id
int64
1
24.5M
paragraph_id
int64
0
159k
content
stringlengths
1
571k
paper_arxiv_id
stringlengths
10
13
paper_section
stringlengths
1
262
section_id
float64
4
360k
paragraph_in_paper_id
int64
1
29.8k
22,512,582
1
\begin{quote} We are given a graph $G = (V, E)$ and a multiset $B$ of drivers. Each driver is specified by a tuple $(u, v, b)$ consisting of two nodes in the graph $u, v \in V$ and a non-negative real number $b$, specifying the driver's budget. The goal is to find a pricing function $p : E \to \mathbb{R}_{\geq 0}$ ...
2502.11979
Introduction
228,320
2
22,512,583
2
Equivalently, the problem can be interpreted as a game between the algorithm setting the prices and the drivers: first, the algorithm observes $G$G and $B$B, and selects the pricing $p$p; then, each driver $(u, v, b) \in B$(u, v, b) \in B computes a lowest-cost $u$u-$v$v path and either pays its cost if it is at most $...
2502.11979
Introduction
228,320
3
22,512,584
3
So far, a variety of special cases of this problem have been investigated: \begin{enumerate}[nosep] \item For $G$ being a path (\emph{``the highway problem''}), Grandoni and Rothvo{\ss} \cite{grandoni2016pricing} give a polynomial-time approximation scheme, while Briest and Krysta \cite{Briest06} show that finding ...
2502.11979
Introduction
228,320
4
22,512,585
4
The case of general graphs $G$G is arguably less understood. If an approximation ratio that depends on the sizes of both $G$G and $B$B is sufficient, then a polynomial-time $O(\log |E| + \log |B|)$O(\log |E| + \log |B|)-ap\-prox\-i\-ma\-tion algorithm follows by the more general results of Balcan, Blum and Mansour \cit...
2502.11979
Introduction
228,320
5
22,512,586
5
\textbf{Our Contribution.} We study the problem on the class of grid graphs, i.e., Manhattan-like networks, and prove that for any fixed (constant) width $\omega$\omega of the grid, there is a polynomial-time $\OO{\log |E|}$\OO{\log |E|}\log |E|-ap\-prox\-i\-ma\-tion algorithm for the maximum achievable revenue. While ...
2502.11979
Introduction
228,320
6
22,512,587
6
\subsection{More Related Work}
2502.11979
Introduction
228,320
7
22,512,588
7
\textbf{Pricing for Envy-Free Revenue Maximization.} The problem we study in this paper can be seen in the broader context of \emph{pricing for envy-free revenue maximization}. To explain the connection, let us quickly lay the foundations of the latter: there, one is given $k$k goods to sell and a multiset $B$B of buye...
2502.11979
Introduction
228,320
8
22,512,589
8
Armed as such, our problem can be phrased as pricing for envy-free revenue maximization as follows: in the unlimited-supply setting, consider the goods to be the edges of the graph $G$G, and the buyers to be the drivers. The valuation of each driver $(u, v, b) \in B$(u, v, b) \in B for a set of edges $P$P is defined as...
2502.11979
Introduction
228,320
9
22,512,590
9
Envy-free pricing has been studied in a variety of settings, not only for revenue maximization, but also for social welfare maximization \cite{largeMarkets, priceDoubling, impreciseDistribution}. Assuming unlimited supply, Demaine et al.~\cite{demaine2008combination} proved a conditional lower bound of $\OMG{\log k}$\...
2502.11979
Introduction
228,320
10
22,512,591
10
The case of limited supply (where each good is available in a certain number of copies) is also interesting and has yielded a number of polynomial-time approximation results for the maximum revenue. For the case of single-minded buyers, i.e., each buyer is only interested in a single set of goods, Cheung and Swamy \cit...
2502.11979
Introduction
228,320
11
22,512,592
11
\noindent \textbf{Stackelberg Pricing.} In \emph{Stackelberg games}, the \emph{leader} takes an action, and then, with knowledge of the action, the \emph{follower} replies with an action of their own. In \emph{Stackelberg pricing games}, there are $m$m items, some of whose prices are set in advance; the leader sets the...
