GRAPH AUGMENTATION PROBLEMS FOR A SPECIFIED SET OF VERTICES

Toshimasa Watanabe Yasuhiko Higashi and Akira Nakamura

Faculty of Engineering, Hiroshima University
Saijo-cho, Higashi-Hiroshima, 724 Japan

1. Introduction

The paper d i s c u s s e s the k-edge-connectivity or k - v e r t e x -

c o n n e c t i v i t y or s t r o n g - c o n n e c t i v i t y a u g m e n t a t i o n problem for a
specified set of vertices: "Given a complete undirected or directed
graph G=(V,E), a s p a n n i n g subgraph G 0=(V,E'), a cost f u n c t i o n
c : E - , Z + ( n o n n e g a t i v e integers) and a specified subset So_V, find a set
Ac-E-E' of the minimum total cost c(A) such that, for any given pair of
vertices in S, the graph G 0 = ( V , E ' w A ) has at least k edge-disjoint or at
least k internally-disjoint paths between them or at least one directed
cycle containing them, where c(e)=0 ( V e ~ E ' ) and adding multiple
edges to a graph is prohibited." They are abbreviated as k - E C A - S V , k -
V C A - S V and S C A - S V , respectively, If we set S=V then these problems
are usual augmentation problems discussed in [1,2,4-8].
In this paper we consider the most fundamental problems 2-ECA-
SV and 2-VCA-SV, that is, the case with k=2. The NP-completeness of
these problems with G O restricted to a tree and S=V have been shown
in [2]. Three O(IVt 2) approximation algorithms STC, BRC and BIC for
SCA-SV, 2-ECA-SV and 2-VCA-SV, all with S=V, were proposed in [2],
respectively. The idea of STC is as follows: First choose a vertex r and
finds a m i n i m u m cost spanning reverse arborescence Tin with the
sink r, with respect to the initial cost c. Modify c into c' by setting
c'<u,v>=0 for V < u , v > ~ E ( T i n ) and c'<w,v>=~ for any <w,r> entering into r.
Again finds a m i n i m u m cost spanning arborescence T o u t with r as the
root, with respect to c'. Clearly E"=(E(Tin)wE(Tout))-E' is a solution to
SCA-SV with S=V. Time complexity of STC is O(IVI2), and c(E") is no
more than twice the optimal. The idea of combining Tin and T o u t is
used in both BRC and BIC, since a strongly connected graph will be 2-
e d g e - c o n n e c t e d if directedness is neglected, where the handling of
cutvertices are incorporated in BIC.
We propose an O ( t V l 2) approximation algorithm BRA-SV for 2-
ECA-SV, and an O(IV31) one BIA-SV for 2-VCA-SV. It is shown that the
 379

worst approximation of BRA-SV (BIA-SV, respectively) is no more than

twice (less than four times) the optimal if G O is connected. Although
these algorithms run on G O which is disconnected, the theoretical
estimate of their worst approximations in this case is left for future
research. It should be noted that if c(e)=c(e') for any e, e'E E then BRA-
SV (BIA-SV, respectively) can be m o d i f i e d into an O ( I V I + I E ' I )
algorithm proposed in [1] ([1,4]) for 2-ECA-SV (2-VCA-SV) with S=V,
where an optimum solution is obtained in each case.
For SCA-SV we show that it is NP-complete even if every edge e of
E has c(e)=l. An approximation algorithms STA-SV and its variations
based on either shortest paths or m i n i m u m arborescens, as well as
their experimental results, are presented.
Let 2-ECA[G0=H, S=W] denote the 2-ECA-SV with G O and S set to a
graph H and a subset Wc_V, respectively. A similar notation will be
used for 2-VCA-SV.

