Convolution and Knapsack in Higher Dimensions

In the Knapsack problem, one is given the task of packing a knapsack of a given size with items in order to gain a packing with a high profit value. In recent years, a connection to the $(\max,+)$-convolution problem has been established, where knapsack solutions can be combined by building the convolution of two sequences. This observation has been used to give conditional lower bounds but also parameterized algorithms. In this paper we want to carry these results into higher dimension. We consider Knapsack where items are characterized by multiple properties - given through a vector - and a knapsack that has a capacity vector. The packing must now not exceed any of the given capacity constraints. In order to show a similar sub-quadratic lower bound we introduce a multi-dimensional version of convolution as well. Instead of combining sequences, we will generalize this problem and combine higher dimensional matrices. We will establish a few variants of these problems and prove that they are all equivalent in terms of algorithms that allow for a running time sub-quadratic in the number of entries of the matrix. We further develop a parameterized algorithm to solve higher dimensional Knapsack. The techniques we apply are inspired by an algorithm introduced by Axiotis and Tzamos. In general, we manage not only to extend their result to higher dimension. We will show that even for higher dimensional Knapsack, we can reduce the problem to convolution on one-dimensional sequences, leading to an $\mathcal{O}(d(n + D \cdot \max{\Pi{i=1}^d{ti}, t{\max}\log t{\max}} ))$ algorithm, where $D$ is the number of different weight vectors, $t$ the capacity vector and $d$ is the dimension of the problem. Finally we also modify this algorithm to handle items with negative weights to cross the bridge from solving not only Knapsack but also Integer Linear Programs (ILPs) in general.