2502.11979
Introduction
228,320
12
22,512,593
12
\subsection{Our Result and Technical Overview}
2502.11979
Introduction
228,320
13
22,512,594
13
For our result, we assume that the graph $G$G is a complete $m \times \omega$m \times \omega grid, where $\omega$\omega is considered to be fixed, and $m$m can vary as part of the problem instance. Moreover, for brevity, write $n := |B|$n := |B| for the number of drivers. Our main result is stated below:
2502.11979
Introduction
228,320
14
22,512,595
14
\begin{restatable}{theorem}{gridgraph}\label{thm:gridgraph} There exists a polynomial-time approximation algorithm for the maximum revenue for the class of graphs $G$ consisting of width-$\omega$ complete grids with approximation ratio $\OO{\log m}$. \end{restatable}{theorem}theorem{gridgraph}gridgraph\label{thm:gridg...
2502.11979
Introduction
228,320
15
22,512,596
15
The $\OO{\log m}$\OO{\log m}\log m approximation ratio is achieved as follows. First, the algorithm partitions the drivers into $\OO{\log m}$\OO{\log m}\log m subsets. Taking advantage of the additional structure of those subsets, we find a constant factor approximation of the optimal solution for each of the resulting...
2502.11979
Introduction
228,320
16
22,512,597
16
Thanks to the properties of the partition of drivers into $\OO{\log m}$\OO{\log m}\log m subsets (denoted $B_1, B_2, \ldots$B_1, B_2, \ldots), each multiset (we allow multiple buyers with the same tuple $(u, v, b)$(u, v, b)) $B_j$B_j can be further subdivided into groups, yielding independent instances of the problem w...
2502.11979
Introduction
228,320
17
22,512,598
17
The `assume-implement' dynamic programming technique is showcased in \cref{sec:rooted}. Normally, in a dynamic program for an optimization problem, in each state one considers solutions to corresponding subproblems and chooses the one maximizing a certain objective function. This becomes more complicated when the objec...
2502.11979
Introduction
228,320
18
22,512,599
18
One interesting contribution concerns grid graph compression (\cref{lemma:compression}). We show that, given any weighted width-$\omega$\omega grid, another weighted width-$\omega$\omega grid of length bounded by a function of $\omega$\omega exists such that for any $1 \leq i \leq j \leq \omega$1 \leq i \leq j \leq \om...
2502.11979
Introduction
228,320
19
22,512,600
0
\label{sect:rooted} \label{sec:rooted}
2502.11979
Rooted Case
228,321
20
22,512,602
2
In this section, we define a simpler variant of our problem and present a polynomial-time constant factor approximation algorithm for it, which will be a building block of the main algorithm.
2502.11979
Rooted Case
228,321
22
22,512,603
3
\begin{definition} An instance of our problem on a grid graph $G$ with driver multiset $B$ is \emph{rooted} if one of the vertices in the top row of $G$ (the \emph{root}, denoted $r$) is an endpoint of every path desired by drivers in the multiset $B$. \end{definition}\begin{definition} An instance of our probl...
2502.11979
Rooted Case
228,321
23
22,512,604
4
We will prove the following:
2502.11979
Rooted Case
228,321
24
22,512,605
5
\begin{lemma} \label{rooted_rev} Assume a fixed grid width $\omega$. For any rooted instance with maximal revenue $\mathrm{OPT}_r$, a~price assignment generating at least $\frac{\mathrm{OPT}_r}{\Crounding}$ in revenue can be found in polynomial time. \end{lemma}\begin{lemma} \label{rooted_rev} Assume a fixed gr...
2502.11979
Rooted Case
228,321
25
22,512,606
6
\subsection{Rounding} \label{sect:rounding}
2502.11979
Rooted Case
228,321
26
22,512,607
7
One of the crucial techniques used throughout the algorithm is price rounding. Let us formalize the underlying observation.
2502.11979
Rooted Case
228,321
27
22,512,608
8
\newcommand{\roundedSet}{\mathcal{P}}\mathcal{P}
2502.11979
Rooted Case
228,321
28
22,512,609
9
\begin{lemma}{\textbf{(Rounding)}} \label{rounding_lemma} For any instance ($G, B$) of the problem on a grid, there exists a pricing of edges which results in revenue of at least $\frac{1}{\Crounding}$ of the optimal one such that the price of each edge belongs to the set: $$ \roundedSet = \longset{\frac{b_{max}}{2...
2502.11979
Rooted Case
228,321
29
22,512,610
10
\begin{proof} Consider any price assignment $p$ generating the optimal revenue $\mathrm{OPT}$. If a price of any edge is greater than $b_{max}$, it can be lowered to $b_{max}$ without loss of revenue. We further round the price of each edge down to the nearest value from $P$. Let us consider any edge $e$, its in...