2. Preliminaries

Technical terms not specified here can be identified in [3]. A

graph G=(V,E) consists of a set V of vertices and a set E of edges. V and
E are also written as V(G) and E(G), respectively. Definitions are given
mostly for the undirected case; those for the directed case is similar.
Two paths P and P' are e d g e - d i s j o i n t (internally disjoint,
r e s p e c t i v e l y ) if E(P)C'~E(P)=O (they have no vertices except their
endvertices in common). The edge-connectivity e c ( G ) ( t h e vertex-
c o n n e c t i v i t y vc(G), respectively) of G is the m i n i m u m n u m b e r of
edges (vertices) whose d e l e t i o n result in a graph with m o r e
components than G or in a single vertex, where a c o m p o n e n t is a
maximal c o n n e c t e d subgraph. G is k - e d g e - c o n n e c t e d ( k - v e r t e x -
c o n n e c t e d , respectively) if ec(G)___k (vc(G)_>k). It is known that
e c ( G ) _ k (vc(G)>k, respectively) if and only if, for any two vertices u,v
of G, G has at least k edge-disjoint (k internally-disjoint) (u,v)-paths,
(where a (u,v)-path is a path connecting u and v). Let G be a directed
graph. A vertex r is called a root of (a sink, respectively) of G if
there in a <r,v>-path, a directed path from r to v, (a <v,r>-path) for any
vertex v of G. A spanning arborescence with the root r (A spanning
reverse arborescence with the sink r) is a directed acyclic spanning
subgraph of G with r, as the root (as the sink), having no incoming
(outgoing) edges, and all other vertices having exactly one incoming
(outgoing) edge.
 380

3. An O(IV[2) approximation algorithm BRA-SV

We propose an O(IVt 2) approximation algorithm BRA-SV for 2-

ECA-SV and show that the worst approximation is no greater than
twice the optimal if G O is connected.
We provide an example of transforming graphs to help the reader
understand our idea. Let G O be an initial graph with edges written in
solid lines in Fig. 1.1. We first construct a complete graph G s = ( V s , E s )
and a tree T=(Vs,Es' ) (the graph with edges written in bold solid or fine
solid lines in Fig. 1.2) from G and G O, respectively, by shrinking each
bridge connected component of G O into a different vertex, and then
define a cost function c l : E s ~ Z + b y
c l(u,v)=min { {~,} u {c(u',v')l(u',v')e E-E', u'e S 1 , v'e $2} },
where bridge connected components S 1 and S 2 are shrunk into u and
v, respectively. The edge(u',v') so used is kept by a backpointer
bl(U,V)=(u',v'), where if Cl(U,V)=~ then bl(u,v)=(u,v). Let VaC-V s be the
set of vertices that correspond to bridge c.onnected components
including vertices of S. Each edge on a (u,v)-path of T with u,ve V a is
called a p a t h - e d g e , and each connected component consisting of
edges that are not path-edges is called an e x t e r n a l - p a t h - c o m p o n e n t
of T. Let G c = ( V c , E c ) and T c = ( V c , E c ' ) (Tc is written in bold solid lines in
Fig. 1.2) be the graphs determined from Gs and T s by shrinking each
external-path-component to a distinct vertex, respectively. We call T c
a p a t h - t r e e . A path-tree is a minimal subtree including all bridges
that exist in T s between any two specified vertices of Va, and every
leaf of Tc is in V a. We define a cost function c 2 : E c ~ Z + by
c 2 ( u , v ) = m i n {c 1 (u',v')lu's S',v'e S" },
where external-path-components S' and S" are shrunk into u and v,
respectively. The edge(u',v') so used is kept by a backpointer

w ,,,~* ~ ~ ~ ~ ~ ~ '. . . . . . O

Fig. 1.1. G0=(V,E'), Sc-V(denoted Fig. 1.2. Ts=(Vs,Es'), T c (bold

by e ) and A (broken bold lines), lines), VaC_V s (black circles) and a
solution to 2-ECA[G0=Ts,
S=Va],(broken fine lines).
 381

