3/2-Approximation for the Matching Augmentation Problem

Abstract

We describe a$\frac{3}{2}$-approximation algorithm for the Matching Augmentation Problem, which is a special case of the weighted 2-edge-connected spanning subgraph problem. This improves upon the previous best ratio$\frac{13}{8}$.

1Introduction

The following is a well-studied problem in network design: Given an undirected simple graph$G=(V,E)$,find a 2-edge-connected spanning subgraph (2-ECSS) of$G$with minimum number of edges. We denote this problem briefly as2-ECSS.It remainsNP-hard andAPX-hard even for subcubic graphs[10].After a series of improvements beyond the trivial approximation factor$2$[8,18,21,26],the current best approximation factor for the problem is$\frac{4}{3}-\epsilon$for some constant$\frac{1}{130}>\epsilon>\frac{1}{140}$[14].The generalization of this problem in which there is a cost function$c:E\rightarrow\mathbb{Q}_{\geq 0}$,which we denote byWeighted 2-ECSS,admits$2$-approximation algorithms[19,21].Intermediate problems between2-ECSSandWeighted 2-ECSShave received tremendous attention in the last decade. The most difficult of them is the Forest Augmentation Problem (FAP) in which the cost function is defined as$c:E\rightarrow\{0,1\}$,and the zero-cost edges form a forest. A recent result[16]improves the approximation ratio to$1.9973$forFAP.

Improving the factor$2$forWeighted 2-ECSSis a major open problem. In particular, a result by Cheriyan et al.[7]implies that the integrality gap of the natural LP relaxation for the problem is lower bounded by$\frac{3}{2}$,and it is likely that there is an approximation algorithm with the same ratio. As the problem is wide open however, even further special cases beyondFAPhave been considered. One of them is the Matching Augmentation Problem (MAP) in which the zero-cost edges form a matching. This problem admits approximation ratios$\frac{7}{4}$[4],$\frac{5}{3}$[3],and$\frac{13}{8}$[15].A further special case is the Tree Augmentation Problem (TAP)in which the zero-cost edges form a tree. Several results with ratios better than$2$have appeared in the literature including[1,5,6,9,11,12,13,17,20,22,23,24,25,27,28].The current best approximation ratio attained is$1.393$[2].The purpose of this paper is to prove the following theorem.

Theorem 1.

There exists a polynomial-time$\frac{3}{2}$-approximation algorithm forMAP.

2Preliminaries

We will use the lower bound derived from the dual of the natural LP relaxation forMAP.Here,$\delta(S)$denotes the set of edges with one end in the cut$S$and the other not in$S$.

minimize $$\sum_{\begin{subarray}{c}e\in E\end{subarray}}c(e)x_{e}$$ (MAP)

subject to $$\sum_{\begin{subarray}{c}e\in\delta(S)\end{subarray}}x_{e}\geq 2,\qquad \forall\,\emptyset\subset S\subset V,$$

$$1\geq x_{e}\geq 0, \forall\,e\in E.$$

The following is the dual of (MAP).

maximize $$\sum_{\begin{subarray}{c}\emptyset\subset S\subset V\end{subarray}}2y_{S}-\sum_{e\in E}z_{e}$$ (MAP-D)

subject to $$\sum_{\begin{subarray}{c}S:e\in\delta(S)\end{subarray}}y_{S}\leq c(e)+z_{e},\qquad \forall e\in E,$$

$$y_{S}\geq 0, \forall\,\emptyset\subset S\subset V,$$

$$z_{e}\geq 0, \forall\,e\in E.$$

We assume that the input graph$G$is 2-connected, since the value of an optimal solution forMAPis the sum of those of blocks (maximal 2-connected subgraphs), and one can argue the approximation ratio only within a block. Given a set of edges$F\subseteq E$,we define$c(F):=\sum_{e\in F}c(e)$.Given a vertex$v\in V$and a 2-ECSS$F$,the degree of$v$on$F$is denoted by$deg_{F}(v)$.The vertex$v$is called adegree-$d$vertexon$F$if$deg_{F}(v)=d$,and ahigh-degree vertexon$F$if$deg_{F}(v)\geq 3$.For a path$P=v_{1}v_{2}\ldots v_{k-1}v_{k}$,$v_{1}$and$v_{k}$are theend verticesof$P$,and all the other vertices are theinternal verticesof$P$.In particular,$v_{2}$and$v_{k-1}$are theside verticesof$P$.A path whose internal vertices are all degree-$2$vertices on$F$is called aplain pathon$F$.A maximal plain path is called asegment.The length of a segment is the number of edges on the segment. If the length of a segment is$\ell$,it is called an$\ell$-segment.A$1$-segment is also called atrivial segment.If a 2-ECSS remains feasible upon removal of the edge of a trivial segment, the edge is calledredundant.Given a 2-vertex-connected spanning subgraph (2-VCSS)$F$,if the removal of a segment from$F$violates feasibility, it is called aweak segmenton$F$,otherwise astrong segmenton$F$.A set of edges$H\subseteq F$is called aspecial maximal setin$F$if it satisfies the following: (1) It consists of zero-cost edges; (2) Contracting the edges in$H$does not introduce a trivial segment on$F$;(3)$H$is maximal.

