Distributionally Robust Inventory with Advance Contracts

Distributionally Robust Inventory Management
with Advance Purchase Contracts

We propose a distributionally robust inventory model for finding an optimal ordering policy that attains the
minimum worst-case expected total cost. In a classical stochastic setting, this problem is typically addressed
by dynamic programming and is solved by the famous base-stock policy. This approach however crucially
relies on two controversial assumptions: the demands are serially independent and the demand distribution
is perfectly known. Aiming to address these issues, inspired by the seminal work of Scarf (1958), we adopt a
mean-variance ambiguity set that imposes neither the shape of each marginal demand distribution nor their
independence structure, and we focus on the case of advance purchase agreements which are prevalent in the
robust inventory literature and have drawn renewed attention because of the Covid-19 vaccine procurement.
The proposed distributionally robust inventory model provably reduces to a finite conic optimization problem
with however an exponential number of constraints. To gain tractability and to err on the safe side, we
propose two conservative approximations. The first approximation is obtained by recognizing the problem
as an artificial two-stage robust optimization problem and then by restricting each adaptive decision to a
linear decision rule. The second approximation, on the other hand, is obtained by a constraint partitioning
and by upper bounding each resultant maximum sum with a sum of maxima. We then present a progressive
approximation based on a scenario reduction technique to gauge the quality of the proposed conservative
approximations. We prove that this progressive approximation is exact when the inventory problem consists
of two periods, and besides we use it to show that our conservative solutions are still close to being optimal
when the planning horizon is longer. All of our exact and approximate inventory models are expressed as
standard conic programs which allow for the incorporation of additional distributional information. The
extensions are readily obtained by deriving a new cone that corresponds to the restricted ambiguity set and
embedding it in the original problems. We analytically derive the worst-case demand distribution from the
mean-variance ambiguity set and numerically use it to show that our robust inventory policy is more resilient
to the misspecification of the demand distribution than the state-of-the-art non-robust policies.
Keywords : Inventory management; distributionally robust optimization; conic programming.
1. Introduction
We consider an inventory management problem where the manager is monetarily liable for holding
a non-empty stock and/or for not being able to meet the demand in addition to the purchasing
cost.
[Type here]
Soumen Bhatta
2 Distributionally Robust Inventory Management with Advance Purchase Contracts
We assume that the problem consists of T P N periods, and we denote the amount of the product
to be delivered in period t P t1, . . . , T u by xt ě 0. In the simplest case where T “ 1, the problem
contains a single decision variable x1 and is intrinsically similar to a newsvendor problem (Shapiro
et al. 2014). Typically, the solution to the risk-neutral newsvendor problem could be characterized
by the inverse of the cumulative distribution function of the unknown demand ξ˜ . However, the1
reliance on the demand distribution, which itself is difficult to accurately empirically estimate, may
limit the practical relevance of this formula. Scarf (1958) addressed this issue by seeking to find an
optimal x1‹ that performs best in view of the worst-case probability distribution amongst all those
that share the same mean µ and variance σ2. This seminal work has been extended in several direc-
tions. For instance, Gallego and Moon (1993) allowed the newsvendor to make a second purchase
even after the demand is realized, whereas Ben-Tal and Hochman (1976) and Das et al. (2021)
adopted a different ambiguity set of plausible demand distributions by replacing the assumption of
a known variance with the assumption of a known mean absolute deviation and a known nth, n ą 1,
moment of the demand distribution, respectively. Besides, Gallego and Moon (1993) emphasized
the practical relevance of the mean-variance ambiguity set by highlighting that, if the demand
distribution is normal, then the expected value of distributional information (EVDI), which is
defined as the gap between the expected costs attained by the stochastically optimal solution and
the robustly optimal solution due to Scarf (1958), is seemingly negligible. We remark that EVDI
has also been used as a decision making criterion; see e.g. Yue et al. (2006). For the sake of com-
pleteness, while our focus is on the mean-variance ambiguity set, which has an advantage of being
characterized by only a few statistics, there are other non-parametric newsvendor models which
directly incorporate the historical observations of the demand and/or other influential features and
simultaneously account for the demand distribution ambiguity, such as Chen and Xie (2021) and
Lee et al. (2021) as well as Fu et al. (2021).
For a multi-period problem (T ě 2), we again consider a risk-neutral inventory manager whose
objective is still to minimize the total expected cost. When the demands are serially independent
and follow a known distribution, the purchasing costs consist of a fixed and a linearly variable part,
and the holding and the backlogging costs are linear, Scarf (1960) formulated this problem as a
dynamic program and showed that there exists an optimal base-stock policy, which is characterized
by a collection of reordering points s P RT and order-up-to levels S P RT (S ě s). In the absence
of
the fixed ordering cost, it can be imposed without loss of optimality that s “ S, and thus the optimal
policy simplifies; see e.g. Bertsekas (1995). Despite the breakthrough in this characterization of the
optimal policy, the computation is provably difficult when T ă 8 and when the demand distribution
is discrete (Halman et al. 2009), and only a few specific variants of the dynamic inventory problem
admit a reasonably efficient exact solution approach (Veinott and Wagner 1965, Federgruen and
[Type here]
Soumen Bhatta
Distributionally Robust Inventory Management with Advance Purchase Contracts 3
Zipkin 1984, Zheng and Federgruen 1991; all of which assume T “ 8), and otherwise one may need
to resort to a simulation-based method (Fu 1994) or to interpolating the value functions (Halman
et al. 2009).
In reality, the demand distribution is hardly available, but it is required for the computation of
the optimal policy. Incorporating inaccurate information (such as, using an empirical distribution in
lieu of the true demand distribution) in a dynamic program systematically induces error
propagation, which is increasingly difficult to keep track of with larger T ; see e.g. Zhang et al.
(2021). Because of this, several authors have turned towards robust (e.g. Ben-Tal et al. 2005)
and distributionally robust (e.g. See and Sim 2010, Bertsimas et al. 2019) optimization. To gain
tractability, these authors solved their respective variant of the inventory problem using a decision
rule approximation. Though, Lu and Sturt (2022) showed that the majority of linear decision rule
coefficients which arise in different variants of the robust inventory problem optimally vanish.
Solyalı et al. (2016) relied on a different formulation which is inspired by the facility layout and
knapsack problem to improve the computational efficiency in the presence of the fixed ordering
cost. As an alternative solution approach, Xin and Goldberg (2022) assumed that the product
demands are a bounded martingale and derived the optimal policy which can be interpreted as a
base-stock policy where each order- up-to level is dependent on the most recent demand.
“Unless commitment is made, there are only promises and hopes; but no plans.”
— Drucker (1974)
Moreover, despite the undoubted usefulness of a dynamic inventory policy, such as the base-
stock policy, it may be difficult to bring this forth as a supply chain contract. Indeed, Mamani
et al. (2017) argued that, when products have short life cycles (e.g. electronic gadgets and
apparels), the amount of purchasing opportunities may not be aplenty. Likewise, Basciftci et al.
(2021) further argued that, in order to make a better arrangement of their operations and to
minimize disruptions, suppliers typically prefer decisive contractual partners who will make their
requirements (e.g., how much and when the products should be delivered) known early. All of
these arguments show the significance and the relevance of an advance purchase contract, where
the inventory manager is to submit a single purchase order to the supplier with however an option
to specify different delivery dates. To give a contemporary example, the global population is
currently rampaged by Covid-19, and virtually every nation competes to acquire sufficient doses of
protective vaccines. Advance purchase agreements have repeatedly been signed by government
representatives and biotechnology companies; see e.g. Jalelah (2021) and Andres (2022).
In this paper, we will focus on the optimal management of inventory levels which evolve in accor-
dance to the advance purchase agreement, which is primarily suitable in the aforementioned
context
[Type here]
Soumen Bhatta
and when the planning horizon T is only moderately large. When no other distributional informa-
tion besides the support of the uncertain demand is available, Bertsimas and Thiele (2006)
identified a robustly optimal policy and interpreted it as a base-stock policy, and subsequently,
Mamani et al. (2017) determined a closed-form solution of the same problem when the uncertainty
set satisfies a certain symmetrical criterion. It should however be noted that the robust model
adopted in these
two papers is superfluously conservative as it imposes that both the delivery quantities txtuTt“ and
T 1
the per-period worst-case holding and the backlogging cost incurred tptut“ are static variables in
1
the sense that
# # ˜ pxτ ´ ξτ ˜ ¸++
t t
pt “ max ¸ ,
ξ max h y0 τ ´b y0 τ pxτ ´ ξτ q ,
ÿ ÿ
` q `
“1 “1
where y0 denotes the initial inventory level and ph, bq contains the unit cost of holding and backlog-
ging, respectively. Here, the maximization optimizes over all possible realizations of the demands ξ
within the uncertaintyřset, and the inventory manager will seek to minimize the sum of the total
T
purchasing cost and p . In mathematical actuality, the per-period p could and should be
t“ t
1 t
adaptive to the previously revealed demands tξτ ut in the sense that

τ
“1
pt pξ1 , . . . , ξt q “ t pxτ ´ ξτ ˜ t ¸+
# ÿ ÿ
max y0 ` ¸ ,
˜
h τ ´b y0 τ pxτ ´ ξτ q ,
“1 “1
q `
and the inventory manager should instead minimize the summation of the total purchasing cost and
řT
max ξ t“ pt pξ1 , . . . , tξ q as this latter approach will give a single worst-case scenario. We refer our
1
readers to Gorissen and den Hertog (2013) for a more detailed discussion on this subtle intricacy,
which was overcome by tighter approximations from Gorissen and den Hertog (2013) and
Ardestani-
Jaafari and Delage (2016) or by iterative heuristics from Bienstock and O¨ zbay (2008) and Rodrigues
et al. (2021). To our knowledge, these techniques are not directly applicable to the distributionally
robust variant of the problem considered in this paper, which takes a viewpoint that nature
chooses the worst-case demand distribution as opposed to the worst-case demand scenario.
Despite these exciting developments in developing and solving various robust inventory models,
they are ignorant of any possible distributional information, such as moments, and are thus
arguably overly conservative. To alleviate this conservativeness, we will choose to develop a
distributionally robust inventory model in this paper, where in line with the above literature we will
[Type here]
assume that the purchase orders are static decisions. In particular, we aim at minimizing the worst-
case expected total cost of managing an inventory, which is obtained by maximizing the expected
total cost over the planning horizon across all possible demand distributions that are consistent
with a marginal mean µ and variance σ2. We remark that the mean-variance ambiguity set has been
extensively studied in the past, see e.g. Delage and Ye (2010), Zymler et al. (2013), and similar
results are
[Type here]
Soumen Bhatta
also available for a more general moment-based ambiguity set, see e.g. Bertsimas and Popescu
(2005). Besides, we note that in this paper the holding and the backlogging costs for all periods are
modelled faithfully in the sense that they can be adaptive to the unfolding demand observations.
Relying on the duality theory for generalized moment problems, we formulate our distributionally
robust inventory problem as a finite convex optimization problem, however, with an exponential
number of constraints. In order to gain tractability, we propose two conservative approximations of
the problem. The first conservative approximation is due to us interpreting the problem as a
concise two-stage robust optimization problem and then approximating each second-stage
decision with a linear decision rule (Ben-Tal et al. 2004). This approximation is inspired by a
clever trick of Roos et al. (2020) who observed that wait-and-see decisions can be artificially
introduced to a robust optimization problem that does not admit an exact tractable robust
counterpart in order to exploit approximation techniques tailored for solving adaptive robust
problems. The second conservative approximation is due to a meticulous partitioning of constraints
and a series of interchanges between summation and maximization. In the remainder of the paper,
we will henceforth refer to these two approximations as ‘L-conservative’ and ‘Q-conservative’,
respectively. In addition, we also come up with a scenario reduction technique (Hadjiyiannis et al.
2011) that helps us develop a progressive approximation of the proposed model, from which
we can calculate an optimality gap of the conservative solutions that are due to the L- and Q-
approximations. To our knowledge, computing the optimality gap is an issue of significant
importance, but it has been often overlooked in the
robust inventory management literature.
All of our exact and approximate models are representable as standard conic optimization prob-
lems. Adopting the plain vanilla mean-variance ambiguity set, the cone therein is an intersection of
several second-order cones. It is worth mentioning that these models, thanks to their conic
represen- tation, can be readily extended in several dimensions. We give two explicit examples in
the paper:
(i) when the demands are almost surely non-negative (i.e., when the inventory manager
implements a no-return policy) and (ii) when the demands are uncorrelated. In both cases, the
exact model still appears intractable and its tractable approximations, both progressive and
conservative, can be effortlessly derived by changing the cone. We find that extension (ii)
considerably lowers the worst- case total expected cost. This observation is line with our
anticipation as the worst-case demand distribution we derive from the plain vanilla model shows
a strong dependency pattern.
Finally, as the dynamics of the inventory level is a cornerstone of many queuing models, our
techniques could have a far-reaching impact on other operational problems of similar nature, such
[Type here]
as an appointment scheduling problem; see Mak et al. (2015) and Padmanabhan et al. (2021).
These papers aim at providing second-order cone and semidefinite programs for minimizing the
worst-case
[Type here]
Soumen Bhatta
expected waiting and overtime cost when the service times are ambiguously random. This appoint-
ment scheduling problem involves a convex maximization subproblem, which had been thought to
be intractable, until Mak et al. (2015) ingeniously formulated this subproblem as a mixed-integer
linear program with a totally unimodular constraint matrix and is consequently equivalent to its lin-
ear program relaxation. Notwithstanding that we are unable to extend this exact solution method,
our proposed conic approximations are numerically accurate and have the flexibility to be used
with other types of the ambiguity set with relative ease as discussed above.
For succinctness, we summarize the main contributions of the paper below.
• To the best of our knowledge, we are the first to consider the distributionally robust inventory
problem, where our decisions consist of the delivery amount xt for each period t in the plan-
ning horizon t1, . . . , T u, under the ambiguity set that is specified by the marginal means and
variances of the demands. We then show that this problem admits an equivalent
reformulation as a robust second-order cone program with a discrete uncertainty set which is
characterized by the unit holding and backlogging costs.
• We then quantify the optimality gaps of our L- and Q- conservative solutions with the
proposed progressive approximation. Numerically, these gaps appear insignificant, and thus
the studied distributionally robust inventory problem can be solved almost exactly.
Additionally, we give
a rationale behind the progressive approximation by proving that it is exact when T “ 2.
• We demonstrate that both the exact and the approximate reformulations of the distribution-
ally robust inventory problem can be presented compactly using a cone that is representable
as an intersection of several standard second-order cones. By means of a cone replacement,
these
conic representations are highly flexible in that they can incorporate additional distributional
information of the demands, such as non-negativity and pairwise uncorrelatedness.
• Besides, we identify the worst-case demand distribution from the ambiguity set that attains
the maximum expected total cost. We utilize it in a stress test to show that our distributionally
robust inventory policy is more resilient than the stochastic policy that assumes an incorrect
demand distribution. In addition, an adaptive version of our policy which can be obtained by
re-solving the problem repeatedly in a shrinking horizon fashion, is almost on par with the
optimal base-stock policy even when the probability distribution of the demand is known, is
correct, and is serially independent.
The remainder of the paper is structured as follows. Section 2 details the distributionally robust
inventory problem we are interested in as well as several other representations of the problem.
One of them is then used to derive the worst-case distribution that plays an important role in a
stress test analysis. Sections 3 and 4 discuss the proposed conservative and progressive
approximations of the exact model, respectively. The conic extensions of our model are
[Type here]
presented in Section 5,
[Type here]
Soumen Bhatta
and finally Section 6 reports the results of our numerical studies which measure the quality of the
proposed approximations and compare our approach with the state-of-the-art benchmarks.
Notation: We use boldface lowercase letters (e.g. v) and uppercase letters (e.g. M) to represent
vectors and matrices, respectively. Specifically, 1 (0) denotes a vector or a matrix of all ones (zeros),
and 1i denotes the ith canonical basis vector. The set of all real numbers is denoted by R and
its subset of non-negative and strictly positive numbers are denoted by R` and R``, respectively. Sn
represents a set of symmetric matrices in Rnˆn. For any A, B P Sn, A ľ B indicates that A ´
B
is positive semidefinite, and ⟨A, B⟩ denotes their trace scalar product. Besides these blackboard
letters, we use capital caligraphic letters to denote other sets (including cones). If K is a cone, then
K‹ denotes its dual. For any vector v P Rn, diagpvq is a diagonal matrix in Sn that has elements of
v sitting on its main diagonal, and #pv, mq counts the number of occurrences of m in v, and v m ,
1 ď m ď n, is a subvector of v containing its first m elements, i.e., v m “ pv1 , . . . , vm qJ .
Throughout, a division by zero is allowed and is defined as
$
’& `8 if a ą 0,
a “ 0 if a “
0 ’%
0,´8 if a ă 0.
2. Mean-variance inventory model and its finite representations