b2(u,v)=(u',v'). G c, T c and c 2 can be constructed in O(IWl 2) time. The

O(IVI 2) approximation algorithm BRC proposed in [2] for 2-ECA-SV with
S=V can be applied to 2-ECA[G0=T c, S=V c] with the cost function c 2. Let
A c be an optimal solution and A c' be an approximate one given by this
BRC, both to 2-ECA[G0=T c, S=Vc], respectively. Then [2] shows that
c2(Ac')<2c2(Ac). Neither A c or A c' has multiple edges, and
transformation by means of b 2 or b 1 does not produce multiple edges:
none of the four sets A s = b 2 ( A c ) , A = b l ( A s ) , A s ' = b 2 ( A c ' ), A ' = b l ( A s ' )
cotains multiple edges. A s and A s ' (A and A', respectively) are an
optimal solution and an approximate one to each problem. It is easy to
see that c 2 ( A c ) = C l ( A s ) = c ( A ) and c 2 ( A c ' ) = c l ( A s ' ) = c ( A ' ), showing that
c(A')<2c(A). Hence we have the following theorem.
T h e o r e m 1. The algorithm BRA-SV finds a approximate solution
for 2-ECA-SV in O(IV21) time and the worst approximation is no more
than twice the optimal if G O is connected.

4. An O([V[3) approximation algorithm BIA-SV

We propose an O(IVI 3) approximation algorithm BIA-SV for 2-

VCA-SV and show that the worst approximation is less than four times
the optimal if G O is connected.
The problem is reduced to a 2-VCA[G0=T c, S=Vc] to which the
O(IVI 2) approximation algorithm BIC proposed in [2] can be applied,
where T c is a tree with the vertex set V c defined later. An approximate

O
U

Fig. 2.2. T
s -- ( V s , E s )' and V a , (black
Fig. 2.1. G0=(V,E'), Sc-V(black spots).
circles and squares) and a solution
A (bold broken lines), where open
circles and open squares denote
block-vertices and cutvertices,
respectively.
 382

solution to the original 2-VCA-SV is obtained from a solution given by

this BIC. Since the reduction seems hard to understand at the first
reading, we provide a schematic explanation to outline our idea.
Suppose that a connected initial graph G O of Fig. 2.1 with edges
written in solid lines are given. We construct the b t o c k - c u t v e r t e x -
tree Ts=(V s, Es') from G O (Fig.2.2), where V s = V c u t w V b l k , Vcut is the set
of all cutvertices of G O and Vblk is a set of block-vertices of G O. Es' is a
set of edges that connect a block-vertex u to a cutvertex v if and only
if v is in the block Bu corresponding to u. Let G s = ( V s , E s ) be the
complete graph with Es={(u,v)l u,v~ Vs,u~v}, and define a cost function
ct:Ec~Z+w{oo} as follows, where u and v correspond to blocks B u and
Bv, respectively:
(c(u,v) if u,vE Vcu t,

Cl(U,V)=~min{{oo}w{c(u',v) l u ' e B u and u'~Vcut}} i f u e V b l k and v e V c u t ,