2502.11979
Rooted Case
228,321
30
22,512,611
11
Let $d$ and $d'$ be the distances (costs of a cheapest path) between any two vertices in $G$ with respect to the initial and rounded prices, respectively. We have: $d \geq d' \geq \frac{d}{2} - \frac{b_{max}}{4 \cdot \abs{B}}$. As the prices only decrease, no driver will be priced-out of the market by the roun...
2502.11979
Rooted Case
228,321
31
22,512,612
12
Let $d$ and $d'$ be the distances (costs of a cheapest path) between any two vertices in $G$ with respect to the initial and rounded prices, respectively. We have: $d \geq d' \geq \frac{d}{2} - \frac{b_{max}}{4 \cdot \abs{B}}$. As the prices only decrease, no driver will be priced-out of the market by the roun...
2502.11979
Rooted Case
228,321
32
22,512,613
13
Let $d$d and $d'$d' be the distances (costs of a cheapest path) between any two vertices in $G$G with respect to the initial and rounded prices, respectively. We have: $d \geq d' \geq \frac{d}{2} - \frac{b_{max}}{4 \cdot \abs{B}}$d \geq d' \geq \frac{d}{2} - \frac{b_{max}}{4 \cdot \abs{B}}. As the prices only ...
2502.11979
Rooted Case
228,321
33
22,512,614
14
In the above proof we chose the constant scaling factor of $2$2 for the sake of clarity and ease of exposition. It can be replaced with $1 + \epsilon$1 + \epsilon for any constant $\epsilon > 0$\epsilon > 0. This, together with adjusting the smallest element in $\roundedSet$\roundedSet, would allow us to prove \cref{ro...
2502.11979
Rooted Case
228,321
34
22,512,615
15
By this lemma we can restrict our attention to prices from $\mathcal{P}$\mathcal{P}. An important consequence of this is that the costs of paths in $G$G will always be multiples of $p_{min} = \frac{b_{max}}{2^{\ceil{\log \bra{4 m \cdot \abs{B}}}}}$p_{min}min = \frac{b_{max}}{2^{\ceil{\log \bra{4 m \cdot \abs{B}}}}}. An...
2502.11979
Rooted Case
228,321
35
22,512,616
16
\subsection{Assume-Implement Dynamic Programming}
2502.11979
Rooted Case
228,321
36
22,512,617
17
Before we proceed to the description of the algorithm, assume we index rows of $G$G top to bottom and introduce some notation: \begin{itemize}[nosep] \item $R_i$ -- the set of vertices in the $i$-th row of $G$. \item $G_i$ -- the subgraph of $G$ consisting of vertices from rows $i$ and $i+1$ and edges in the $i$-th row...
2502.11979
Rooted Case
228,321
37
22,512,618
18
Let us consider a subgrid $H$H consisting of some consecutive rows of $G$G. For the purpose of paths that only pass through $H$H (have endpoints either outide of $H$H or on its boundary), we do not need to consider all the edges inside $H$H. It is enough to consider the distances between the vertices on the boundary of...
2502.11979
Rooted Case
228,321
38
22,512,619
19
\begin{definition} \label{def:distance_matrix} Given a grid $H$ and a subset $S$ of its vertices, a \emph{distance matrix} for $S$ is an $\abs{S} \times \abs{S}$ matrix $D$ indexed by vertices from $S$, where $D_{a, b}$ is the assumed distance, i.e. sum of weights along a cheapest path, between $a$ and ...
2502.11979
Rooted Case
228,321
39
22,512,620
20
We will say that a certain edge-weighting \emph{realizes} a distance matrix if it satisfies the above condition (for all $a, b \in S$, $D_{a,b}$ is actually equal to the distance between $a$ and $b$ in $H$).
2502.11979
Rooted Case
228,321
40
22,512,621
21
\textbf{Note:}~We will frequently consider $H$ to be a subgraph of the (weighted) grid $G$. In that case the distance matrix describes the cheapest paths in $H$ and not $G$. Even though under some weight assignments $G$ may contain shorter paths between vertices in $S$. \end{definition}\begin{definition} \l...
2502.11979
Rooted Case
228,321
41
22,512,622
22
We will say that a certain edge-weighting \emph{realizes} a distance matrix if it satisfies the above condition (for all $a, b \in S$, $D_{a,b}$ is actually equal to the distance between $a$ and $b$ in $H$).