3The Algorithm forMAP

Figure 1:An example of an improvement operation

Figure 2:An example of an improvement operation

Thefirst stepof the algorithm computes an inclusion-wise minimal 2-VCSS$F$(Recall that we assume$G$is$2$-connected). In thesecond step,the algorithm contracts all the zero-cost edges in$E\setminus F$and a special maximal set of edges in$F$to form the graph$G^{\prime}=(V^{\prime},E^{\prime})$.The rest of the algorithm is assumed to run on the blocks of$G^{\prime}$,as we did for$G$.Given this,$F$is updated to an inclusion-wise minimal 2-VCSS of$G^{\prime}$computed as follows: Start from$F=E^{\prime}$and perform deletion of edges one by one starting from the unit-cost edges, followed by zero-cost edges, as long as feasibility is maintained. In thethird step,the algorithm recursively modifies the running solution$F$viaimprovement processes.Given a strong$2$-segment$S$on$F$and its internal vertex$u$,let$N_{E}(u)$denote the set of edges incident to$u$in$E$.An improvement process first tries to replace$F$by$(F\setminus B)\cup A$while maintaining feasibility, where$A\subseteq E\setminus F$is a set of$k$edges called acritical edge set,and$B\subset F$is a set of$k+1$edges,$1\leq k\leq 2$.We seek such$A$to be a subset of$N_{E\setminus F}(u)$.If there is such$A$,the described operation is called animprovement operation.Two improvement operations and the corresponding critical edge sets are given in Figure1and Figure2,where all the included and the excluded edges are of unit-cost.

1 // First step

2 Let$F$be an inclusion-wise minimal 2-VCSS of$G$

3

4// Second step

5 Contract all the zero-cost edges in$E\setminus F$and a special maximal set in$F$to obtain$G^{\prime}=(V^{\prime},E^{\prime})$

6 Let$F$be an inclusion-wise minimal 2-VCSS of$G^{\prime}$,which prioritizes zero-cost edges over unit-cost edges

7

8// Third step: Improvement operations

9 whilethere is a strong$2$-segment$S$on$F$such that no improvement process has been called on its internal vertex$u$do

10 Improvement-Process($G^{\prime},F,S,u$)

11// Fourth step: Clean-up

12 Delete redundant edges from$F$to obtain$\overline{F}$

return$(F,\overline{F})$

Algorithm 1MAP$(G(V,E))$

1 ifthere is an improvement operation that can be performed on$u$then

2 Apply the improvement operation on$u$

3 return

4foreach critical edge set$A$incident to$u$do

5 Let$\mathcal{S}$be the set of strong$2$-segments on$F\cup A$that do not exist on$F$

6 foreach strong$2$-segment$T$in$\mathcal{S}$and the internal vertex$v$of$T$do

7 ifno improvement process has been called on$(T,v)$then

8 Improvement-Process($G^{\prime},F\cup A,T,v$)

9 ifthere is an improvement operation performed inImprovement-Process($G^{\prime},F\cup A,T,v$)then

10 Perform deletion operation on$F\cup A$in the order$F$,$A$

11 return

12ifthere is no improvement operation performed in any of the recursive calls abovethen

13 Restore$F$to the original set considered before the function call

Algorithm 2Improvement-Process($G^{\prime},F,S,u$)