min{ {oo}w{c(u',v') I u'e Bu,v'e Bv and u',v'~ Vcut}} if u,ve Vblk,

A block-vertex and a cutvertex are denoted by a circle and a
square in Figures 2.1-2.4, respectively. A black spot denotes a specified
vertex of S. Let
Va'= (VcutC~S)k){u~ Vbl k I ( B u - V c u t ) ~ S ¢ O } .
We first define a tree T c = ( V c , E c ' ) , called the p a t h - t r e e , from Ts as
follows. Let Tc' be a maximal subtree determined by the edges on (u,v)-
paths of T s with u , v ~ V a ' . Let L = V c u t ~ { l e a v e s of Tc' } and delete all
vertices of L from Ts. Let Tc be the resulting tree (Fig,2.3). Replace any
cut-vertex u ~ V c of L by the block-vertex u ' ~ V c such that u is adjacent
to u' in Tc', and let Va denote the resulting set. (See u and u' in Fig.2.2,
and black spots are elements of Va in Fig.2.3) Note that every leaf of Tc
is in V a n V b l k. Let G c = ( V c , E c ) denote a complete graph.
Another tree T b = ( V b , E b ' ) is constructed from T s by means of T c.
An approximate solution to 2-VCA-SV is obtained from 2-VCA[G0=Tc,
S=Vc], and Tb is not necessary in this stage. The estimate of that
approximate solution, however, is done by way of Tb. Delete every
vertex of V c u t ~ V c from Ts. and we call each of resulting connected
c o m p o n e n t s an o u t e r - c o m p o n e n t of Ts. The term "outer-component"
is also means its vertex set for notational simplicity. Let K be any
outer-component (a vertex set) of Ts with IKI_>2. Then Ts has either a
block-vertex v K such that { v } = K t ~ V ( T c ) or a bridge ( U K , V K )
connecting a block-vertex u K ~ K (that is, U K f f V ( T c ) ) and a out-vertex
v K ~ V ( T c ) . In the latter case K is called an a t t a c h m e n t of v K . We
define the g u a s i - p a t h - t r e e T b by shrinking each o u t e r - c o m p o n e n t K
with IKI_>2 into VK if VK is a block-vertex or into UK if VK is a cut-
 383

f ~'x

L ~ - 1

Fig. 2.3. Tc=(Vc,Ec'), Va(black Fig. 2.4. Tb=(Vb,Eb' ).

spots) and a solution to 2-
VCA[G0=Tc, S=Vc] (broken fine
lines).

vertex.(Fig.2.4). If an attachment K of some cut-vertex is shrunk into a

block-vertex u then u is called an outer-point. Let K u denote the set of
vertices shrunk into outer-point u. Note that every leaf of Tb is a
block-vertex. Let G b = ( V b , E b ) be a complete graph on V b, and define a
cost function c2:Eb--*Z+w{ ~ } (a backpointer b2:Eb---)Es: is defined
similarly):
c 1 (u,v) ,if u , w V(Ts),
f
c2(u'v)=]min{c(u'v)/ I u'~, Ku} ,if u is an outer-point and w V ( T s ) ,
~min{c(u',v') I u ~ Ku,v ~ Kv} if u, v are outer-points.
We define another cost function c3:Eb---)Z*u{~} and a back-pointer
b 3 : E b ~ {paths of Gb represented by sequences of vertices} as follows.
First set c3e--c 2 and b3(u,v)e--u'v'(a sequence of vertices) for each
( u , v ) ~ E b, where b2(u,v)=(u',v'). Then, for each outer-point v f ~ V b - V c,
excute the following procedures (a),(b) and (c) in this order, where
v ~ V c denotes the cut-vertex adjacent to vf.
(a) For every pair u',v'~V b such that T b has a (u',v')-path
containing v, set
c3(u',v')e--c3(u',vf)+c3(vf,v') and b3(u',v')e--b3(u',vf)b3(vf, v')
if c3(u',vf)+c3 (vf, v')<c3(u',v'),
where b 3 ( u ' , v f ) b 3 ( v f , v ' ) denotes the concatenation of the two
sequences with the first vertex of b3(vf,v') deleted.
(b) For each edge (vf, w ) ~ E b with w~V c and w;~v, set
c 3 ( v , w ) ~ - c 3 ( v f , w) and b3(v,w)~---b3(vf,w) if c3(vf, w)<c3(v,w).
(c) For each (vf, w f ) ~ E b with w f ~ V b - V c and wf-~vf, set
c3(v,w)e--c3(vf, wf) and b3(v,w)~--b3(vf, w f ) i f c3(vf, wf)<c3(v,w),
 384