We consider an uncapacitated inventory system that stores a single product for sale, and we
denote by c, h and b the unit ordering, holding and backlogging costs, respectively. We assume that
these cost parameters are positive and that they are time-invariant, although these are
assumptions that could readily be lifted. Supposing that the inventory manager is risk-neutral, that
the demand distribution is perfectly known, and that any unfulfilled order could be indefinitely
backlogged, facing the uncertain demands ξ “ pξ1 , . . . , ξT qJ , the inventory manager would be
interested in minimizing the total expected cost, i.e., solving
«
f
ÿT ´
minimize cxt `max hyt pξt q, ´byt pξt q ( f¯
t“
1
(1)
subject to x P X , yt : Rt ÞÑ R @t P t1, . . . , T u
yt pξt q “ yt´1 pξ t´1 q ` xt ´ ξt P-a.s. @t P t1, . . . , T u,
where x “ px1 , . . . , xT qJ and y “ py1 , . . . , yT qJ collect all decision variables representing the
order quantities and the end-of-period inventory levels, respectively, and y0 P R denotes the initial
inventory level, which is given. For the ease of exposition, we assume that all orders are
instantaneous, i.e., the lead times are zero. Besides, we follow Bertsimas and Thiele (2006) and
Mamani et al.
[Type here]
(2017) in imposing that x is chosen here-and-now and the feasible set X Ď is uncertainty-free.
`
RT
Unless stated otherwise, X “ R`T .
[Type here]
Soumen Bhatta
In the actual reality, not only ξ is uncertain but its probability distribution P, which needs to
be empirically estimated, is ambiguous. Therefore, to hedge against such an estimation risk, the
inventory manager may choose to robustify Problem (1) and solve
« f
ÿT ćx `max hy pξ t q, ´by pξt q f¯
minimize max EP t t t (
PPP
t“
1
t (2)
subject to x P X , yt : R ÞÑ R @t P t1, . . . , T
u
yt pξ t q “ yt´1 pξt´1 q ` xt ´ ξt @ξ P RT , @t P t1, . . . , T u.
Note that in both Problems (1) and (2), the decision variables yt pξt q could be easily eliminated by
replacing each yt pξt q with y0 ř t
τ pxτ ´ ξτ q. Therefore, although x is decided here-and-now, the
` “1
per-period holding and backlogging costs depend on the realization of past demands. In contrast,
Mamani et al. (2017) ignored such randomness in exchange for a closed-form solution.
We henceforth express Problem (2) compactly as
minimize f pxq
subject to x P X ,
where the objective function f characterizes the worst-case expected cost incurred from x, i.e.,
«T ˜ t pxτ ´ ξτ ˜ t pxτ ´ ξτ
# ÿ ÿ
f px q max EP ÿ y0 ` ¸ , ¸+¸ff . (3)
PPP ˜
“ t“ cxt `max h τ ´b y0 τ
1 “1 “1
q ` q
We note that f is convex but not necessarily easy to evaluate because it involves solving an infinite
optimization problem over the probability distribution P P P. Following the seminal work of Scarf
(1958), we choose to work with the following mean-variance ambiguity set:
(
T T
P “ P P M` pR q : P pξ P R q “ 1, EP rξt s “ µ, EPtrξ 2 s 2
“µ `σ 2
@t “ 1, . . . , T , (4)
where M` denotes the cone of all non-negative measures supported on the input set, because
these two summary statistics can be inferred relatively easily from empirical data. This choice of
P is equivalent to us assuming that the demands ξ have a stationary mean µ ą 0 and a stationary
variance σ2 ą 0 but nothing else is known. We remark that this basic ambiguity set allows demands
to be negative to capture possible product returns. When returns are disallowed (perhaps because
of a hygienic reason or the product’s short shelf life), we will consider a restricted ambiguity set in
Section 5.
Under this P, we will first show that the worst-case expected total cost f pxq, with x fixed, can
be determined by solving a finite-optimization problem or its dual. Before presenting these results,
we introduce a cone Kt, t P t1, . . . , T u, that is instrumental to our subsequent analyses.
+
` β
[Type here]
Kt “ #
ÿ
t 2
: 4α τ
.
γτ
pα, β, γq P R` ˆR t ě τ “1
ˆRt
[Type here]
Soumen Bhatta
Note that Kt is a proper cone, that is, it is closed, solid, pointed and convex. Specifically, we can
express Kt as an intersection of multiple second-order cones, i.e.,
Kt # t +
` ` ÿ
“ : Dθ P Rt , α
pα, β, γq P R` ˆR t ě τ
“1
θτ , }pβτ , θτ ´ γτ qJ }2 ď θτ ` γτ @τ “ 1, . . . , t .
t
ˆR
Moreover, we also introduce a discrete set E “ th, ´buT and note that its cardinality is 2T .
We will use it to capture the 2 T underlying linear lower bounds of the total holding and backlogging
cost, which appears as a part of the objective function of Problems (1) and (2), i.e.,
ÿT ( ˜ ¸
T t
ÿe ÿ .
max hyt pξt q, ´byt pξt q “ max t y0 ` pxτ ´ ξτ
t“ ePE
1 t“ τ q
1 “1
This exponential complexity arises as a result of that, at the end of each period t P t1, . . . , T u,
the inventory manager will have to pay for either the holding cost if the end-of-period inventory
level
is positive or the backlogging cost otherwise.
THEOREM 1. We have that
f pxq “ minimize c1J x ` α ` µ1J β ` pµ2 `

σ 2 q1J γ
(5)
T T
subject to α P R, β P R , γ P R
pαpx, eq, βpeq, γq ľKT 0 @e P E ,

where αpx, eq “ α ´ y 1Je řT řT
0 x eτ , β tpeq “ ` eτ , for all x P R`T .
řT t“
´
1
t
β t
τ τ
“t “t
We can interpret Problem (5) as a robust conic program with an auxiliary uncertain vector e that
belongs to a discrete uncertainty set E.
Proof of Theorem 1. Starting from its definition in (3), we may characterize the function f as
the optimal objective value of a generalized moment problem, i.e.,
f pxq ´ c1Jx “ T
ÿ ˜ t pxτ ´ ξτ
ÿ
max ż max et ¸ Ppdξq
ePE
t“1 τ
y0 “1
T
R ` q
T
s.t. P P M` pR q
[Type here]
ż Ppdξq “ 1 (6)
żRT ξt Ppdξq “ µ @t “ 1, . . . , T
żRT ξ2tPpdξq “ µ2 ` σ2 @t “ 1, . . . ,
T.
By assigning the dual variables α to the normalization constraint, each βt to each of the first-order
R T
moment constraints, and each γt to each of the second-order moment constraints and invoking
[Type here]
Soumen Bhatta
strong duality due to Shapiro (2001, Proposition 3.4), we find a minimization problem that attains
the same optimal objective value as the above maximization problem:
min α ` µ1J β ` pµ2 ` σ 2 q1J γ

s.t. α P R, β P RT , γ P RT
T T T ˜ t pxτ ´ ξτ
ÿ ÿ ÿ ÿ
α` βtξt ` γtξ t 2
ě ¸ @e P E @ξ P RT .
t“ t“ y0 τ
1 1 “1
et ` q
t“1
Note that
« the inequality constraint
ff in the above minimization problem could be rearranged as
T t
«T T t T
J ÿ ÿ ÿ ÿ ÿ
α ´ y01 e et xτ ÿ βtξt ` et ff ` γtξ2t ě 0 @e P E @ξ P RT
t“1 τ t“ t“1 τ t“
“1 1 “1 ξτ 1
´ ÿ ` ÿ
ðñ αpx, eq T T
βt peqξt ` γtξ2 ě 0 @e P E @ξ P RT .
` t“ t“
1 1 t
For any fixed e P E, as the above quadratic inequality holds for every ξ P RT , it follows that
γt must be non-negative. Furthermore, if γt vanishes, then so does βt peq. Denoting by T the
index set tt P t1, . . . , T u : γt ą 0u, we can re-express the considered quadratic inequality as
ÿ ÿ
αpx, eq ` βt peqξt ` γttξ 2 ě 0 @ξ P RT
«ˆ ff
tPT tPT
ðñ αpx, eq ˙ 2
γt ξt ` t ´ βt peq2 ě 0 @ξ P RT
` β peq
tÿ 2γ t 4γt2
ÿ
PT
ðñ 4αpx, eq βt peq2 ÿ
T βt peq
2
ě γ “ γ ,
t t
tPT t“1
and the proof is completed by recalling the definition of the cone KT . l

REMARK 1. As a by-product of the proof of Theorem 1, γ ě 0 is imposed as an explicit constraint
in Problem (5). As γt, t P t1, . . . , T u, is a dual variable of the constraint EPrξt 2s “ µ2 ` σ2, the
optimal objective value of Problem (5) coincides with the worst-case expected total cost across all
probability distributions of the demand from the relaxed ambiguity set
[Type here]
(
1 T T 2 2 2
P “ P P M` pR q : P pξ P R q “ 1, EP rξt s “ µ, EPtrξ s ď µ ` σ @t “ 1, . . . , T Ą P.
In other words, the worst-case distribution achieves maximum variances. This observation is con-
sistent with various moment-based distributionally robust optimization problems; see e.g. Delage
and Ye (2010) and Rujeerapaiboon et al. (2016). l
REMARK 2. The proof of Theorem 1 reveals that, for any fixed y0 P R and x +
P X at optimality,
# T t T T t T
ÿ ÿ ÿ ÿ ÿ ÿ γ ξ2
α “ max max y01Je et xτ ´ βtξt ´ et ξτ ´ t
t , (7)
ePE ξ
` t“1 τ t“ t“1 τ t“
1 1
“1 “1
[Type here]
Soumen Bhatta
where the optimization problem over e P E on the right-hand side constitutes a convex maximiza-
tion problem. Mak et al. (2015) and Padmanabhan et al. (2021) recently encountered a similar
optimization problem when analyzing a robust appointment scheduling problem, where they have
α max # T max gt pet , ξt , βt , # T max g pe , ξ , β ,

t t t t
ePE
ÿ ξt “ max ÿ ξt
“ + e + (8)
PconvpE
t“ t“
1 1
γt q γtq
for suitably defined functions g1, . . . , gT that are convex in e. Subsequently, they express (8) as a
mixed-integer binary program with a totally unimodular constraint matrix so that the binarity
requirement can be lifted. We cannot however directly express (7) as an instance of (8) because
řT
of the term t“ et řτt “1 ξτ ř t“1 T
ξt řτ et. Although we can use a variable transformation ηt “
T
ř 1 “ “t
T
τ et, t P t1, . . . , T u, and T “ 0 and replace each with ηt ´ so that
“t `1
η et ηt`1
α max T +
η
“ #
ÿt“ max gst pηt , ξt , βt , γt q : ´b ď ηt ´ ηt`1 ď `h, ηT `1 “ 0
1 ξt
for suitable functions g1, . . . , gT that are convex in e, such a transformation appears to prevent us
s
from following the steps in Mak et al. (2015) to obtain a polynomial-sized exact reformulation of
Problem (5). s l
In the next step, we will describe another way of computing f pxq based on the dual reformu-
lation of Problem (5). As one would expect, there is certain semblance between Problem (6) and
the optimization problem in Theorem 2 below. Subsequently, an analysis of this latter formulation,
which is still exact, will be provided. This will eventually lead us to the characterization of the worst-
case probability distribution within the ambiguity set P.

ÿ ˆ ˙¸
J
f pxq “ maximize c1 x ` T ˜ t
ÿ ÿ
αpeq et xτ ´ βτ peq
αpeq
e :
αpeq‰0
t“1 y0 τ
“1
`
PE
subj
ect
to
` for all x P RT . α:E
ÞÑ
R,
β:E
[Type here]
ÞÑ RT , γ : E , βpeq, γpeq ľKT‹ 0 @e P E
T
ÞÑ R
αpeq “ 1
` (9)
α ePE
ÿ
p ÿ βpeq “ µ1
˘ePE
e ÿ
q γpeq “ pµ2 ` σ2q1,
ePE
We note that Problems (5) and (9) are similarly intractable because they contains an exponential
number of constraints and decision variables, respectively, as E is exponentially-sized.
[Type here]
Soumen Bhatta
Proof of Theorem 2. Leveraging the exact characterization of the worst-case expected cost f pxq
` to the conic constraint
from Theorem 1, we assign a set of dual variables αpeq, βpeq, γpeq P
KT‹
˘
in Problem (5) that is associated with each e P E:
pαpx, eq, βpeq, γq ľKT 0

ðñ T T T T eτ , . . . , βT ` eT , γ1, γ2, . . . ,
ÿ ÿ ÿ ÿ
˜ α ´ y01Je xt eτ , β1 eτ , β2 ¸ ľKT 0,
´ t“1 τ ` τ ` τ
“t “1 “2
γT
and obtain the following dual formulation of Problem (5):

e ˜
T T T T
ÿ ÿ ÿ ÿ
f pxq “ maximize c1 x ` ÿ αpeq
J
x t eτ ´ÿ β t peq eτ
y01Je ` t“1 τ ePE τ “t
“t ¸ t“1
PE
subject to α : E ÞÑ R, β : E ÞÑ RT , γ : E ÞÑ RT
`
αpeq, βpeq, γpeq ľKT‹ 0 @e P E
ÿ
˘ αpeq “ 1
ePE
ÿ
βpeq “ µ1
ePE
ÿ
γpeq “ pµ2 ` σ2q1.
ePE
By noting that the objective function of the above optimization problem, when evaluated at any
feasible pα, β, γq, could be re-expressed as
ÿ ˜ ÿ
c1Jx ` ˆ
T
ÿ ˙ T T
ÿ T
ÿ
J tÿ
αpeq y01 e ` β ¸´ β t peq eτ
e :
αpeq‰0 t“ xt αpeq e : t“
αpeq“0 τ “t
ÿ 1 peq τ
eτ 1
´ “t
˙¸
T t
J PE
“ c1 x ÿ ÿˆ βτ peq PE
αpeq ˜ et
` eÿ:
αpeq‰0 ˜ t“1 τ αpeq
J “1 x τ ´
y1e`
ÿ0 ˙¸¸
PE T t
“ c1Jx ÿ β τ p eq
αpeq et ˜ ˆ
,
` e : t“1 τ αpeq
y0 “1 xτ ´
αpeq‰0
PE
where the first equality holds because βpeq “ 0 whenever αpeq “ 0 (see Proposition 1 just
below), the proof is completed. l
[Type here]
The characterization of Problem (9) involves the cone that is dual to KT . In order to solve
KT‹
this problem, we thus need the exact description of the dual cone, which is given next.
PROPOSITION 1. We have that