2502.11979
Rooted Case
228,321
42
22,512,623
23
\textbf{Note:}~We will frequently consider $H$ to be a subgraph of the (weighted) grid $G$. In that case the distance matrix describes the cheapest paths in $H$ and not $G$. Even though under some weight assignments $G$ may contain shorter paths between vertices in $S$. \end{definition} \label{def:distance_...
2502.11979
Rooted Case
228,321
43
22,512,624
24
We will say that a certain edge-weighting \emph{realizes} a distance matrix if it satisfies the above condition (for all $a, b \in S$a, b \in S, $D_{a,b}$D_{a,b}a,b is actually equal to the distance between $a$a and $b$b in $H$H).
2502.11979
Rooted Case
228,321
44
22,512,625
25
\textbf{Note:}~We will frequently consider $H$H to be a subgraph of the (weighted) grid $G$G. In that case the distance matrix describes the cheapest paths in $H$H and not $G$G. Even though under some weight assignments $G$G may contain shorter paths between vertices in $S$S.
2502.11979
Rooted Case
228,321
45
22,512,626
26
In the algorithm we will often have to merge two graphs into a single one. We will be interested in knowing the distances in the union of the two graphs based on assumptions about the distances in the original graphs. The following formalizes this operation.
2502.11979
Rooted Case
228,321
46
22,512,627
27
\begin{definition} \label{def:dist_product} Let $G_A$ and $G_B$ be two same-width edge-disjoint weighted grids with (possibly) some edges missing that share a common row with vertex set $R$ that is the top row of one and the bottom row of the other. Let $A$ and $B$ be distance matrices for vertices $S_A$ in graph $G_A$...
2502.11979
Rooted Case
228,321
47
22,512,628
28
Sometimes, in order to limit the dimensionality, we will want the product of $A$ and $B$ to be defined only for a certain subset $S \subseteq S_A \cup S_B$, in which case we write $A \dpro{S} B$. \end{definition}\begin{definition} \label{def:dist_product} Let $G_A$ and $G_B$ be two same-width edge-disjoint weighted gri...
2502.11979
Rooted Case
228,321
48
22,512,629
29
Sometimes, in order to limit the dimensionality, we will want the product of $A$ and $B$ to be defined only for a certain subset $S \subseteq S_A \cup S_B$, in which case we write $A \dpro{S} B$. \end{definition} \label{def:dist_product} Let $G_A$G_A and $G_B$G_B be two same-width edge-disjoint weighted grids with (pos...
2502.11979
Rooted Case
228,321
49
22,512,630
30
Sometimes, in order to limit the dimensionality, we will want the product of $A$A and $B$B to be defined only for a certain subset $S \subseteq S_A \cup S_B$S \subseteq S_A \cup S_B, in which case we write $A \dpro{S} B$A \dpro{S}S B.
2502.11979
Rooted Case
228,321
50
22,512,631
31
Note that the product of two distance matrices is well-defined: because $R$R is a separator of $G_A$G_A and $G_B$G_B, any shortest path between two vertices from $S_A \cup S_B$S_A \cup S_B can be partitioned into segments whose lengths are determined by either $A$A or $B$B. Thus, $A$A and $B$B already determine all pai...
2502.11979
Rooted Case
228,321
51
22,512,632
32
Each state of the dynamic program corresponds to a certain ($i$i-th) row of the grid and two distance matrices (\cref{def:distance_matrix}): \begin{itemize}[nosep] \item \emph{lower matrix} $L$ -- a distance matrix for $R_i$ in the part of $G$ below and including row $i$, \item \emph{upper matrix} $U$ -- a distance mat...
2502.11979
Rooted Case
228,321
52
22,512,633
33
Thanks to the rounding technique, we can restrict out attention to distance matrices with values equal to $0$0, $\infty$\infty or a multiple of $p_{min}$p_{min}min as argued in \cref{sect:rounding}.
2502.11979
Rooted Case
228,321
53
22,512,634
34
The value of $dp_{i, L, U}$dp_{i, L, U}i, L, U is defined as the maximum revenue generated by the drivers whose non-root endpoint lies on or below row $i$i under a price assignment to the edges in $i$i-th row and below that realizes the matrix $L$L and assuming that the price assignment to the edges above row $i$i real...
2502.11979
Rooted Case
228,321
54
22,512,635
35
\end{figure} \centering \includegraphics[width=0.65\textwidth]{dp_colors} \caption{ Dynamic programming states for row $i+1$ (left) and $i$ (right). For reference, here $i = 2$. Black dotted edges represent the graph $G_i$. Black vertices marked with asterisks are non...