where w ~ V c is the cut-vertex adjacet to wf. Let c4:Ec--)Z+w{oo} be a cost

function on Gc defined by c 4 ( u , v ) = c } ( u , v ) for each (u,v)~Ec. We use
the backpointer b3 for c4. The O(IVI z) approximation algorithm BIC
proposed in [2] for 2-VCA-SV with S=V can be applied to 2-VCA [G0=Tc,
S = V c ] with the cost function c4. Let A c' be an approximate solution
given by this BIC. Let A, As, Ab and Ac be minimum solutions to the
original 2-VCA-SV, 2-VCA [G0=T s, S=Va], 2-VCA [G0=Tb, S=Vc] and 2-VCA
[ G 0 = T c, S=V c ], respectively. Then [21 shows that
c 4 ( A c ' ) = c 3 ( A c ' ) _ < 2 c 3 ( A c ) = 2 c 4 ( A c ) . Let A' be an approximate solution to
the original 2-VCA-SV determined from A c' by using backpointers b 3,
b 2 and b 1 in this order, where multiplicity is deleted if any. Then
c ( A ' ) < c 3 ( A c ' ) = c 4 ( A c ' ). It is easy to see that c(A)=cl(As)=C2(Ab). Now we
consider Ab and Ac. The crucial point of the estimate is to show that a
solution Ac"_c E c - E c' to 2-VCA[G0=T c, S=Vc] with c 4 ( A c " ) = c 3 ( A c " )
< 2 c 2 ( A b ) can be determined from Ab. Before proving this statement we
show that if this is true then the desired estimate follows: since
c4(Ac)_<c4(Ac" ), c4(Ac')_<2c4(Ac) (by [21) and c(A')<_c4(Ac'), we have
c(A')-<-c4(Ac')-<2c4(Ac)-<2c4(Ac")<4c2(Ab)=4c 1(As)=4c(A).
We provide Fig. 2.5 to explain our idea in proving the crucial point
m e n t i o n e d above. Suppose that Ab consists of edges denoted by
halftone lines, and let H b denote the subgraph induced by Ab, where
we can assume that H b is a tree if A b is nonempty. The bold solid lines
shows the tree Tc, and black spots on H b corresponds to leaves of T c.
Let Ruv denote the path connecting u and v in Hb. The point is that we
can find a path P (shown by dotted fine lines) as follows: P consists of

Fig. 2.5. A schematic explanation in

constructing Ac" to 2-VCA[G0=Tc,S=Vc]
from Ab to 2-VCA[G0=Tb, S=Vc].
 385

edges not in Hb and containing only black spots on Hb, each

appearing only once. Furthermore for each edge e of H b, P has at most
two edges (u,v) and (u',v') such that both Ruv and Ru' v, pass through
e. For the edge ( u , v ) ~ E ( H b ) of Fig.2.5, P has two edges (u,w) and (u,x)
for which Ruw and Rux exist. On the other hand Hb has at least one
edge, say (y,z), for which P has only one edge, (z,w), for which Rz w
exists. Let Ac" be the set of edges denoted by dotted bold lines in the
figure. Note that there is one-to-one correspondence between E(P) and
Ac". Clearly Ac" is a solution to 2-VCA[G0=T c, S=Vc]. The difinition of c 3
means that c3(a,b)-<c3(E(Rab)) for each (a,b)~E(P). Hence
c3 ( E ( P ) ) = E ( u , v ) ~ E(p)c3(u,v) <]~(u,v)~ E ( p ) c 3 ( E ( R u v ) )
< 2 ]~(u',v')~ A bC3 (u',v')=2c 3 (Ab)_<2c2(Ab).
Since the definition of c3 also implies that c 3 ( A c " ) < c 3 ( E ( P ) ) , we have
c 3 ( A c " ) < 2 c 2 ( A b ) . The O(IWl 3) time complexity comes from the
definition of c 3. Thus we obtain the next theorem.
Theorem 2. The algorithm BIA-SV finds an approximate solution
for 2-VCA-SV in O(IVI 3) time and the worst approximation is less than
four times the optimal if G O is connected.