# +
2
‹ βτ
Kt pα, β, γq P R` ˆR t
` : α ě max . (10)
τPt1,...,t u
“ ˆRt τ
[Type here]
Soumen Bhatta
Proof. We denote the set given on the right-hand side of (10) by Kt1 . So as to show that Kt‹ “
Kt1 , it suffices to prove that Kt1 Ď Kt‹ (step 1) and Kt‹ Ď Kt1 (step 2).
For the first part, we consider an arbitrary pα, β, γq P Kt and pα, β, γq P Kt1 . It follows that
˜
1 ÿ t
β 2 ÿ t
¸
γτ ě
τ τ
` γ ě ´ ÿ tβ τ “ ´βJβ,
ÿt 2 τ τ
β
γ β2β γ
αα ` γJγ α
4 γττ `
ě τ “1
τ τ
4γτ γτ
“1 τ 2 “1 τ
β
τ “1
where the last inequality holds because ` γτ γτ ě |βτ βτ |, @τ . Indeed, due to the relationship
τ τ
4γτ γτ
between arithmetic and geometric means this latter inequality holds whenever γτ and γτ are both
strictly positive. If γτ “ 0 then βτ “ 0, and if γτ “ 0 then βγ “ 0. In both cases, the same inequality
is ‹still valid. The above derivation implies that contains inside. Conversely, consider an
Kt Kt1
arbitrary pα, β, γq P Kt‹ . It holds that
αα ` βJβ ` γJγ ě 0 @pα, β, γq P Kt.
As p1, 0, 0q P Kt, it necessarily follows from the above inequality that α ě 0. Similarly as p0, 0, 1τ q
P Kt, @τ P t1, . . . , tu, it necessarily follows that γτ ě 0. Moreover, for any fixed τ P t1, . . . , tu, if βτ ‰ 0,
then γ τ ą 0. Suppose otherwise for the sake of a contradiction that there exists pα, β, γq P Kt‹
such that βτ ‰ 0 and γτ “ 0, then as
˜ 2 ¸
1τ βτ 1 τ P Kt @α ą 0,
4α
α, ´βτ ,
it must follow that αα ´ β2τ ě 0 for any α ą 0. This resultant inequality however cannot hold
true because α could be arbitrarily close to zero. Assuming now that βτ ‰ 0 and γτ ą 0, we hence
find ˜
2
2γ τ 1τ , 1τ
γτ
, P Kt,
β¸
2
β
´
τ τ
and consequently,
γ2 2γ β2
τ
α ´ τ τ ` γ τ ě 0 ùñ α ě τ .
2
βτ γτ
β βτ
As α ě 0, the same conclusion can also be reached even if βτ “ 0. Since the above inequality holds
for any τ P t1, . . . , tu, we may conclude that Kt‹ is contained in Kt1 . Combining both halves
of the argument yields the desired result. l
We close this section by visualizing K2 and K2‹ in Figure 1. The three axes of Figure 1 (left) repre-
sent the value of β1, β2 and (the minimum) γ1 `γ2, respectively, when α is set to one. Algebraically,
this surface is obtained from the following observation that
[Type here] ˙
2 1
γ`γ
ě p|β | `|β | ˆ 2 β22 ´1 ě 1 p|β | `|β |q 2
q β
1 2 1 2 1 2
γ1 ` γ2 4
for any pβ, γq such that p1, β, γq P K2. Analogously, the three axes of Figure 1 (right) represent
the value of β1, β2 and (the minimum) γ1 ` γ2, respectively, when α is set to one. The surface
2
shown 2
can be obtained by noting that γ1 ` ě β1 β2 for any pβ, γq such that p1, β, γq P K2‹ .
γ2 `
[Type here]
Soumen Bhatta
1 2
0.8
1.5
0.6
1
0.4
0.5
0.2
0 0
1 1
0.5 1 0.5 1
0 0.5 0 0.5
0 0
-0.5 -0.5
-0.5 -0.5
-1 -1 -1 -1
‹
Figure 1 Visualization of K2 X pt1u ˆ R4 q and K2 X pt1u ˆ R4 q.
2.1. Worst-case demand distribution

This section is devoted to the determination of a probability distribution from the ambiguity set P
that solves the maximization problem on the right-hand side of (3) for any fixed inventory decision
x P X. The construction of the worst-case distribution is often used in a stress test (Bertsimas et al.
2010), and the main ingredients are the dual formulation derived in Theorem 2 and the following
technical lemmas which are pertain to the optimal values of the decision variables of Problem (9).
‹ ‹
LEMMA 1. There exists an optimal solution pα‹ , β , γ ‹ q of Problem (9) such that β peq “ γ ‹ peq “
0, @e P E : α‹ peq “ 0.
:
Proof. Denote by pα:, β , γ:q an arbitrary optimal solution of Problem (9) and by E 0 a subset
of E which contains all scenarios e such that α:peq “ 0. Note that E 0 must be strictly contained in
E and there must exist a scenario e: P EzE 0 such that α:pe:q is strictly positive. Then, we construct
‹ ‹ :
a new solution pα‹ , β , γ ‹ q with α‹ “ α: , β “ β , and
ř
γ ‹ peq “
γ : peq ` e1 PE 0 γ : pe1 q if e “ e: ,
0 if e P E 0,
% ’
γ : peq if e P E zpE 0 Y te: uq.
ř ř ‹ and that
It is readily seen that e
PE γ ‹ peq “ e PE γ : peq “ pµ2 ` σ 2 q1, pα‹ pe: q, β pe: q, γ ‹ pe: qq P
KT‹
all other constraints remain satisfied by this newly constructed solution. Since γ does not appear in
‹
the objective function of Problem (9), we may thus conclude that pα‹ , β , γ ‹ q is feasible and optimal.
‹
The proof is then completed by noting that β peq automatically vanishes whenever α‹ peq does. l
‹
LEMMA 2. Any optimal solution of pα‹ , β , γ ‹ q of Problem (9) satisfies
t
“1
˜
¸¸ ´ ‹ q ax ¸¸ t ˜ xτ ´ ¸¸+
‹
ÿ τ ÿ # ˜ τ
˜ 0 y ` β peq “ y
β peq0 ` β peq
xτ α‹ pe α‹ peq
et τ
m h τ “1
[Type here]
, ´b ˜ ` t ˜ ´ ‹ q
τ ÿ
‹
y0 τ “1 xτ α pe
for any t P t1, . . . , T u and any e P E such that α‹ peq ą 0.
[Type here]
Soumen Bhatta
For any fixed e P E, where α‹ pe q ą 0, if ‹

ξτ β peq
“ τ‹ , @τ , then the inventory level at the end of
α peq
period
ř
t is given by y0 ` t ´ ¯
‹
τ β τ peq
“1 xτ ´ , and the right-hand side of the above equation
α‹peq
characterizes
either the holding or the backlogging cost that would incur. Lemma 2 therefore provides a simpler
expression of such a cost. This lemma is particularly useful as the worst-case distribution, which
we will later construct, is a mixture distribution comprising of components whose mean demands
‹
βτ peq
at period τ , @τ , are equal to α‹peq
.
Proof of Lemma 2. Consider any scenario e P E such that α‹ peq ą 0. To simplify the
exposition, we abbreviate (and slightly abuse the notation) the dual optimal solution α‹ peq,
‹ ‹
β peq, γ ‹ peq by α‹ , β , γ ‹ , respectively.1 Besides, we denote by pα‹ , β ‹ , γ ‹ q an optimal
solution of the primal problem (5). The KKT optimality condition of this primal-dual pair includes
a complementary slackness condition which necessarily holds and reads
˜ ‹ T
ÿ T
ÿ ¸ T
ÿ ˜ T
ÿ ¸ T
ÿ
α ´ y01Je xt ` ‹t βt‹ ` γ ‹t γt‹ “ 0
‹ t“1 τ t“ τ t“
α ´ “t eτ 1 β ` “t eτ 1
ÿ ÿ
‹
ðñ α‹ ˜ T T T T
ÿ ¸ ÿ
xt ´ βt eτ “ α ‹ α ‹ ` pγ ‹ qJ γ ‹
J t“1 τ t“ τ ‹
y01 e ` ˜ eτ
“t 1 “t
` pβ‹ qJ β
ÿ
ðñ α‹ ÿ
T ˜ t ‹
et xτ ´ τ “ α‹ α‹ ` pγ ‹ qJ γ ‹ .
t“1 y0 τ α‹
¸¸ ‹
“1 ` pβ‹ qJ β
` β
For any time period t1 P t1, . . . , T u, it thus follows that

˜ t1 ˜ ‹
ÿ ¸¸
α‹ et1 xτ ´ τ β
y0 τ α‹ ˜ ¸¸
t ‹
`‹ ‹ ‹
J ÿ ÿ β
“ α α “1 ` pγ ‹ q γ ‹ ´
J ˜
` pβ‹ q β α‹ et (11)
1
t‰t τ
y0 ` xτ ´
τ α‹
“1
ÿ eτ
‹
¨ t T
ÿ ÿ ˛ ÿT
β T
ÿ
ÿ et xt ˛ γ ‹t γt‹ .
¨
˚ ‹ ´ t“1 τ t
‹
t“ τ t t“
‹ τ 1 τ 1
“α ˝α ‰t
1 eτ ‰t
1
‹
˚ ‚
´ y0 ‰t1 ` ˝β‹
`
`
[Type here]
‚ t
Next, we collect ¨ ˛
the coefficients of the dual solution on the right-hand side of the above
J equation
ÿ ÿT ÿ ÿ ÿ
‹
t‰t1 eτ , γ1‹ , . . . , γT‹ ‚
ζ“ α ´ ‹
et xt ě
eτ , β1‹ eτ , . . . , βT
ě ‹ ě
t“1 τ t τ 1 τ T
y0 ´ τ 1 ` τ 1 ` τ 1
and observe that ζ can be expressed as a convex combination of the
‰t ‰t ‰t following two vectors
ζ“ T T T T
¸J
ÿ ÿ ‹ ÿ ‹ ÿ
eτ , . . . , βT ` eT , γ1 , γ2 , . . . , γT‹
‹ ‹ ‹
‹ J
α ´ y01 e xt eτ , β 1 eτ , β 2
˜
t“1 τ τ τ
` `
T “t “1 “2 ¸J
‹ ÿ ÿ ÿ ÿ ‹ ‹ ‹ ‹
´α ´ xt T T T e ,...,β
τ T ` eT , γ1 , γ2 , . . . , γT ,
ζ“ t“1
eτ , β1‹ eτ , β2‹
τ τ τ
˜ J “t ` “1 ` “2
y01 e ´
1
That is we drop their dependency on e.
[Type here]
Soumen Bhatta
where the scenarios e and e are characterized by
et “ et “ et @t P t1, . . . , T uztt1 u and et1 “ `h as well as et1 “ ´b.
In particular, we can express ζ as h`bb ζ `h`bh ζ . Note that the feasibility of pα‹ , β ‹ , γ ‹ q in view
of
Problem (5) implies that ζ, ζ P KT . As KT constitutes a convex cone, we have that ζ P KT and
˜ 1 ‹
t
ÿ τ
e t1 xτ ´ ě0
y0 τ ˜ α‹
¸¸
“1
` β
‹
because pα‹ , β , γ ‹ q P KT‹ and because of (11). The above inequality and the fact that et1 P t`h,
´bu, together, finally completes the proof. l
Next, we propose a candidate for the worst-case probability distribution that attains the maxi-
mum expected cost for a given inventory decision x P X. In particular, we solve Problem (9) and
‹
determine an optimal solution pα‹ , β , γ ‹ q that satisfies the requirement of Lemma 1. Note that if
a solver gives a solution that does not meet the requirement of Lemma 1, we can use the construc-
tion provided in the proof of Lemma 1 to determine one that does. Then, for any e P E such that
α‹ peq ‰ 0, we construct a probability distribution Pe,ε P M` pRT q, ε P p0, 1q, via
` ˘
Pe,ε ξ “ ξe,ε “ 1 ´ ε and Pe,ε
´ ¯
e,ε b e,ε
ξ“ ξ “ ε, ‹
(12)
‹
where the atoms ξe,ε P RT and ξ P RT are
α peqγ peq
chosen as ´pβ peqqt
‹ ‹
β peq ´ c ε
t
e,ε “ ‹
ξt α peq 1´
2
t
@t “ 1, . . . , T,
ε ‹
α peq
‹ b t
ξ e,ε β peq c @t “ 1, . . . , T.
“ 1´ ε α‹ peqγ ‹t peq
t ‹
t ‹
` ´ pβ peqq2
b α peq ε α‹ peq
‹
Note that α‹ peqγ t‹ peq ´ pβt peqq2 is well-defined because term inside is non-negative, which is due
to the characterization of KT‹ from Proposition 1. Next, we construct a mixture distribution
Pε “ ÿ
α‹ peqPe,ε (13)
ePE:α‹peqą0
[Type here]
and assert that this demand distribution attains the worst-case inventory cost in (3) as ε Ó 0.
THEOREM 3. The demand distribution Pε, ε P p0, 1q, has the following properties:
• Pε P P for all ε P p0, 1q,
” !
• lim εÓ0 P řTt“ ř ¯ řt ¯)ı
E ε 1 max h 0 ` t 0` τ
´ τ τ τ τ τ
“1 px ´ ξ q , ´b “1 px ´ ξ q “ f pxq ´ c1J x.
y ´
y
[Type here]
Soumen Bhatta
Proof. To establish that the mixture distribution Pε belongs to the ambiguity set P, we first
compute from (12) the first two moments of each probability distribution component and find
‹
‹
β e “ 2‰ γ
tp q
EP rξt s “ ‹
e,ε
and EPe,ε ξ t “ @t P t1, . . . , T u.
α peq
peq
t
α‹ peq
Subsequently, from (13) we obtain
ÿ ÿ ÿ ÿ
‹ ‹
“ ‰
EPε rξts “ β peq “ β peq “ µ and 2
EPε ξ “ ‹
γ peq “ γ ‹ peq “ µ2 ` σ 2
p qą
t ePE t t t t
eP E: ePp E:
qą ePE
‹ ‹
α e 0 α e 0
for any t P t1, . . . , T u because of Lemma 1, and we can conclude that Pε P P.
By construction, it similarly follows that
«T ˜ t pxτ ´ ξτ ˜ t ¸+ff
max ÿ ÿ
EPe,ε ÿ # ¸ ,
t“ τ τ
1 y0 “1 ´b y0 “1 pxτ ´ ξτ q
t
h ` q `¸ ˜ ¸+
# ˜ e,ε e,ε
ÿ
T
t ÿ
ÿ
“ p1 ´
εq max h y0 ` px ,´b y0 `
t“ τ q τ ´ ξ¸τ ˜
τ pxτ ´ ¸+
ξτ q `
1 # ˜ ÿ
“1
ÿ
“1
ÿT e,ε
t e,ε t
ε max h y0 ` pxτ ´ ξ τ , y0 ` pxτ ´ ξ τ q .

