows again that stochas-
tic programming problems are usually more complicated than their deterministic
counterparts.
We conclude with a section containing bibliographical notes on approximation
in stochastic programming as reviewed in this chapter. It also addresses some
interesting open problems and new research directions in this field, major parts of
which are still unexplored.

2 Complexity of two-stage stochastic programming
problems
In this section we study the complexity of two-stage stochastic programming prob-
lems. The complexity of a problem, in terms of time or space to solve it, is related
to input size. For each instance a bound on the number of elementary computer
operations or on the number of computer storage units required to solve the prob-
lem instance as a function of the size of its input indicates, respectively, the time
or space complexity of the problem. We will see that the way in which the random
parameters in stochastic programming problems are described has a crucial impact
on the complexity.
To illustrate this we start by studying problem (2), the deterministic equivalent
LP formulation of the two-stage stochastic programming problem.
If in the input of the problem each scenario (q k , T k , hk ) and its corresponding
probability pk is specified separately, then the input size of the problem is just the
size of the binary encoding of all the parameters in this (large-scale) deterministic
equivalent problem and hence the problem is polynomially solvable in case the de-
cision variables are continuous and NP-complete if there are integrality constraints
on decision variables.
However, consider another extreme in which all parameters are independent
identically distributed random variables. For example, if in this case each param-
eter has value a1 with probability p and a2 with probability 1 − p, then there are
K = 2n1 +mn+m possible scenarios. Hence, the size of the deterministic equivalent
problem is exponential in the dimension of the parameter space, which is essentially
the size required to encode the input. The complexity changes correspondingly, as
will become clear below.
Let us consider models wherein all random (second-stage) parameters are in-
dependently and discretely distributed. We will establish P -hardness of the eval-
uation of the second-stage expected value function Q(x) for fixed x. The class
P consists of counting problems, for which membership to the set of items to be
counted can be decided in polynomial time. We notice that strictly following this
definition of P , none of the stochastic programming problems can belong to this
complexity class. We will use the term P -hard for an optimization problem in
the same way as N P -hardness is used for optimization problems, whose recognition
version is N P -complete. For an exposition of the definitions and structures of the

7

various complexity classes we refer to [28].
To prove P -hardness of the evaluation of the second stage expected value func-
tion Q(x) we use a reduction from the P -complete problem graph reliability.
Definition 2.1 Graph reliability. Given a directed graph with m arcs and n
vertices, what is the probability that the two given vertices u and v are connected
if all edges fail independently with probability 1/2 each.
This is equal to the problem of counting the number of subgraphs, from among all
2m possible subgraphs, that contain a path from u to v.

Theorem 2.1 Two-stage stochastic programming with discretely distributed pa-
rameters is P -hard.

Proof. That the problem is P -easy can be seen from the fact that for any re-
alization of the second-stage random parameters a linear program remains to be
solved.
To prove P -hardness, take any instance of graph reliability, i.e., a network
G = (V, A) with two prefixed nodes u and v in V . Introduce an extra arc from v to
u, and introduce for each arc (i, j) ∈ A a variable yij . Give each arc a random weight
qij except for the arc (v, u) that gets weight 1. Let the weights be independent and
identically distributed (i.i.d.) with distribution P r{q = −2} = P r{q = 0} = 1/2.
Denote A = A ∪ (v, u). Now define the two-stage stochastic programming problem
max{−cx + Q(x) | 0 ≤ x ≤ 1}
with Q(x) = Eq [v(x, q)] and

v(x, q) = max qij yij + yvu
(i,j)∈A

s.t. yij − yjk = 0 ∀j ∈ V
i:(i,j)∈A k:(j,k)∈A
yij ≤ x ∀(i, j) ∈ A.
The event {q = −2} corresponds to failure of the arc in the Graph reliability
instance. For a realization of the failures of the arcs, the network has a path
from u to v if and only if in the corresponding realization of the weights there
exists a path from u to v consisting of arcs with weight 0. The latter accounts
for an optimal solution value x of the corresponding realization of the second-stage
problem, obtained by setting all yij ’s corresponding to arcs (i, j) on this path and
yvu equal to x, whereas yij = 0 for all (i, j) not on the path. If for a realization the
graph does not have a path from u to v, implying in the reduced instance that on
each path there is an arc with weight −2 and vice versa, then the optimal solution of
the realized second-stage problem is 0, by setting all yij ’s equal to 0, and henceforth
also yvu = 0). Therefore, the network has reliability R if and only if Q(x) = Rx
and hence the objective function of the two-stage problem is (R − c)x.
Thus, if c ≤ R then the optimal solution is x = 1 with value (R − c), and if
c ≥ R then the optimal solution is x = 0 with value 0. Since R can take only 2m
possible values, bisection allows to solve only m two-stage stochastic programming
problems to know the exact value of R.

By total unimodularity of the restriction coefficients matrix in the proof, the
same reduction shows that two-stage integer programming problem with discretely
distributed parameters is P -hard.
Given a P -oracle for evaluating Q in any point x, solving two-stage stochastic
linear programming problems (with discretely distributed random variables) will

8

require a polynomial number of consultations of the oracle, since Q is a concave
function in x, and maximizing a concave function over a convex set is known to be
easy [26]. Thus, two-stage stochastic linear programming is in the class P P = P .
Assuming a P -oracle for evaluating Q in any point x of a two-stage stochastic
integer programming problem, makes the decision version of this problem a member
of N P . The function Q is not convex in this case, but there are a finite number of
points x that are candidate for optimality. Thus, the decision version of two-stage
stochastic integer programming is in the class N P P .
In case the random parameters of the two-stage stochastic programming problem
are continuously distributed, the evaluation of the function Q in a single point of
its domain requires the computation of a multiple integral. Most of the stochastic
programming literature on this subclass of problems is concerned with how to get
around this obstacle. We give the complexity of this class of problems without
proof.

Theorem 2.2 Two-stage stochastic programming problems with continuously dis-
tributed parameters is P -hard, even if all stochastic parameters have the uniform
[0, 1] distribution.

The membership of this problem in P requires additional conditions on the
input distributions, since exact computation may not even be in PSPACE.

3 Convex approximations for integer recourse prob-
lems
In this section we consider convex approximations for pure integer recourse models.
For such problems, the second-stage problem is necessarily defined using only in-
equalities. Moreover, in all models considered here only the right-hand side vector
h is random. The second-stage value function is thus
v(x, h) := min qy
y
s.t. W y ≥ h − T x, x ∈ Rn , h ∈ Rm
y ∈ Zn2 ,
+

where the components of W are assumed to be integers. Assuming complete and
sufficiently expensive recourse as before, v is a finite, discontinuous, piecewise con-
stant function; in particular, v is non-convex. It follows from Theorem 1.2 that
the integer recourse function Q(x) = Eh [v(x, h)], x ∈ Rn , is continuous if h is
continuously distributed, but in general Q is non-convex.
However, for certain integer recourse models, characterized by their recourse
matrices W , a class of distributions of h is known such that the corresponding
recourse function Q is convex. Thus, for such integer recourse models we can
construct convex approximations of the function Q by approximating any given
distribution of h by a distribution belonging to this special class.
Below we first apply this approach to the simple integer recourse model. Sub-
sequently, we consider general complete integer recourse models, starting with the
case of totally unimodular recourse matrices.

9

3.1 Simple integer recourse
The simple integer recourse second-stage problem is defined as
min q+ y+ + q− y−
y
s.t. y + ≥ h − T x,
y − ≥ −(h − T x),
y + , y − ∈ Zm ,
+

where the indices + and − are conventionally used to indicate surplus and shortage,
respectively. This recourse structure is obviously complete, and it is sufficiently
expensive if q + ≥ 0 and q − ≥ 0 (componentwise), as will be assumed from now on.
It is trivial to find a closed form for the simple integer recourse value func-
tion. Due to the simple recourse structure, this function is separable in the tender
variables z := T x:
m
v(z, h) = vi (zi , hi ), z, h ∈ Rm ,
i=1

where
+ + − −
vi (zi , hi ) = qi hi − zi + qi hi − zi , (3)
with s + := max{0, s } and s − := max{0, − s }, s ∈ R. Since all functions
vi have the same structure, we restrict the presentation to one such function, and
drop the index. It is straightforward to translate the results below back to the
full-dimensional case.
Given the closed form (3), it follows that the one-dimensional generic simple
integer recourse function Q equals
Q(z) = q + Eh h−z +
+ q − Eh h−z −
, z ∈ R, (4)
where h ∈ R is a random variable. Throughout we assume that Eh [|h|] is finite,
which is necessary and sufficient for finiteness of the function Q.

