Question 1.1
ZD0FCan the optimal objective value of (?) ever be strictly greater than the expected revenue of the firm’s best past assortment?
The Value of Robust Assortment Optimization Under Ranking-based Choice Models
- arXiv
- 2112.05010v4 · PDF · 26 February 2024
- Categories
- math.OC
- Licence
- CC BY 4.0 The site shows the text.
Abstract
We study a class of robust assortment optimization problems that was proposed by Farias, Jagabathula, and Shah (2013). The goal in these problems is to find an assortment that maximizes a firm’s worst-case expected revenue under all ranking-based choice models that are consistent with the historical sales data generated by the firm’s past assortments. We establish for various settings that these robust optimization problems can either be solved in polynomial-time or can be reformulated as compact mixed-integer optimization problems. To establish our results, we prove that optimal assortments for these robust optimization problems have a simple structure that is closely related to the structure of revenue-ordered assortments. We use our results to show how robust optimization can be used to overcome the risks of estimate-then-optimize and the need for experimentation with ranking-based choice models in the overparameterized regime.
Items 85
Question 1.2
DM5TIf the answer to Question 1.1 is yes, then what structural properties of a firm’s past assortments are necessary for the optimal objective value of (?) to be strictly greater than the expected revenue of the firm’s best past assortment?=-1
Question 1.3
9WQJWhat structural properties of a firm’s past assortments are sufficient for designing efficient algorithms for solving the robust optimization problem (?)?
Question 1.4
D9NNIf the optimal objective value of (?) is not strictly greater than the expected revenue of the firm’s best past assortment for a particular problem instance, then what assortment should the firm offer to its customers?
Theorem 4.1
N6UYThere exists an assortment S \in \widehat{\mathcal{S}} that is optimal for (?).
Lemma 4.2
XTF4If \mathscr{M} = \bar{\mathcal{S}}, then \widehat{\mathcal{S}} = \mathscr{M}.
Corollary 4.3
V83GIf \mathscr{M} = \bar{\mathcal{S}} and \eta = 0, then \max \limits_{S \in \mathcal{S}} \min \limits_{\lambda \in \mathcal{U}} \mathscr{R}^{\lambda}(S) = \max \limits_{m \in \mathcal{M}} r^\intercal v_m.
Lemma 4.4
ATPZIf M = 2, then
\begin{aligned}
\widehat{\mathcal{S}} = \left \{ S \in \mathcal{S}: \quad \begin{aligned} &\textnormal{there exists }i_1 \in S_1 \setminus S_2 \textnormal{ and } i_2 \in S_2 \setminus S_1 \textnormal{ such that}\\
&S = \left(S_1 \cap S_2 \right) \cup \left \{j \in S_1 \setminus S_2: j \ge i_1 \right \} \cup \left \{j \in S_2 \setminus S_1: j \ge i_2 \right \}\end{aligned} \right \}.
\end{aligned}
Lemma 4.5
I1F1| \widehat{\mathcal{S}} | \le (n+2)^{2^M}.
Lemma 4.6
STCHThe collection of assortments \widehat{\mathcal{S}} can be constructed in \mathcal{O}(n^{2} (M+| \widehat{\mathcal{S}}|)) time.
Definition 4.7
QO0AWe say that the firm’s past assortments are reverse revenue-ordered if S_1 \triangleq \{0,n\}, S_m \triangleq \left \{ 0,1,\ldots,m-1,n \right \} for each m \in \{2,\ldots,n\}, and \mathscr{M} = \tilde{\mathcal{S}} \triangleq \left \{S_1,\ldots,S_n\right \}.
Lemma 4.8
M6JEIf \mathscr{M} = \tilde{\mathcal{S}}, then \widehat{\mathcal{S}} = \left \{ S \subseteq \mathcal{N}_0: \{0,n\} \subseteq S \right \}.
Theorem 4.9
96JHSuppose that \mathscr{M} = \tilde{\mathcal{S}}, that \eta = 0, and that the revenues r_1 < \cdots < r_n are fixed. Then for every assortment \bar{S} \in \widehat{\mathcal{S}}, there exists a realization of the historical data v \equiv (v_{m,i}: m \in \mathcal{M}, i \in S_m) such that \bar{S} is the unique optimal solution of the robust optimization problem (?).
Definition 4.10
9QV2\mathcal{D}_i(S) \triangleq \left \{ \sigma \in \Sigma: \; i = \mathop{\SFrm arg\,min}_{j \in S} \sigma (j) \right \}.
Definition 4.11
U4UR\mathcal{L} \triangleq \left \{(i_1,\ldots,i_M) \in S_1 \times \cdots \times S_M: \; \bigcap_{m \in \mathcal{M}} \mathcal{D}_{i_m}(S_m) \neq \emptyset \right \}.
Definition 4.12
F98K\rho_{i_1 \cdots i_M}(S) \triangleq \min \limits_{i \in S: \; \cap_{m \in \mathcal{M}} \mathcal{D}_{i_m}(S_m) \cap \mathcal{D}_i(S) \neq \emptyset} r_i.
Proposition 4.13
1F4CFor each S \in \mathcal{S}, \min_{\lambda \in \mathcal{U}} \mathscr{R}^\lambda(S) is equal to the optimal objective value of the following linear optimization problem:
\begin{aligned}
& \; \underset{\lambda, \epsilon}{\textnormal{minimize}} && \sum_{(i_1,\ldots,i_M) \in \mathcal{L}} \rho_{i_1 \cdots i_M}(S) \lambda_{i_1\cdots i_M}\\
&\textnormal{subject to}&& \sum_{(i_1,\ldots,i_M) \in \mathcal{L}: \; i_m = i} \lambda_{i_1 \cdots i_M} - \epsilon_{m,i}= v_{m,i} \quad \forall m \in \mathcal{M}, \; i \in S_m\\
&&& \sum_{(i_1,\ldots,i_M) \in \mathcal{L}} \lambda_{i_1 \cdots i_M} = 1 \\
&&& \| \epsilon \| \le \eta\\
&&& \lambda_{i_1 \cdots i_M} \ge 0 \quad \forall (i_1,\ldots,i_M) \in \mathcal{L}.