t“ τ τ
1 “1
q ´b “1
e,ε
Since limεÓ0 εξτ “ 0 for all τ P t1, . . . , T u, we further find
«T t pxτ ´ ξτ ˜ t ¸+ff
# ÿ ÿ
lim EPe,ε ÿ max y0 ` ¸ ,
εÓ0 t“ ˜ τ τ
1
h “1 ´b y0 “1 pxτ ´ ξτ q
t
q ¸ ˜` ¸+
# ˜ e,ε e,ε
T ÿ ¸¸+
ÿ t ÿ
ÿ
β peq
pxτ ´ ξ τ q
li
t“
max h y0 ` τ px
q
τ ´ ξ
τ , y0 ` τ
1 “1 “1
˜ ˜ ´b ˜ ˜
“m #
εÓ0
ÿ max
h `y0 ÿ ´ xτ y
, ` 0 ´ xτ
T t
¸¸
‹ ´b t ‹
“ÿ τ β peq τ
“1 τ “1 τ
t“
1 α‹ peq α‹ peq
[Type here]
˜ ˜
¸¸
`y0 ´ xτ
β peq
ÿ
T t ‹
,
τ
“
τ
“1
α‹ peq
et
t“1
where the last equality is due to Lemma 2. Utilizing the fact that Pε is a mixture of Pe,ε and invoking
Theorem 2 complete the proof. t“ l
1
Under the constructed worst-case distribution, the demands tξtuT appear to be highly depen-
dent across time periods, and we ascertain this claim in Section 6. An ambiguity-averse inventory
[Type here]
manager is therefore advised to take extra caution when it comes to stochastic independence, which
is typically assumed (e.g. Scarf 1960).
3. Efficiently solvable conservative approximation
Theorems 1 and 2 ensure that we can evaluate f pxq by solving a finite convex optimization problem.
[Type here]
Nevertheless, the difficulty arises when T gets large as the uncertainty set E is exponential-sized. In
[Type here]
Soumen Bhatta
this section, we provide two distinct ways of approximating f pxq from above. The first
approximation is a result of interpreting Problem (5) as an artificial two-stage robust optimization
problem and restricting the second-stage decisions using affine adaptation, whereas the second is due
to constraint categorization and a series of interchanges between maximization and summation.
To this end, we first leverage Theorem 1 to provide a new characterization of f pxq as the optimal
objective value of a two-stage robust optimization problem.

f pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ
subject to α P R, β P RT , γ P RT , φ : R2T ÞÑ RT , ψ : R2T ÞÑ RT
` `
J J J J
α ` β u ` γ v ě b1 φpu, vq ` h1 ψpu, vq
(14)
@pu, vq : p1, u, vq P
t
KT‹
y0 ` φt pu, vq ´ ψt pu, vq
ÿ
“ τ puτ ´ xτ q @t “ 1, . . . ,
T
“1
for all x P R`T .
@pu, vq : p1, u, vq P KT‹ ,
In a nutshell, as soon revealed Theorem 4 is similar to Theorem 2 in that they both characterize
the worst-case expected total cost f pxq by applying a duality technique to Problem (5).
Observe
that the dual cone KT‹ shows up in both of the theorems.
Proof of Theorem 4. Considering Problem (5), which characterizes f pxq, we could re-express
the robust constraints therein as:
min max
ePE tω : pαpx, eq, βpeq, γq ľKT pω, 0, 0qu ě 0.
ωPR
Then, by dualizing the maximization problem on the left-hand side of this inequality, we attain:
! )
min minT
ePE u,vPR αpx, eq ` uJ βpeq ` γ J v : p1, u, vq ľKT‹ 0 ě 0
! ( )
min J J
ð u,vPRT γ v ` min αpx, eq ` u βpeq : p1, u, vq ľK ‹ 0 ě 0
T (15)
ePE
ñ min " *
ð (
u,vPRT
ñ γ J v ` convp
min αpx, eq ` uJ βpeq : p1, u, vq ľKT‹ 0
eP Eq
ě 0,
where the second equivalence is because αpx, eq and βpeq are linear in e and because any solvable
linear program has a vertex solution. Next, we focus on the inner minimization problem, which we
could expand as
minimize
[Type here]
« T T « βJu ` T T
ÿ ÿ ÿ ÿ ff
xt ff ut
α ´ y01Je ´ t“1 τ “t
eτ ` t“1 τ
eτ
“t
subject to e ě ´b1, e ď h1
[Type here]
Soumen Bhatta
thanks to definition of αpx, eq, βpeq and E. Assigning the dual variables φ P
RT and ψ P to the `
T
R `
two constraints of this linear program, respectively, and rewriting the objective function as
allows us to derive the dual: T ˜ t ¸

J
ÿ
α` βu et
` t“1 ý0 ` τ puτ ´ xτ q
“1
maximize α ` β u ´ b1 φ ´ h1J ψ
J J
subject to φ P RT , ψ P RT
` `
t
φ t ´ ψt “ ř τ puτ ´ xτ q @t “ 1, . . . , T.
` “1
ý0
Replacing the inner minimization problem in (15) with its dual finally completes the proof. l
We consider Problem (14) as a two-stage robust optimization that contains first-stage decision
variables pα, β, γq as well as second-stage decision variables pφ, ψq whose values could depend
on yet another auxiliary uncertain vector pu, vq which belongs to a projection of the dual
cone KT‹ . Although the number of constraints in Problem (14) grows linearly with T , the main
difficulty of solving it lies in the fact that it is a two-stage problem with wait-and-see decision
variables φ and ψ. Our next result describes how to use a linear decision rule approximation
(Ben-Tal et al. 2004)
in conjunction with Theorem 4 to derive the following upper bound on f pxq:
f L pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ
subject to α P R, β P RT , γ P RT , κ P R2T `1 , pπt , π u , π v q P R ˆ RT ˆ RT @t “ 1, . . . , T
` t t
u v
pπt ´ κt , π , π q ľK 0 @t “ 1, . . . , T
t t T
˜ ¸
t u J
πt ` y0 ÿ xτ ´ κT `t , π ´ p1, . . . , 1, 0, . . . , 0q , ľK
v
` π
0 @t “ 1, . . . ,
τ “1 t
˜ T
T
t
T T
ÿ ÿ
α ´ hTy0 ´ xt pT ´ t ` 1q ´ pb ` πt ´ κ2T `1,
h t“
hq t“
1
ÿ
1 ¸
π
β ´ pb ` hq T
J
T
ÿ v
u
πt ` hpT, T ´ 1, . . . , 1q , γ ´ pb ` ľK 0,
t“ t“
1 hq 1
t T
where the vector p1, . . . , 1, 0, . . . , 0qJ in the second line of constraints has t components equal to
[Type here]
one and T ´ t components equal to zero. Perhaps interestingly, unlike Problem (14) this
minimization problem involves only the primal cone KT . From a computational perspective, we
only use the dual
cone KT‹ to construct the worst-case demand distribution; see Section 2.1.
We emphasize that the linear decision rule approximation adopted here is different than those
used in the earlier papers, such as Ben-Tal et al. (2005), See and Sim (2010), and Bertsimas et al.
(2019), because unlike theirs our main operational decisions x P X are to be chosen here and now.
We only apply the affine restriction to the artificial wait-and-see decisions that appear in Prob-
lem (14), which we introduce as a means to capture the worst-case expected total cost f pxq.
[Type here]
Soumen Bhatta
PROPOSITION 2 (L-conservative approximation). We have that f pxq ď f Lpxq for all x P R`T
. Proof. Our first step entails simplifying the exact characterization of f pxq from Theorem 4 by
řt
replacing the adaptive decisions ψ tpu, vq by y0 ` φt pu, vq ` τ pxτ ´ uτ q, 1 ď t ď T . Hence,
“1
f pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ
subject to α P R, β P RT , γ P RT , φ : R2T ÞÑ RT
φt pu, vq ě 0 @pu, vq : p1, u, vq P KT‹ , t “ 1, . . . , T

ÿt
y0 ` φt pu, vq pxτ ´ uτ q ě 0 @pu, vq : p1, u, vq P KT‹ , t “ 1, . .
t
τ
` “1 . , T T
ÿ ÿT
ÿ pxτ ´ uτ q
α ` β J u ` γ J v ě hT y0 ` pb ` hq φt pu, vq ` h
t“1 t“1 τ “1
@pu, vq : p1, u, vq P KT‹ .
We subsequently obtain a conservative approximation by restricting each φt pu, vq to an affine form:

φt pu, vq “ πt ` uJ πt u ` v J πt v for some linear decision rule coefficients πt P R, πt u P RT , πt v P RT . As
a result, we have the following upper bound of f pxq that involves only here-and-now
decisions. minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ
subject to α P R, β P RT , γ P RT , pπt , π u
t ,π
v T T
t q P R ˆ R ˆ R @t “ 1, . . . , T
πt ` uJπu ` vJπv ě 0 @pu, vq : p1, u, vq P K‹ , t “ 1, . . . , T

t t T
t t
πt ` y0 ` ÿ xτ ` uJπu ´ ÿ uτ ` vJπv ě 0 @pu, vq : p1, u, vq P K‹ , t “ 1, . . . , T
τ “1 t t T
τ “1
ÿT ˜ ¸
T T
ÿ J ÿ π
α
h ´ hTy0 ´ x
hqt pT ´ t ` 1q ´ pb ` πt ` u β
hq´ pb `
t“ t“ tu
1 ˜ 1 ¸ t“1
ÿT
` ÿ ut pT ´ t ` 1q ` γ ´ pb ` π ě tv0 @pu, vq : p1, u, vq P
t“
hT vJ hq 1 KT‹ .
t“
1
Leveraging the primal-dual pair involving cone KT familiar from the proof of Theorem 4, we can
derive the robust counterpart of the first constraint as follows.
πt ` uJπut ` vJπvt ě 0 @pu, vq : p1, u, vq P KT‹

(
J tu J tv ‹
ðñ min
u,v πt ` u π ` v π : p1, u, vq P K T
ě0
ðñ max tκt : pπt ´ κt , π u , π v q P KT u ě 0
κt t t
u v
ðñ pπt ´ κt , πt , πt q 0 Dκt P R`.
T
ľK
[Type here]
Applying the same routine for the remaining constraints, which are similarly linear in the conically
constrained uncertain parameters u and v, completes the proof. l
[Type here]
Soumen Bhatta
Next, we derive an alternative conservative approximation of f pxq which is obtained by

constraint categorization and a series of interchanges between maximization and summation.
f Q pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ 2 q1J γ
subject to α P R, β P RT , γ P RT , λ0,..., λT P RT
α ` by0pT ´ kq ´ hy0k ě 1Jλk @k “ 0, . . .
(16)
´, T ¯
řT
λk ´t x t eqkτ, β tpeqk q, γt ľK 0@k “ 0, . . . , T @t “ 1, . . . , T
τ 1
“t
´ ¯
tř t τ τ t t K1
k T k k
λ ´ xk Tep ,
β pep
k
“t q, γ ľ 0 @k “ 0, . . . , T @t “ 1, . . . , T,
where the auxiliary vectors P and P T are defined through
J
ep “ p´b, . . . k, ´b, `lohoo,oo.m.R.o,ooèq
oohnq R and
k
eq “ p`looho,oo.m. .o,ooòohn, ´b, . . . , ´bqJ .
k components k components
The correctness of this upper bound is validated in the next theorem.

THEOREM 5 (Q-conservative approximation). We have that f pxq ď f Qpxq for all x P RT .`
Proof. It is sufficient to show that, for any pα, β, γ, λ0, . . . , λT q that is feasible in Problem
(16),
pα, β, γq is feasible in Problem (5). To achieve this, we first observe that, for each k P t0, . . . , T u
and t P t1, . . . , T u, the two K1-inequalities of Problem (16) imply that
#
k
ÿ β ppe ÿ
4λ ě max k 2 T k 2 T +
q βt pe q qt
` eq
t
γt k
t
` k
4xt τ
τ, γ
4x
τ ep τ .
“t t “t
Next, fixing the index k P t0, . . . , T u, summing up the above inequality over t P t1, . . . , T u and
finally using the remaining (linear) constraint of Problem (16) results in
#
4α ` 4by0pT ´ kq ´ 4hy0k T T
k
k 2 T
k +
ÿ β pek q2 ÿ ÿ
ě max q t ` eq βt ppe `
t“ γt τ q γt τ
1
# 4xt “t
τ, 4xt “t epτ
T
T +
“ ÿ max βt peq2
`
ÿ
t“
1 ePE γt τ
# 4xt “t eτ : #pe, hq “ k
ÿ +
ÿT 2 T
¸
˜ βt peq `
ě max
ePE γt
t“ 4xt τ eτ : #pe, hq “ k ,
1 “t
2 řT
βtpeq
where the equality holds because the quadratic expression `4x t τ “t is convex in e
γt ře
T
τ τ
τ
“t
and hence its maximum value over the discrete feasible set te P E : #pe, hq “ ku is attained
when
e “ ep
k
or e “keq , and the last inequality follows as a result of the interchange between
summation
[Type here]
and maximization. A slight rearrangement of terms further yields
# T T T : #pe, hq “
ÿ ÿ ÿ βt peq2
min xt eτ ´ + ě 0 @k “ 0, . . . , T
ePE γt
4α ´ 4y01Je ´ t“1 τ “t t“
1
4 k
ðñ pαpx, eq, βpeq, γq ľKT 0 @e P E : #pe, hq “ k @k “ 0, . . . , T
ðñ pαpx, eq, βpeq, γq ľKT 0 @e P E

and correspondingly the feasibility of pα, β, γq in Problem in (5). The proof is now completed. l
[Type here]
Soumen Bhatta
4. Efficiently solvable progressive approximation

Previously, we have proposed two conservative approximations for f pxq, x P`RT , with however
no way of quantifying how accurate they are. In this section, we will complement our earlier results
with a progressive approximation of f pxq. This approximation is based on a scenario reduction
technique for a robust optimization problem, whereby we choose to remove purportedly less
important scenarios from the uncertainty set. Tractability of the approximation is ensured if the
number of retained scenarios grows linearly with the problem size.
Specifically, the proposed lower bound is
f pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ 2 q1J γ
subject to α P R, β P RT , γ P RT
` ˘
αpx, eqk q, βpeqk q, γ ľKT 0 @k “ 0, . . . , T,
`
where, for any x P RT and e P E, αpx, eq and βpeq are defined as in Theorem 1.
T
THEOREM 6 (Progressive approximation). We have that f pxq ě f pxq for all x P R
` .
Proof. This is an immediate consequence of Theorem 1 as we can interpret f pxq as the

optimal objective value of Problem (5) with the uncertain vector e P E and f pxq as the optimal
objective
0
value of the same robust program with, however, a smaller uncertainty set teq , . . . ,Teq u Ď E. l
To motivate the rationale behind this progressive approximation, we consider a stylised example
of two-period distributionally robust uncapacitated inventory problem where the inventory is ini-
tially empty. Thanks to Theorem 1, the minimization problem minxě0 f pxq could be expanded to
minimize cpx1 ` x2 q ` α ` µpβ1 ` β2 q ` pµ2 ` σ 2 qpγ1 ` γ2 q (17a)
subject to x1 P R`, x2 P R`, α P R, β1 P R, β2 P R, γ1 P R`, γ2 P R` (17b)
α ´ 2hx1 ´ hx2 ´ pβ1 ` 2hq2 {4γ1 ´ pβ2 ` hq2 {4γ2 ě (17c)

0
α ` 2bx1 ` bx2 ´ pβ1 ´ 2bq2 {4γ1 ´ pβ2 ´ bq2 {4γ2 ě 0 (17d)
α ` pb ´ hqx1 ´ hx2 ´ pβ1 ` h ´ bq2 {4γ1 ´ pβ2 ` hq2 {4γ2 ě 0 (17e)
α ` pb ´ hqx1 ` bx2 ´ pβ1 ` h ´ bq2 {4γ1 ´ pβ2 ´ bq2 {4γ2 ě 0, (17f)