2502.11979
Rooted Case
228,321
55
22,512,636
36
The left term denotes the best revenue that can be achieved by the drivers from rows below $i$i assuming the distance matrices $L$L and $U$U are realized. It is a maximum over all possible lower and upper distance matrices $L'$L' and $U'$U' for the previous row and each possible way $W$W to price the edges in $G_i$G_i....
2502.11979
Rooted Case
228,321
56
22,512,637
37
Note that $E(G_i) = E(L) \setminus E(L')$E(G_i) = E(L) \setminus E(L') and $E(G_i) = E(U') \setminus E(U)$E(G_i) = E(U') \setminus E(U), where $E(M)$E(M) denotes the edges of the graph corresponding to the matrix $M$M. Thus, it is clear that by pricing $E(G_i)$E(G_i) we extend the lower matrix $L'$L' to $L$L and the up...
2502.11979
Rooted Case
228,321
57
22,512,638
38
The right term is the revenue from the drivers in $B_{R_i}$B_{R_i}R_i (each driver contributes the distance between her endpoints provided it does not exceed her budget). Note that $U \dpro{} \mathit{L}$U \dpro{} \mathit{L} is a distance matrix describing the whole grid $G$G with respect to $R_i \cup \set{r}$R_i \cup \...
2502.11979
Rooted Case
228,321
58
22,512,639
39
The algorithm obtains the revenue $\max_{L} dp_{1, L, U_{\infty}}$\max_{L}L dp_{1, L, U_{\infty}}1, L, U_{\infty}\infty, i.e. chooses the biggest revenue among the lower matrices and the infinite upper matrix $U_{\infty}$U_{\infty}\infty which has $0$0's on the main diagonal and $\infty$\infty's elsewhere. It is becaus...
2502.11979
Rooted Case
228,321
59
22,512,640
40
The price assignment itself can retrieved by a simple backtracking procedure, because the algorithm stores $W$W and the best previous state from \cref{eq:dp} alongside the values of $dp$dp.
2502.11979
Rooted Case
228,321
60
22,512,641
41
\begin{lemma} The above algorithm runs in polynomial time. \end{lemma}\begin{lemma} The above algorithm runs in polynomial time. \end{lemma} The above algorithm runs in polynomial time.
2502.11979
Rooted Case
228,321
61
22,512,642
42
\begin{proof} The number of states in the dynamic programming is bounded by a product of $m$ (the number of rows in $G$) and the numbers of possible lower and upper matrices. The upper and lower matrices are respectively of size $(\omega + 1) \times (\omega + 1)$, and $\omega \times \omega$, both of which a...
2502.11979
Rooted Case
228,321
62
22,512,643
43
Calculating the value of each state is also done in polynomial time. As we have already shown, the number of different matrices $U'$ and $L'$ is polynomially bounded. $\abs{\mathcal{W}_i}$ is also polynomially bounded, because we assign a polynomial number of weights (from $\roundedSet$) to a constant ($2\omega...
2502.11979
Rooted Case
228,321
63
22,512,644
44
Since the number of states is polynomially bounded and the time to calculate the value of each state is polynomial, the algorithm runs in polynomial time. \end{proof}\begin{proof} The number of states in the dynamic programming is bounded by a product of $m$ (the number of rows in $G$) and the numbers of possible l...
2502.11979
Rooted Case
228,321
64
22,512,645
45
Calculating the value of each state is also done in polynomial time. As we have already shown, the number of different matrices $U'$ and $L'$ is polynomially bounded. $\abs{\mathcal{W}_i}$ is also polynomially bounded, because we assign a polynomial number of weights (from $\roundedSet$) to a constant ($2\omega...
2502.11979
Rooted Case
228,321
65
22,512,646
46
Since the number of states is polynomially bounded and the time to calculate the value of each state is polynomial, the algorithm runs in polynomial time. \end{proof} The number of states in the dynamic programming is bounded by a product of $m$m (the number of rows in $G$G) and the numbers of possible lower and up...
2502.11979
Rooted Case
228,321
66
22,512,647
47
Calculating the value of each state is also done in polynomial time. As we have already shown, the number of different matrices $U'$U' and $L'$L' is polynomially bounded. $\abs{\mathcal{W}_i}$\abs{\mathcal{W}_i}\mathcal{W}_i is also polynomially bounded, because we assign a polynomial number of weights (from $\...