Lemma 3.1 Consider the one-dimensional simple integer recourse function Q de-
fined in (4).
(a) For all z ∈ R,
∞ ∞
Q(z) = q + Pr{h > z + k} + q − Pr{h < z − k}.
k=0 k=0

(b) Assume that h has a pdf f that is of bounded variation. Then the right deriva-
tive Q+ exists everywhere:
∞ ∞
Q+ (z) = −q + f+ (z + k) + q − f+ (z − k), z ∈ R,
k=0 k=0

where f+ is the right-continuous version of f .

Theorem 3.1 The one-dimensional simple recourse function Q is convex if and
only if the underlying random variable h is continuously distributed with a pdf f
that is of bounded variation, such that
f+ (s) = G(s + 1) − G(s), s ∈ R, (5)
where G is an arbitrary cdf with finite mean value.

10

Sufficiency of (5) is easy to see, since it implies that
Q+ (z) = −q + 1 − G(z) + q − G(z + 1), z ∈ R, (6)
is non-decreasing. Below we will make extensive use of the following special case.

Corollary 3.1 Assume that h is continuously distributed with a pdf f whose right-
continuous version is constant on every interval [α + k, α + k + 1), k ∈ Z, for some
α ∈ [0, 1). Then the function Q is piecewise linear and convex, with knots contained
in {α + Z}.

Proof. Immediate from Theorem 3.1 and (6), since f+ (s) = G(s + 1) − G(s) where
G is the cdf of a discrete distribution with support contained in α + Z.
To arrive at convex approximations of the function Q, we will use Corollary 3.1
to construct suitable approximations of the distribution of the random variable h.
For future reference, we present the multivariate definition of the approximations
that we have in mind.

Definition 3.1 Let h ∈ Rm be a random vector with arbitrary continuous or dis-
crete distribution, and choose α = (α1 , . . . , αm ) ∈ [0, 1)m . Define the α-approximation
hα as the random vector with joint pdf fα that is constant on every hypercube
m
Cα := i=1 (αi + ki − 1, αi + ki ], k ∈ Zm , such that Pr{hα ∈ Cα } = Pr{h ∈ Cα },
k k k

k∈Z . m

Returning to the one-dimensional case, it is easy to see that the α-approximations
hα , α ∈ [0, 1), of an arbitrary random variable h, satisfy the assumptions of
Corollary 3.1. It follows that the α-approximations of the function Q, defined
for α ∈ [0, 1),
Qα (z) := q + Ehα hα − z +
+ q − Ehα hα − z −
, z ∈ R,

are piecewise linear convex approximation of Q, with knots contained in {α + Z}.
Moreover, it follows from Lemma 3.1 (a) and Definition 3.1 that
Qα (z) = Q(z), z ∈ {α + Z}.
We conclude that, for each α ∈ [0, 1), Qα is the piecewise linear convex function
generated by the restriction of Q to {α + Z}. See Figure 1 for an example of the
function Q and one of its α-approximations.
In the discussion above, no assumptions were made on the type of distribution
of h. However, to establish a non-trivial bound on the approximation error, we need
to assume that h is continuously distributed. This loss of generality is acceptable,
because for the case with discretely distributed h it is possible to construct the
convex hull of the function Q.

Theorem 3.2 Assume that h is continuously distributed with a pdf f that is of
bounded variation. Then, for all α ∈ [0, 1),
|∆|f
Qα − Q ∞ ≤ (q + + q − ) ,
4
where |∆|f denotes the total variation of f .

Proof. We will sketch a proof for the special case that q + = 1 and q − = 0. The
proof for the general case is analogous.
Assume that q + = 1 and q − = 0. Then the function Q reduces to the expected
surplus function g(z) := Eh [ h − z + ], z ∈ R, with α-approximations gα (z) :=

11

7

6

5

4

3

2

1

0
−4 −3 −2 −1 0 1 2 3 4

Figure 1: The function Q and its α-approximation Qα (dashed) in case h is expo-
nentially distributed with parameter 5, q + = 1, q − = 1.5, and α = 0.5.

Ehα [ hα − z + ], α ∈ [0, 1). Since g(z) = gα (z) if z ∈ {α+ Z}, consider an arbitrary
ﬁxed z ∈ {α + Z}, and let z ∈ {α + Z} be such that z < z < z + 1.
Using Lemma 3.1 (b) we ﬁnd that
z ∞
g(z) − g(z) = f (t + k)dt.
z k=0

It follows from Lemma 2.5 in [20] that
∞
|∆|f |∆|f
1 − F (z) − ≤ f (t + k) ≤ 1 − F (z) + , t ∈ (z, z + 1),
2 2
k=0

so that
|∆|f |∆|f
1 − F (z) − (z − z) ≤ g(z) − g(z) ≤ 1 − F (z) + (z − z). (7)
2 2
On the other hand, using Lemma 3.1 (a) we see that
g(s + 1) = g(s) − (1 − F (s)), s ∈ R.
Since the function ga coincides with g on {α + Z}, and moreover gα is linear on the
interval [z, z + 1], it follows that
g(z) − gα (z) = (1 − F (z))(z − z). (8)
Together, (7) and (8) imply
|∆|f
|gα (z) − g(z)| ≤ (z − z) , z ∈ [z, z + 1]. (9)
2
Similarly, by comparing g(z) and gα (z) to g(z + 1), one obtains
|∆|f
|gα (z) − g(z)| ≤ (z + 1 − z) , z ∈ [z, z + 1]. (10)
2

12

For α-approximations of expected surplus function g, the claimed error bound now
follows from (9) and (10) on the observation that min{(z − z), (z + 1 − z)} ≤ 1/2.
Analogously, the same error bound can be derived for the special case with
q + = 0 and q − = 1. The claim for the general case then follows trivially.
The uniform error bound of Theorem 3.2 can be reduced by a factor 2 if the
following combination of α-approximations is used. For α ∈ [0, 0.5) and β = α+0.5,
define the pdf
fα (s) + fβ (s)
fαβ (s) = , s ∈ R,
2
where fα and fβ are density functions of α-approximations as before. The resulting
convex approximations Qαβ of Q satisfy
|∆|f
Qαβ − Q ∞ ≤ (q + + q − ) . (11)
8
It can be shown that this error bound can not be reduced by using other convex
combinations of pdf of type fα .
The error bound presented above is proportional to the total variation of the pdf
f of h. For many distributions, e.g. with unimodal densities, the total variation of a
pdf decreases as the variance of the distribution increases. We may therefore expect
that the approximation Qα becomes better as the variance of such distributions
becomes higher.
Finally, we remark that convex approximations of the function Q can be rep-
resented as (one-dimensional) continuous simple recourse functions. The latter
functions are defined like (4), except that no rounding operations are involved.
In the case of α-approximations, the corresponding modification of the underlying
distribution is known in closed form [19].
Lemma 3.2 Let h be a continuous random variable with cdf F with finite mean
value, and α ∈ [0, 1). Then
q+ q−
Qα (z) = q + E(ϕα − z)+ + q − E(ϕα − z)− + , z ∈ R,
q++ q−
where ϕα is a discrete random variable with support in α + Z and, for k ∈ Z,
q+ q−
Pr{ϕα = α + k} = Pr{h ∈ Cα } + +
k
Pr{h ∈ Cα }.
k+1
q+ + q− q + q−

We conclude that simple integer recourse functions can be approximated by
continuous simple recourse functions with discretely distributed right-hand side pa-
rameters, simply by dropping the integrality restrictions and a modification of the
distribution according to Lemma 3.2. The resulting convex problem can be solved
using existing algorithms for continuous simple recourse problems with discrete un-
derlying distributions.

3.2 Complete integer recourse
We now turn to the much more general class of complete integer recourse models.
In addition to completeness and sufficiently expensive recourse, so that v is finite,
we assume that the recourse matrix W is integer (or rational, so that integrality of
W can be obtained by scaling). We will see that also in this case α- approximations
of the distribution of h lead to convex approximations of the recourse function Q.