\end{aligned}
Remark 4.14
IFDAThe assortment S does not appear in any of the constraints of (4.13).
Remark 4.15
UUWL(4.13) has | \mathcal{L}| = \mathcal{O}(n^M) decision variables and \mathcal{O}(nM) constraints.
Proposition 4.16
PQT5Let S \in \mathcal{S} and i \notin S. If there exists i^* \in S that satisfies r_{i^*} < r_i and \mathcal{M}_{i^*} \subseteq \mathcal{M}_i, then \rho_{i_1 \cdots i_M}(S) \le \rho_{i_1 \cdots i_M}(S \cup \{i\}) for each (i_1,\ldots,i_M) \in \mathcal{L}.
Remark 5.1
LS4TWe assume throughout §? that every product i \in \mathcal{N}_0 has been offered by the firm to its customers in at least one of the past assortments, that is, \cup_{m \in \mathcal{M}} S_m = \mathcal{N}_0. This assumption simplifies our notation and can always be satisfied by redefining \mathcal{N}_0 to be equal to \cup_{m \in \mathcal{M}} S_m. Such a redefining of \mathcal{N}_0 does not impact the optimal objective value of the robust optimization problem (?); see Appendix ? for further details.
Theorem 5.2
XQ3KIf M = 2 and \eta = 0, then (?) can be solved in \mathcal{O}(n^5 \log (n r_n)) time.
Lemma 5.3
CYSZIf M = 2, then \mathcal{L} = \left((S_1 \setminus S_2) \times S_2 \right) \cup \left( S_1 \times (S_2 \setminus S_1) \right) \cup \left\{(i,i): i \in S_1 \cap S_2 \right \}.
Lemma 5.4
VK17If M = 2 and \eta = 0, then \min_{\lambda \in \mathcal{U}} \mathscr{R}^\lambda(S) can be evaluated for each assortment S \in \mathcal{S} by solving a minimum-cost network flow problem in a directed acyclic graph with \mathcal{O}(n) vertices and \mathcal{O}(n^2) directed edges.
Theorem 5.5
X53R(?) can be solved in weakly \mathcal{O}(\textnormal{poly}(n)) time for every fixed M.
Lemma 5.6
9KG1|\mathcal{L}| = \mathcal{O}(n^M), and \mathcal{L} can be constructed in \mathcal{O}(Mn^{M+1}) time.
Lemma 5.7
MYYYFor each assortment S \in \widehat{\mathcal{S}} and each tuple of products (i_1,\ldots,i_M) \in \mathcal{L}, the quantity \rho_{i_1 \cdots i_M}(S) can be computed in \mathcal{O}(M^2 n) time.
Assumption (Nested)
CWXCThe past assortments \mathscr{M} \equiv \{S_1,\ldots,S_M\} \subseteq \mathcal{N}_0 satisfy S_1 \subset \cdots \subset S_M.
Lemma 5.8
N9G1If Assumption Assumption holds, then
\mathcal{L} = \left \{(i_1,\ldots,i_M) \in S_1 \times \cdots \times S_M:\textnormal{ for each }m \in \{1,\ldots,M-1\}, \; i_{m+1} \in \{i_m \} \cup \mathcal{B}_{m+1} \right \}.
Uses Assumption
Corollary 5.9
ATK8If Assumption Assumption holds, then |\mathcal{L}| \ge 2^{M-1}.
Uses Assumption
Lemma 5.10
I1FHIf Assumption Assumption holds and the assortment S \subseteq \mathcal{N}_0 satisfies S \cap S_1 \neq \emptyset,(The condition that S \cap S_1 \neq \emptyset in Lemma 5.10 is satisfied, for example, when S, S_1 \in \mathcal{S}.) then the following equality holds for each tuple (i_1,\ldots,i_M) \in \mathcal{L} and m \in \mathcal{M}:
\begin{aligned}
\rho_{i_1 \cdots i_M}(S \cap S_m) &= \begin{cases}
r_{i_m},&\textnormal{if } i_m \in S, \\
\min \limits_{i \in \mathcal{B}_m \cap S} r_i, &\textnormal{if } i_m \notin S \textnormal{ and } m = 1,\\
\min \left \{ \min \limits_{i \in \mathcal{B}_m \cap S} r_i ,\; \rho_{i_1 \cdots i_M} \left(S \cap S_{m-1} \right) \right \}, &\textnormal{if } i_m \notin S \textnormal{ and } m \ge 2.