Figure 3:An example of an improvement process of recursion depth$2$

If no improvement operation can be performed, fixing a critical edge set$A$incident to$u$,the algorithm checks if$F\cup A$contains new strong$2$-segments that do not exist on$F$.If it does, it calls the procedure described above for$S$and$u$recursivelyon the internal vertices of the newly appearing strong$2$-segments provided that no improvement process has been previously called on the internal vertex of a given segment. These calls are performed for all$A$on$u$and for each$u\in S$.If there is an improvement operation in one of the recursive calls, the called function returns and the caller performs a specificdeletion operationas follows. It attempts to delete the edges from$F\cup A$in the order$F$,$A$.Specifically, it deletes an edge as long as the residual graph remains feasible. This enforces to keep the edges in$A$in the solution. An examples of this operation is given in Figure3,where all the included and excluded edges are of cost$1$,and the depth of the recursion tree is$2$After the reverse-delete operation, the current function call returns. If after all the recursive calls from$u$there is no improvement operation performed, the solution$F$is restored back to the original one before the function call on$u$.The main iterations continue until there is no$S$and$u$on which we can perform an improvement process.

The fourth step of the algorithm excludes all the redundant edges, i.e., edges whose removal does not violate 2-edge-connectivity. The result of this operation is denoted by$\overline{F}$.

Proposition 2.

There exists a feasible solution$F^{\prime}$for$G^{\prime}$such that

• $c(F^{\prime})=c(\overline{F})$.

• $F^{\prime}$can be obtained from$\overline{F}$in linear time.

• The union of$F^{\prime}$and the set of contracted edges in the second step of the algorithm is a feasible solution for$G$.

Proof.

This follows from the fact that the zero-cost edges form a matching in the input graph. In particular, one can select the appropriate elements from all double edge pairs in$\overline{F}$to obtain$F^{\prime}$,so that their union with the deleted zero-cost edges is feasible for$G$.It is clear that this can be performed in linear time, and$c(F^{\prime})=c(\overline{F})$. ∎

Proposition 3.

Algorithm 1 can be implemented in polynomial-time.

Proof.

It is clear that the the second and the fourth step of the algorithm can be implemented in polynomial-time. The computation of a minimal 2-VCSS in the first step is also obvious. The operations on strong$2$-segments followed by this takes$O(|E|)$steps, since an excluded edge of cost$1$is never included again. To see that the third step also takes polynomial time, it suffices to see that the main loop ofImprovement-Processterminates in polynomial number of operations. There are polynomially many critical edge sets$A$,since$|A|$is constant. Starting from the internal vertex$u$of a strong$2$-segment$S$,consider the recursion tree in which each node represents a recursive function call. By definition, each node of this tree is associated to the internal vertex of a strong$2$-segment. A vertex can be an internal vertex of a single strong$2$-segment. This implies that the number of nodes in the tree is polynomially bounded. So the main loop ofImprovement-Processterminates in polynomial number of operations. ∎

4Proof of Theorem1

Let$opt(G)$denote the value of an optimal 2-ECSS on$G$,and$opt(G^{\prime})$denote the value of an optimal 2-ECSS on$G^{\prime}$,the result of the second step of the algorithm. Let$(F,\overline{F})$be a solution returned by Algorithm 1,$F_{0}$be the result of the first step of the algorithm, and$F_{0}^{\prime}$be the result of the second step of the algorithm.

Lemma 4.

$opt(G)\geq opt(G^{\prime})$.

Proof.

Take an optimal 2-ECSS$O$on$G$.Contract the zero-cost edges in$O\cap(E\setminus F_{0})$and the intersection of$O$with the special maximal set in$F_{0}$used by the algorithm. Let the result be$O^{\prime}$.By definition then,$O^{\prime}$is a feasible solution for$G^{\prime}$,which implies the result. ∎

Lemma 5.

For any vertex$v\in V^{\prime}$,there is no zero-cost edge in$E^{\prime}\setminus F_{0}^{\prime}$.

Proof.

Recall that the second step of the algorithm contracts the zero-cost edges in$E\setminus F_{0}$and a special maximal set of edges in$F_{0}$.Given this, a zero-cost edge$f\in E^{\prime}\setminus F_{0}^{\prime}$must belong to$F_{0}$.If$f$is an edge of a$2$-segment on$F_{0}$,then it remains in$F_{0}^{\prime}$by the reverse-delete procedure in computing$F_{0}^{\prime}$,which keeps zero-cost edges. The edge$f$cannot belong to another segment in$F_{0}$either, since this implies that it belongs to a special maximal set. Thus, we have that there is no zero-cost edge$f\in E^{\prime}\setminus F_{0}^{\prime}$. ∎

Lemma 6.

Let$S$be a strong$\ell$-segment with$\ell\geq 3$,or a weak segment on$F$.Then none of the edges of$S$can be zero-cost.