13

In fact, if the recourse matrix is totally unimodular (TU) then this approach leads
to the convex hull of Q. Below we first derive the results for this special case.
Because W is TU, the extreme points of the feasible set {y ∈ Rn2 : W y ≥ h}
are integral for any integer right-hand side h. However, in our recourse problem the
right-hand side h − T x is not an integer vector in general. But since W y is integral
for all y ∈ Zn2 we may round up the right-hand-side. Due to the assumption that
W is TU, we may now relax the integrality restrictions on y, without changing the
optimal value of the problem. That is,
v(x, h) := min qy
y
s.t. W y ≥ h − T x, y ∈ Zn2
+
= min qy (12)
y
s.t. Wy ≥ h − Tx , y ∈ Rn2
+
= max λ h − Tx (13)
λ
s.t. λW ≤ q, λ ∈ Rm ,
+

where the last equality follows from (strong) LP duality.
Since the recourse structure is complete and sufficiently expensive, it follows
that the dual feasible region Λ := {λ ∈ Rm : λW ≤ q} is a bounded, non-empty
+
polyhedral set. Hence,
v(x, h) = max λk h − T x , x ∈ Rn , h ∈ Rm , (14)
k=1,...,K

where λk , k = 1, . . . , K, are the finitely many extreme points of the dual feasible
set Λ.
Thus, v is the maximum of finitely many round up functions, and hence non-
convex. However, as we will see below, the recourse function Q is convex if the
underlying distribution of h is of a certain type. Analogous to the simple recourse
case, this allows the construction of convex approximations of Q by means of special
purpose approximations of the distribution.
To set the stage, we first study the expected round up function
R(z) := λEh [ h − z ] , z ∈ Rm ,
defined for any fixed λ ∈ Rm .
If m = 1, λ = 1, and h is continuously distributed, then
+ −
R(z) = Eh h−z − Eh h−z+1 , z ∈ R, (15)
+ − − −
since s = s − s , s ∈ R, and s = s + 1 for all s ∈ Z. The right-hand
side of (15) is very similar to the one-dimensional simple recourse function with
q + = 1 and q − = −1. Hence, in view of Corollary 3.1 it is not surprising that this
one-dimensional function R is convex if h has a piecewise constant pdf of the type
specified in that lemma. This result can be generalized to m-dimensional round up
functions.
Lemma 3.3 Let h ∈ Rm be a continuous random vector with joint pdf fh that
is constant on every hypercube Cα := m (αi + ki − 1, αi + ki ], k ∈ Zm , for an
k
i=1
arbitrary but fixed α = (α1 , . . . , αm ) ∈ [0, 1)m . Then
Eh [ h − z ] = Eϕα [ϕα − z] = µα − z, z ∈ Rm ,
where ϕα = h−α +α is a discrete random vector with mean value µα and support
in α + Zm , with
Pr{ϕα = α + k} = Pr{h ∈ Cα },
k
k ∈ Zm .
Hence, in this case the round up function R(z) = λEh [ h − z ], z ∈ Rm , is affine
with gradient −λ.

14

Proof. We use that
Eh [ h − z ] = Pr{h ∈ Cα }Eh
k
h−z h ∈ Cα ,
k
z ∈ Rm . (16)
m
k∈Z

For each fixed k ∈ Zm , Pr{h ∈ Cα } is either zero or the conditional distribution
k

of h given h ∈ Cα is uniform on Cα . In that case, the components of the vector
k k
k
h are independent random variables on Cα , with each hi uniformly distributed on
(αi + ki − 1, αi + ki ], i = 1, . . . , m. Hence, writing each component as in (15) and
applying Lemma 3.2 to each term individually, it follows that
Eh h−z h ∈ Cα = α + k − z,
k
z ∈ Rm . (17)
Substitution of (17) in (16) proves the first claim.
The second claim follows trivially from the first one.
Based on Lemma 3.3, we define α-approximations of the function R: for α ∈
[0, 1)m ,
Rα (z) := λEhα [ hα − z ] , z ∈ Rm .
In general, an α-approximation is neither a lower bound nor an upper bound. How-
ever, since R(z + k) = R(z) − λk, k ∈ Zm , for every z, we see that R(z) + λz is a
k
periodic function, which repeats itself on every set Cα . Thus, defining
α ∈ argmin {R(z) + λz : z ∈ [0, 1)m } , (18)
Rα is a lower bound for R, which is sharp at every z ∈ α + Zm . By construction,
the affine function Rα is actually the convex hull of R.
The components αi , i = 1, . . . , m, of the parameter vector α can independently
be determined analytically in almost all practical cases. If the marginal distribution
of hi is continuous, one-sided derivatives of the function Ri (zi ) := λi Ehi [ hi − zi ]
(analogous to Lemma 3.1) are used; if it is discrete with finitely many different
fractional values in its support, the computation of αi is based on the direct rela-
tion between these fractional values and discontinuities of the lower semicontinuous
function Ri .
Now we are ready to prove the main result for this class of models with TU
recourse matrix. Using the dual representation (14) of the value function v, we
have

Q(x) = Eh max λk h − T x , x ∈ Rn .
k=1,...,K

Note that Q is not simply the pointwise maximum of a number of expected round
up functions R. However, the results above for the function R play a major role in
the proof of Theorem 3.3.

Theorem 3.3 Consider the integer recourse expected value function Q, defined as

Q(x) = Eh min qy : W y ≥ h − T x, y ∈ Zn2 ,
+ x ∈ Rn1 , (19)
y

Assume that
(i) the recourse is complete and sufficiently expensive, and
(ii) the recourse matrix W is totally unimodular.
If
(iii) the matrix T is of full row rank,

15

then the convex hull of Q is the continuous recourse expected value function Qα ,
defined as

Qα (x) = Eϕα min qy : W y ≥ ϕα − T x, y ∈ Rn2 ,
+ x ∈ Rn1 , (20)
y

where α is defined by (18), and ϕα is the discrete random vector ϕα = h −
α + α with support in α + Zm , and
Pr{ϕα = α + k} = Pr{h ∈ Cα },
k
k ∈ Zm .
If condition (iii) is not satisfied, then Qα is a lower bound for Q.

Proof. We will prove that Qα is the convex hull of Q if T is of full row rank.
The other case then follows from Theorem 2.2 in [17].
Assuming that T is of full row rank, we may conveniently consider Q as a
function of the tender variables z := T x ∈ Rm .
First we will prove that Qα is a lower bound for Q, and subsequently that
Qα (z) = Q(z) for all z ∈ α + Zm . This completes the proof, since all vertices of
the polyhedral function Qα are contained in α + Zm .
Using the dual representation (14) of the value function v, we have

Q(z) = Eh max λk h − z , z ∈ Rm ,
k=1,...,K

and, analogously,

Qα (z) = Eϕα max λk (ϕα − z) , z ∈ Rm .
k=1,...,K

Conditioning on the events h ∈ Cα , l ∈ Zm , we obtain, for z ∈ Rm ,
l

Q(z) = Pr{h ∈ Cα } Eh
l
max λk h − z h ∈ Cα
l
k=1,...,K
l∈Z
m

≥ Pr{h ∈ Cα } max λk Eh
l
h−z h ∈ Cα
l
k=1,...,K
l∈Z
m

≥ Pr{h ∈ Cα } max λk Ehα
l
hα − z hα ∈ Cα
l
k=1,...,K
l∈Z
m

= Pr{h ∈ Cα } max λk (α + l − z)
l
k=1,...,K
l∈Z
m

= Pr{ϕα = α + l} max λk (α + l − z) = Qα (z).
k=1,...,K
l∈Z
m

The second inequality is valid because each λk is nonnegative, so that the α-
approximation λk Ehα hα − z hα ∈ Cα is a lower bound for λk Eh h − z h ∈ Cα
l l

by the choice of α . The subsequent equality holds by Lemma 3.3.
It remains to prove that Qα = Q on α + Zm . Consider a fixed z ∈ α + Zm
¯
and a fixed l ∈ Zm . Then h − z = l − z is constant for all h ∈ Cα , so that there
¯ ¯ l

exists a λ(¯, l) satisfying
z
λ(¯, l) ∈ argmax λk h − z
z ¯ ∀h ∈ Cα .
l
k=1,...,K

