Polylogarithmic Approximation for Generalized Minimum Manhattan Networks

Given a set of $n$ terminals, which are points in $d$-dimensional Euclidean space, the minimum Manhattan network problem (MMN) asks for a minimum-length rectilinear network that connects each pair of terminals by a Manhattan path, that is, a path consisting of axis-parallel segments whose total length equals the pair’s Manhattan distance. Even for $d=2$, the problem is NP-hard, but constant-factor approximations are known. For $d \ge 3$, the problem is APX-hard; it is known to admit, for any $\eps > 0$, an $O(n^\eps)$-approximation. In the generalized minimum Manhattan network problem (GMMN), we are given a set $R$ of $n$ terminal pairs, and the goal is to find a minimum-length rectilinear network such that each pair in $R$ is connected by a Manhattan path. GMMN is a generalization of both MMN and the well-known rectilinear Steiner arborescence problem (RSA). So far, only special cases of GMMN have been considered. We present an $O(\log^{d+1} n)$-approximation algorithm for GMMN (and, hence, MMN) in $d \ge 2$ dimensions and an $O(\log n)$-approximation algorithm for 2D. We show that an existing $O(\log n)$-approximation algorithm for RSA in 2D generalizes easily to $d>2$ dimensions.

💡 Research Summary

The paper tackles the Generalized Minimum Manhattan Network (GMMN) problem, a natural extension of both the Minimum Manhattan Network (MMN) and the Rectilinear Steiner Arborescence (RSA) problems. In GMMN we are given a set R of n terminal pairs (s_i, t_i) in d‑dimensional Euclidean space and asked to construct a shortest rectilinear network such that each pair is connected by a Manhattan‑distance‑preserving path (i.e., a path consisting of axis‑parallel segments whose total length equals the L₁ distance between the terminals). The problem is NP‑hard already for d = 2, and for d ≥ 3 it is APX‑hard; previously only an O(n^ε) approximation (for any ε > 0) was known in higher dimensions, while constant‑factor approximations exist only in the planar case.

The authors present a new algorithmic framework that yields a polylogarithmic‑factor approximation for GMMN in any fixed dimension d ≥ 2. Their main results are:

1. An O(log^{d+1} n)‑approximation algorithm for GMMN in d dimensions.
2. An O(log n)‑approximation algorithm for the planar case (d = 2), which matches the best known bound for MMN.
3. A simple extension of the classic O(log n)‑approximation for 2‑D RSA to arbitrary dimensions, showing that RSA also admits an O(log n)‑approximation for any fixed d.

The core technical contribution is a recursive partitioning scheme that reduces GMMN to a series of stabbing problems. A stabbing problem asks for a minimum set of axis‑parallel hyperplanes that intersect (or “stab”) all the axis‑aligned boxes induced by the terminal pairs. The authors observe that each stabbing instance can be solved by invoking an RSA approximation algorithm: the hyperplanes correspond to the edges of a rectilinear arborescence that connects the terminals to a common root. By applying the known O(log n)‑approximation for RSA (which they also generalize to higher dimensions), they obtain an O(log n)‑approximation for each stabbing subproblem.

The recursive decomposition proceeds as follows. Choose a coordinate axis and split the point set at the median coordinate value, thereby creating two sub‑instances that lie on opposite sides of the median hyperplane. For each sub‑instance, the algorithm recursively solves the GMMN problem. The crucial observation is that the total cost contributed by the hyperplanes that cross the median can be bounded by the RSA‑based stabbing solution, which adds only an O(log n) factor. Since the recursion depth is at most d (the number of dimensions, because each level can be associated with a different axis), the approximation factor multiplies across levels, yielding the final O(log^{d+1} n) bound. In two dimensions the recursion depth is one, so the factor collapses to O(log n).

The paper also includes a detailed analysis of the RSA extension. By sorting terminals along each axis and constructing independent 1‑dimensional spanning trees, the authors combine these trees into a feasible rectilinear arborescence. This construction preserves the logarithmic approximation guarantee irrespective of the dimension, because the cost of each axis‑wise tree is bounded by the optimal RSA cost and the union of the trees introduces only a logarithmic overhead.

Complexity-wise, the algorithm runs in polynomial time; the dominant operations are sorting (O(n log n)) and solving the RSA subproblems (each also polynomial). The dependence on the dimension appears only in the exponent of the logarithmic factor, not in the polynomial term, making the approach practical for moderate‑dimensional instances such as 3‑D logistics networks or 4‑D VLSI routing problems.

The authors discuss implementation considerations, noting that the median splits and the RSA subroutines are highly parallelizable, which can further accelerate the method on modern multicore or GPU platforms. They also point out that the recursive structure naturally yields a hierarchical network, which may be advantageous for applications requiring multi‑level routing or incremental updates.

In the concluding section, the paper outlines several open directions: achieving a dimension‑independent constant‑factor approximation for GMMN, improving the exponent from d + 1 to d or lower, and exploiting special geometric configurations (e.g., grid‑like or clustered terminals) to obtain tighter bounds. The authors suggest that a deeper combinatorial understanding of the relationship between GMMN and RSA could lead to unified frameworks applicable to a broader class of rectilinear network design problems. Overall, the work represents a significant step forward by breaking the long‑standing barrier of only polynomial‑factor approximations in higher dimensions and delivering the first polylogarithmic‑approximation algorithm for the generalized problem.