5. Strong-connectivity augmentation problem

The problem is formulated as a decision problem as follows:

(SCA-SV with u n i t y weights)
Instance: A directed complete graph G=(V,E), a spanning subgraph
G0=(V,E'), a specified subset S~V, and an integer k.
Question: Is there a set AC-E-E ' with [Al_<k such that, for any pair of
vertices in S, G0+A has at least one directed cycle containing them ?
The reduction is from the set covering problem, a well-known
NP-complete problem, defined as follows:
(Set C o v e r )
Instance: A ,collection C of sets, all subsets of the universal set U, and
an integer k.
Question: Is there a subcollection C' c_ C with tC'l<k such that C' covers
U, that is, the union of all sets in C' is equal to U ?
We can ]prove the following theorem.
Theorem 3. SCA-SV with unity weights is NP-complete.

We propose an O(IVI 3) approximation algorithm STA-SV and its

variations. Only one of them, STA-SV, is explained: it has two stages. It
first constructs a spanning arborescence Tou t rooted at a vertex re S.
This is done by finding a shortest <r,v>-path for each vertex w V. In
the second stage set c<u,v>=0 for any edge <u,v> of Tout, and STA-SV
 386

again c o n s t r u c t s another reverse a r b o r e s c e n c e T i n rooted at r by

finding <v,r>-shortest path for each ve V with respect to the current
cost c. We obtain an approximate solution A=E(Tout)tOE(Tin)-E(G0).
STA-SV
Input: G=(V,E) a complete directed graph with a cost function
c:E--)Z+(,where if <u,v>~E then c<u,v>=0), G0=(V,E') a subgraph of G
and S ~ V a specified set of vertices.
O u t p u t : an approximation solution A~E-E'.
1. For each pair of vertices u , v e V , find a shortest path P<u,v> and the
distance d<u,v> from u to v.
{the first stage}
2. Choose any vertex r in S as a candidate for the root. Set S:={r}, S':=S-
{r}, and A:=~.
3. If S ' : = ~ t h e n go to step 6.
4. Find U l e S , v l e S' with d<ul,Vl>=min{{d<u,v> I ue S,ve S'}u{d<w,r> I
weS'}}.
If d<ul,Vl>=d<w,r> is greater than d<r,w> t h e n set r:=u 1.
5. Set S':=S'-{v 1 }, A:=AuE(P<ul,Vl>), S:=StoV(P<u 1,v 1 >) and go to step 3.
{the second stage}
6. Set c<u,v>:=0 for each edge <u,v> in A.
7. For each pair of vertices u , v e V , find a shortest path P<u,v> and the
distance d<u,v> from u to v.
8. Set S:={r}, S:=S-{vertices which are not leaves of Tout}, S':=S-{r},
A':=A.
9. If Sv'=O, then set A:=A'-E', and Stop.
10. Find u l e S ' , v l e S such that d<ul,Vl>=min{d<u,v> I u e S ' , v e S}.
11. Set S':=S'-{u 1 }, A':=A'~E(P<Ul,Vl>), S : = S ~ V ( P < u t,v 1>), and go to step
9.
5.4. E x p e r i m e n t a l r e s u l t s , There is an example of the problem
whose approximation solution given by the algorithm STA-SV may
close to iSI times the optimal. (The theoretical analysis of worst case
a p p r o x i m a t i o n s is, h o w e v e r , left for future reserch.) S o m e of
experimental results are shown in Tables 1 and 2. A L l denotes SCA-SV,
AL2 does an algorithm in which ALl is repeated ISI times by taking
each vertex of S as a starting candidate for the root, and AL3, AL4 are
other approximation ones not explained here. OPT is an optimal cost.
Each upper integer is an approximate total cost and the lower denotes
the CPU time in second. All algorithms are implemented on NEC EWS-
4800(CPU:68020, 16MHz) by means of C language.