Since this is true for every z ∈ α + Zm and l ∈ Zm , it follows that, for z ∈ α + Zm ,
¯

Q(z) = Pr{h ∈ Cα } λ(z, l) Eh
l
h−z h ∈ Cα
l

l∈Z
m

16

= Pr{h ∈ Cα } λ(z, l) Ehα
l
hα − z hα ∈ Cα
l

l∈Z
m

= Pr{h ∈ Cα } λ(z, l) (α + l − z)
l

l∈Z
m

= Pr{h ∈ Cα } max λk (α + l − z) .
l
k=1,...,K
l∈Z
m

The second equality follows from the fact that each α-approximation is sharp on
α+ Zm . The last equality follows from the definition of λ(z, l) and ϕl − z = l − z ,
α
z ∈ α + Zm .
We conclude that if the recourse matrix W is totally unimodular, then the integer
complete recourse problem with recourse function Q can be approximated by the
continuous complete recourse problem with recourse function Qα . To construct this
approximation, the integer restrictions on the second-stage variables are dropped,
and the distribution of the right-hand side parameters is modified according to
Theorem 3.3. The resulting continuous complete recourse problem with discretely
distributed right-hand side parameters can be solved by existing special purpose
algorithms [2, 16].
In particular, if the matrix T is of full row rank, then solving the approximating
problem will yield the true optimal solution, at least if the first-stage constraints
are not binding.
Finally, we drop the assumption that W is TU. In this case, we will prove that
Qα is a strictly better convex approximation than the one obtained using the LP
relaxation of the second-stage problem. The latter convex function will be denoted
by QLP , defined as

QLP (x) := Eh min qy : W y ≥ h − T x, y ∈ Rn2
+ , x ∈ Rn1 . (21)
y

Theorem 3.4 Consider the functions Qα and QLP , defined by (20) and (21)
respectively, which both are convex lower bounds for the integer recourse expected
value function Q, defined by (19).
(a) Qα ≥ QLP
(b) Assume
(i) q ≥ 0, so that 0 is a trivial lower bound for v and Q;
(ii) there exists a subset L of Zm such that the support Ω ⊂ l∈L {h : h ≤ α + l}
and Pr{h < α + l h ∈ Cα } > 0 for all l ∈ L.
l

Then the function Qα is a strictly better convex approximation of Q than
QLP , in the sense that Q(x) > 0 implies Qα (x) > QLP (x).

Proof. As before, we condition on the events h ∈ Cα , l ∈ Zm , to obtain, for
l

x∈R ,
n1

Qα (x) = Pr{h ∈ Cα } max λk (α + l − T x)
l
(22)
k=1,...,K
l∈Z
m

and

QLP (x) = Pr{h ∈ Cα } Eh
l
max λk (h − T x) h ∈ Cα
l
. (23)
k=1,...,K
l∈Z
m

17

m
For each l ∈ Zm it follows from the definition of Cα = i=1 (αi + li − 1, αi + li ]
l

that α + l ≥ h for all h ∈ Cα . Thus, for k = 1, . . . , K, λk (α + l − T x) ≥
l

λk (h − T x) for all h ∈ Cα , since λk ≥ 0. Substitution in (22) and (23) proves
l

that Qα ≥ Q . LP

To prove (b), we first show that Q(x) > 0 implies Qα (x) > 0. To this end,
define
N (x) := {t ∈ Rm : v(t − T x) > 0} , x ∈ Rn1 .
Then Q(x) > 0 if and only if Pr{h ∈ N (x)} > 0, which is equivalent to Pr{h ∈
int N (x)} > 0 since N (x) is an open set. By Definition 3.1, it follows that then also
Pr{hα ∈ N (x)} > 0, which implies Qα (x) > 0 for all α ∈ [0, 1)m .
Let x be such that Q(x) > 0, implying Qα (x) > 0. Then, since each term of (22)
is non-negative by assumption (i), there exists an ¯ ∈ L such that maxk=1,...,K λk α + ¯ − T x >
l l
¯ ¯
0; obviously, any optimal solution λ of this problem satisfies λ = 0. For an arbitrary
¯
but fixed h ∈ Cα such that h < α + ¯ it holds
l ¯ l,
¯
λ h − T x ≤ λ α + ¯− T x
l ∀λ ≥ 0,
ˆ ¯
with strict inequality unless λ = 0. Let λ be an optimal solution of maxk=1,...,K λk h − T x .
Then there are two possibilities:
ˆ ˆ ¯ ¯
(i) λ = 0, so that λ h − T x = 0 < λ α + ¯ − T x ;
l
ˆ ˆ ¯ ˆ
(ii) λ = 0, so that λ h − T x < λ α + ¯ − T x ≤ λ α + ¯ − T x .
l ¯ l
¯
¯
We conclude that, for all h ∈ Cα with h < α + ¯
l ¯ l,
max λk α + ¯ − T x >
l ¯
max λk h − T x . (24)
k=1,...,K k=1,...,K

¯
Since Pr{h < α + ¯| h ∈ Cα } > 0 by assumption (ii), and (24) holds with weak
l l
¯
inequality for all h ∈ Cα , it follows that
l

¯
max λk α + ¯ − T x > Eh
l max λk (h − T x) h ∈ Cα
l
. (25)
k=1,...,K k=1,...,K

Finally, using that (25) holds with weak inequality for all l ∈ Zm , we see from (22)
and (23) that Qα (x) > QLP (x).
For example, condition (b) (ii) of Theorem 3.4 is satisfied if h follows a non-
degenerated continuous distribution.
Note that the distribution of ϕα as defined in Theorem 3.3 is always discrete, no
matter what kind of distribution h follows. Thus, in particular if h is continuously
distributed, Qα is not only a better approximation of Q, it is also computationally
more tractable than QLP which in this case is defined as an m-dimensional integral.

4 Hierarchical planning models
Consider a two-level decision situation. At the higher level, aggregate decisions are
made concerning acquisition of resources. At the lower level, one has to decide on
the actual allocation of the resources. The time horizon for aggregate decisions in
such hierarchical decision problems may range from several months to a year. At the
time aggregate decisions are made much detailed information of what will ultimately
be required of the resources is not yet known with certainty. As mentioned in the
introduction, two-stage stochastic programming is the right tool to model the lower
level of hierarchical planning problems accurately, using stochastic parameters for
which probability distributions are specified. The objective at the higher level is

18

to minimize known costs at that level plus the expected objective value of optimal
lower level decisions.
We focus on hierarchical planning problems with detailed-level problems which
are of a combinatorial nature. This class of problems includes hierarchical schedul-
ing problems, hierarchical vehicle routing problems, and hierarchical knapsack prob-
lems (capital budgeting problems). We will consider the design and analysis of ap-
proximation algorithms for such problems. In management science literature such
algorithms are often called hierarchical planning systems. We diverge here from
the previous section by not aiming at approximation algorithms for the stochastic
integer programming problem in which all hierarchical combinatorial optimization
problems can be formulated, but instead at algorithms tailored to the speciﬁc par-
ticular hierarchical combinatorial optimization problems. We use as an example
a hierarchical scheduling problem and apply probabilistic analysis to measure the
performance quality of an approximation algorithm for this problem
Consider the following hierarchical scheduling problem. At the time machines
are to be installed only probabilistic information is available on the jobs to be
processed. The two-stage stochastic programming model of this problem has to
select the number or the types of the machines so as to minimize the installation
costs of the machines plus the expected cost of processing the jobs optimally on the
installed machines.
In this problem, the machines to be installed at the aggregate level are identical
and work in parallel. Installation of each machine costs c. A decision is required
on the number of machines to be installed. If x denotes this number, then the
installation costs are cx.
There are N jobs to be processed and each job j requires processing for a time
tj , j = 1, . . . , N . At the time the machines are purchased, there is only probabilistic
information about the processing times of the jobs. A schedule of the jobs on the
available machines is feasible if each job gets assigned one time interval of length
equal to its processing time on one of the machines, and each machine processes
only one job at a time. The makespan of a set of jobs is the time by which the
last job is completed in a feasible schedule. The objective is to minimize the sum
of installation costs of the machines and the expected makespan of the jobs on
the available machines. (To make dimensions compatible assume, without loss of
generality, that the cost per time unit in the second-stage schedule is equal to 1.)
Let v ∗ (x, t) denote the optimal second-stage costs, which is a random variable,
a function of the random processing times of the jobs t = (t1 , . . . , tN ). Let Q(x) =
Et [v ∗ (x, t)] denote its expectation. Then the objective is min z(x) = cx + Q(x).
Let x∗ denote the optimal solution.
From Section 2 we know that computing Q(x) is a formidable task. Even if the
distribution of t would be given by discrete probabilities over a set of vectors, the
deterministic equivalent problem is NP-hard, since computing the optimal makespan
of a set of jobs on more than one machine is NP-hard. The approximation algorithm
H consists of replacing Q(x) by a simple function QH (x) as an approximate.
N
Obviously, given a realization t of t, j=1 tj /x, the makespan of the schedule
in which all x machines have equal workload, is a lower bound on v ∗ (x, t):
N
v ∗ (x, t) ≥ tj /x. (26)
j=1