Proof.

If there is such a zero-cost edge in$F_{0}$,then it belongs to a special maximal set in$F_{0}$,and hence contracted in the second step of the algorithm. This implies that there is no such edge in$F_{0}^{\prime}$.Since there is no zero-cost edge in$E^{\prime}\setminus F_{0}^{\prime}$by Lemma5,it follows that there is no such edge in$F$either. ∎

Lemma 7.

There exists$G_{1}$,a 2-VCSS$F_{1}\subseteq E(G_{1})$,and a 2-ECSS$\overline{F_{1}}\subseteq E(G_{1})$such that the following hold:

1. $\overline{F_{1}}$is obtained by deleting redundant edges from$F_{1}$.

2. Given the internal vertex$s$of a strong$2$-segment on$F_{1}$,there is no edge$e\in E(G_{1})\setminus F_{1}$incident to$s$.

3. $F_{1}$is minimal with respect to inclusion.

4. $\frac{c(\overline{F_{1}})}{opt(G_{1})}\leq\frac{3}{2}\Rightarrow\frac{c(\overline{F})}{opt(G^{\prime})}\leq\frac{3}{2}$.

Proof.

We reduce$G^{\prime}$to$G_{1}$and$F$to$F_{1}$by performing a series of operations. Let$S$be a strong$2$-segment on$F$,and$s$be its internal vertex. Let$O$be an optimal 2-ECSS on$G^{\prime}$.Then$O$contains two edges incident to$s$,say$e_{1}$and$e_{2}$.Assume it contains a third edge$e_{3}$incident to$s$.Let the other end vertices of these edges be$w_{1},w_{2}$,and$w_{3}$,respectively. If$O$contains all the edges incident to$w_{i}$that are in$F$,we call$w_{i}$aspecial vertex,for$i=1,2,3$.Note that none of$e_{1}$,$e_{2}$,and$e_{3}$is a zero-cost edge by the construction of$G^{\prime}$.To finish the proof of the lemma, we need the following two claims.

Claim 8.

We can switch to an optimal solution$O$such that there is at most one special vertex in the set$\{w_{1},w_{2},w_{3}\}$.

Proof.

Assume without loss of generality that$w_{1}$and$w_{2}$are special vertices. Then by the structure of a 2-ECSS, we can discard$e_{1}$or$e_{2}$from$O$without violating feasibility. ∎

Claim 9.

There exists an optimal 2-ECSS$O^{\prime}$on$G^{\prime}$such that$O^{\prime}$contains$2$edges incident to$s$.

Proof.

By Claim8,there are at least two vertices in the set$\{w_{1},w_{2},w_{3}\}$that are not special. Let two of them be without loss of generality$w_{2}$and$w_{3}$.By the structure of a 2-ECSS, one of these vertices, say$w_{2}$,satisfies the following. There is a neighbor$w_{2}^{\prime}$of$w_{2}$such that$f=(w_{2},w_{2}^{\prime})\in F\setminus O$,and$O^{\prime}=O\cup\{f\}\setminus\{e_{2}\}$is another optimal solution. In this case the degree of$s$on$O^{\prime}$is$2$,which completes the proof. ∎

Let$O^{\prime}(S)$be the set of edges in this solution incident to the internal vertex of$S$.Let$F^{\prime}=F\cup O^{\prime}(S)\setminus P$be a minimal 2-VCSS on$G$,where$P\subseteq F\setminus O^{\prime}(S)$.Let$E(S)$denote the set of edges incident to the internal vertex of$S$on$E(G^{\prime})$,which excludes the edges in$O^{\prime}(S)$,and note that it contains$P$.Delete the edges in$E(S)$from$G^{\prime}$to obtain$G^{\prime\prime}$.Perform these operations, including the switch to an optimal solution implied by Claim9,recursively on the new strong$2$-segments that appear on$F^{\prime}$,which we callemerging segments.Note that since none of the aforementioned vertices$w_{1}$,$w_{2}$,and$w_{3}$can be an internal vertex of a strong$2$-segment due to an improvement operation, the switch from$O$to$O^{\prime}$cannot be reversed. After the recursion starting from$S$terminates, continue performing the described operations on the strong$2$-segments on the residual solution and the graph. Let the results be$F_{1}$and$G_{1}$.Let$\overline{F_{1}}$be a 2-ECSS obtained by deleting the redundant edges from$F_{1}$,so that the first claim of the lemma holds. Note that the second claim of the lemma also holds, since there is no edge in$E(G_{1})\setminus F_{1}$incident to the internal vertex of a strong$2$-segment on$F_{1}$by construction. The third claim of the lemma follows from Claim9.

