Parallelized approximation algorithms for minimum routing cost spanning trees

Let $G=(V,E)$ be an undirected graph with a nonnegative edge-weight function $w.$ The routing cost of a spanning tree $T$ of $G$ is $\sum_{u,v\in V}d_{T}(u,v)$ where $d_{T}(u,v)$ is the weight of any shortest $u$-$v$ path in $T,$ or equivalently, the weight of the simple $u$-$v$ path in $T.$ The Minimum Routing Cost Spanning Tree (MRCT) problem WLB⁺ [00] asks for a spanning tree $T$ of $G$ with the minimum routing cost. It is also known as the Shortest Total Path Length Spanning Tree problem. The MRCT problem is first proposed by Hu Hu [74], who referred to the problem as the Optimum Distance Spanning Tree Problem. In Hu’s formulation of the more general Optimum Communication Spanning Tree (OCT) problem Hu [74], an additional value $\tau_{u,v}\geq 0$ is given for each pair $(u,v)$ of vertices. The communication cost Hu [74] of a spanning tree $T$ of $G$ is $\sum_{u,v\in V}\tau_{u,v}\,d_{T}(u,v).$ The OCT problem asks for a spanning tree of $G$ with the minimum communication cost. When $G$ is a complete graph and the edge-weight function $w$ obeys the triangle inequality, a randomized $O(\log{|V|})$-approximation algorithm is known for the OCT problem Bar [98], CCGG [98], WLB⁺ [00], FRT [03]. The MRCT problem is the special case of the OCT problem when $\tau_{u,v}=1$ for all $u,v\in V.$

The MRCT problem has applications in network design Hu [74], JLK [78] as well as multiple sequences alignment in computational biology FD [87], Pev [92], Gus [93], BLP [94], WLB⁺ [00]. Unfortunately, it is shown to be ${\cal NP}$-hard JLK [78], and it is ${\cal NP}$-hard even when all edge weights are equal JLK [78], GJ [79] or when the edge-weight function obeys the triangle inequality WLB⁺ [00].

Exact and approximation algorithms for the MRCT problem have been extensively researched BFW [73], Hoa [73], DF [79], Won [80], WCT00b , WLB⁺ [00], FLS [02]. Boyce et al. BFW [73], Hoang Hoa [73] and Dionne and Florian DF [79] study branch-and-bound algorithms as well as heuristic approximation algorithms for the Optimal Network Design problem BFW [73], which includes the MRCT problem as a special case. Fischetti et al. FLS [02] give exact algorithms for the MRCT problem while avoiding exhaustive search. Wong Won [80] gives a polynomial-time $2$-approximation algorithm for the MRCT problem. That is, he gives a polynomial-time algorithm that, given a graph $G=(V,E)$ with a nonnegative edge-weight function $w,$ outputs a spanning tree of $G$ whose routing cost is at most $2$ times the minimum. Subsequent work by Wu et al. WCT00b shows a different polynomial-time $2$-approximation algorithm as well as polynomial-time $15/8,3/2$ and $(4/3+\epsilon)$-approximation algorithms for the MRCT problem, where $\epsilon>0$ is an arbitrary constant. Their results are later improved by Wu et al. WLB⁺ [00] to give a polynomial-time approximation scheme (PTAS) CLRS [01] for the MRCT problem. That is, a polynomial-time $(1+\epsilon)$-approximation algorithm is given for any constant $\epsilon>0.$

There are other variants of the MRCT problem that also have applications in network design WLB⁺ [00], WCT00a , WCT00c , Wu [02]. In the Sum-Requirement Optimal Communication Spanning Tree (SROCT) problem WCT00a, each vertex $u$ is associated with a requirement $r(u)\geq 0.$ The objective is to find a spanning tree $T$ of $G$ minimizing $\sum_{u,v\in V}(r(u)+r(v))\,d_{T}(u,v).$ The Product-Requirement Optimal Communication Spanning Tree (PROCT) WCT00a problem is to find a spanning tree $T$ of $G$ minimizing $\sum_{u,v\in V}r(u)\,r(v)\,d_{T}(u,v).$ The SROCT and PROCT problems are clearly generalizations of the MRCT problem.

Wu et al. WCT00a give a $2$-approximation algorithm for the SROCT problem. They also propose a $1.577$-approximation algorithm WCT00a for the PROCT problem. The result is improved by Wu et al. WCT00c to yield a polynomial-time approximation scheme (PTAS) for the PROCT problem.