We choose to take QH (x) = Et [ N tj ]/x as an approximate value for Q(x). If we
j=1
assuming for simplicity that t1 , . . . , tN have equal mean µ then QH (x) = N µ/x.
We solve the approximate problem
min z H (x) = cx + QH (x). (27)

19

H
z H (x) is a convex function and dz = 0 at x = N µ/c. Since the number of ma-
dx
chines must be integer we choose xH equal to the value x ∈ { N µ/c , N µ/c }
that minimizes z H (x). The outcome of the approximation algorithm is then z(xH ).
Taking expectations in (26) yields Q(x) ≥ QH (x). Hence z(x) ≥ z H (x) for all
x, and therefore
min z(x) = z(x∗ ) ≥ z H (xH ) = cxH + QH (xH ) ≥ 2 cN µ. (28)
To estimate the quality of the approximation we aim to find an appropriate upper
bound on z H (xH ) in terms of z(x∗ ). It is well-known [14] that the list scheduling
rule, which assigns the jobs in an arbitrary order to the machines and each next
job is assigned to the earliest available machine, yields the following upper bound
on the makespan for any number of machines x given any realization t of t:
N
v ∗ (x, t) ≤ tj /x + max tj .
j=1,...,N
j=1

In particular for xH , denoting tmax = maxj=1,...,N tj and taking expectations yields
Q(xH ) ≤ QH (xH ) + Et [tmax ]. (29)
Hence,
z(xH ) ≤ cxH + QH (xH ) + Et [tmax ]. (30)
Together, (28) and (30) give a bound on the worst-case performance ratio of the
approximation algorithm.
Lemma 4.1
z(xH ) Et [tmax ]
∗)
≤1+ √ .
z(x 2 cN µ
Probability theory (see e.g. [34]) tells us that
Lemma 4.2 If t1 , . . . , tN have finite second moments, then
√
lim Et [tmax ]/ N = 0.
N →∞

This probabilistic result applied to Lemma 4.1 yields asymptotic optimality of
the approximation algorithm.
Theorem 4.1 If t1 , . . . , tN have finite second moments then
z(xH )
lim = 1.
N →∞ z(x∗ )

z(xH )
Proof. The combination of Lemmas 4.1 and 4.2 implies limN →∞ z(x∗ ) ≤ 1.
z(xH )
Clearly, z(x∗ ) ≥ 1.

The above shows something more than asymptotic optimality. In replacing the
expected second-stage optimal costs by an estimate, the NP-hardness of the second-
stage problem is not taken into consideration. This could imply that we take the
estimate of a quantity that we are unable to compute efficiently once we obtain a
realization of the second-stage parameters (the processing times of the jobs).
However, as we have seen, asymptotic optimality is proved by comparing the so-
lution given by the algorithm to a solution based on a polynomial time approxima-
tion algorithm for the second-stage problem. It implies that optimality is retained
even if we use a simple approximation algorithm to solve the second-stage problem.

20

We could also assess the quality of an algorithm that uses xH as the number of
machines and list scheduling for the second-stage problem on xH machines. The
solution value is a random variable z LS (xH , t) = cxH + v LS (xH , t). One could
wonder how close this value is to the solution value of an algorithm that selects
x∗ machines and upon a realization of the processing times is able to select the
optimal schedule on those x∗ machines. Let denote this solution value by z ∗ (x∗ ) =
cx∗ + v ∗ (x∗ , t).
Or one could even wonder how z LS (xH , t) compares to the solution value of an
optimal clairvoyant algorithm that is able to know the realization of the processing
times before deciding the number of machines to be installed. In this case the
optimal number of machines becomes a random variable denoted by x0 (t). Let us
denote the optimal solution value in this case by z 0 (x0 (t), t) = cx0 (t) + v ∗ (x0 (t), t).
This is the solution of the model that in stochastic programming is called the
distribution model. In more popular terms it is called the wait and see model for
obvious reasons opposed to here and now model used for the two-stage model. The
approximation algorithm presented above appears to have the strong asymptotic
optimality property that, again under the assumption that the random processing
times have finite second moments,
z LS (xH , t)
lim = 1,
N →∞ z 0 (x0 (t), t)

with probability 1, or almost surely. A sequence of random variables y1 , . . . , yN is
said to converge almost surely to a random variable y if P r{limN →∞ yN = y} = 1.
The proof is similar to the proof of Theorem 4.1. Under some mild extra conditions
on the random variables, which are satisfied for example if the distributions of all
random variables have bounded support, this asymptotic quality guarantee implies
directly that
z LS (xH , t)
lim = 1,
N →∞ z ∗ (x∗ , t)

almost surely. It also implies the result of Theorem 4.1.
The ideas used above in constructing an asymptotically optimal approximation
algorithm for the two-stage stochastic scheduling problem are applicable more gen-
erally. Given a two-stage combinatorial optimization problem replace the value
function by an estimate that is asymptotically accurate and use an asymptotically
optimal approximation algorithm for the second-stage problem in case this problem
is NP-hard.

5 Worst-case performance analysis
As an example of worst-case performance analysis of approximation algorithms for
stochastic optimization problems we consider a service provision problem. Actually,
to the best of our knowledge it is the only example of this type of analysis in
stochastic programming. In that sense, a rich research area lies nearly unexplored.
The problem we study concerns provision of services from a resource. For each
of a given set of services there are requests from customers. In order to meet a
request for a service, the service has to be installed and once installed, the request
has to be served. Both installation and provision of a service requires capacity from
the same resource. The resource has limited capacity. Each request served yields
a given profit. The problem is to select a subset of the services to be installed and
to decide which customer requests to serve, such as to maximize the total profit by
serving requests.

21

If all demands for services are known in advance, the problem is NP-hard in the
ordinary sense and a fully-polynomial time approximation scheme exists.
We study the problem with uncertain demand for services. The uncertainty
is represented by a discrete probability distribution over the demands. The two-
stage stochastic programming problem is to select services to be installed such as
to maximize expected profit of serving requests for services. We will show that this
problem is strongly NP-hard. Thus, the complexity of the problem increases by
introducing stochasticity.
We analyse the performance of an approximation algorithm for this problem
under the restriction that the resource has enough capacity to install all services.
It may not be optimal to install all of them since it may leave too little capacity for
serving the requests.
We start with formulating the problem as a two-stage stochastic integer pro-
gramming problem. Let n be the number of services and s the capacity of the
single resource. Let qj be the profit obtained from allocating one resource unit
to meeting demand for service, j. Each service j requires a resource capacity rj
to be installed, which is independent of the demand met. Demand is denoted by
the random vector d ∈ Rn , with dj denoting the demand for service j. Binary
decision variables zj are used to indicate whether service j is installed (zj = 1), or
not (zj = 0), j = 1, . . . , n. Decision variable xj gives the amount of resource used
to meet demand for service j. The two-stage stochastic programming formulation
becomes:
max Ed [v(z, d)]
n
s.t. rj zj ≤ s
j=1
zj ∈ {0, 1} j = 1, . . . , n,
with
n
v(z, d) = max qj xj
j=1
n n
s.t. xj ≤ s − rj zj
j=1 j=1
xj ≤ dj zj j = 1, . . . , n,
xj ≥ 0 j = 1, . . . , n.
The second-stage problem is to set the values of the variables xj under two con-
straints: the capacity constraint ensuring that resource capacity is not exceeded
and the demand constraint ensuring that demand is not exceeded and met only for
services that have been installed. The constraint in the first stage ensures relatively
complete recourse; i.e. for every first stage solution that is feasible with respect to
the first stage constraints, the resulting second-stage problem is feasible for every
realization of the random parameters.
Let K be the number of scenarios describing the probability distribution on
demand, pk the probability that scenario k occurs, and dk the demand for service
j
j in scenario k. Given the scenarios the the following deterministic equivalent
linear mixed integer program can be formulated, in which we use xjk to denote the
resource allocated to providing service j in scenario k (we use a subscript instead

22

of superscript for k here because of notational convenience later on).
K n
max pk qj xjk
k=1 j=1
n
s.t. (rj zj + xjk ) ≤ s k = 1, . . . , K, (31)
j=1
dk zj − xjk ≥ 0
j j = 1, . . . , n, k = 1, . . . , K,
zj ∈ {0, 1}, xjk ≥ 0 j = 1, . . . , n, k = 1, . . . , K.
Though integrality conditions only hold for the first stage variables zj , if the
data, resource capacity, installation requirements, and demands are integral, the
second stage will have an integer solution in every scenario.