J 2 J 0 J
where the four constraints correspond to e “ ph, hq “ eq , e “ p´b, ´bq “ eq , e “ p´b, hq and
J 1 J
e
ph,“´bq “ eq , respectively. Theorem 6 suggests that the scenario e “ p´b, hq is least
important.
Indeed, we show below that the shadow price of the constraint (17e) is equal to zero.
PROPOSITION 3. The dual of Problem (17) is given by
[Type here]
á ¯
?
maximize ph ` bqσ 4q1 q2 ` q1 q3 ` q1 q4 ` q2 q3 ` q2 q4 ` q1 q2 ` q1 q4 ` q2 q3 ` q3 q4 ´
[Type here]
Soumen Bhatta
µ p3hq1 ´ 3bq2 ` p2h ´ bqq3 ` ph ´ 2bqq4 q (18a)
subject to q1, q2, q3, q4 ě 0 (18b)
q1 ` q2 ` q3 ` q4 “ 1 (18c)
c ` 2hq1 ´ 2bq2 ` ph ´ bqq3 ` ph ´ bqq4 ě 0 (18d)
c ` hq1 ´ bq2 ` hq3 ´ bq4 ě 0. (18e)
Proof. This dual formulation is obtained by assigning Lagrangian multipliers q1, q2, q3, q4 ě 0 to
the constraints (17c)–(17f), respectively. The exact derivation is lengthy and is hence relegated to
the appendix. l
THEOREM 7. Any optimal solution q ‹ of Problem (18) satisfies q3‹ “ 0.
As Problem (17) constitutes a convex minimization problem and as q3 represents the shadow price
of the constraint (17e). The theorem implies that the constraint (17e) is redundant and can thus
be removed without any impact on the optimal objective value. It could nonetheless be possible
that, after the constraint elimination, the relaxed problem admits multiple optimal solutions and
some of those may not be feasible in Problem (17).
Proof of Theorem 7. We divide our analysis into two cases depending on the value of q4‹. Our
goal is to show that q3‹ “ 0.
Case 1 (q4‹ ą 0): Suppose for the sake of a contradiction that q3‹ ą 0 and consider an
alternative
solution q : “ pq1‹ ` δ, q2‹ ` δ, q3‹ ´ δ, q4‹ ´ δq for some sufficiently small δ ą 0 such that q :
remains in the nonnegative orthant. It can be directly verified that q: is feasible in Problem (18).
Besides, it
holds that
: : : : : :
q1 q2 ` q1 q4 ` q2 q3 “ pq1‹ ` δqpq2‹ ` δq ` pq1‹ ` δqpq4‹ ´ δq ` pq2‹ ` δqpq3‹ ´ δq ` pq3‹ ´
: :
` q3 q4
δqpq4‹ ´ δq “ q1‹ q2‹ ` q1‹ q4‹ ` q2‹ q3‹ ` q3‹ q4‹
and that
: : : : : : : :
4q1 q2 ` q1 q3 ` q1 q4 ` q2 q3 ` “ 4pq ‹ ` δqpq ‹ ` δq ` pq ‹ ` δqpq ‹ ´ δq ` pq ‹ ` δqpq ‹ ´ δq`
1 2 1 3 1 4
: :
q2 q4
pq2 ` δqpq3 ´ δq ` pq2 ` δqpq4‹
‹ ‹ ‹
´ δq “ 4q1‹ q2‹ ` q1‹ q3‹ ` q1‹ q4‹ ` q2‹ q3‹ ` q2‹ q4‹ ` 2δ
as well as that
“ 3hpq1‹ ` δq ´ 3bpq2‹ ` δq ` p2h ´ bqpq3‹ ´ δq ` ph ´
: : :
3hq1 ´ 3bq2 ` p2h ´ bqq3 ` ph ´
: 2bqpq4‹ ´ δq “ 3hq1‹ ´ 3bq2‹ ` p2h ´ bqq3‹ ` ph ´ 2bqq4‹ .
2bqq4
Together, the above three observations imply that q: attains an objective function value that is
[Type here]
strictly larger than q‹, which in turns contradicts with the supposed optimality of q‹ and therefore
renders q3‹ ą 0 impossible to materialize.
[Type here]
Soumen Bhatta
Case 2 (q4‹ “ 0): Next, we consider the other case with vanishing q4‹ . Under this
condition, Problem (18) simplifies to a trivariate optimization problem:
á ¯
?
maximize ph ` bqσ 4q1 q2 ` q1 q3 ` q2 q3 ` q1 q2 ` q2 q3 ´ µ p3hq1 ´ 3bq2 ` p2h ´ bqq3 q
subject to q1, q2, q3 ě 0
q1 ` q2 ` q3 “ 1
c ` 2hq1 ´ 2bq2 ` ph ´ bqq3 ě 0
c ` hq1 ´ bq2 ` hq3 ě 0
and subsequently to a bivariate optimization problem (barring the shifting and positive scaling of
the objective function):
á a ¯
maximize σ q1 ` q2 ´ pq1 ´ q2 q2 ` q2 p1 ´ q2 q ` µ p2q2 ´ q1 q (19a)
subject to q1, q2 ě 0 (19b)
q1 ` q2 ď 1 (19c)
q2 ´ q1 c` h´
(19d)
bh`
ď
c` h b
q2 ď . (19e)
h` b
We remark that Problem (19) constitutes a concave maximization problem because square root
function is concave and increasing and a quadratic function with a non-positive leading coefficient
is also concave. In view of Problem (19), our original goal to establish that q3‹ “ 0 is
tantamount to showing that the constraint (19c) is binding at optimality. We begin by arguing
that the constraint (19e) cannot be binding because otherwise it would hold that
q‹ ` q‹
c` h´ 2p c ` h q c ` h > 1,
b ‹
ě 2q2 ´ ´
“ b ´
1 2
h` h` h` b
b b
where the first inequality is due to (19d) and hence a contradiction. Therefore, the constraint (19e)
can be omitted without any loss of optimality. Suppose next that the constraint (19d) is binding,
Problem (19) simplifies (barring again the shifting and positive scaling of the objective function)
to the following univariate concave
d maximization problem:
c` h´ b a
maximize 2q2 ´ c ` h ´ ˆ
bh` 2 ` q2 p1 ´ q2 q ` (20)
b ´ rq2
` ˙ h` b
`
c h b´
subject to ˆ q h` b
˙ 2
[Type here] ď
ď
2h ` c 2h ` 2b
,
µ
where r is a positive constant equal to . In Problem (20), the lower bound of q2 ensures that both
σ
q1 and q2 are nonnegative, whereas the upper bound ensures that q1 ` q2 ď 1, which is an explicit
[Type here]
Soumen Bhatta
constraint in Problem (19). To show that (19c) is a binding constraint, it suffices to show that the
2h`c
upper bound
2h`2b is attained by q2‹, and to achieve this we will show that the objective function
of (20) is increasing in q2 over its admissible range. Observe that Problem (20) can only be feasible
when 2b ě c, which we shall assume, and the derivative of the objective function denoted by g(20) is:
Bg(20) 1 1 ´ 2q2
Bq2 “ ` `r
c`h´b `c`h´b ˘2
b 2q2 ´ h`b h`b a
2 q2p1 ´
´
h` b q2 q
“ a ` a1 ´2q2 ` r.
2q ph ` bq ´ pc ` h ´ bqp2h ` cq 2 q p1 ´ q q
2
2 2 2
2h`c
When q2 “ 2h`2b ,
Bg(20) h` b pb ´ h ´ cq{ph ` bq
ˇ “ a ` a `
2
r
2h`2b
Bq2 ˇq “ “ a
h` b
à
b´ h´ c
`
ph ` bqp2h ` cq ´ pc ` h ´ bqp2h ` cq rp2h ` cqp2b ´ cq{ph ` bq2
2h`c
p2h ` cqp2b ´ cq p2h ` cqp2b ´ cq

c 2b ´ c
2h ` c r
> “0.
”` ˘ ı
c`h´b `
By its concavity, g(20) is increasing in q2 P h`b , 2h`c
2h`2b and as a result q‹ 3indeed vanishes.
Henceforth, we can assume without any loss of optimality that constraints (19d) and (19e) are
not binding. If at optimality the constraint (19c) is also not binding (i.e., if q3‹ ą 0), it must be
possible to identify q‹ that is optimal in Problem (19) by solving
a a
maximize q1 ` q2 ´ pq1 ´ q2 q ` q2 p1 ´ q2 q ` r p2q2 ´ q1 q
2
subject to 0 ď q1 ď 1, 0 ď q2 ď (21)
1.
Note that the square roots are well-defined over the feasible region as q1 ` q2 ě |q1 ´ q2 | ě pq1
´ q2 q2 . By a familiar argument that we use to analyze Problem (19) earlier, Problem (21) remains a
concave
maximization problem. Then, by expressing its objective function g(21) as
a a
g(21) pq1 , q2 q “ q1 p1 ´ q1 q ` q2 p1 ´ q2 q ` 2q1 q2 ` q2 p1 ´ q2 q ` r p2q2 ´ q1 q
and noting that the mapping q2 ÞÑ q2p1 ´ q2q is strictly increasing over the interval r0, 12 s, it
necessarily holds that q‹ ě 1 . Next, we derive the partial derivative of g(21) with respect to q1:
2 2 Bg(21)
[Type here] 2
1 ´2q1 `2q2 “ a ´ r.
Bq1 ` ´ pq1 ´ q q2
2 q1
q2
[Type here]
Soumen Bhatta
When q2 is chosen optimally as q2‹ , we can determine the values of q1 such that the above
partial derivative vanishes as follows:
` ˘
p1 ´ 2q1 ` 2q2‹ q2 “ 4r 2 q1 ` q2‹ ´ pq1 ´ q2 ‹ q2
2
ðñ 4pq ‹ ´ ‹ 2 ‹ 2 ‹ 2
2 q1 q ` 4pq 2´ q1 q ` 1 “ 8r q2 ´ 4r pq2 ´ q1 q ´ 4r pq
2
‹
´ q1 q2
ðñ 4p1 ` r2 qpq ‹ ´ q1 q2 ` 4p1 ` r2 qpq ‹ ´ q1 q ` 1 ´ 8r 2 q ‹ “ 0
‹ 2 c 2 2
ðñ
´ ‹
1 1 r p1 ` 8q2 q
2
q2 q‹1 “ ´ ˘ 1 ` r2
ðñ 2 “ 2 c
1
1 r2 p1 ` 8q2‹ q
q1 q2 ` ¯
2 1 ` r2 .
The above two values2 of q1 together with the boundary points (i.e., 0 and 1) are the candidate values
for q‹. However, as q‹ ě 1 and as q‹ ` q‹ ď 1, we can rule out the values of q‹ that are greater
1 2 2 1 2 b r2 p1`8q q 1
1 ‹ :
than . Hence, we find that q P t0, q u, where q “ q ` ´ : ‹ 1 2
‹
. We can now divide
2 1 1 1 2 2 2 1`r2
1 our
remaining analysis into two subcases.
:
Case 2a (q1‹ “ 0): In this case, as g(21) is concave in q1 we have q1 ď 0. We will assume here
: :
that q1 ă 0 and treat the remaining possibility of q1‹ “ q1 “ 0 in Case 2b. Here, Problem
(21) reduces to a univariate optimization problem in q2. By a slight abuse of notation, we denote
the resulting objective g(21) p0, q2 q by g(21) pq2 q, and we find
Bg(21) 1 ´ 2q
2
a
Bq2 “ ` 2r.
q p1 ´ q q
2 2
This gradient simply evaluates to 2r ą 0 when q2 “ 1 2, and it evaluates to a negative number when
` ˘
1
q2 approaches one from below. By the intermediate value theorem, there exists a value of q2 P 2, 1
such that the gradient vanishes. Since g(21) is concave in q2, it then follows that
1´ c
2q2‹ ´ 1
a 2q2 ‹ r2 ‹
` 2r “ 0 ùñ r “ 2 2
q ‹ p12 ´ q ‹ q a
2
“ 2q2 ´
2 q ‹ p1 ´ q ‹ q 1
1.` r
2
As a result, we obtain that

1 1 a
q2‹ ` ´ 2q2‹ ´ 1q 1 ` 8q2 ‹ “ 1q : ă 0 ùñ p2q2‹ ` 1q2 ă p2q2‹ ´ 1q2 p1 ` 8q 2 ‹ q.
p 2 2
The latter inequality implies that 32pq2‹ q2 pq2‹ ´ 1q ą 0, which is an impossible consequence in
view of the feasibility of Problem (21).
:
Case 2b (q1‹ “ q1 ): Regardless of the value of q1 , we find
Bg(21) 1 `2q1 ´2q2 1 ´2q2
“ a ` a ` 2r.
Bq 2
2 q ` q ´ pq ´ q q2 2 1 q p12 ´ q 1q 2 2 2
` ˘
Similarly to the above case, for any fixed q1 ď 12, there exists a value of q2 P 1
,1 such that this
gradient vanishes, and thus at optimality, we have
1 ` 2q1‹ ´ 2q2‹
[Type here]
2
1 ´ 2q2‹
a ` a ` 2r “ 0. (22)
2 q1‹ ` q2‹ ´ pq1‹ ´ 2q ‹ q2 2 2q ‹ p1 ´ 2q ‹ q
[Type here]
Soumen Bhatta
:
Note that, as q1‹ “ q1 , it follows that ˜
c c c
a 1 1 ‹
r p1 ` 8q q 1
2
r ¸p1 `
2
1 ` 8q‹
2 2
2
8q ‹ q
e ´ 1´ 2
“
2 q ‹ ` q ‹ ´ pq ‹ ´ q ‹ q2 “ 2 2q ‹ ` ´
1 2 1 2 2
2 2 1 ` r2 4 1` 1 ` r2
c r2
and that
‹ ‹
1 ` 8q2‹
1 ` 2q ´ 2q “ 2´ r .
1 2
1 ` r2
Substituting
d these observations into (22), we obtain
1` 1 ´‹ 2q2‹ ´ c
2q2 1 r2 ‹
r 2
2 ùñ ‹ ` a 2 ` r “ 0 ùñ ră a 2
1 ` 8q ‹2 ‹
2 q p1 ´ q q 2 q ‹ p12 ´ q ‹ q
2
ă 2q2 ´ 1,
1 ` r2
b
r 2
where the rightmost implication holds because the mapping r ÞÑ 1` 2r
is increasing in r P r0, 8q.
As the inequality q1 ` q2 ď 1 is an explicit constraint of Problem (20), which is supposed to be
non-binding, it must strictly
c hold when q1 “ q1‹ and q2 “ q2‹ . Hence,
‹ 1 1 r2 p1 ` 8q2‹ q ‹ ‹ ‹ 2 ‹ 2 ‹
2q2 ` ´ ă p2q2 ´ 1q p1 ` 8q2q.