Another variant of the MRCT problem is the $2$-MRCT problem Wu [02]. In this problem, except for $G=(V,E)$ and $w:E\to\mathbb{R}_{0}^{+},$ we are given two source vertices $s_{1},s_{2}\in V.$ The objective is to find a spanning tree $T$ of $G$ minimizing $\sum_{v\in V}d_{T}(s_{1},v)+d_{T}(s_{2},v).$ This problem is ${\cal NP}$-hard even when $w$ obeys the triangle inequality Wu [02]. Wu Wu [02] shows a $2$-approximation algorithm as well as a PTAS for this problem. A variant of the $2$-MRCT problem is the weighted $2$-MRCT problem Wu [02] where an additional $\lambda\geq 1$ is given as input. The objective is to find a spanning tree $T$ of $G$ minimizing $\sum_{v\in V}\lambda\,d_{T}(s_{1},v)+d_{T}(s_{2},v).$ Wu Wu [02] proposes a $2$-approximation algorithm for the weighted $2$-MRCT problem. When the edge-weight function $w$ obeys the triangle inequality, there is a PTAS for the weighted $2$-MRCT problem Wu [02].

In this paper, however, we will focus on parallelizing the approximation algorithms for the above problems. We first describe our results concerning the MRCT problem. For an arbitrary $\epsilon>0$ and when the edge-weight function $w$ is given in unary, we show that the $(4/3+\epsilon)$-approximation algorithm proposed by Wu et al. WCT00b can be implemented by an ${\cal RNC}$ circuit. That is, the approximation algorithm can be performed by a uniform polynomial-size circuit Pap [94] with random gates and poly-logarithmic depth. Indeed, with a small probability our parallelized algorithm may fail to find a $(4/3+\epsilon)$-approximate solution, in which case it outputs “fail.” Thus, our algorithm does not fail (to find a $(4/3+\epsilon)$-approximate solution) without ever knowing that it fails, which is a desirable property for randomized algorithms with a small probability of failure.

We now turn to describe our results concerning the SROCT problem. When the edge-weight and the vertex-requirement functions are given in unary, we parallelize the $2$-approximation algorithm WCT00a by realizing it using ${\cal RNC}$ circuits, with a slight degradation in the approximation ratio (from the currently best $2$ to our $2+o(1)$). Still, with a small probability our algorithm may fail to output a $(2+o(1))$-approximate solution, in which case it knows the failure and outputs “fail.”

Finally, for the weighted $2$-MRCT problem with the edge-weight function given in unary, we parallelize the $2$-approximation algorithm Wu [02] into ${\cal RNC}$ circuits, with a slight degradation in the approximation ratio (from the currently best $2$ to our $2+o(1)$). Again, there is a small probability that our algorithm fails to find a $(2+o(1))$-approximate solution, in which case it outputs “fail.”

To the best of our knowledge, our results are the first efforts towards parallelized approximation algorithms for the MRCT problem and its variants. Our results open up new opportunities to compute approximate solutions to the above problems in parallel poly-logarithmic time. In the applications of the MRCT problem to network design Hu [74], JLK [78] as well as the applications of the SROCT and PROCT problems to network design WCT00a , WCT00c, the network is often modeled as a graph with a nonnegative edge-weight function representing the distances between pairs of nodes. Although approximate solutions to the aforementioned problems (MRCT, SROCT, PROCT, weighted $2$-MRCT) are attainable in polynomial time, in any real networking environment, however, the cost of traffic between any pair of nodes may vary over time. Thus, it is highly desirable to be able to compute approximate solutions to these problems as fast as possible, so as to reduce the risk that the traffic costs change during the computation. Our results imply that approximate solutions to the MRCT, SROCT and weighted $2$-MRCT problems can indeed by computed in parallel poly-logarithmic time on multiprocessors.

For other applications of the MRCT problem where the data does not change quickly over time, for example multiple sequences alignment in computational biology FD [87], Pev [92], Gus [93], BLP [94], WLB⁺ [00], being able to compute approximate solutions to the MRCT problem in parallel sublinear time is still beneficial. Indeed, Fischer Fis [01] argues that in many practical applications today, the input size is so large that even performing linear-time computations is too time-consuming. Certainly, multiple sequences alignment in computational biology constitutes a good example where the input size is usually too large. It is therefore a desirable property that our algorithms operate in parallel sublinear time, and in fact poly-logarithmic time.