Theorem 5.1 The stochastic single resource service provision problem is strongly
NP-hard.

Proof. The natural recognition version of this problem obtained by introducing
a number Λ and asking if there is a feasible solution with objective value at least
Λ is in NP, following directly from the deterministic equivalent formulation. To see
that it is strongly NP-Complete consider a reduction from the well-known strongly
NP-Complete vertex cover problem (see [13]):

Given a graph G = (V, E) with |V | vertices and |E| edges and a constant
κ, does there exist a subset V of the vertices, such that each edge in E
is incident to at least one vertex in V , and such that |V | ≤ κ?

For every vertex j ∈ V introduce a service j with installation requirement α =
1
κ|E| . For every edge introduce a scenario with demand 1 for the two services incident
1
to it and demand 0 for all other services. Let all scenarios have probability |E| .
Define qj = |E| ∀j ∈ V , s = κα + 1 (resource capacity), and Λ = |E|.
If there exists a vertex cover of size at most κ then there is a solution to the
instance of the stochastic service provision problem with total expected profit at
least |E|. Install the services corresponding to the vertices in the vertex cover. Then
for each scenario (edge) at least one of the services with demand 1 is installed. The
total capacity used by the installation of the services is at most κα leaving at least
capacity 1 in each scenario to satisfy demand.
The other direction is a bit more complicated. Suppose there does not exist a
vertex cover of size κ or less. Then installing all services corresponding to a vertex
cover would use node capacity strictly greater than κα leaving strictly less than 1
for meeting demand in each of the |E| scenarios, making a total expected profit of
at least |E| unattainable. Installing any set of services of size L < κ would leave
(κ − L)α + 1 node capacity for meeting demand in each scenario. However, at
least one edge will remain uncovered, implying that there is at least one scenario
in which both services with a positive demand are not installed. With at most
|E| − 1 scenarios the expected profit will be at most (|E| − 1)((κ − L)α + 1) ≤
1
(|E| − 1)(κα + 1) = (|E| − 1)( |E| + 1) < |E| = Λ.

As announced, we assume that n rj ≤ s. Moreover, to facilitate the expo-
j=1
sition the assumption is made that no demand is higher than the node capacity
minus the corresponding installation requirement: For any service j in any sce-
nario k, dk ∈ [0, s − rj ]. If necessary, this can be ensured by preprocessing.
j
The approximation algorithm that we will present is based on rounding the
optimal solution of the LP-relaxation of problem (31), obtained by replacing the

23

binary restrictions on the z-variables by 0 ≤ zj ≤ 1, j = 1, . . . , n. To facilitate
the exposition we assume, without loss of generality, that the resource capacity s is
equal to 1.
Let (z lp , xlp ) be an optimal basic solution of the LP relaxation. Let be the
lp
number of fractional zj and let w of these services have rj ≤ w for some 0 < w < 1
lp
to be chosen later. Let Z be the set of services with zj = 1. By renumbering the
services if necessary, assume that 0 < zj lp < 1 and r ≤ w for j = 1, . . . ,
j w and
lp
0 < zj < 1 and rj > w for j = w + 1, . . . , . Write the optimal LP value as
lp lp
π lp = π0 + π1 + π2
lp (32)
where
K
lp
π0 = pk qj xlp ,
jk
j∈Z k=1

w K
lp
π1 = pk qj xlp ,
jk
j=1 k=1

and
K
lp
π2 = pk qj xlp .
jk
j= w +1 k=1

Feasible solutions generated from the LP solution constitute the approximation
algorithm, which selects from those solutions the best one. The algorithm is there-
fore a kind of rounding algorithm and we denote its solution value by π R . Let π OP T
denote the optimal solution value of the stochastic integer program.
lp
The first feasible solution is obtained by installing service j if and only if zj = 1;
i.e., install all services j ∈ Z. The remaining capacity is then allocated to serve
demand for the installed services in a greedy way, in order of non-increasing qj
values. Denote the resulting solution by (z G , xG ) and its value by π G . Then,
obviously,
lp
π0 ≤ π G ≤ π R . (33)
lp
The next set of feasible solutions is used to bound π1 . Define A = j=1 rj zj lp
and note that w
xlp ≤ 1−A for each k = 1, . . . , K. Partition the set {1, . . . , }
j=1 jk w
into I subsets, {Si }I ,
i=1 such that

rj ≤ β + w i = 1, . . . , I
j∈Si

and
rj ≥ β i = 1, . . . , I − 1, (34)
j∈Si

for some constant β > 0 to be chosen later, such that β + w < 1. Notice that
j∈SI rj is allowed to be smaller than β. In the algorithm this partition is made in
the most simple way, starting filling set S1 until addition of the next service would
make the sum of installation requirements exceed β + w. This service is then the
first one of S2 , etc.
In the optimal solution of the LP relaxation at most 1 − A units of capacity
are available for the x variables. Installing only the services in one of the sets Si

24

will leave at least 1 − β − w units of capacity available. The x-variable values from
the LP relaxation solution corresponding to services in Si may be scaled down, if
necessary, to use a total of no more than 1− β − w units of capacity in each scenario.
H
For each i = 1, . . . , I we obtain a feasible solution (z Hi , xHi ) with zj i = 1 for
H
/ lp
j ∈ Si , zj i = 0 for j ∈ Si , xHi = γxjk for j ∈ Si , k = 1, . . . , K and xHi = 0 for
jk jk
j ∈ Si and all k, where
/

 1−β−w
if β + w ≥ A,
γ= 1−A (35)
 1 otherwise.

The objective value of solution (z Hi , xHi ) is
K K
π Hi = pk qj xHi = γ
jk pk qj xlp .
jk
j∈Si k=1 j∈Si k=1

Hence,
I K I
lp 1 I R
π1 = pk qj xlp =
jk π Hi ≤ π (36)
i=1 j∈Si k=1
γ i=1
γ
n
By the assumption j=1 rj ≤ s and the deﬁnition of the sets Si (speciﬁcally (34))
we have
n w I
1≥ rj ≥ rj = rj ≥ (I − 1)β. (37)
j=1 j=1 i=1 j∈Si

Thus, I ≤ 1 + 1/β which inserted in (36) implies that
lp β+1 R
π1 ≤ π . (38)
βγ
The last set of feasible solutions considered by the algorithm consists of installing
each service j = w +1, . . . , (having rj ≥ w) individually. Since A = j=1 rj zj ≥lp
r z lp ≥ w z lp , we have
w +1 w +1
j= j j j= j

lp A
zj ≤ .
w
j= w +1

Just installing service j has objective value qj E[δj ], since we have assumed that for
any service j in any scenario k. dk ∈ [0, s − rj ]. Satisfying the demand constraints
j
implies that k=1 pk xlp ≤ E[δj ]zj . Altogether this yields the following bound.
lp
K
jk