\end{cases}
\end{aligned}
Uses Assumption
Proposition 5.11
X13VLet Assumption Assumption hold. Then for each S \subseteq \mathcal{N}_0 that satisfies S \cap S_1 \neq \emptyset, \min_{\lambda \in \mathcal{U}} \mathscr{R}^\lambda(S) is equal to the optimal objective value of the following linear optimization problem: 911 0.8em 0.3em 0.5em 0.3em 0.6em 0.3em 0.3em 0.5em 1em 0 2 11pt plus 0.5pt minus 0.35pt =11pt=9pt
\begin{aligned}
& \; \underset{f, g,\epsilon}{\textnormal{minimize}} && \sum_{i,\kappa: (M,i,\kappa) \in \mathfrak{V}} \kappa g_{M,i,\kappa} \\
&\textnormal{subject to}&&
\sum_{\kappa: (m,i,\kappa) \in \mathfrak{V}} g_{m,i,\kappa} - \epsilon_{m,i} = v_{m,i} && \forall m \in \mathcal{M}, \; i \in S_m \\
&&& \sum_{i',\kappa': ((m,i,\kappa),(m+1,i',\kappa')) \in \mathfrak{E}} f_{m,i,\kappa,i',\kappa'} = g_{m,i,\kappa} && \forall (m,i,\kappa) \in \mathfrak{V}: m \in \{1,\ldots,M-1\} \\
&&& \sum_{i',\kappa': ((m-1,i',\kappa'),(m,i,\kappa)) \in \mathfrak{E}} f_{m-1,i',\kappa',i,\kappa} = g_{m,i,\kappa} && \forall (m,i,\kappa) \in \mathfrak{V}: m \in \{2,\ldots,M\} \\
&&& g_{m,i,\kappa} = 0 && \forall (m,i,\kappa)\in \mathfrak{V}: i \in S,\; i \in \mathcal{B}_m, \; \kappa \neq r_i \\
&&& g_{m,i,\kappa} = 0 && \forall (m,i,\kappa)\in \mathfrak{V}: i \in S,\; i \notin \mathcal{B}_m, \; \kappa \in \{r_j: j \in \mathcal{B}_m \} \\
&&& g_{m,i,\kappa} = 0 && \forall (m,i,\kappa)\in \mathfrak{V}: \kappa \in \{ r_j: j \in \mathcal{B}_m \setminus S \} \\
&&& \sum_{i,\kappa: (M,i,\kappa) \in \mathfrak{V}} g_{M,i,\kappa} = 1 \\
&&& \| \epsilon \| \le \eta \\
&&& f_{m,i,\kappa,i',\kappa'} \ge 0 && \forall ((m,i,\kappa),(m+1,i',\kappa')) \in \mathfrak{E} \\
&&& g_{m,i,\kappa} \ge 0 && \forall (m,i,\kappa) \in \mathfrak{V}.
\end{aligned}
Uses Assumption
Theorem 5.12
5TWFLet Assumption Assumption hold and S_1,\ldots,S_M \in \mathcal{S}. Then the robust optimization problem (?) can be reformulated as a mixed-integer linear optimization problem with \mathcal{O}(n) binary decision variables, \mathcal{O}(\textnormal{poly}(n,M)) continuous decision variables, and \mathcal{O}(\textnormal{poly}(n,M)) constraints.
Remark 5.13
QOXJThe reformulation in Appendix ? is presented for cases in which the firm does not have any business constraints on the assortments S \in \mathcal{S}. In many applications of assortment optimization, firms have additional business constraints (e.g., due to limited shelf space or bundling requirements) that limit the space of possible new assortments; see [mivsic2016data, p.113]. Because our algorithm from Theorem 5.12 consists of solving a mixed-integer optimization problem with a binary decision variable x_i \in \{0,1\} for each product i \in \mathcal{N}_0 that satisfies x_i = 1 if and only if product i is included in the assortment S, such business constraints can be easily integrated directly into our mixed-integer optimization problem. In particular, as long as the number of such business constraints is not exponential in n or M, the size of our mixed-integer optimization reformulation of (?) will remain polynomial in the number of products and the number of past assortments.
Uses Theorem 5.12
Theorem 6.1
06EX(?) can be solved in weakly \mathcal{O}(\textnormal{poly}(n)) time for every fixed M.
Theorem 6.2
G0G2If Assumption Assumption holds and S_1,\ldots,S_M \in \mathcal{S}, then (?) can be reformulated as a mixed-integer linear optimization problem with \mathcal{O}(n) binary decision variables, \mathcal{O}(\textnormal{poly}(n,M)) continuous decision variables, and \mathcal{O}(\textnormal{poly}(n,M)) constraints.
Theorem 6.3
CC0ULet Assumption Assumption hold and S_1,\ldots,S_M \in \mathcal{S}. Then the optimization problem (?) can be reformulated as a mixed-integer linear optimization problem with \mathcal{O}(n) binary decision variables, \mathcal{O}(\textnormal{poly}(n,M)) continuous decision variables, and \mathcal{O}(\textnormal{poly}(n,M)) constraints.
Lemma B.1
Q39W(i_1,\ldots,i_M) \in \mathcal{L} if and only if \mathcal{G}_{i_1 \cdots i_M} is acyclic.
Definition B.3
47AILet (i_1,\ldots,i_M) \in S_1 \times \cdots \times S_M and i, j \in \mathcal{N}_0. We say that vertex j is reachable from vertex i, denoted by j \prec_{i_1 \cdots i_M} i, if there exists a directed path in \mathcal{G}_{i_1 \cdots i_M} from i to j.
Lemma B.4
ZDRWLet (i_1,\ldots,i_M) \in \mathcal{L} and i,j \in \mathcal{N}_0. If j \prec_{i_1 \cdots i_M} i, then there exists a past assortment m \in \mathcal{M} that satisfies j = i_m.
Lemma B.6
OWYJLet (i_1,\ldots,i_M) \in \mathcal{L} and i,j \in \mathcal{N}_0. Then, there exists a ranking \sigma \in \cap_{m \in \mathcal{M}} \mathcal{D}_{i_m}(S_m) that satisfies \sigma(i) < \sigma(j) if and only if j \nprec_{i_1 \cdots i_M} i.