We now show that the fourth claim holds.

Claim 10.

$c(\overline{F_{1}})\geq c(\overline{F})$.

Proof.

Let$S$be a strong$2$-segment on which we start the recursive operations above or an emerging segment. The inequality$c(P)>c(O^{\prime}(S))$derives a contradiction to the algorithm and the construction of$F_{1}$,since there is no improvement process performed on$S$that has improved the cost of the solution. In particular, by all the listed improvement operations we cannot have the configurations on the left hand sides of Figure1-Figure3.We thus have$c(P)\leq c(O^{\prime}(S))$,which implies$c(F_{1})\geq c(F)$.Note next that if$e\in F_{1}\setminus\overline{F_{1}}$,then we also have$e\in F\setminus\overline{F}$by construction. This implies$c(F)-c(\overline{F})\geq c(F_{1})-c(\overline{F_{1}})$,which is equivalent to$c(\overline{F_{1}})-c(\overline{F})\geq c(F_{1})-c(F)\geq 0$. ∎

We next note that$opt(G_{1})\leq opt(G^{\prime})$.This follows from our construction ensuring that there is an optimal solution$O$such that for any strong$2$-segment$S$on$F_{1}$,$E(S)$does not contain any edge from$O$.Combining this with Claim10,we obtain$\frac{c(\overline{F_{1}})}{opt(G_{1})}\geq\frac{c(\overline{F})}{opt(G^{\prime})}$,which implies the fourth claim of the lemma, and completes the proof. ∎

Remark 11.

Let$S$be a strong$2$-segment on$F_{1}$.Then at most one edge of$S$can be zero-cost.

Proof.

Follows from the fact that zero-cost edges form a matching in the input graph. ∎

Lemma 12.

Let$S$be a strong$\ell$-segment with$\ell\geq 3$,or a weak segment on$F_{1}$.Then none of the edges of$S$can be zero-cost.

Proof.

Follows from Lemma6and the fact that the operations performed in the proof of Lemma7do not introduce a zero-cost edge on$F_{1}$. ∎

Lemma 13.

$$\frac{c(\overline{F_{1}})}{opt(G_{1})}\leq\frac{3}{2}.$$

Proof.

We construct a feasible dual solution in (MAP-D) with total value at least$\frac{2}{3}c(\overline{F_{1}})$.Given an internal vertex$u$of a strong$2$-segment on$F_{1}$,we assign$y_{\{u\}}=1$.At most one of the edges of a strong$2$-segment can be zero-cost by Remark11.In order to maintain feasibility, if such a segment has a zero-cost edge$e$,we set$z_{e}=1$.For any internal vertex$v$of a strong$\ell$-segment on$F_{1}$with$\ell\geq 3$or a weak segment, we assign$y_{\{v\}}=1/2$.Note that these assignments form a feasible solution in (MAP-D) by Lemma5,Lemma12,and the second claim of Lemma7.We will be tacitly assuming these facts in the rest of the proof while enlarging the dual assignment.

We distinguish a dual value we assign and its contribution in the objective function of (MAP-D), which is twice the dual value. The latter is called thedual contribution.We use a cost sharing argument, so that the cost of a specific set of edges is countered with a unique set of dual contributions with ratio at least$\frac{2}{3}$,which establishes the main result. In doing so, we will also make sure that we count all the$z$dual values of edges exactly once. For a given specific set of edges, we call the ratio of the dual value with the cost thecover ratio.If the cover ratio is at least$1$,we say that the set isoptimally covered.

We first describe the argument on the strong segments. Given a strong segment$S$on$F_{1}$and a side vertex$s$of$S$:

• If$S$is a$2$-segment, then both of its edges might be of cost$1$.In this case the dual contribution$2$of$y_{\{s\}}$results in a cover ratio of$1$.If one of the edges has cost$1$,and the other edge$e$is a zero-cost edge, then$2y_{\{s\}}-z_{e}=1$,again optimally covering the edges of the segment.

• If$S$is an$\ell$-segment with$\ell\geq 3$,then the dual contribution of$y$-dual variables of the internal vertices of$S$is$\ell-1$,which results in a cover ratio$\frac{\ell-1}{\ell}\geq\frac{2}{3}$.