References

[1] K. P. Eswaran and R. E. Tarjan, Augmentation problems, SIAM J.

 387

Table 1. Experimental results for G with IVI=16. Optimum

solutions are known.
DATA AL I AL 2 AL 3 AL 4 OPT DATA ALl AL2 AL3 AL4 OPT
~zo1 10 I0 10 23 10 106 22 22 23 25 22
O. 188 O. 538 O. 168 O. 433 0.183 1.049 0.316 0,418
102 13 11 11 23 11 107 24 24 24 24 24
O. 168 O. 583 O. 166 O. 466 0.188 1.086 0.316 0.416
103 17 15 15 20 15 108 15 14 14 18 14
0. 183 O. 549 O. 188 O. 449 0.183 1.086 0.299 0.483
104 16 16 16 18 16 109 26 25 26 25 25
0. 183 0. 598 0, 149 0. 516 0.183 1.066 0.816 0.433
105 11 8 8 II 8 II0 19 19 19 23 19
O. 183 0. 549 O. 186 O. 433 0.183 1.049 0.299 0.499

Table 2. Experimental results for G with IVI=50. G and costs are

randomly generated. Optimum solutions are unknown.
.,data AL 1 A L 2,,,,, A L 3 AL 4 ....A,t, 1 AL2 AL 3 AL 4 AL,I AL 2 AL 3 ,AL4
Ist = 2o Is{ = 30 ,ill S I = 40
sol 44 3S 36 51 53 48 48 59 64 61 63 61
4.583 49.981 7.016 10.599 4.583 77.998 14. 316 10.416 4.85 112.56 26.82 9.48
502 20 16 19 24 22 18 23 24 25 22 24 24
4.416 49.484 6.438 18.115 4. 566 76.848 12. 749 18. 549 4.79 107.88 23,15 18.62
503 24 24 24 32 30 29 29 34 36 34 37 34
4.433 49.298 7.183 13,099 4. 583 78.046 13.732 13. 249 4.85 1 1 0 . 9 9 25.08 12,68
504 2 3 23 23 30 49 47 47 45 47 45 42 45
4.416 48.814 6.899 12.382 4.599 77.863 15.016 11.816 4.82 111.40 26.27 11.77
505 21 20 20 23 25 22 22 23 32 28 28 25
4.433 48.998 6.968 9.818 4.618 77.863 14.649 11.532 4.78 110.40 27.22 11.79
506 23 21 23 32 35 28 30 32 39 33 37 30
i 4.383 48.214 8.888 16.482 4,586 76.080 13.199 16.432 5,03 114.41 23.60 15.97

Comput. , 5(1976), 653-655.

[2] G. N. Frederickson and Ja'Ja', Approximation algorithms for several
graph augmentation problems, SIAM J. Comput. , 10(1981), 270-283.
[3] F. Harary, "Graph Theory", Addison-Wesley, Reading, MA (1969)
[4] A. Rosenthal and A. Goldner, Smallest augmentations to biconnect a
Graph, SIAM J. Comput. , 6(1977), 55-66.
[5] T. Watanabe and A. Nakamura, Edge-connectivity augmentation
problems, J. Comput. and System Sci. , 35(1987-08), 96-144.
[6] T. Watanabe and A. Nakamura, 3-connectivity augmentation
problems, Proceedings of 1988 IEEE Int. Sympo. on Circuits and
Systems(1988.-6), 1847-1850.
[7] T. Watanabe, T. Narita and A. Nakamura, 3-Edge-connectivity
augmentation problems, 1989 IEEE Int. Sympo. on Circuits and Systems
(1989-05), 335-338.
[8] T. Watanabe and A. Nakamura, A Smallest augmentation to 3-
connect a graph, Discrete Applied Mathematics, to appear.