Lemma B.8
ZYB6Let (i_1,\ldots,i_M) \in \mathcal{L}, S \in \mathcal{S}, and i \in S. Then, there exists a ranking \sigma \in \cap_{m \in \mathcal{M}} \mathcal{D}_{i_m}(S_m) that satisfies i = \mathop{\SFrm arg\,min}_{j \in S} \sigma(j) if and only if j \nprec_{i_1 \cdots i_M} i for all j \in S.
Definition B.10
HUEZ\mathcal{I}_{i_1 \cdots i_M}(S) \triangleq \left \{ i \in \mathcal{N}_0:\; \textnormal{for all } m \in \mathcal{M}, \; \textnormal{if } i_m \in S, \text{ then } i_m \nprec_{i_1 \cdots i_M} i \right \}.
Proposition B.11
ZILLFor all S \in \mathcal{S} and (i_1,\ldots,i_M) \in \mathcal{L}, \rho_{i_1 \cdots i_M}(S) = \min_{i \in S \cap \mathcal{I}_{i_1 \cdots i_M}(S)} r_i.
Claim 1
LIEVFor all S \in \mathcal{S}, (i_1,\ldots,i_M) \in \mathcal{L}, and i \notin S,
\begin{aligned}
&\rho_{i_1 \cdots i_M}(S \cup \{i\}) = \begin{cases}
\rho_{i_1 \cdots i_M}(S),&\textnormal{if } i \notin \mathcal{I}_{i_1 \cdots i_M}(S),\\
\min \left \{ \min \limits_{j \in S \cap \mathcal{I}_{i_1 \cdots i_M}(S) \cap \left\{ j' \in \mathcal{N}_0: i \nprec_{i_1 \cdots i_M} j' \right\}} r_j, r_i \right \}, &\textnormal{if } i \in \mathcal{I}_{i_1 \cdots i_M}(S).\\
\end{cases}
\end{aligned}
Procedure
HT91Initialize the vertex set \mathscr{V} \leftarrow \emptyset and edge set \mathscr{E} \leftarrow \emptyset.
For each product i \in \mathcal{N}_0:
Update \mathscr{V} \leftarrow \mathscr{V} \cup \{i \}.
Construct the set \mathcal{M}_i of past assortments which offered product i.
For each pair of products (i^*,i) \in \mathcal{N}_0 \times \mathcal{N}_0:
If r_{i^*} < r_{i} and \mathcal{M}_{i^*} \subseteq \mathcal{M}_i:
Update \mathscr{E} \leftarrow \mathscr{E} \cup \{(i^*,i) \}
Output \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) and terminate.
A procedure for constructing the directed acyclic graph \mathscr{G}.
Procedure
6O3EConstruct the directed acyclic graph \mathscr{G} \leftarrow \textsc{Construct-G(M,r)}.
Compute the collection of assortments \widehat{\mathcal{S}} \leftarrow \textsc{RecursiveStep(G)}.
For each S \in \widehat{\mathcal{S}}:
If 0 \notin S:
Update \widehat{\mathcal{S}} \leftarrow \widehat{\mathcal{S}} \setminus \{ S \}.
Output \widehat{\mathcal{S}} and terminate.
Procedure
SVZ7If \mathscr{V} = \emptyset:
Output the collection \mathscr{A} \equiv\{ \emptyset \} and terminate.
Otherwise:
Claim 2
XPZDLet \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}. Then for each vertex i \in \mathscr{V}, we have that \tilde{\mathscr{G}} \equiv (\mathscr{V} \setminus \{i \}, \mathscr{E} \setminus \mathscr{E}_i) is a directed acyclic graph that is a transitive closure.
Claim 3
JZSPLet \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}. Then for each subset of vertices \mathscr{B} \subseteq \mathscr{V}, we have that \tilde{\mathscr{G}} \equiv (\mathscr{V} \setminus \mathscr{B}, \mathscr{E} \setminus (\cup_{i \in \mathscr{B}} \mathscr{E}_i)) is a directed acyclic graph that is a transitive closure.
Claim 4
PTPILet \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}. For each vertex i \in \mathscr{V},
\begin{aligned}
&\left \{S \subseteq \mathscr{V} \setminus \{i \}: \textnormal{ if $k \in S$ and $(k,j) \in \mathscr{E}$, then $j \in S$} \right\} \\
& = \left \{ S \subseteq \mathscr{V} \setminus (\{ i \} \cup \{\ell: (\ell,i) \in \mathscr{E} \}): \textnormal{if } k \in S \textnormal{ and } (k,j) \in \mathscr{E} \setminus \left( \mathscr{E}_i \cup \bigcup_{\ell: (\ell,i) \in \mathscr{E}} \mathscr{E}_\ell \right), \textnormal{ then } j \in S \right \}.
\end{aligned}
Claim 5
IZ7DLet \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}. For each vertex i \in \mathscr{V},
\begin{aligned}
&\left \{S \subseteq \mathscr{V}: \left[ i \in S \right] \textnormal{ and } \left[ \textnormal{if $k \in S$ and $(k,j) \in \mathscr{E}$, then $j \in S$} \right] \right\} \\
& = \left \{ \begin{aligned}
&S' \cup \{i \} \cup \left\{ \ell: (i,\ell) \in \mathscr{E} \right \} :\\
&\quad S' \in \left \{ S \subseteq \mathscr{V} \setminus \left( \{i \} \cup \left\{ \ell: (i,\ell) \in \mathscr{E} \right \} \right): \textnormal{if } k \in S \textnormal{ and } (k,j) \in \mathscr{E} \setminus \left( \mathscr{E}_i \cup \bigcup_{\ell: (i,\ell) \in \mathscr{E}} \mathscr{E}_\ell \right), \textnormal{ then } j \in S \right \} \end{aligned} \right \}.