The main idea underlying our proofs is that many of the previously proposed approximation algorithms for the MRCT, SROCT and weighted $2$-MRCT problems rely heavily on finding shortest paths between pairs of vertices in a graph. This motivates applying the well-known result that ${\cal NL}\subseteq{\cal NC}$ to parallelize these algorithms since we can guess a path (possibly the shortest one) between two vertices of a graph in nondeterministic logarithmic space. There is the complication that, in our proofs, we will often need to generate the same shortest path between two vertices $u,v,$ whenever a shortest $u$-$v$ path is needed. For this purpose, we use the isolation technique Wig [94], GW [96], RA [00] to slightly modify the edge-weight function of the input graph, so that there is exactly one shortest path between each pair of vertices with high probability. We then apply the double-counting technique RA [00] to decide whether the input graph (with the modified edge-weight function) exhibits a unique shortest path between each pair $u,v$ of vertices. If so, we are able use the double counting technique to generate the unique shortest $u$-$v$ path whenever it is needed. The whole procedure runs in unambiguous logarithmic space and our results follow by ${\cal UL}\subseteq{\cal NL}\subseteq{\cal NC}.$ The approximation ratio would be slightly degraded. The degradation comes from randomly modifying the edge-weight function when we apply the isolation technique.

Our paper is organized as follows. Section 2 provides the basic definitions. Section 3 presents the parallelized $(4/3+\epsilon)$-approximation algorithm for the MRCT problem. Section 4–5 describe our parallelized approximation algorithms for the SROCT and the weighted $2$-MRCT problems, respectively. Section 6 concludes the paper. Proofs are given in the appendix for references.

Throughout this paper, graphs are simple undirected graphs Wes [01]. That is, we disallow parallel edges and self-loops. There will always be a nonnegative edge-weight function mapping each edge to a nonnegative real number. For a graph $G,$ $V(G)$ is its vertex set and $E(G)$ is its edge set. Let $R$ be a subgraph of $G.$ A path $P$ connects a vertex $v$ to $R$ (or $V(R)$) if one endpoint of $P$ is $v$ and the other is in $V(R).$ An edge connecting two vertices $u$ and $v$ is denoted $uv.$ A path connecting two vertices $u$ and $v$ is said to be a $u$-$v$ path. A path $(v_{0},\ldots,v_{k})$ is one which traverses $v_{0},\ldots,v_{k},$ in that order. A simple path is a path that traverses each vertex at most once Wes [01]. A graph $G$ contains another graph $G^{\prime}$ if $G^{\prime}$ is a subgraph of $G.$ The set of nonnegative real numbers is denoted $\mathbb{R}_{0}^{+}.$

In Definition 1, it is not hard to show that $G=(V,E)$ is strongly min-unique if $\left|\text{SP}^{(k)}_{u,v}\right|\leq 1$ for all $k\in\{0,\ldots,|V|-1\}$ and $u,v\in V,$ provided $|V|\geq 3.$

The MRCT of a graph, standing for its Minimum Routing Cost spanning Tree, is defined below.

The MRCT problem asks for $\text{MRCT}(G)$ on input $G,w.$ The following fact shows that the routing cost of a tree can be computed efficiently.

To ease the description, we introduce the following definition.

Throughout this paper, when a nondeterministic Turing machine $A$ runs or simulates another nondeterministic machine $B,$ it means that $A$ runs $B$ and make nondeterministic branches as $B$ does. It does not mean that $A$ enumerates all computation branches of $B$ and simulate them deterministically. For convenience, $A$ does not necessarily have to output $B$’s output. Instead, it may extract portions of $B$’s output for output.

We will need the notion of a general star to introduce the approximation algorithms for the MRCT problem.

Given any subtree $S$ of $G=(V,E),$ a general star with core $S$ exists WCT00b. This follows by observing that for any shortest path $P$ connecting $u\in V$ and $S,$ the part of $P$ from any vertex $x\in V(P)$ to $S$ constitutes a shortest path connecting $x$ and $S.$

The notion of a metric graph is defined below.

We begin with the following form of the famous isolation lemma. It is implicit in some previous works Wig [94], GW [96], RA [00].

The following theorem is implicit in RA [00]. It uses the double counting technique RA [00] similar to the inductive counting technique used to prove the Immerman-Szelepcsényi theorem Imm [88], Sze [88].

The following theorem is due to Wu et al. WCT00b.

That $R_{i,S}$ in Theorem 9 is a subtree of $G$ for $2\leq i\leq k$ is easily shown because $w(e)>0$ for each $e\in E$ by the strong min-uniqueness of $G.$ The core $R_{k,S}$ in Theorem 9 is unambiguously computable in logarithmic space on strongly min-unique connected graphs. To show this, we need the following lemma.

With Lemma 10, we are able to compute the core $R_{k,S}$ in Theorem 9 unambiguously in logarithmic space on strongly min-unique connected graphs.

With Theorem 9 and Lemma 11, it is not hard to show the following fact.

The general star with a core in Fact 12 can be computed unambiguously in logarithmic space, as the next lemma shows.

The following lemma allows unambiguous logarithmic-space computation of the routing cost of a tree.

Combining Fact 12 and Lemmas 13–14 gives the following lemma.

The following lemma will be useful.

Combining Theorem 7 and Lemma 15–16 yields the following theorem.