“ q1 ` q2 ă 1 ùñ p4q2 ´
2 1q
1 ` r2
The latter can hold only if 4q2‹ pq2‹ ´ 1qp8q2‹ ´ 3q ą 0 and hence a contradiction.
We thus conclude that, in all cases, it is impossible for q3‹ to be a strictly positive number and
complete the proof. l
As a progressive approximation, its optimal solution x‹ is not necessarily optimal in Problem (2),
i.e., f px‹ q could be strictly larger than f px‹ q, and this unfounded optimism could jeopardize
the operations of the inventory manager. This statement holds even when T “ 2 because minxPX f pxq
involves optimizing over px, α, β, γq and it may admit multiple optimal solutions and some of those
may not be feasible in minxPX f pxq. We thus propose to use a conservative approximation which is
due to Theorem 5 or Proposition 2 to obtain an approximate optimal solution x‹ and approximate
the associated worst-case cost f px‹ q by rf px‹ q, f px‹ qs as well as upper bound the optimality
gap f px‹ q ´ f px‹ q by f px‹ q ´ f px‹ q. To avoid overloading the notation, we will henceforth use
x‹L and x‹Q to denote the L- and the Q-conservative solutions, respectively, in the remainder of
the paper.
REMARK 3. It is possible to extend our inventory management model and its approximations
to cater for fixed ordering costs which are incurred whenever a strictly positive order is placed.
Denot- ing by k P R`` a fixed ordering cost and introducing a decision vector z P t0, 1uT , the worst-
case expected cost minimization problem could be written down as
[Type here]
minimize k1Jz ` f pxq
subject to x P X , z P t0, 1uT
x ď M z,
[Type here]
Soumen Bhatta
for a sufficiently large M ą 0. As all of the previously proposed approximation methods are con-
cerned with estimating f pxq for any fixed x P X, they could still be utilized here. However, due
to the presence of z, the corresponding exact and approximation models will become mixed-integer
second-order cone programs. Although non-convex, to some extent these problems can numerically
be handled by off-the-shelf solvers, such as CPLEX and Gurobi. l
5. Plug-and-play conic extensions

We now consider how to reduce the conservatism of the proposed distributionally robust inventory
model and its approximations by injecting additional distributional information besides the mean
and the variance. The results in this section primarily rely on the conic representations of our base
and approximate models, which are the consequences of Theorem 1, Proposition 2, and Theorem
6. The main idea is that when the ambiguity set is shrunk, a new cone is derived, and it is then
used in lieu of KT in our earlier results. We refer to this cone replacement as a plug-and-play
feature of the studied distributionally robust inventory model.
5.1. Non-negative support

When returns are not allowed, naturally demands are non-negative with probability one, and to
incorporate this information, we may consider the shrunk ambiguity set
` ˘ (
P : “ P P M` pR`T q : P ξ P R
` T “ 1, EP rξt s “ µ, EPtrξ 2 s “ µ2 ` σ 2 @t “ 1, . . . , T ,
and the less conservative cost function

«T ˜ # ˜ t pxτ ´ ξτ ˜ t ¸+¸ff
: ÿ ÿ
f px q max ÿ ¸ ,
EP“ t“ cxt `max h y0 τ ´b y0 τ pxτ ´ ξτ q .
:
PPP 1 “1 “1
` q `
It turns out that the majority of main results (Theorems 1, 4 and 6 as well as Proposition 2) readily
extend by replacing the original cone KT therein with
:
K “ pα, β, γq P R` ˆ RT ˆ RT : Dpδ, θq P RT ˆ RT , α ě 1J θ, pθt , βt ´ δt , γt q P K1 @t “ 1, . . . , T
T ` ` `
(
.
:
One could verify that KT Ă KT , which could be perceived as a sign of reduced conservatism.
It is
:
a routine exercise to verify that the cone KT is proper.

f : pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ 2 q1J γ
[Type here]
pαpx, eq, βpeq, γq ľ : 0 @e P E ,
K T
where αpx, eq and βpeq are defined as in Theorem 1, for all x P R`T .
[Type here]
Soumen Bhatta
Proof. From an argument that is widely parallel to the proof of Theorem 1, we have that
f : pxq ´ c1J x “ minimize α ` µ1J β ` pµ2 ` σ 2 q1J γ
T T
αpx, eq `ÿ βt peqξt ` ÿ γt ξ 2 ě 0 @e P E @ξ P RT .
t“1 t `
t“1
For each fixed e P E, the robust constraint holds for all ξ P RT iff there exists θ P RT such that
`
J 2
αpx, eq ě 1 θ and θt ` βt peqξt ` γt ξ t ě 0 @ξt ě 0 @t “ 1, . . . , T.
Note that it immediately holds that each θt must be non-negative for otherwise the robust
quadratic constraint fails to hold when ξt “ 0. Furthermore, by the virtue of S-lemma, each of these
robust quadratic constraints holds iff there exists a δt ě 0, t “ 1, . . . , T , such that
„  „  "
γt 1
βt peq 0 2 1
γt ě 0, θt ě 0
1
2
ľ Sδ2 t 1 ðñ 2
2 βt peq θt ` 4γ t θt ě pβt peq ´ δt q
0
2
" γ ě 0, θ ě 0 t t
ðñ pθt , βt peq ´ δt , γt q ľK1 0.

:
By leveraging the definition of KT , we finally complete the proof. l
:
It appears in our experiments however that the difference between min xPX f pxq and minxPX f pxq is
negligible. Hence, the non-negativity of the demands ξ may not be the most significant
requirement to be embedded in the ambiguity set.
5.2. Uncorrelated demands

Next, we consider a scenario where the demands are uncorrelated, which itself could be an approx-
imation of independent demands. Analogously to Section 5.1, we focus on the following restricted
ambiguity set
$ P pξ P RT q “ 1,
PK “ EP rξt s “ µ, @t “ 1, . . . , ,
’
TRT : /
’& p q /
t
PP .
EP rξs ξt s “ µ2 , @ps, tq : 1 ď s ă t ď
T
M `2
µ EPrξ2s “ ` σ2, @t “ 1, . . . , T /
-
« ˜ # pxτ ´ ξτ ˜ ¸+¸ff
T t t
ÿ ÿ
K
f px q max ÿ ¸ ,
h y0 `
EP “ t“ cxt `max τ ´b y0 τ pxτ ´ ξτ q .
PPPK 1 ˜ “1 “1
q `
For the subsequent results, we use the following cone
[Type here]
" „  *
KKT α, β, γ, q P `
ˆR T ˆR T ˆST diagpγq ` Θ 1
β
`
“ pΘ R : 1
2 ľ0 .
βJ
2
α
It follows from the Schur Complement Lemma that pα, β, γ, 0q P KTK iff pα, β, γq P KT . Besides,
one could readily show that KTK is proper.
[Type here]
Soumen Bhatta
f K pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ ` µ2 1J Θ1
subject to α P R, β P RT , γ P RT , Θ P ST
pαpx, eq, βpeq, γ, Θq ľK K 0
T
@e P
E diagpΘq “ 0,
where αpx, eq and βpeq are defined as in Theorem 1, for all x P R`T .
Proof. From an argument that is widely parallel to the proof of Theorem 1, we have that
f K pxq ´ c1J x “ minimize α ` µ1J β ` pµ2 ` σ2 q1J γ ` µ2 ÿ θst

1ďsătďT
subject to α P R, β P RT , γ P RT , Θ P RTˆT
ÿ
ÿ
T
ÿ
T
αpx, eq ` t
βt peqξt ` γtξ2 ` θstξsξt ě 0 @e P E @ξ P RT ,
1ďs ă
t“ t“
1 1
tďT
where each θst is a dual variable assigned to the constraint that ensures ξs and ξt are uncorrelated.
Note that θst with s ě t does not enter the above optimization problem.
Without any loss of optimality, we replace each θst, s ă t, with 2θst and assign θss “ 0 for all s
and θst “ θts for all pairs ps, tq such that s ą t resulting in an equivalent problem of the form
f K pxq ´ c1J x “ minimize α ` µ1J β ` pµ2 ` σ 2 q1J γ ` µ2 1J Θ1
subject to α P R, β P RT , γ P RT , Θ P ST
ÿ ÿ
αpx, eq ` T T
βt peqξt ` γtξ2 ` ξJΘξ ě 0 @e P E @ξ P RT
t“ t“ t
1 1
diagpΘq “ 0.
For any fixed e P E, it can be recognized that the robust quadratic constraint (which has to
hold for all ξ P RT ) is equivalent to
„ J „ „  „ 
ξ diagpγq ` Θ 12 βpeq ξ 0 ξ T diagpγq ` Θ 1
2 βpeq
0.
1 1
βpeqJ
αpx, 1 ě @ PR ðñ 1
βpeqJ αpx, ľ
2 2
eq eq
[Type here]
This latest observation together with the definition of KTK completes the proof. l
The above (exact) formulation is a correlation-aware counterpart of Theorem 1, which readily leads
us to a progressive approximation akin to Theorem 6. For the conservative approximation, we can
have the following results, which respectively are the counterparts of Theorem 4 and Proposition 2.
[Type here]
Soumen Bhatta

f K pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ ` µ2 1J Θ1
2 2
subject to α P R, β P RT , γ P RT , Θ P ST , φ : R2T`T ÞÑ RT , ψ : R2T`T ÞÑ RT
` `
α ` β J u ` γ J v ` ⟨Θ, W⟩ ě b1J φpu, v, Wq ` h1J ψpu, v, Wq
@pu, v, Wq : p1, u, v, Wq P (23)

t
ÿ
y0 ` φt pu, v, Wq ´ ψt pu, v, Wq pKTK q‹ puτ ´ xτ q @t “ 1, . .
τ
“
.,T @pu, v, Wq : p1, u, v,
“1
Wq P pKTK q‹
diagpΘq “ 0.
Proof. The proof widely parallels to that of Theorem 4 and is therefore omitted for brevity. l
PROPOSITION 4. We have that

f K pxq ď minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ ` µ2 1J Θ1
subject to α P R, β P RT , γ P RT , Θ P ST , κ P `R2T`1,
pπt , πtu , πtv , Πtw q P R ˆ RT ˆ RT ˆ ST @t “ 1, . . . , T

pπt ´ κt, πu, πv, Πwq ľKK 0 @t “ 1, . . . , T
t t t T
˜ ¸
ÿt
πt ` y0 xτ ´ κT `t , π u ´ p1, . . . , 1, 0, . . . , 0qJ , π v , ľKK
` τ “1
Π w t t t 0 T
@t “ 1, . . . , T
˜ T
ÿ
α ´ hTy0 ´ h ÿT
t“1 tx pT ´ t ` 1q ´ pb ` πt ´ κ2T `1,

t“
ÿhq 1
β ´ pb ` hq T
πtu ` hpT, T ´ 1, . . . , 1qJ ,
t“
1
ÿ
Πw ¸
γ ´ pb ` hq T T
ÿ
πtv , Θ ´ pb ` t ľK K 0
T
t“ t“
diagpΘq “ 0, 1
hq 1
where the vector p1, . . . , 1, 0, . . . , 0qJ in the second line of constraints has t components equal to
one and T ´ t components equal to zero, for all x P RT .
`
Proof. The proof widely parallels to that of Proposition 2, where the additional decision
variables
t Πw, t “ 1, . . . , T , characterize the sensitivity of φ in Problem (23) with respect to W.
l It could numerically be observed that both the progressive and the L-
conservative approximation of minxPX f Kpxq are still close and that minxPX f Kpxq could be
[Type here]
significantly smaller than minxPX f pxq. Hence, if there is evidence supporting that demands are
serially uncorrelated, the inventory manager may wish to adopt these extended models with the
cone KK instead of K in
[Type here]
Soumen Bhatta
order to avoid being overly conservative in his/her ordering policy. Note also that the upper
approximation provided in Proposition 4 involves only the primal cone KTK and not the dual
cone pKTK q‹ .
6. Numerical experiments
In this section, we will compare the solutions of our distributionally robust inventory model with
those of the state-of-the-art benchmarks, where the latter have access to the demand distribution,
under which each marginal demand follows a two-point or a normal distribution. In addition, we
also assess the quality of our conservative and progressive approximations, both of which are
scalable, by synthetically determining how close they are to the original problem. Finally, we
perform a sensitivity analysis with respect to the cost parameters: c, b, and h.
6.1. Comparison with benchmarking approaches

We compare the exact minimizer x‹ of minxPX f pxq, which due to Theorem 1 is representable as
a finite second-order cone program, with the solutions of two stochastic inventory problems. First,
we refer to the minimizer of
«T ˜ t pxτ ´ ξτ ˜ t ¸+¸ff
# ÿ ÿ
minimize ÿ y0 ` ¸ ,
˜
t“ cxt `max h τ ´b y0 τ pxτ ´ ξτ q
1 “1 “1
EP q `
subject to x P X
as a stochastic solution and that of
«T ˜ t ¸
# ÿ
minimize ÿ y0 `
t“ ˜ τ
1 cxt pξt´1 q ` max h “1pxτ pξτ ´1 q ´ ξτ q ,
EP ÿ
˜ t ¸+¸ff
y0 `
´b τ pxτ pξ τ ´1 q ´ ξτ q
“1
`
x1 , x2 pξ1 q, . . . , xT pξ T ´1 q J P X P-a.s.
subject to
˘
as an adaptive solution. In this experiment, we set aside computational tractability and are
only concerned with the quality of different solutions. Note that tractability will be the focus of
our next experiment concerning the proposed approximations. Throughout, it is assumed that X
`
“ RT
and the demands are serially independent and identically distributed with
` ˘ ` ˘
P ξt “ ξ L “ pL and P ξt “ ξ H “ pH @t P t1, . . . , T u
for some ppL , ξ L , pH , ξ H q such that ξ H ą ξ L , pL and pH are both non-negative, and they sum up
[Type here]
to one. It is well-known that the adaptive solution follows a base-stock policy (Scarf 1960), whereas
the stochastic solution has been recently studied by Basciftci et al. (2021), for instance. As a result,
both stochastic and adaptive solutions can be characterized by vectors in RT which we shall denote
by x˜ ‹ and S ‹ , respectively. When P is really the true data-generating distribution, the base-
stock
[Type here]
Soumen Bhatta
policy S ‹ (despite its impracticality) yields a lower expected total cost than that of x˜ ‹ , and
the stochastic solution x˜ ‹ yields a lower expected total cost than that of the proposed robust
solution x‹. As this premise however is likely untrue, inspired by Bertsimas et al. (2010), we perform
a stress test by determining the expected total cost corresponding to x‹ and x˜ ‹ under the
contaminated probability distribution
Pλ “ p1 ´ λqP ` λPwc , (24)
where λ P r0, 1s represents the contamination level and Pwc is the demand distribution that (approx-
imately) attains the worst-case expected cost when x “ x˜ ‹ (ε “ 10´4 is used). Note that, unlike
P, Pwc does not necessarily generate serially independent demands.
We consider two plausible scenarios: (i) when it is likely for the demand to be small but there can
be a rare surge and (ii) when it is likely for the demand to be large but there might be a rare
drop. As a representative of scenario (i), we set pL “ 0.7, pH “ 0.3, ξL “ 30 and ξH “ 70.
Even
though it is more likely for the demands to be small, the inventory manager should still prepare
for a sudden surge in demand, which may force him or her to backlog the unmet demands. This is
particularly crucial when backlogging is expensive, and to investigate this effect we set b “ 3 ą 1 “ h
and c “ 8. As a representative of scenario (ii), we similarly set pL “ 0.3, pH “ 0.7 and retain ξL “
30, ξH “ 70. For this case, we are particularly concerned with the possible drop in the demands
and high holding cost. We therefore set h “ 3 ą 1 “ b and c “ 3. As none of these optimization
problems are tractable, we choose to work with a relatively small T “ 6, and we further assume
that the initial inventory is y0 “ 0. Results for these two scenarios are reported in Figure 2 (left) and
Figure 2 (right), respectively, from which it can be observed that the robust policy x‹ is more
resistant to the misspecification in the demand distribution. Besides, the contamination level at
which the robust solution starts to outperform the stochastic solution is given by λ “ 11.85% for
scenario (i) and by λ “ 34.78% for scenario (ii), respectively.
Next, we similarly compare our robust solution with the base-stock policy S‹. However, since
the base-stock policy is adaptive and thus at an advantageous position as future ordering deci-
sions may depend on the realizations of the previous demands, we execute our robust policy in
a shrinking horizon fashion to facilitate a fairer comparison. Put differently, for the first period
we solve minxPX f pxq to determine x‹ but only implement the here-and-now decision x‹1 . For
the second period, we resolve the same optimization problem with the updated initial inventory
and with the number of periods reduced by one, and we only implement the here-and-now
ordering decision again, and so on and so forth; see e.g. Solyalı et al. (2016) and Mamani et
al. (2017). We evaluate both the adaptive robust and the base-stock policies under different
contaminated demand distributions Pλ, λ P r0, 1s, where we substitute for Pwc the worst-case
[Type here]
demand distribution
[Type here]
Soumen Bhatta
1400
ROBUST ROBUST
2300 STOCHASTIC STOCHASTIC
1350
1300
2200
1250
2100
Expected Cost
Expected Cost
1200
1150
2000
1100
1900
1050
1000
1800
950
1700
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Contamination Level Contamination Level
Figure 2 Comparison between our distributionally robust and the stochastic solutions under different
contaminated demand distributions in terms of the expected total cost (solid line), the expected total cost`12
standard deviation (upper dashed line), and the expected total cost´12standard deviation (lower dashed line).
« ‹ « ‹
when x “ EP rx s with x representing (adaptive) ordering decisions equivalent to S ‹ . For
both scenarios (i) and (ii), when λ “ 0 (and thus Pλ “ P), the base-stock policy (which is known to be
optimal under this distributional setting) is superior to the adaptive robust policy. Nonetheless, as
the out-of-sample distribution deviates further from the in-sample distribution, Figure 3 shows
that the adaptive robust policy becomes increasingly competitive and eventually outperforms the
base-stock policy, which highlights the resilience of the proposed robust policies even in the adap-
tive setting. We also remark that the differences between the two policies are more prominent in
scenario (ii), see Figure 3 (right), where the adaptive robust policy seems to have an
inadvertent variance reduction effect for the distribution of the total cost. Finally, we can also
relate Figure 3 to its non-adaptive counterpart, i.e., Figure 2. While we indeed observe that the
adaptive robust and the base-stock solutions yield a smaller expected cost than the robust and the
stochastic solutions, respectively, the reduction is not monumental. Thus, when T is only
moderately large, determining the adaptive robust and/or the base-stock policy may not be a
worthwhile pursuit considering the other managerial benefits of agreeing to an advance
purchase.
6.2. Quality of the proposed approximations