\end{aligned}
Claim 6
WIA2If \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) is a directed acyclic graph that is a transitive closure with | \mathscr{V}| \ge 1, then lines (?), (?), (?), and (?) of Algorithm ? can be performed in \mathcal{O}\left(| \mathscr{V}| \times \left|\textsc{RecursiveStep}(\mathscr{G}) \right| \right) time.
Claim 7
OBWZIf \mathscr{G} \equiv (\mathscr{V},\mathscr{E}) is a directed acyclic graph that is a transitive closure with | \mathscr{V}| \ge 1, then the computation time for Algorithm ? is \mathcal{O}(|\mathscr{V}|^2 \times \left|\textsc{RecursiveStep}(\mathscr{G}) \right| ).
Lemma E.1
FFNHThe system (?) is nonempty and has a unique solution \lambda which satisfies the following equality for each (i_1,\ldots,i_n) \in \mathcal{L}:
\begin{aligned}
\lambda_{i_1 \cdots i_n} &= \begin{cases}
\frac{1}{n},&\textnormal{if there exists } j \in \bar{S}\setminus \{0,n\} \textnormal{ such that } (i_1,\ldots,i_n) = \left(\text{\b{$i$}}^j_1,\ldots,\text{\b{$i$}}^j_n \right),\\
\frac{1}{n},&\textnormal{if there exists } j \in \mathcal{N} \setminus \bar{S} \textnormal{ such that }(i_1,\ldots,i_n) = \left(\bar{i}^j_1,\ldots,\bar{i}^j_n \right) ,\\
\frac{1}{n},&\textnormal{if } (i_{1}, \ldots, i_n) = (n,\ldots,n),\\
0,&\textnormal{otherwise}.
\end{cases}
\end{aligned}
Claim 8
X4GGSuppose that \lambda is a feasible solution of the system (?), and let (i_1,\ldots,i_n) \in \mathcal{L}. If \lambda_{i_1 \cdots i_n} > 0, then either (i_1,\ldots,i_n) = (n,\ldots,n) or there exists j \in \{1,\ldots,n-1\} such that i_{j+1} = j.
Claim 9
DWKCSuppose that \lambda is a feasible solution of the system (?), and let (i_1,\ldots,i_n) \in \mathcal{L}. If \lambda_{i_1 \cdots i_n} > 0 and if (i_1,\ldots,i_n) \neq (n,\ldots,n), then there exists j \in \{1,\ldots,n-1\} and \dagger \in \{0,n\} such that i_1 = \cdots = i_j = \dagger and i_{j+1} = \cdots = i_n = j.
Proposition F.1
CFWD\max \limits_{S \in \mathcal{S}} \min \limits_{\lambda \in \mathcal{U}}\mathscr{R}^{\lambda}(S) = \max \limits_{S \in \mathcal{S}} \min \limits_{\lambda \in \mathcal{U}}\mathscr{R}^{\lambda}\left(S \cap \left( \bigcup_{m \in \mathcal{M}} S_m \right) \right).
Remark G.1
MBQOThe inequality of the form \sum_{i_1 \in S_1 \setminus S_2} \lambda_{i_1 i} \le v_{2,i} for each vertex i \in S_1 \cap S_2 can be interpreted as an upper bound on the incoming flow entering into vertex i. These inequalities can be converted into edge capacities by splitting each vertex i \in S_1 \cap S_2 into two vertices, i^{\text{in}} and i^{\text{out}}, where the incoming edges to vertex i^{\text{in}} are from the vertices in S_1 \setminus S_2, the outgoing edges from i^{\text{out}} are to the vertices in S_2 \setminus S_1, the only outgoing edge from vertex i^{\text{in}} is the only incoming edge to vertex i^{\text{out}}, the flow through the edge from vertex i^{\text{in}} to vertex i^{\text{out}} must satisfy the constraints 0 \le g_{i^{\text{in}} i^{\text{out}}} \le v_{2,i}, and the supplies on the vertices i^{\text{in}} and i^{\text{out}} are 0 and v_{1,i} - v_{2,i}, respectively.
Procedure
KH4RConstruct the directed acyclic graph \mathcal{G}_{i_1 \cdots i_M}.
For each past assortment m \in \mathcal{M}:
If i_m \in S:
Mark each unmarked vertex in \mathcal{G}_{i_1 \cdots i_M} that has a directed path to vertex i_m.
Output the minimum r_i among all unmarked vertices i in \mathcal{G}_{i_1 \cdots i_M} and terminate.
Claim 10
XZP6Consider any i \in S \cap \mathcal{B}_m and m' \in \{1,\ldots,m-1\}. If i_{m'} \in S, then i_{m'} \nprec_{i_1 \cdots i_M} i.
Lemma I.2
ACL0Let Assumption Assumption hold, let S \subseteq \mathcal{N}_0 satisfy S \cap S_1 \neq \emptyset, and let (\lambda,\epsilon) be a feasible solution for the linear optimization problem (4.13). For each vertex (m,i,\kappa) \in \mathfrak{V}, let
\begin{aligned}
g_{m,i,\kappa} &\triangleq \sum_{(i_1,\ldots,i_M) \in \mathcal{L}: \; i_m = i, \rho_{i_1 \cdots i_M}(S \cap S_m) = \kappa} \lambda_{i_1 \cdots i_M},
\end{aligned}
and for each directed edge ((m,i,\kappa),(m+1,i',\kappa')) \in \mathfrak{E}, let
\begin{aligned}
f_{m,i,\kappa,i',\kappa'} &\triangleq \sum_{(i_1,\ldots,i_M) \in \mathcal{L}: \; i_m = i, \rho_{i_1 \cdots i_M}(S \cap S_m) = \kappa, i_{m+1} = i', \rho_{i_1 \cdots i_M}(S \cap S_{m+1}) = \kappa' } \lambda_{i_1 \cdots i_M}.
