[Paper Review] Approximating $k$-Median via Pseudo-Approximation

Li Shi, Ola Svensson|arXiv (Cornell University)|Nov 1, 2012

Complexity and Algorithms in Graphs22 references4 citations

TL;DR

This paper presents a novel approximation algorithm for the $k$-median problem that achieves a $1+\sqrt{3}+\epsilon$ approximation ratio, improving upon the previous best of $3+\epsilon$. The approach leverages a pseudo-approximation framework: it first shows that any algorithm producing a $\alpha$-approximate solution using $k+O(1)$ facilities can be transformed into a $\alpha$-approximation using exactly $k$ facilities, and then designs such a pseudo-approximation algorithm with the improved ratio.

ABSTRACT

We present a novel approximation algorithm for $k$-median that achieves an approximation guarantee of $1+\sqrt{3}+ε$, improving upon the decade-old ratio of $3+ε$. Our approach is based on two components, each of which, we believe, is of independent interest. First, we show that in order to give an $α$-approximation algorithm for $k$-median, it is sufficient to give a \emph{pseudo-approximation algorithm} that finds an $α$-approximate solution by opening $k+O(1)$ facilities. This is a rather surprising result as there exist instances for which opening $k+1$ facilities may lead to a significant smaller cost than if only $k$ facilities were opened. Second, we give such a pseudo-approximation algorithm with $α= 1+\sqrt{3}+ε$. Prior to our work, it was not even known whether opening $k + o(k)$ facilities would help improve the approximation ratio.

Motivation & Objective

• To overcome the longstanding $3+\epsilon$ approximation barrier for the NP-hard $k$-median problem.
• To develop a new algorithmic framework that relaxes the hard constraint of opening exactly $k$ facilities, enabling improved approximation guarantees.
• To establish a general connection between pseudo-approximation algorithms (allowing $k+O(1)$ facilities) and true $k$-median approximation algorithms.
• To analyze the integrality gap of the natural LP relaxation under relaxed facility constraints, particularly when $k+O(1)$ facilities are allowed.
• To explore the potential of soft constraints and hierarchical LP relaxations (e.g., Sherali-Adams) in designing better approximation algorithms for clustering problems.

Proposed method

• Introduces the concept of a pseudo-approximation algorithm that opens $k+c$ facilities for a constant $c$, and proves that such an algorithm can be transformed into a true $k$-median approximation with the same approximation ratio.
• Develops a bi-point solution framework where the cost is analyzed as a convex combination of two extreme solutions: one opening all facilities ($d_1$) and one opening only a subset ($d_2$).
• Uses a randomized rounding scheme over stars in the solution graph, where facilities are selected with probabilities dependent on their degree and a parameter $b$, to bound expected connection costs.
• Applies a balancing argument between the bi-point solution and the trivial all-facility solution to derive a $\frac{1+\sqrt{3}+\epsilon}{2}$-approximation for the bi-point cost, which is then used to bound the overall approximation ratio.
• Employs a parameter $a = \frac{\sqrt{3}-1}{4}$ and $b = \frac{2}{1+\sqrt{3}}$ to optimize the trade-off between the two cost components in the bi-point analysis.
• Uses a $\delta$-relaxation of the LP to allow for $O(1/\epsilon)$-additive error, enabling efficient computation of the bi-point solution within the desired approximation factor.

Experimental results

Research questions

• RQ1 Can a pseudo-approximation algorithm that opens $k+O(1)$ facilities be transformed into a true $k$-median approximation without losing approximation ratio?
• RQ2 What is the maximum ratio between the cost of an optimal solution with $k+O(1)$ facilities and the value of the natural LP relaxation with $k$ facilities?
• RQ3 Does allowing $k+o(k)$ facility violations help close the integrality gap of the $k$-median LP relaxation?
• RQ4 Can the Sherali-Adams or Lasserre hierarchies be used to design better approximation algorithms for $k$-median by relaxing hard constraints?
• RQ5 Is the $1+\sqrt{3}+\epsilon$ approximation ratio tight, or can it be further improved using the proposed framework?

Key findings

• The paper achieves a $1+\sqrt{3}+\epsilon$ approximation ratio for the $k$-median problem, improving upon the previous best of $3+\epsilon$.
• It proves that any $\alpha$-approximate pseudo-approximation algorithm using $k+O(1)$ facilities can be transformed into an $\alpha$-approximation algorithm using exactly $k$ facilities.
• The integrality gap of the natural LP relaxation for $k$-median is shown to be at most $1+\sqrt{3}+\epsilon$ when $k+O(1)$ facilities are allowed, suggesting the barrier of 2 may be circumvented.
• The bi-point solution framework achieves a $\frac{1+\sqrt{3}+\epsilon}{2}$-approximation for the combined cost, which is optimal under the given parameterization.
• The algorithm runs in time $O(n^{O(1/\epsilon^2)})$, making it efficient for constant $\epsilon$.
• The analysis shows that the worst-case integrality gap instances for the $k$-median LP do not persist when $k+1$ facilities are allowed, indicating a potential path to closing the gap.