To validate the performance of our proposed approximations, we experiment with randomly gen-
erated problem instances. Without loss of generality, we set the unit ordering cost c to be one, and
we independently choose h and b from a uniform distribution on the unit interval r0, 1s. Similarly,
we normalize the mean demand µ to be one without any loss, and then randomly generate its
standard deviation σ from the uniform distribution U pr0, 2sq. For each T P t10, 20, 30, 40, 50u,
we set y0 “ 0 and run fifty random experiments in total. Note that, when T ě 20, minxPX f pxq
cannot be solved exactly within a time limit of 12 hours using MOSEK ApS (2019) and YALMIP
[Type here]
interface
[Type here]
Soumen Bhatta
ADAPTIVE ROBUST 1080 ADAPTIVE ROBUST BASE-STOCK

2100 BASE-STOCK
1060
1040
2000
1020
1900
Expected Cost
Expected Cost
1000
980
1800
960
1700
940
920
1600
900
1500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Contamination Level Contamination Level
Figure 3 Comparison between our adaptive distributionally robust and the adaptive (i.e., base-stock) solutions
under different contaminated demand distributions in terms of the expected total cost (solid line), the expected
total cost`12standard deviation (upper dashed line), and the expected total cost´1 2standard deviation (lower
dashed line).
(Löfberg 2004) on a 2.90GHz i7-10700 CPU machine with 16GB RAM, which implies the need
for the approximate solution methods.
We have three objectives here. Firstly, we would like to numerically show that the Q-
conservative approximation that is due to Theorem 5 is (marginally) better than the L-conservative
approximation that is due to Proposition 2. To this end, we measure the improvement in
conservativeness by
f L‹px‹L q ´ ‹f
Q px Q q f Q px Q q
,
where xL‹ and xQ‹ denote the L- and the Q-conservative solution, respectively. Secondly, we want
to quantify how close xQ‹ is to being optimal in view of the problem minxPX f pxq. We define the
relative optimality gap as pf pxQ‹ q ´ f px‹ qq{f px‹ q and propose to upper bound it by
f px‹ q f px‹ q f px‹ q

and the rightmost expression, dependent only on the approximate worst-case expected total costs,
could be computed rather efficiently even for a large T . Finally and similarly to the above, we can
also measure the precision of our Q-conservative solution, x‹Q , via the following metric
f px‹ q ´ f
px‹QQ q Qf px‹Q q
.
Note that by construction, our precision metric is necessarily smaller than the upper bound of the
relative optimality gap in (25). These three measurements are computed for each simulation run
and reported collectively in Figure 4. On average, the solution x‹Q is less conservative than x‹L ,
[Type here]
and
[Type here]
Soumen Bhatta
6 4
2
3.5
Precision of Our Conservative Solution (%)

5
Improvement in Conservativeness (%)
Relative Optimality Gap (%)

1.5
4
2.5
3 2
1
1.5
2
1
0.5
1
0.5
0 0
0
T=10 T=20 T=30 T=40 T=50 T=10 T=20 T=30 T=40 T=50 T=10 T=20 T=30 T=40 T=50
Figure 4 Measurements of our approximation quality: (left) improvement in conservativeness of x‹Q over
x‹L , (middle) our conservative bound on the relative optimality gap of x‹Q , (right) precision of x‹Q .
it is also close to being robustly optimal. While x‹L appears inferior in this experiment, it can
cater for additional requirements (see Section 5) more readily and is thus more flexible.
Given the quality of the proposed progressive approximation, we can use the underlying
scenario- reduction idea to construct an approximate worst-case distribution by solving a
variant of Prob- lem (9) with the exponentially-sized uncertainty setq E replaced
q by E “ te0, . . .
, eT u. In this case, q
the construction of the approximate worst-case distribution is dependent on the optimally cho-
‹ ‹ T
sen α‹ : Eq ÞÑ R, β : Eq ÞÑ RT , γ : Eq ÞÑ R . While the construction of the distribution
described
in (12) and (13) indicates that demands should be statistically dependent in the worst-case, we
provide evidence asserting a stronger statement that they are in fact highly correlated (i.e., they
are strongly linearly dependent). As an illustrative example, we set c “ 1, h “ 1, b “ 1{3, T “
? ? ?
20,
y0 “ 0, µ “ 1 and consider different σ P t 0.001, 0.005, 0.025u. For each of the parameter set-
‹
tings, we compute the xQ and the corresponding (approximate) worst-case distribution
accordingly to the described procedure. We visualize the resultant correlation matrices with
heatmaps in Fig- ure 5 and find that there exists a breakpoint t‹ P t1, . . . , T u such that the
demands ξ1, . . . , ξt‹ are almost perfectly positively correlated and so are the remaining demands
ξt‹`1, . . . , ξT . One could think of t‹ as the period with a transitional shift, e.g., a sudden surge
or a sudden drop in the
demand pattern, which reminds us of several relatable reasons, such as the availability of the sub-
stitute products. Recent examples include energy supplies and Covid-19 vaccines. This observation
remains unchanged with different parameter settings, and it sheds light on the potential benefit of
using the uncorrelated ambiguity set PK instead of P when appropriate.
6.3. Normality and independence

Here, we revisit the stochastic solution but consider a demand distribution P which is serially
independent and is characterized by ξt „ Npµ, σ2q, @t, as opposed to the two-point distribution pre-
viously considered. For the ease of exposition, we introduce the cumulative demands ζt ř t
τ ξτ ,
“ “1
@t. It follows that ζt is a normal random variable with the following density function
ˆ ˙
[Type here]
ft pzq “ pz ´ tµq2
? exp .
2tσ2
σ 2tπ
[Type here]
Soumen Bhatta
1 1 1
1 1 1
2 2 2
0.9 3
3 3 0.9
4 4 4 0.995
5 0.8 5 5
6 6 0.8 6
7 0.7 7 7
0.99
8 8 8
9 0.6 9 0.7
9
10 10 10
11 0.5 11 0.6 11 0.985
12 12 12
13 0.4 13 13
14 14 0.5 14
0.98
15 0.3 15 15
16 16 16
0.4 17
17 0.2 17
18 18 18 0.975
19 19 0.3
19
0.1
20 20 20
Figure 5 Correlation matrices of the (approximate) worst-case distributions.
Besides, we denote the cumulative distribution function of ζt by Ft. By its definition, the stochastic
‹
solution
# x˜ minimizes an objective gpxq which equals ˜ +
ÿt ˜
ż y0`řt xτ t ż `8 t
τ ÿ
cxt ` “1
y0 ` x ¸
τ ´ y0 ` xτ¸´ ft pζt q dζt .
´8
ft pζt q dζt ´ b
y0` řtτ “1
τ
“1 xτ ÿ τ
“1
h ζt ζt
t“
1
Note that, even though the above objective function g is convex in x, its explicit characterization
?ş
2 z
involves the Gauss error function erfpzq “ π ´8
exppú2 q du; see e.g. Glaisher (1871). Despite
this complication, however, the gradient of g can be easily determined as
Bg ÿT ż y0`x1`...`xτ
B
Bxt “ c ` py0 ` x1 ` . . . ` xτ ´ ζτ q fτ pζτ q dζτ
Bxt ´8
h τ
“t
T
ÿ B `8
´b py ` x1 ` . . . ` xτ ´ ζτ q fτ pζτ q dζτ
τ Bxt ży0`x1`...`xτ 0
“t
ÿ"
“ c` T
Bx ż y0`x1`...`xτ
B py ` x ` . . . ` x q F py ` x ` . . . ` x q *
h Bxt
τ B ´8 ζτ fτ pζτ q dζτ
“t
" ´ ż
0 1 τ τ 0 1 τ
T
ÿ B py ` x ` . . . ` x q p1 ´ F py ` x ` . . . ` x qq
`8 *
´b 0 1 τ τ 0 1 τ
τ Bx t B
“t
y0`x1`...`xτ ζτ fτ pζτ q dζτ
ÿ ´
Bx
“ c` T T
ÿ
Fτ py0 ` x1 ` . . . ` xτ q p1 ´ Fτ py0 ` x1 ` . . . ` xτ qq
h
τ τ
“t
´b “t
ÿT
“ c ´ bpT ´ t ` 1q ` ph ` Fτ py0 ` x1 ` . . . ` xτ
bq τ q,
“t
where the third equality holds because of the Fundamental Theorem of Calculus. As a result,
[Type here]
the stochastic solution can be obtained by using a projected gradient method (Alber et al. 1998),
on the feasible set X “ ` when necessary.
which
RT iteratively constructs a new solution by following the gradient descent direction and
projecting it normality and independence assumption, we then proceed to compare the stochastic
Under this
solution x˜ ‹ with our conservative solution xQ‹ . Figure 6 (left) shows the total order of both
solutions, i.e., 1J x˜ ‹ and 1J xQ‹ when T “ 20, c “ 1, b “ 0.5, y0 “ 0, µ “ 1 and σ “ 0.5 for varying
h P t0, . . . , 1u.
[Type here]
Soumen Bhatta
22 22
ROBUST ROBUST
20 STOCHASTIC 20 STOCHASTIC
18 18
16 16
Total Order Amount
Total Order Amount

14 14
12 12
10 10
8 8
6 6
4 4
2 2
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
h b
Figure 6 Total order comparison between our robust and the stochastic solutions under the normality and
independence assumption.
Similarly, Figure 6 (right) shows the total order of both solutions for the same set of parameters
with the exception that now h is fixed at 0.5 and b is varied in t0, . . . , 1u. As expected, when
the holding cost increases, both solutions order less, and when it is expensive to backlog, both
solutions order more. From the figure, in comparison to the stochastic solution, it can be seen that
the robust solution is more receptive to the changes in h and b, which we attribute to the fact that
the change in b and h has an impact on the worst-case demand distribution but not on the nominal
normal distribution. Note that each bar consists of two parts: the darker-shaded part represents
the total order made in the first half of the planning horizon (t P t1, . . . , T {2u), whereas the
lighter-shaded part represents that made in the second half (t P tT {2 `1,..., T u). For both
solutions, at least half of the total order is made in the first half of the planning horizon.
Nevertheless, the ratio between the first-half and the second-half order of the robust solution
appears significantly greater than that of the stochastic solution.
Acknowledgements: This research is supported by the Ministry of Education, Singapore, under
its 2019 Academic Research Fund Tier 3 grant call (Award ref: MOE-2019-T3-1-010).
References
Alber, Y., A. Iusem, M. Solodov. 1998. On the projected subgradient method for nonsmooth convex opti-
mization in a Hilbert space. Mathematical Programming 81 23–35.
Andres, G. 2022. Novavax’s Nuvaxovid COVID-19 vaccine granted interim authorisation in Singapore.
https://www.channelnewsasia.com/singapore/novavax-nuvaxovid-covid-19-vaccine-interim
-authorisation-singapore-hsa-2496476. Accessed on 2nd May 2022.
Ardestani-Jaafari, A., E. Delage. 2016. Robust optimization of sums of piecewise linear functions with
application to inventory problems. Operations research 64(2) 474–494.
[Type here]
Soumen Bhatta
Basciftci, B., S. Ahmed, N. Gebraeel. 2021. Adaptive two-stage stochastic programming with an application
to capacity expansion planning. Available on arXiv #1906.03513.
Ben-Tal, A., B. Golany, A. Nemirovski, J. Vial. 2005. Retailer-supplier flexible commitments contracts: A
robust optimization approach. Manufacturing & Service Operations Management 7(3) 248–271.
Ben-Tal, A., A. Goryashko, E. Guslitzer, A. Nemirovski. 2004. Adjustable robust solutions of uncertain
linear programs. Mathematical programming 99(2) 351–376.
Ben-Tal, A., E. Hochman. 1976. Stochastic programs with incomplete information. Operations Research
24(2) 336–347.
Bertsekas, D. 1995. Dynamic programming and optimal control , vol. 1. Athena Scientific.
Bertsimas, D., X. Doan, K. Natarajan, C. Teo. 2010. Models for minimax stochastic linear optimization
problems with risk aversion. Mathematics of Operations Research 35(3) 580–602.
Bertsimas, D., I. Popescu. 2005. Optimal inequalities in probability theory: A convex optimization approach.
SIAM Journal on Optimization 15(3) 780–804.
Bertsimas, D., M. Sim, M. Zhang. 2019. Adaptive distributionally robust optimization. Management Science
65(2) 604–618.
Bertsimas, D., A. Thiele. 2006. A robust optimization approach to inventory theory. Operations Research
54(1) 150–168.
Bienstock, D., N. O¨ zbay. 2008. Computing robust basestock levels. Discrete Optimization 5(2) 389–414.
Chen, Z., W. Xie. 2021. Regret in the newsvendor model with demand and yield randomness. Production
and Operations Management 30(11) 4176–4197.
Das, B., A. Dhara, K. Natarajan. 2021. On the heavy-tail behavior of the distributionally robust newsvendor.
Operations Research 69(4) 1077–1099.
Delage, E., Y. Ye. 2010. Distributionally robust optimization under moment uncertainty with application
to data-driven problems. Operations research 58(3) 595–612.
Drucker, P. 1974. Management: Tasks, Responsibilities, Practices. Harper & Row.
Federgruen, A., P. Zipkin. 1984. An efficient algorithm for computing optimal ps, Sq policies.
Operations Research 32(6) 1268–1285.
Fu, M. 1994. Sample path derivatives for ps, Sq inventory systems. Operations Research 42(2) 351–364.
Fu, M., X. Li, L. Zhang. 2021. Data-driven feature-based newsvendor: A distributionally robust approach.
Available at SSRN#3885663.
Gallego, G., I. Moon. 1993. The distribution free newsboy problem: review and extensions. Journal of the
Operational Research Society 44(8) 825–834.
Glaisher, J. 1871. XXXII. On a class of definite integrals. The London, Edinburgh, and Dublin Philosophical
Magazine and Journal of Science 42(280) 294–302.
[Type here]
Soumen Bhatta
Gorissen, B., D. den Hertog. 2013. Robust counterparts of inequalities containing sums of maxima of linear
functions. European Journal of Operational Research 227(1) 30–43.
Hadjiyiannis, M., P. Goulart, D. Kuhn. 2011. A scenario approach for estimating the suboptimality of linear
decision rules in two-stage robust optimization. In Proceedings of the 50th IEEE Conference on Decision
and Control and European Control Conference. Florida, United States.
Halman, N., D. Klabjan, M. Mostagir, J. Orlin, D. Simchi-Levi. 2009. A fully polynomial-time