2502.11979
Rooted Case
228,321
67
22,512,648
48
Since the number of states is polynomially bounded and the time to calculate the value of each state is polynomial, the algorithm runs in polynomial time.
2502.11979
Rooted Case
228,321
68
22,512,649
49
The above algorithm can be easily extended to work with incomplete grids (i.e. grids where some of the edges are missing). Such a grid can be modeled as a complete grid, where we require some of the edges to have infinite weight ($b_{max}+1$b_{max}max+1). Because no driver will be able to afford such edges, they will n...
2502.11979
Rooted Case
228,321
69
22,512,650
0
\newcommand{\MaxGadgetDepth}{\frac{\omega^5}{4}}\frac{\omega^5}{4} \newcommand{\TwoMaxGadgetDepth}{\frac{\omega^5}{2}}\frac{\omega^5}{2} \newcommand{\FourMaxGadgetDepth}{\omega^5}\omega^5
2502.11979
Decomposition
228,322
70
22,512,652
2
The algorithm first partitions the drivers using a recursive decomposition of the grid $G$G. At each level of the decomposition, the grid is split into blocks, which are $\omega$\omega-wide subgrids (continuous subsequences of rows). At the first level $\mathcal{L}_1$\mathcal{L}_1, the decomposition is trivial: the who...
2502.11979
Decomposition
228,322
72
22,512,653
3
Note that, although the blocks at a single level of the decomposition do not necessarily cover the whole grid, they are disjoint. Also, all blocks across all $\mathcal{L}_j$\mathcal{L}_j's form a laminar family, i.e., any two blocks are either disjoint or one is fully contained in the other. Furthermore, at the last le...
2502.11979
Decomposition
228,322
73
22,512,654
4
\begin{remark} Block lengths at the same level of the decomposition differ by at most one. \end{remark}\begin{remark} Block lengths at the same level of the decomposition differ by at most one. \end{remark} Block lengths at the same level of the decomposition differ by at most one.
2502.11979
Decomposition
228,322
74
22,512,655
5
\begin{proof} At the first level, we only have one block, so the claim holds naturally. At each of the following levels, assuming that we have blocks of sizes belonging to the set $\set{s, s+1}$ for some $s$, in the next level we will have blocks of sizes ranging from $\cefrac{s}{2}$ to $\flfrac{s-1}{2}$. It is because...
2502.11979
Decomposition
228,322
75
22,512,656
6
\begin{definition} A driver $(u,v,b)$ is \emph{assigned} to a block $H$ iff $H$ is the smallest block containing vertices $u$ and $v$. We will denote a multiset of those drivers as $B_H$. \end{definition}\begin{definition} A driver $(u,v,b)$ is \emph{assigned} to a block $H$ iff $H$ is the smallest block containing ver...
2502.11979
Decomposition
228,322
76
22,512,657
7
For each driver such a block is unique (from laminarity) and must exist because the whole grid is one of the blocks. Also, for a driver $(u, v, b)$(u, v, b) assigned to $H$H, $u$u and $v$v must lie on opposite sides of the \emph{middle row} of $H$H (or on the middle row itself).
2502.11979
Decomposition
228,322
77
22,512,658
8
All drivers assigned to blocks from $\mathcal{L}_j$\mathcal{L}_j form $B_j$B_j. From now on we focus on a single level of the decomposition $j$j and present a constant factor approximation algorithm for the problem restricted to $B_j$B_j. In other words, we will arrive at a price assignment to all edges in $G$G whose r...
2502.11979
Decomposition
228,322
78
22,512,659
9
\subsection{Separating Subproblems}
2502.11979
Decomposition
228,322
79
22,512,660
10
Intuitively, we would like to split the problem restricted to $B_j$B_j into subproblems corresponding to the blocks of $\mathcal{L}_j$\mathcal{L}_j (fragmentation of $G$G) and drivers assigned to them (partition of $B_j$B_j). Ideally, we would solve those subproblems independently and each of the solutions would yield ...
2502.11979
Decomposition
228,322
80
22,512,661
11
This, however, is not possible as the blocks of $\mathcal{L}_j$\mathcal{L}_j together with assigned drivers do not form independent subproblems. Although, for any $H$H, both endpoints of all paths desired by a driver in $B_H$B_H are in block $H$H, a cheapest path desired by her may contain edges outside of $H$H. The so...