K
lp
π2 = pk qj xlp
jk
j= w +1k=1

≤ lp
qj E[δj ]zj
j= w +1

lp A R A
≤ πR zj ≤ π ≤ π opt (39)
w w
j= w +1

Combining (33), (38), and (39) gives
β+1 A
π lp ≤ 1+ + πR (40)
βγ w

25

n
Theorem 5.2 Under the assumption that j=1 rj ≤ 1, the approximation algo-
rithm has worst-case performance ratio
π OP T √
≤ (5 + 2 3).
πR
√
Proof. The choice of w and β depends on A in (40). When A < 1 take w = 1− 1 3
√ 2 2
and β = − 1 + 1 3 and when A ≥ 1 take w = β = 1 A. In both cases w + β ≥ A,
2 2 2 2
and therefore γ = 1−β−w . In the former case (40) leads to
1−A
√
2(1 + 3)(1 − A) A
π opt ≤ π lp ≤ 1+ √ + √ πR
−1 + 3 1− 1 3
2
√ 2 1√ √
= 1 + (1 + 3) (1 − A) + 4(1 + 3)A π R = (5 + 2 3)π R .
2
In the latter case (40) leads to
2 √
π opt ≤ π lp ≤ 4+ π R ≤ 8π R ≤ (5 + 2 3)π R .
A

We notice that so far tightness of the bound has not been established. There
exist an instance in which the ratio between the LP-bound and the optimal value
is 4 and an instance for which the algorithm has ratio 2. The results show the pos-
sibilities of achieving worst-case performance results for approximation algorithms
for stochastic integer programming problems. It is worthwhile to stress once more
that the deterministic counterpart of the problem, having the same number of bi-
nary decision variables, is weakly NP-hard. Thus, the complexity of the problem
increases by introducing stochasticity, even if it only means adding continuous de-
cision variables for each scenario of the problem.

6 Notes
Stochastic programming models date back to the ﬁfties [5, 3]. Several surveys on
stochastic programming have appeared of which we mention here the introductory
book of Kall and Wallace [16] and the comprehensive books by Prekopa [29] and by
Birge and Louveaux [2]. For surveys speciﬁcally on stochastic integer programming
we refer to the chapter by Louveaux and Schultz in the Handbook on Stochas-
tic Programming [31], and the survey papers Klein Haneveld and Van der Vlerk
[21], R¨misch and Schultz [30], and Stougie and Van der Vlerk [37]. Resources on
o
the Internet are the Stochastic Programming Community Home Page [4] and the
bibliography [42].
The focus in this chapter is on the two-stage recourse model. For a detailed
discussion of the multistage model and generalizations (including random recourse
matrices and nonlinear models) we refer to the Handbook on Stochastic Program-
ming [31] or to [2, 16, 29].
More about the important class of chance-constrained problems and the related
(conditional) value at risk models can be found in the Handbook on Stochastic
Programming [31]. This class is of problems is very well surveyed in [29] and [39].
The mathematical properties of two-stage stochastic linear programming prob-
lems have been derived by various people and at a rather early stage in the research
activities on stochastic programming. In particular we refer to the overview by
Wets [45] and the monograph Kall [15].

26

The mathematical properties of two-stage stochastic integer programming prob-
lems have been established much more recently [36, 41, 32]. Schultz [32] proved
the properties of the mixed-integer recourse function presented in Theorem 1.2. In
addition, Schultz presented rather technical conditions for Lipschitz continuity of
the function Q.
The results in Section 2 are selected from [12]. P -completeness of the prob-
lem graph reliability has been proved in [40]. That exact evaluation of the
second-stage expected value function may not even be in P SP ACE in case random
parameters are continuously distributed follows from a result in [23].
Dyer and Stougie [12] also prove P SP ACE-hardness of a specific non-standard
version of a multi-stage stochastic programming problem if the number of stages
is considered to be part of the input. The complexity of standard multi-stage
stochastic programming remains unsolved.
Kannan et al. [11] have designed a polynomial randomized approximation scheme
for the two-stage stochastic programming problem with continuously distributed
parameters and continuous decision variables, when the input distributions are re-
stricted to be log-concave. Their scheme relies heavily on the convexity of Q, and
therefore cannot be applied to the two-stage stochastic integer programming prob-
lem.
The idea in Section 3 of approximating the expected value function of a stochas-
tic programming problem with integer recourse by a convex function through per-
turbing the distributions of the random right-hand sides is due Klein Haneveld et
al. [20, 19]. They implemented this idea for the case of simple integer recourse. See
Van der Vlerk [44] for a generalization to multiple simple recourse models, allowing
for piecewise-linear penalty cost functions. The extension to the compete integer
recourse case was done by Van der Vlerk [43].
For the problem with simple integer recourse, the formula and properties in
Lemma 3.1 have been derived by Louveaux and Van der Vlerk [25], while the char-
acterization of all probability distributions that lead to convex expected value func-
tions in Theorem 3.1 is due to Klein Haneveld et al. [20].
The uniform error bounds on the α-approximation in Theorem 3.2 and on the
αβ-approximation in (11) are from [19]. There it is also shown that the latter error
bound can not be reduced by using other convex combinations of probability density
functions of type fα . The error bounds are derived in case the distributions of the
random right hand sides are continuous. For the case with discretely distributed h
it is possible to construct the convex hull of the function Q, see [18].
Algorithms for continuous simple recourse problems with discretely distributed
right-hand side parameters can be found in e.g. [2, 16]. Using the structure of such
problems, they can be represented as relatively small deterministic LP problems.
If the matrix W is complete but not TU, then the function Qα defined in The-
orem 3.3 can be used as a convex lower bounding approximation of the function Q,
allowing to approximately solve the integer recourse problem by solving a continu-
ous complete recourse model. Although this approach is easy to implement and in
many cases will give better results than using the LP lower bound QLP , no (non-
trivial) bound on the approximation error is known. Indeed, in most applications
the approximation will not be good enough for this purpose. On the other hand,
because of the properties discussed in Section 3, the function Qα is well-suited as a
building block in special-purpose algorithms for integer complete recourse models;
several of these algorithms [1, 22, 27, 33] use the LP relaxation QLP for bounding
purposes.
Hierarchical planning problems appear in many applications in management sci-
ence. Usually the solution methods consist of solving the problems at the different
levels separately and glue them together. Dempster et al. [6, 7] gave the first math-
ematically rigorous analysis of such a hierarchical planning system. They presented

27

the result on the hierarchical scheduling problem exposed in Section 4. Their result
has been extended to other hierarchical scheduling problems with different types of
machines and common deadlines for the jobs by Stougie [36].
The notion of asymptotic optimality with respect to an optimal clairvoyant
algorithm was introduced by Lenstra et al. [24]. In the same paper the authors
investigated a general framework for the probabilistic analysis of approximation
algorithms for hierarchical planning problems. They show implications between the
various asymptotic quality statements. Applications of this framework on routing
and location problems appeared in [36], where also an survey of the above mentioned
research can be found.
The probabilistic value analysis of combinatorial optimization problems which
are used in the estimates for the second-stage costs form a body of literature on its
own (see for a survey [35]).
Section 5 is extracted from work by Dye et al. [9]. In the same paper a pseudo-
polynomial time dynamic programming algorithm is derived if the number of sce-
narios is fixed. The existence of a fully polynomial time approximation scheme for
this case is open. NP-hardness in the ordinary sense of the deterministic coun-
terpart of the problem was proved in [8]. In the same paper a fully-polynomial
time approximation scheme has been presented for this deterministic problem. All
versions of the problem with multiple resources are strongly NP-hard [10, 8].
The setting of the problem is inspired by an application in telecommunication
dealing with provision of processing based services on a computer network with
distributed processing capabilities [38].
Worst-case performance analysis in stochastic integer programming with dis-
cretely distributed second-stage parameters like the one presented in Section 5 is
an almost unexplored rich research field with many challenging questions.

References
[1] S. Ahmed, M. Tawarmalani, and N.V. Sahinidis. A finite branch and bound
algorithm for two-stage stochastic integer programs. Stochastic Programming
E-Print Series, http://www.speps.info, 2000.
[2] J.R. Birge and F.V. Louveaux. Introduction to Stochastic Programming.
Springer Verlag, New York, 1997.
[3] A. Charnes, W.W. Cooper, and G.H. Symonds. Cost horizons and certainty
equivalents: An approach to stochastic programming of heating oil. Manage-
ment Science, 4:183–195, 1958.