Given a weak$\ell$-segment, the dual contribution of its internal vertices is also$\ell-1$.We impose that this contribution pays for the cost of$\ell-1$of the weak segment, thus covering the cost$\ell-1$of the segment optimally. Since the cover ratio of the strong segments is at least$\frac{2}{3}$,it suffices to show that the remaining cost of$1$of each weak segment is covered with cover ratio at least$\frac{2}{3}$.

We argue by induction on the number of weak segments$k$.We consider the base case$k=2$in which the two weak segments do not share a common end vertex. Let$u$be an end vertex of a weak segment. If$u$is not shared by any strong$2$-segment, we assign$y_{\{u\}}=1/2$.Figure4(a)shows this configuration with all the end vertices applied this assignment. Otherwise, let$S$be a set satisfying the following:

1. $u\in S$,

2. $S$consists of vertices of strong$2$-segments,

3. $S$induces a connected subgraph,

4. $S$is maximal.

We call$S$anaugmented set on$u$.Assign$y_{S}=1/2$.Note that this is also feasible by the stated properties of$S$and the second claim of Lemma7.We call the dual variables$y_{\{u\}}$and$y_{S}$augmented dual variables.It is clear that in the base case, there are at least$2$augmented dual variables, thereby optimally covering the weak segments.

(a)

(b)

(c)

Figure 4:

In the inductive step one may introduce one, two, three, or four new weak segments by extending the graph in the induction hypothesis. All the cases are given in Figure4,Figure5,and Figure6,where we depict the extending subgraphs in their simplest form. In Figure4(b)and Figure4(c)one new weak segment is introduced. Let$u$be a newly introduced high-degree vertex. If there is no strong$2$-segment with an end vertex$u$,we define$y_{\{u\}}=1/2$.Otherwise, we define$y_{S}=1/2$for an augmented set$S$on$u$.In either case, the new weak segment is optimally covered. In Figure5(a)two new weak segments are introduced. Let$u$and$v$be two newly introduced high-degree vertices, which are also end vertices of weak segments. If neither$u$nor$v$is an end vertex of a strong$2$-segment, we define$y_{\{u\}}=y_{\{v\}}=1/2$,which optimally covers the new weak segments. Assume without loss of generality that both of them are the end vertices of the same strong$2$-segment. Then assign$y_{S}=1/2$for an augmented set$S$on$u$.In this case we incorporate the$2$-segment into the analysis. Recall that at most one edge of a strong$2$-segment can be zero-cost. The total cost of the$2$-segment and the new weak segments is then at least$\ell+2$,where$\ell\geq 1$,and the total dual contribution including$y_{S}$and the dual value defined for the$2$-segment is$\ell+1$.This leads to a cover ratio of$\frac{\ell+1}{\ell+2}\geq\frac{2}{3}$.We do not depict the generalization of Figure5(a),analogous to the one from Figure4(b)to Figure4(c),which does not change the analysis.

In Figure5(b)two new weak segments are introduced together with at least two strong segments. The analysis is identical to that of Figure5(a),resulting in a cover ratio of at least$\frac{2}{3}$.In Figure5(c),Figure6(a),and Figure6(b)three new weak segments are introduced. In all these cases the analysis essentially reduces to that of the previous case, since there is a high-degree vertex we can assign the dual value$1/2$on, which optimally covers the extra new segment, and hence only betters the cover ratio of the previous case. Figure6(c)is also a straightforward generalization of Figure6(a),where there are four new weak segments. We do not depict the generalizations of Figure6(c),analogous to the one from Figure6(a)to Figure6(b),which does not change the analysis. This completes the induction and the proof. ∎

(a)

(b)

(c)

Figure 5:

(a)

(b)

(c)

Figure 6:

By Lemma13and the fourth claim of Lemma7,we have$\frac{c(\overline{F})}{opt(G^{\prime})}\leq\frac{3}{2}$.Theorem1now follows from Proposition2,Proposition3,and Lemma4.

5A Tight Example

(a)

(b)

(c)

Figure 7:A tight example for the algorithm: (a) Input graph; (b) A solution returned by the algorithm; (c) An optimal solution. The bold lines represent the edges of cost$1$,the other lines represent zero-cost edges.

A tight example for the algorithm is given in Figure7.The bold lines represent the edges of cost$1$,and the other lines represent zero-cost edges. The solution returned by the algorithm has cost$3k-2$,where$k$is an even integer. Note that there is no improvement that can be performed on the depicted solution. The optimal solution has cost$2k$.