\end{aligned}
Then (f,g,\epsilon) is a feasible solution for the linear optimization problem (5.11), and the objective value associated with (f,g,\epsilon) in the linear optimization problem (5.11) is equal to the objective value associated with (\lambda,\epsilon) in the linear optimization problem (4.13).
Claim 11
OLK0For each (i_1,\ldots,i_M) \in \mathcal{L}, we have that ((1,i_1,\rho_{i_1 \cdots i_M}(S \cap S_1)),\ldots, (M,i_M,\rho_{i_1 \cdots i_M}(S \cap S_M))) \in \mathscr{P} and that (m,i_m,\rho_{i_1,\ldots,i_M}(S \cap S_m)) \notin \mathfrak{N}(S) for all m \in \mathcal{M}.
Lemma I.5
AER1Let Assumption Assumption hold, let S \subseteq \mathcal{N}_0 satisfy S \cap S_1 \neq \emptyset, and let (f,g,\epsilon) satisfy the following constraints:
\begin{aligned}
&&& \sum_{\kappa: (m,i,\kappa) \in {\mathfrak{V}}} g_{m,i,\kappa} - \epsilon_{m,i}= v_{m,i} && \forall m \in \mathcal{M}, \; i \in S_m \\
&&& \sum_{i',\kappa': ((m,i,\kappa),(m+1,i',\kappa')) \in {\mathfrak{E}}} f_{m,i,\kappa,i',\kappa'} = g_{m,i,\kappa} && \forall (m,i,\kappa) \in {\mathfrak{V}}: m \in \{1,\ldots,M-1\} \\
&&& \sum_{i',\kappa': ((m-1,i',\kappa'),(m,i,\kappa)) \in {\mathfrak{E}}} f_{m-1,i',\kappa',i,\kappa} = g_{m,i,\kappa} && \forall (m,i,\kappa) \in {\mathfrak{V}}: m \in \{2,\ldots,M\}\\
&&& \| \epsilon \| \le \eta \\
&&& \sum_{i,\kappa: (M,i,\kappa) \in \mathfrak{V}} g_{M,i,\kappa} = 1 \\
&&& f_{m,i,\kappa,i',\kappa'} \ge 0 && \forall ((m,i,\kappa),(m+1,i',\kappa')) \in {\mathfrak{E}} \\
&&& g_{m,i,\kappa} \ge 0 && \forall (m,i,\kappa) \in {\mathfrak{V}}.
\end{aligned}
Then there exists a vector \lambda such that (\lambda,\epsilon) is a feasible solution for the linear optimization problem (5.11) and
\begin{aligned}
\sum_{(i_1,\ldots,i_M) \in \mathcal{L}} \rho_{i_1 \cdots i_M}(S) \lambda_{i_1 \cdots i_M} \le \sum_{i,\kappa: (M,i,\kappa) \in \mathfrak{V}} \kappa g_{m,i,\kappa} + \sum_{(m,i,\kappa) \in \mathfrak{N}(S)} r_n g_{m,i,\kappa}.
\end{aligned}
Claim 12
CEK1If ((1,i_1,\kappa_1),\ldots, (M,i_M,\kappa_M)) \in \mathscr{P}, then (i_1,\ldots,i_M) \in \mathcal{L}.
Claim 13
ASZ3If ((1,i_1,\kappa_1),\ldots, (M,i_M,\kappa_M)) \in \mathscr{P} and (m,i_m,\kappa_m) \notin \mathfrak{N}(S) for all m \in \mathcal{M}, then \kappa_m \ge \rho_{i_1 \cdots i_M}(S \cap S_m) for all m \in \mathcal{M}.
Claim 14
M46HConsider any assortment S \in \mathcal{S}, and let \mathfrak{M}(S) \in \mathbb{R}^{{\mathfrak{V}}} be any vector that satisfies the inequality \mathfrak{M}_{m,i,\kappa}(S) \ge r_n for all (m,i,\kappa) \in \mathfrak{N}(S) and satisfies the equality \mathfrak{M}_{m,i,\kappa}(S) = 0 for all (m,i,\kappa) \in {\mathfrak{V}} \setminus \mathfrak{N}(S). Then the optimal objective value of the linear optimization problem (5.11) is equal to the optimal objective value of the following linear optimization problem:
\begin{aligned}
& \; \underset{f, g,\epsilon}{\textnormal{minimize}} && \sum_{i,\kappa: (M,i,\kappa) \in {\mathfrak{V}}} \kappa g_{M,i,\kappa} + \sum_{(m,i,\kappa) \in {\mathfrak{V}}} \mathfrak{M}_{m,i,\kappa}(S) g_{m,i,\kappa} \\
&\textnormal{subject to}&& \sum_{\kappa: (m,i,\kappa) \in {\mathfrak{V}}} g_{m,i,\kappa} - \epsilon_{m,i}= v_{m,i} && \forall m \in \mathcal{M}, \; i \in S_m \\
&&& \sum_{i',\kappa': ((m,i,\kappa),(m+1,i',\kappa')) \in {\mathfrak{E}}} f_{m,i,\kappa,i',\kappa'} = g_{m,i,\kappa} && \forall (m,i,\kappa) \in {\mathfrak{V}}: m \in \{1,\ldots,M-1\} \\
&&& \sum_{i',\kappa': ((m-1,i',\kappa'),(m,i,\kappa)) \in {\mathfrak{E}}} f_{m-1,i',\kappa',i,\kappa} = g_{m,i,\kappa} && \forall (m,i,\kappa) \in {\mathfrak{V}}: m \in \{2,\ldots,M\}\\
&&& \sum_{i,\kappa: (M,i,\kappa) \in \mathfrak{V}} g_{M,i,\kappa} = 1 \\
&&& \| \epsilon \| \le \eta \\
&&& f_{m,i,\kappa,i',\kappa'} \ge 0 && \forall ((m,i,\kappa),(m+1,i',\kappa')) \in {\mathfrak{E}} \\
&&& g_{m,i,\kappa} \ge 0 && \forall (m,i,\kappa) \in {\mathfrak{V}}.