approximation scheme for single-item stochastic inventory control with discrete demand. Mathematics
of Operations Research 34(3) 674–685.
Jalelah, B. 2021. Singapore made advance purchases for COVID-19 vaccines, including Sinovac. https:
//www.channelnewsasia.com/singapore/covid-19-vaccines-singapore-sinovac-advance-purc
hases-277091. Accessed on 24st February 2022.
Lee, S., H. Kim, I. Moon. 2021. A data-driven distributionally robust newsvendor model with a
wasserstein ambiguity set. Journal of the Operational Research Society 72(8) 1879–1897.
Löfberg, J. 2004. YALMIP : A toolbox for modeling and optimization in MATLAB. In Proceedings
of the CACSD Conference. Taipei, Taiwan.
Lu, H., B. Sturt. 2022. On the sparsity of optimal linear decision rules in robust inventory management.
Available on arXiv #2203.10661.
Mak, H., Y. Rong, J. Zhang. 2015. Appointment scheduling with limited distributional information. Man-
agement Science 61(2) 316–334.
Mamani, H., S. Nassiri, M. Wagner. 2017. Closed-form solutions for robust inventory management. Man-
agement Science 63(5) 1625–1643.
MOSEK ApS. 2019. The MOSEK optimization toolbox for MATLAB manual. Version 9.3 . URL http:
//docs.mosek.com/9.3/toolbox/index.html.
Padmanabhan, D., K. Natarajan, K. Murthy. 2021. Exploiting partial correlations in distributionally robust
optimization. Mathematical Programming 186 209–255.
Rodrigues, F., A. Agra, C. Requejo, E. Delage. 2021. Lagrangian duality for robust problems with decom-
posable functions: The case of a robust inventory problem. INFORMS Journal on Computing 33(2)
685–705.
Roos, E., D. den Hertog, A. Ben-Tal, F. de Ruiter, J. Zhen. 2020. Tractable approximation of hard
uncertain optimization problems. Available on Optimization-Online #6679.
Rujeerapaiboon, N., D. Kuhn, W. Wiesemann. 2016. Robust growth-optimal portfolios. Management Science
62(7) 2090–2109.
Scarf, H. 1958. A min-max solution of an inventory problem. Studies in the mathematical theory of inventory
and production 201–209.
[Type here]
Soumen Bhatta
Scarf, H. 1960. The optimality of pS, sq policies in the dynamic inventory problem. K. Arrow, S. Sarlin,
P. Suppes, eds., Mathematical Models in the Social Sciences.
See, C., M. Sim. 2010. Robust approximation to multiperiod inventory management. Operations Research
58(3) 583–594.
Shapiro, A. 2001. On duality theory of conic linear problems. M. Goberna, M López, eds., Semi-
Infinite Programming: Recent Advances . Springer, 135–165.
Shapiro, A., D. Dentcheva, A. Ruszczyn´ski. 2014. Lectures on Stochastic Programming: Modeling and Theory
.
SIAM.
Solyalı, O., J.-F. Cordeau, G. Laporte. 2016. The impact of modeling on robust inventory management
under demand uncertainty. Management Science 62(4) 1188–1201.
Veinott, A., H. Wagner. 1965. Computing optimal ps, Sq inventory policies. Management Science
11(5) 525–552.
Xin, L., D. Goldberg. 2022. Distributionally robust inventory control when demand is a martingale. Math-
ematics of Operations Research In Press.
Yue, J., B. Chen, M. Wang. 2006. Expected value of distribution information for the newsvendor problem.
Operations research 54(6) 1128–1136.
Zhang, X., Z. Ye, W. Haskell. 2021. Asymptotic analysis for data-driven inventory policies. Available on
arXiv #2008.08275.
Zheng, Y., A. Federgruen. 1991. Finding optimal ps, Sq policies is about as simple as evaluating a
single policy. Operations research 39(4) 654–665.
Zymler, S., D. Kuhn, B. Rustem. 2013. Distributionally robust joint chance constraints with second-order
moment information. Mathematical Programming 137(1) 167–198.
Appendix (Proof of Proposition 3)

The proof of Proposition 3 builds on two technical lemmas, which we present below.
LEMMA 3. Given a P R` and b, c P R, it holds that

` bz ` b2
min az2
zPR ( “ c´ .
4a
c
Proof. When a “ 0 and b ‰ 0, the minimization problem is unbounded. When a “ 0 but b “ 0,
the optimal objective value trivially evaluates to c which coincides with the right-hand side because
of our convention that 0{0 “ 0. Finally, when a ą 0, it is a routine exercise to verify that the
b
minimzer z‹ is ´ 2a and hence the statement follows again. l
LEMMA 4. For a, q P Rn such that 1J q “ 1, it holds that
ñ ˜ ¸
qi a2
ÿ i ÿ
i“1
[Type here]
¸ n i“1
2
“
qi ai qi qj pai ´
´ aj q2 .
ÿ
1ďi
ăjďn
[Type here]
Soumen Bhatta
řn
q a2 q
Observe that if q ě 0, a direct consequence of Lemma 4 is that p i“ i qi ai q2 ě 0, which
1 ř
coincides with the result of the Cauchy-Schwarz inequality. ´p
i n i“
1
Proof of Lemma 4. The statement can be verified algebraically as

ñ ¸ ˜ ¸2 ˜ ¸˜ ¸ ˜ ¸2
n n n n
ÿ i ÿ qi ai ÿ qi a i ÿ 2
ÿ qi ai
i“
1 qi a2 i“
1 “
i“
1
i“
1 qi i“
1
´ ´ ff ff
ÿ ÿ
«
« 2 2 2 2
qi qj pa ` a q 2 2 2qi qj ai aj
n q a `
“ ÿ ÿ qa `
n
i i i j ´ i i
i“1 1ďiăjďn i“1 1ďiăjďn
ÿ ÿ
2
“ qi qj pa2i ` aj2 ´ 2ai aj q “ qi qj pai ´ aj q .
1ďiăjďn 1ďiăjďn
The proof is then completed. l

We are now ready to utilize Lemmas 3 and 4 to validate Proposition 3, that is, Problem (18) is
dual to Problem (17).
Proof of Proposition 3. Let q P ` be a collection of Lagrange multipliers of the four constraints
R4
in Problem (17) respectively. We write down the associated Lagrangian function as
Lpx, α, β, γ, qq “ cpx1 ` x2 q ` α ` µpβ1 ` β2 q ` pµ2 ` σ2 qpγ1 ` γ2 q´

` ˘
q1 α ´ 2hx1 ´ hx2 ´ pβ1 ` 2hq2 {4γ1 ´ pβ2 ` hq2 {4γ2 ´
` ˘
q2 α ` 2bx1 ` bx2 ´ pβ1 ´ 2bq2 {4γ1 ´ pβ2 ´ bq2 {4γ2 ´
` ˘
q3 α ` pb ´ hqx1 ´ hx2 ´ pβ1 ` h ´ bq2 {4γ1 ´ pβ2 ` hq2 {4γ2 ´
` ˘
2 2
q4 α ` pb ´ hqx1 ` bx2 ´ pβ1 ` h ´ bq {4γ1 ´ pβ2 ´ bq {4γ2 .
The dual problem is thus

max min Lpx, α, β, γ, qq.
qě0 xě0,γě0,α,β
For the inner minimization to be bounded from below with respect to the choice of x1, the
inequality (18d) must hold. With respect to the choice of x2 and α, the inequality (18e) and the
equality (18c) must similarly hold, respectively. These three conditions show up explicitly as dual
constraints in Problem (18).
After eliminating x and α, the Lagrangian L can be decomposed into two parts:
1 “ 2 2 3 24 2
‰
2 1 2
4γ 4µβ γ ` q pβ ` 2hq ` q pβ ´ 2bq ` q pβ ` h ´ bq ` q pβ ` h ´ bq ` pµ ` σ qγ ,
1 1 1 1 2 1 1 1
1
which depends only on β1 , γ1 , q and is thus abbreviated as L1 pβ1 , γ1 , qq, and
[Type here]
1 “ 2 2 2 2
‰
4µβ2 γ2 ` q1 pβ2 ` hq `2q pβ
2 ´ bq `3 q 2pβ ` hq 4` q2 pβ ´ bq ` pµ2 ` 2σ 2 qγ ,
4γ 2
which depends solely on β2 , γ2 , q and we shall abbreviate it as L2 pβ2 , γ2 , qq.
[Type here]
Soumen Bhatta
„
Next, we invoke Lemma 3 to minimize L1 pβ1 , γ1 , qq over β1 and find that the minimum objective is
pµ2 ` σ2qγ ` 4q h2 ` 4q b2 ` q ph ´ bq2 ` q ph ´ bq2 ´
1
1 2 (26)
1 1 2 3 4
4γ
1 
p4µγ1 ` 4q1h ´ 4q2b ` 2q3ph ´ bq ` 2q4ph ´ bqq .
4
„ 
Utilizing Lemma 3 again to minimize L2 pβ2 , γ2 , qq over β2 yields the minimum objective of
2
pµ2 ` σ2 qγ ` q h2 ` q b2 ` q h2 ` q b2 4µγ ` 2q h ´ 2q b ` 2q h ´ 2q bq . (27)
1 1
2
4γ2
1 2 3 4 ´ 2 1 2 3 4
p 4
We subsequently leverage Lemma 4 with a “ p2h, ´2b, h ´ b, h ´ bqJ to simplify minβ1 L1 pβ1 , γ1 ,
qq from (26) further to
1
γ σ2 ´ µ p2q h ´ 2q b ` q ph ´ bq ` q ph ´ bqq ` ph ` bq2
`q `q `q ` q q q , (28)
p4q q q q q
1 1 2 3 4 1 2 1 3 1 4 2 3 2 4
4γ1
and with a “ ph, ´b, h, ´bqJ to simplify minβ2 L2 pβ2 , γ2 , qq from (27) further to
γb σ2 ´ µ pq h ´ q b ` q h ´ q
1 qh ` bq2 pq `
q q `
q q `q q q. (29)
2
2 1 2 3
4 q` 1 2 1 4 2 3 3 4
p 4γ
From (28) and (29), we respectively use the arithmetic mean–geometric mean inequality to argue
that minγ1 ě0,β1 L1 pβ1 , γ1 , qq is equal to
a
σph ` bq 4q1 q2 ` q1 q3 ` q1 q4 ` q2 q3 ` q2 q4 ´ µ p2q1 h ´ 2q2 b ` q3 ph ´ bq ` q4 ph ´ bqq
and that minγ2 ě0,β2 L2 pβ2 , γ2 , qq is equal to

?
σph ` bq q1 q2 ` q1 q4 ` q2 q3 ` q3 q4 ´ µ pq1 h ´ q2 b ` q3 h ´ q4 bq .
The summation of minγ1 ě0,β1 L1 pβ1 , γ1 , qq and minγ2 ě0,β2 L2 pβ2 , γ2 , qq gives rise to the dual objective
in Problem (18). The proof is now completed. l
[Type here]

Distributionally Robust Inventory with Advance Contracts

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Distributionally Robust Inventory with Advance Contracts

Uploaded by

Copyright:

Available Formats

Distributionally Robust Inventory Management

with Advance Purchase Contracts

Keywords : Inventory management; distributionally robust optimization; conic programming.

adaptive to the previously revealed demands tξτ ut in the sense that

2. Mean-variance inventory model and its finite representations

THEOREM 1. We have that

f pxq “ minimize c1J x ` α ` µ1J β ` pµ2 `

pαpx, eq, βpeq, γq ľKT 0 @e P E ,

min α ` µ1J β ` pµ2 ` σ 2 q1J γ

and the proof is completed by recalling the definition of the cone KT . l

α max # T max gt pet , ξt , βt , # T max g pe , ξ , β ,

THEOREM 2. We have that

pαpx, eq, βpeq, γq ľKT 0

and obtain the following dual formulation of Problem (5):

PROPOSITION 1. We have that

αα ` βJβ ` γJγ ě 0 @pα, β, γq P Kt.

2.1. Worst-case demand distribution

for any t P t1, . . . , T u and any e P E such that α‹ peq ą 0.

For any fixed e P E, where α‹ pe q ą 0, if ‹

For any time period t1 P t1, . . . , T u, it thus follows that

where the scenarios e and e are characterized by

et “ et “ et @t P t1, . . . , T uztt1 u and et1 “ `h as well as et1 “ ´b.

ε max h y0 ` pxτ ´ ξ τ , y0 ` pxτ ´ ξ τ q .

is typically assumed (e.g. Scarf 1960).

3. Efficiently solvable conservative approximation

THEOREM 4. We have that

allows us to derive the dual: T ˜ t ¸

φt pu, vq ě 0 @pu, vq : p1, u, vq P KT‹ , t “ 1, . . . , T

@pu, vq : p1, u, vq P KT‹ .

We subsequently obtain a conservative approximation by restricting each φt pu, vq to an affine form:

decisions. minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ

πt ` uJπu ` vJπv ě 0 @pu, vq : p1, u, vq P K‹ , t “ 1, . . . , T

πt ` uJπut ` vJπvt ě 0 @pu, vq : p1, u, vq P KT‹

Next, we derive an alternative conservative approximation of f pxq which is obtained by

The correctness of this upper bound is validated in the next theorem.

ðñ pαpx, eq, βpeq, γq ľKT 0 @e P E

4. Efficiently solvable progressive approximation

Proof. This is an immediate consequence of Theorem 1 as we can interpret f pxq as the

minimize cpx1 ` x2 q ` α ` µpβ1 ` β2 q ` pµ2 ` σ 2 qpγ1 ` γ2 q (17a)

subject to x1 P R`, x2 P R`, α P R, β1 P R, β2 P R, γ1 P R`, γ2 P R` (17b)

α ´ 2hx1 ´ hx2 ´ pβ1 ` 2hq2 {4γ1 ´ pβ2 ` hq2 {4γ2 ě (17c)

α ` pb ´ hqx1 ´ hx2 ´ pβ1 ` h ´ bq2 {4γ1 ´ pβ2 ` hq2 {4γ2 ě 0 (17e)

α ` pb ´ hqx1 ` bx2 ´ pβ1 ` h ´ bq2 {4γ1 ´ pβ2 ´ bq2 {4γ2 ě 0, (17f)

PROPOSITION 3. The dual of Problem (17) is given by

µ p3hq1 ´ 3bq2 ` p2h ´ bqq3 ` ph ´ 2bqq4 q (18a)

subject to q1, q2, q3, q4 ě 0 (18b)

THEOREM 7. Any optimal solution q ‹ of Problem (18) satisfies q3‹ “ 0.

c ` 2hq1 ´ 2bq2 ` ph ´ bqq3 ě 0

c ` hq1 ´ bq2 ` hq3 ě 0

p2h ` cqp2b ´ cq p2h ` cqp2b ´ cq

As a result, we obtain that

2q2 ` ´ ă p2q2 ´ 1q p1 ` 8q2q.

subject to x P X , z P t0, 1uT

5. Plug-and-play conic extensions

5.1. Non-negative support

and the less conservative cost function

THEOREM 8. We have that

ðñ pθt , βt peq ´ δt , γt q ľK1 0.

5.2. Uncorrelated demands

THEOREM 9. We have that

f K pxq “ minimize c1J x ` α ` µ1J β ` pµ2 ` σ2 q1J γ ` µ2 1J Θ1

f K pxq ´ c1J x “ minimize α ` µ1J β ` pµ2 ` σ2 q1J γ ` µ2 ÿ θst