2502.11979
Decomposition
228,322
81
22,512,663
12
However, it turns out that extending a subproblem by adding only a constant number of rows adjacent to the corresponding block is enough. This is because we can model any finite-length grid using a relatively shallow grid with the same width. In other words, for any block $H$H, it makes no difference to the revenue whe...
2502.11979
Decomposition
228,322
82
22,512,664
13
\begin{restatable}{lemma}{compressionlemma}\label{lemma:compression} For any weighted grid of width $\omega$ there exists another weighted grid of the same width and depth at most $\MaxGadgetDepth$ such that the distances between vertices in the first row of the original grid are preserved. \end{restatable}...
2502.11979
Decomposition
228,322
83
22,512,665
14
For the sake of brevity, here we only present the main idea of the proof, while the full proof can be found in \cref{apx:compression_lemma}. First, we identify a partial grid that has a simple structure, but maintains the distances between the vertices in the first row. We realize this by finding a collection o...
2502.11979
Decomposition
228,322
84
22,512,666
15
\begin{restatable}{lemma}{crossingvtexlemma} \label{lemma:crossing_vertices} For any weighted graph $G$ and a subset $S$ of its vertices, there exists a collection of shortest paths in $G$ between all pairs of vertices from $S$ that results in at most ${\abs{S} \choose 2} \cdot \bra{{\abs{S} \choose 2} ...
2502.11979
Decomposition
228,322
85
22,512,667
16
The proof of the above, which can also be found in \cref{apx:compression_lemma}, is based on the observation that when two paths have multiple crossing vertices, one of them can be rerouted so that the number of crossing vertices is limited to two. Unfortunately, because of the lack of a natural monovariant tha...
2502.11979
Decomposition
228,322
86
22,512,668
17
Based on \cref{lemma:compression}, we create independent subproblems using an odd- and even-step approach. Let us index the blocks of $\mathcal{L}_j$\mathcal{L}_j from top to bottom. In the odd-step (even-step), we extend all the odd-indexed (even-indexed) blocks by $\MaxGadgetDepth$\MaxGadgetDepth up and down. Those e...
2502.11979
Decomposition
228,322
87
22,512,669
18
Note that some of the extensions could be shorter than $\MaxGadgetDepth$\MaxGadgetDepth. This can only happen to the top and bottom blocks which are closer than $\MaxGadgetDepth$\MaxGadgetDepth to the top or bottom end of the grid. Let $H$H be such a top (bottom) block. The extension of $H$H to the top (bottom) is shor...
2502.11979
Decomposition
228,322
88
22,512,670
19
Each of the steps (odd and even) consists of independent and disjoint subproblems, which are solved by the algorithm presented in \cref{sec:single_block}. We ensure the separation of the subproblems within the same step by pricing the edges between them at $\infty$\infty (more specifically $b_{max} + 1$b_{max}max + 1)....
2502.11979
Decomposition
228,322
89
22,512,671
20
All subproblems for blocks in all levels of decomposition apart from the last one have a special structure that we exploit in the algorithm. Because such a block $H$H is divided into two blocks by the middle row, only drivers ($u, v, b$u, v, b) whose vertices $u$u and $v$v are separated by or lying on the middle row of...
2502.11979
Decomposition
228,322
90
22,512,672
21
That does not hold for a block $H'$H' at the last level of the decomposition. However, as we observed earlier, such a block $H'$H' is at most $\FourMaxGadgetDepth + 1$\FourMaxGadgetDepth + 1 long. Even with extensions, it gives us a constant upper bound on the size of the instance associated with $H'$H' ($\frac{3}{2} \...
2502.11979
Decomposition
228,322
91
22,512,673
0
\label{sec:single_block} \newcommand{\locrevab}[2]{\mathrm{#1}^{\mathrm{#2}}_{H, a, b}}\mathrm{#1}^{\mathrm{#2}}\mathrm{#2}_{H, a, b}H, a, b \newcommand{\locreva}[2]{\mathrm{#1}^{\mathrm{#2}}_{H, a}}\mathrm{#1}^{\mathrm{#2}}\mathrm{#2}_{H, a}H, a \newcommand{\locrev}[2]{\mathrm{#1}^{\mathrm{#2}}_{H}}\mathrm{#1}^{\mathr...