\end{aligned}
Definition J.1
AJIS\xi_{i_1 \cdots i_M}(S) \triangleq \max \limits_{i \in S: \; \cap_{m \in \mathcal{M}} \mathcal{D}_{i_m}(S_m) \cap \mathcal{D}_i(S) \neq \emptyset} r_i.
Proposition J.2
V24XFor each S \in \mathcal{S}, \max_{\lambda \in \mathcal{U}} \mathscr{R}^\lambda(S) is equal to the optimal objective value of the following linear optimization problem:
\begin{aligned}
& \; \underset{\lambda, \epsilon}{\textnormal{maximize}} && \sum_{(i_1,\ldots,i_M) \in \mathcal{L}} \xi_{i_1 \cdots i_M}(S) \lambda_{i_1\cdots i_M}\\
&\textnormal{subject to}&& \sum_{(i_1,\ldots,i_M) \in \mathcal{L}: \; i_m = i} \lambda_{i_1 \cdots i_M} - \epsilon_{m,i}= v_{m,i} \quad \forall m \in \mathcal{M}, \; i \in S_m\\
&&& \sum_{(i_1,\ldots,i_M) \in \mathcal{L}} \lambda_{i_1 \cdots i_M} = 1 \\
&&& \| \epsilon \| \le \eta\\
&&& \lambda_{i_1 \cdots i_M} \ge 0 \quad \forall (i_1,\ldots,i_M) \in \mathcal{L}.
\end{aligned}
Proposition J.3
0Y97For all S \in \mathcal{S} and (i_1,\ldots,i_M) \in \mathcal{L}, \xi_{i_1 \cdots i_M}(S) = \max_{i \in S \cap \mathcal{I}_{i_1 \cdots i_M}(S)} r_i.
Proposition J.4
KKSALet S \in \mathcal{S} and i \notin S. If there exists i^* \in \mathcal{S} that satisfies r_{i^*} > r_{i} and \mathcal{M}_{i^*} \subseteq \mathcal{M}_i, then \xi_{i_1 \cdots i_M}(S) \ge \xi_{i_1 \cdots i_M}(S \cup \{i \}) for each (i_1,\ldots,i_M) \in \mathcal{L}.
Claim 15
UUHHFor all S \in \mathcal{S}, (i_1,\ldots,i_M) \in \mathcal{L}, and i \notin S,
\begin{aligned}
&\xi_{i_1 \cdots i_M}(S \cup \{i\}) = \begin{cases}
\xi_{i_1 \cdots i_M}(S),&\textnormal{if } i \notin \mathcal{I}_{i_1 \cdots i_M}(S),\\
\max \left \{ \max \limits_{j \in S \cap \mathcal{I}_{i_1 \cdots i_M}(S) \cap \left\{ j' \in \mathcal{N}_0: i \nprec_{i_1 \cdots i_M} j' \right\}} r_j, r_{i} \right \}, &\textnormal{if } i \in \mathcal{I}_{i_1 \cdots i_M}(S).
\end{cases}
\end{aligned}
Theorem J.6
L0RBThere exists an assortment S \in \widehat{\mathcal{S}}^{\textnormal{OO}} that is optimal for (?).
Lemma J.8
EE7L| \widehat{\mathcal{S}}^{\textnormal{OO}} | \le (n+1)^{2^M}.
Lemma J.10
AG89The collection of assortments \widehat{\mathcal{S}}^{\textnormal{OO}} can be constructed in \mathcal{O}(n^{2} (M+| \widehat{\mathcal{S}}^{\textnormal{OO}}|)) time.
Procedure
VPBPInitialize the vertex set \mathscr{V}^{\textnormal{OO}} \leftarrow \emptyset and edge set \mathscr{E}^{\textnormal{OO}} \leftarrow \emptyset.
For each product i \in \mathcal{N}^{\textnormal{OO}}:
Update \mathscr{V}^{\textnormal{OO}} \leftarrow \mathscr{V}^{\textnormal{OO}} \cup \{i \}.
Construct the set \mathcal{M}_i of past assortments which offered product i.
For each pair of products (i^*,i) \in \mathcal{N} \times \mathcal{N}:
If r_{i^*} > r_{i} and \mathcal{M}_{i^*} \subseteq \mathcal{M}_i:
Update \mathscr{E}^{\textnormal{OO}} \leftarrow \mathscr{E}^{\textnormal{OO}} \cup \{(i^*,i) \}
Output \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) and terminate.
A procedure for constructing the directed acyclic graph \mathscr{G}^{\textnormal{OO}}.
Procedure
5J06Construct the directed acyclic graph \mathscr{G}^{\textnormal{OO}} \leftarrow \textsc{Construct-G^OO(M,r)}.
Compute the collection of assortments \widehat{\mathcal{S}}^{\textnormal{OO}} \leftarrow \textsc{RecursiveStep-OO(G^OO)}.
Output \{ S \cup \{0\}: S \in \widehat{\mathcal{S}}^{\textnormal{OO}}\} and terminate.
Procedure
BLPDIf \mathscr{V}^{\textnormal{OO}} = \emptyset:
Output the collection \mathscr{A}^{\textnormal{OO}} \equiv\{ \emptyset \} and terminate.