[4] COSP. Stochastic programming community home page. World Wide Web,
http://stoprog.org, 2002.
[5] G.B. Dantzig. Linear programming under uncertainty. Management Science,
1:197–206, 1955.
[6] M.A.H. Dempster, M.L. Fisher, L. Jansen, B.J. Lageweg, J.K. Lenstra, and
A.H.G. Rinnooy Kan. Analytical evaluation of hierarchical planning systems.
Oper. Res., 29:707–716, 1981.
[7] M.A.H. Dempster, M.L. Fisher, L. Jansen, B.J. Lageweg, J.K. Lenstra, and
A.H.G. Rinnooy Kan. Analysis of heuristics for stochastic programming: re-
sults for hierarchical scheduling problems. Math. Oper. Res., 8:525–537, 1983.

28

[8] S. Dye, L. Stougie, and A. Tomasgard. Approximation algorithms and relax-
ations for a service provision problem on a telecommunication network. Work-
ing paper 2-98, Department of Industrial Economics and Technology Manage-
ment, Norwegian University of Science and Technology, N-7034 Trondheim,
Norway, 1998 (to appear in Discrete Applied Mathematics).
[9] S. Dye, L. Stougie, and A. Tomasgard. The stochastic single node service
provision problem. COSOR-Memorandum 99-04, Department of Mathematics
and Computer Science, Technische Universiteit Eindhoven, Eindhoven, 1999.
[10] S. Dye, A. Tomasgard, and S.W. Wallace. Feasibility in transportation net-
works with supply eating arcs. Networks, 31:165–176, 1998.
[11] M. Dyer, R. Kannan, and L. Stougie. A simple randomised algorithm for convex
optimisation. SPOR-Report 2002-05, Dept. of Mathematics and Computer
Science, Eindhoven Technical University, Eindhoven, 2002.
[12] M. Dyer and L. Stougie. Computational complexity of stochastic program-
ming problems. SPOR-Report 2003-09, Dept. of Mathematics and Computer
Science, Eindhoven Technical University, Eindhoven, 2003.
[13] M.R. Garey and D.S. Johnson. Computers and Intractability; A Guide to the
Theory of NP-Completeness. Freeman, New York, 1979.
[14] R.L. Graham. Bounds for certain multiprocessing anomalies. Bell System
Technological Journal, 45:1563–1581, 1966.
[15] P. Kall. Stochastic Linear Programming. Springer-Verlag, Berlin, 1976.
[16] P. Kall and S.W. Wallace. Stochastic Programming. Wiley, Chichester, 1994.
Also available as PDF ﬁle at http://www.unizh.ch/ior/Pages/Deutsch/
Mitglieder/Kall/bib/ka-wal-94.pdf%.
[17] W.K. Klein Haneveld, L. Stougie, and M.H. van der Vlerk. Some theorems on
the convex hull of a function with an application to stochastic integer program-
ming. Discussion Paper TI 94-81, Tinbergen Institute, Amsterdam-Rotterdam,
1994.
[18] W.K. Klein Haneveld, L. Stougie, and M.H. van der Vlerk. An algorithm for
the construction of convex hulls in simple integer recourse programming. Ann.
Oper. Res., 64:67–81, 1996.
[19] W.K. Klein Haneveld, L. Stougie, and M.H. van der Vlerk. Convex ap-
proximations for simple integer recourse models by perturbing the underly-
ing distribution. Research Report 97A19, SOM, University of Groningen,
http://som.rug.nl, 1997.
[20] W.K. Klein Haneveld, L. Stougie, and M.H. van der Vlerk. Convex simple inte-
ger recourse problems. Research Report 97A10, SOM, University of Groningen,
http://som.rug.nl, 1997.
[21] W.K. Klein Haneveld and M.H. van der Vlerk. Stochastic integer programming:
General models and algorithms. Ann. Oper. Res., 85:39–57, 1999.
[22] G. Laporte and F.V. Louveaux. The integer L-shaped method for stochastic
integer programs with complete recourse. Oper. Res. Lett., 13:133–142, 1993.
[23] J. Lawrence. Polytope volume computation. Math. Comput., 57:259–271, 1991.

29

[24] J.K. Lenstra, A.H.G. Rinnooy Kan, and L. Stougie. A framework for the
probabilistic analysis of hierarchical planning systems. Ann. Oper. Res., 1:23–
42, 1984.

[25] F.V. Louveaux and M.H. van der Vlerk. Stochastic programming with simple
integer recourse. Math. Program., 61:301–325, 1993.
[26] Yu. Nesterov and A. Nemirovski. Interior point polynomial methods in Convex
Programming: theory and applications. SIAM Series in Applied Mathematics.
SIAM, 1994.
[27] V.I. Norkin, Yu.M. Ermoliev, and A. Ruszczy´ski. On optimal allocation of
n
indivisibles under uncertainty. Oper. Res., 46(3):381–395, 1998.
[28] C. H. Papadimitriou. Computational complexity. Addison-Wesley, 1994.
[29] A. Pr´kopa. Stochastic Programming. Kluwer Academic Publishers, Dordrecht,
e
1995.
[30] W. R¨misch and R. Schultz. Multistage stochastic integer programs: An in-
o
troduction. In M. Gr¨tchel, S.O. Krumke, and J. Rambau, editors, Online
o
Optimization of Large Scale Systems, pages 581–600. Springer, Berlin, 2001.
[31] A. Ruszczynski and A. Shapiro, editors. Stochastic Programming, volume 10 of
Handbooks in Operations Research and Management Science. North-Holland,
to appear (2003).
[32] R. Schultz. Continuity properties of expectation functions in stochastic integer
programming. Math. Oper. Res., 18:578–589, 1993.
[33] R. Schultz, L. Stougie, and M.H. van der Vlerk. Solving stochastic programs
with complete integer recourse: A framework using Gr¨bner Bases. Math.
o
Program., 83(2):229–252, 1998.
[34] R.J. Serﬂing. Approximation theorems of mathematical statistics. Wiley, New
York, 1980.
[35] J.M. Steele. Probability theory and combinatorial optimization. SIAM CBMS-
NSF Regional Conference Series in Applied Mathematics, 69. SIAM, Philadel-
phia, 1997.
[36] L. Stougie. Design and analysis of algorithms for stochastic integer program-
ming, volume 37 of CWI Tract. Centrum voor Wiskunde en Informatica, Am-
sterdam, 1987.
[37] L. Stougie and M.H. van der Vlerk. Stochastic integer programming. In
M. Dell’Amico, F. Maﬃoli, and S. Martello, editors, Annotated Bibliographies
in Combinatorial Optimization, chapter 9, pages 127–141. Wiley, 1997.
[38] A. Tomasgard, J.A. Audestad, S. Dye, L. Stougie, M.H. van der Vlerk, and
S.W. Wallace. Modelling aspects of distributed processing in telecommunica-
tion networks. Ann. Oper. Res., 82:161–184, 1998.
[39] S.P. Uryasev, editor. Probabilistic constrained optimization. Kluwer Academic
Publishers, Dordrecht, 2000. Methodology and applications.
[40] L. G. Valiant. The complexity of enumeration and reliability problems. SIAM
Journal on Computing, 8:410–421, 1979.

30

[41] M.H. van der Vlerk. Stochastic programming with integer recourse. PhD thesis,
University of Groningen, The Netherlands, 1995.
[42] M.H. van der Vlerk. Stochastic programming bibliography. World Wide Web,
http://mally.eco.rug.nl/spbib.html, 1996-2002.
[43] M.H. van der Vlerk. Convex approximations for complete integer recourse
models. Research Report 02A21, SOM, University of Groningen, http://som.
rug.nl, 2002. Also available as Stochastic Programming E-Print Series 2002–
10, http://www.speps.info.
[44] M.H. van der Vlerk. On multiple simple recourse models. Research Report
02A06, SOM, University of Groningen, http://som.rug.nl, 2002. Also avail-
able as Stochastic Programming E-Print Series 2002–7, http://www.speps.
info.
[45] R.J-B. Wets. Stochastic programming: Solution techniques and approximation
schemes. In A. Bachem, M. Gr¨tschel, and B. Korte, editors, Mathematical
o
Programming, The State of the Art, pages 566–603. Springer-Verlag, Berlin,
1982.

31

Approximation in Stochastic Integer Programming

More Related Content

What's hot (19)

Viewers also liked (7)

Similar to Approximation in Stochastic Integer Programming (20)

More from SSA KPI (20)

Recently uploaded (20)

Approximation in Stochastic Integer Programming