2502.11979
Algorithm for a Single Block
228,323
92
22,512,674
1
In this section, we focus on a single block $H$H from a certain level of the decomposition $\mathcal{L}_j$\mathcal{L}_j. Blocks on the last level of the decomposition will be solved naively, as described in \cref{sec:decomp}, so we do not concern ourselves with them here. We consider the instance of our problem consist...
2502.11979
Algorithm for a Single Block
228,323
93
22,512,675
2
To begin, note that, by construction of the decomposition, for each driver $(u, v, b) \in B_H$(u, v, b) \in B_H, we have that one of $u$u and $v$v lies on or above the middle row of $H$H and the other on or below the middle row of $H$H. As a result, each $u$u-$v$v path contains at least one vertex from the middle row o...
2502.11979
Algorithm for a Single Block
228,323
94
22,512,676
3
Consider any $u$u-$v$v path and let $x$x be the first vertex from the middle row of $H$H on the path and $y$y be the last vertex from the middle row of $H$H on the path. Then, the path can be split into three sub-paths $u$u-$x$x, $x$x-$y$y and $y$y-$v$v, which we call the \emph{upper section}, \emph{middle section} and...
2502.11979
Algorithm for a Single Block
228,323
95
22,512,677
4
Consider an optimal solution $S^*$S^* for our instance. For each driver $(u, v, b) \in B_H$(u, v, b) \in B_H, there could be multiple lowest-cost $u$u-$v$v paths under the pricing $S^*$S^* --- choose one such path for each driver. Let $M$M be the set of vertices on the middle row of $H$H. For any two vertices $s, t \in...
2502.11979
Algorithm for a Single Block
228,323
96
22,512,678
5
From now on, consider two fixed vertices on the middle row $s, t \in M$s, t \in M. We want to construct a solution $S_{s, t}$S_{s, t}s, t generating revenue at least $\alpha R_{S^*}(B_{s, t}^*)$\alpha R_{S^*}S^*(B_{s, t}s, t^*). To achieve this, let us more closely analyze the paths of drivers in $B_{s, t}^*$B_{s, t}s...
2502.11979
Algorithm for a Single Block
228,323
97
22,512,679
6
Edges from steps 1-3 are marked with black dotted lines in \cref{fig:pricing_scheme}, whereas the ones from step 4 are marked green.
2502.11979
Algorithm for a Single Block
228,323
98
22,512,680
7
By the previous observations, we have that $R_{S_{s, t}'}(B_{s, t}^*) = R_{S^*}(B_{s, t}^*)$R_{S_{s, t}'}S_{s, t}s, t'(B_{s, t}s, t^*) = R_{S^*}S^*(B_{s, t}s, t^*), i.e., drivers in $B_{s, t}^*$B_{s, t}s, t^* generate exactly the same revenue under $S^*$S^* and $S'_{s, t}$S'_{s, t}s, t. Consequently, it would be enough...
2502.11979
Algorithm for a Single Block
228,323
99
22,512,681
8
\begin{enumerate}[nosep] \item For $i = \mathrm{up}$, note that this is symmetric with $i = \mathrm{low}$, which we treat below. \item For $i = \mathrm{mid}$, note that a revenue of at least $R_{S_{s, t}'}^\mathrm{mid} (B_{s, t}^*)$ can be obtained by a pricing derived from $S_{s, t}'$ by leaving the price of e...
2502.11979
Algorithm for a Single Block
228,323
100
22,512,683
10
\begin{lemma} For the instance defined by drivers $B_H$ and the extended block $H_{\mathrm{ext}}$, let $\mathrm{REV}_H$ be the revenue of the solution returned by the described algorithm and $\mathrm{OPT}_H$ be the revenue of an optimal solution. Then: \[ \mathrm{REV}_H \geq \frac{1}{\Crounding \cdot 3 \cdot \omega^2}...
2502.11979
Algorithm for a Single Block
228,323
102
22,512,684
11
Plugging back into our main algorithm, we gain another factor of $2$2 from the odd-even splitting on each level, and a logarithmic factor from the decomposition, so our algorithm achieves an approximation factor of $24 \omega^2 \log m$24 \omega^2 \log m in poly-time for any fixed $\omega$\omega.
2502.11979
Algorithm for a Single Block
228,323
103
22,512,685
0
Having established our result for grids with bounded width, it would be interesting to see if our ideas extend to the case of bounded-pathwidth or bounded-treewidth graphs, or to grids with some edges removed. Moreover, solving grids without a bound on their width seems like a natural first step to understanding the co...
2502.11979
Conclusion and Future Work
228,324
104