Otherwise:
Claim 16
6OW0Let \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i^{\textnormal{OO}} denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}^{\textnormal{OO}}. Then for each vertex i \in \mathscr{V}^{\textnormal{OO}}, we have that \tilde{\mathscr{G}}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}} \setminus \{i \}, \mathscr{E}^{\textnormal{OO}} \setminus \mathscr{E}_i^{\textnormal{OO}}) is a directed acyclic graph that is a transitive closure.
Claim 17
JI21Let \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i^{\textnormal{OO}} denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}^{\textnormal{OO}}. Then for each subset of vertices \mathscr{B}^{\textnormal{OO}} \subseteq \mathscr{V}^{\textnormal{OO}}, we have that \tilde{\mathscr{G}}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}} \setminus \mathscr{B}^{\textnormal{OO}}, \mathscr{E}^{\textnormal{OO}} \setminus (\cup_{i \in \mathscr{B}^{\textnormal{OO}}} \mathscr{E}_i^{\textnormal{OO}})) is a directed acyclic graph that is a transitive closure.
Claim 18
HO81Let \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i^{\textnormal{OO}} denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}^{\textnormal{OO}}. For each vertex i \in \mathscr{V}^{\textnormal{OO}},
\begin{aligned}
&\left \{S \subseteq \mathscr{V}^{\textnormal{OO}} \setminus \{i \}: \textnormal{ if $k \in S$ and $(k,j) \in \mathscr{E}^{\textnormal{OO}}$, then $j \notin S$} \right\} \\
& = \left \{ S \subseteq \mathscr{V}^{\textnormal{OO}} \setminus \{ i \}: \textnormal{if } k \in S \textnormal{ and } (k,j) \in \mathscr{E}^{\textnormal{OO}} \setminus \mathscr{E}_i^{\textnormal{OO}}, \textnormal{ then } j \notin S \right \}.
\end{aligned}
Claim 19
4ZVBLet \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) be a directed acyclic graph that is a transitive closure. For notational convenience, let \mathscr{E}_i^{\textnormal{OO}} denote the set of incoming and outgoing edges from each vertex i \in \mathscr{V}^{\textnormal{OO}}. For each vertex i \in \mathscr{V}^{\textnormal{OO}},
\begin{aligned}
&\left \{S \subseteq \mathscr{V}^{\textnormal{OO}}: \left[ i \in S \right] \textnormal{ and } \left[ \textnormal{if $k \in S$ and $(k,j) \in \mathscr{E}^{\textnormal{OO}}$, then $j \notin S$} \right] \right\} \\
& = \left \{ \begin{aligned}
&S' \cup \{i \}:\\
&\quad S' \in \left \{ \begin{aligned}
&S \subseteq \mathscr{V}^{\textnormal{OO}} \setminus \left( \{i \} \cup \left\{ \ell: (i,\ell) \in \mathscr{E}^{\textnormal{OO}} \right \} \cup \left\{ \ell: (\ell,i) \in \mathscr{E}^{\textnormal{OO}} \right \} \right): \\
&\quad \textnormal{if } k \in S \textnormal{ and } (k,j) \in \mathscr{E}^{\textnormal{OO}} \setminus \left( \mathscr{E}_i^{\textnormal{OO}} \cup \bigcup_{\ell: (i,\ell) \in \mathscr{E}^{\textnormal{OO}} } \mathscr{E}_\ell^{\textnormal{OO}} \cup \bigcup_{\ell: (\ell,i) \in \mathscr{E}^{\textnormal{OO}} } \mathscr{E}_\ell^{\textnormal{OO}} \right), \textnormal{ then } j \in S \end{aligned} \right \} \end{aligned} \right \}.
\end{aligned}
Claim 20
2KVYIf \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) is a directed acyclic graph that is a transitive closure with | \mathscr{V}^{\textnormal{OO}}| \ge 1, then lines (?), (?), (?), and (?) of Algorithm ? can be performed in \mathcal{O}\left(| \mathscr{V}^{\textnormal{OO}}| \times \left|\textsc{RecursiveStep-OO}(\mathscr{G}^{\textnormal{OO}}) \right| \right) time.
Claim 21
CX8KIf \mathscr{G}^{\textnormal{OO}} \equiv (\mathscr{V}^{\textnormal{OO}},\mathscr{E}^{\textnormal{OO}}) is a directed acyclic graph that is a transitive closure with | \mathscr{V}^{\textnormal{OO}}| \ge 1, then the computation time for Algorithm ? is \mathcal{O}(|\mathscr{V}^{\textnormal{OO}}|^2 \times \left|\textsc{RecursiveStep-OO}(\mathscr{G}^{\textnormal{OO}}) \right| ).
Claim 22
1M1FFor each S \subseteq \mathcal{N}_0 that satisfies S \cap S_1 \neq \emptyset, \max_{\lambda \in \mathcal{U}} \mathscr{R}^\lambda(S) is equal to the optimal objective value of the following linear optimization problem: 911 0.8em 0.3em 0.5em 0.3em 0.6em 0.3em 0.3em 0.5em 1em 0 2 11pt plus 0.5pt minus 0.35pt =11pt=9pt
\begin{aligned}
& \; \underset{f, g,\epsilon}{\textnormal{maximize}} && \sum_{i,\kappa: (M,i,\kappa) \in \mathfrak{V}} \kappa g_{M,i,\kappa} \\
&\textnormal{subject to}&&
\textnormal{ same constraints as \eqref{prob:lp_reform_nested}}\end{aligned}
Source and licence
The text on this page is from arXiv:2112.05010v4 by Bradley Sturt, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.
The record of each item is in analysis/objects/2112/2112.05010.jsonl.