Theoretical Computer Science 410 (2009) 766–775
Contents lists available atScienceDirect
Theoretical Computer Science
journal homepage:www.elsevier.com/locate/tcs
Fault-free longest paths in star networks with conditional link faults
Ping-Ying Tsai
a, Jung-Sheng Fu
b, Gen-Huey Chen
c,∗aDepartment of Computer Science and Information Engineering, Hwa Hsia Institute of Technology, Taipei, Taiwan, ROC bDepartment of Electronics Engineering, National United University, Miaoli, Taiwan, ROC
cDepartment of Computer Science and Information Engineering, National Taiwan University, Taipei, 10764, Taiwan, ROC
a r t i c l e i n f o Article history:
Received 26 May 2007
Received in revised form 15 July 2008 Accepted 6 November 2008 Communicated by D. Peleg
Keywords:
Conditional fault model Embedding
Fault tolerance Hamiltonian laceability Random fault model Star network
a b s t r a c t
The star network, which belongs to the class of Cayley graphs, is one of the most versatile interconnection networks for parallel and distributed computing. In this paper, adopting the conditional fault model in which each node is assumed to be incident with two or more fault-free links, we show that an n-dimensional star network can tolerate up to 2n−7 link faults, and be strongly (fault-free) Hamiltonian laceable, where n≥4. In other words, we can embed a fault-free linear array of length n!−1 (n!−2) in an n-dimensional star network with up to 2n−7 link faults, if the two end nodes belong to different partite sets (the same partite set). The result is optimal with respect to the number of link faults tolerated. It is already known that under the random fault model, an n-dimensional star network can tolerate up to n−3 faulty links and be strongly Hamiltonian laceable, for n≥3.
© 2008 Elsevier B.V. All rights reserved.
1. Introduction
The star network [1], which belongs to the class of Cayley graphs [2], has been recognized as an attractive alternative to the hypercube. It possesses many favorable topological properties such as recursiveness, symmetry, maximal fault tolerance, sublogarithmic degree and diameter, and strong resilience (see [1]). They are all desirable when we are building an interconnection topology for a parallel and distributed system. Besides, the star network can embed rings [28], grids [18], trees [5], and hypercubes [27]. Efficient communication algorithms for shortest-path routing [28], multiple-path routing [9], broadcasting [26] and scattering [13] are also available.
A linear array, which is one of the most fundamental networks for parallel and distributed computation, is suitable for developing simple and efficient algorithms. Numerous algorithms that were designed on linear arrays for solving various algebraic problems and graph problems can be found in [21]. A linear array can be also used as a control/data flow structure for distributed computation in a network (refer to [3] for an example).
Since node faults and/or link faults may occur in networks, it is important to consider faulty networks. Previously, communication problems (e.g., routing [5,12], broadcasting [32], multicasting [25], and gossiping [10]), embedding problems [4,6,16,17,24,30], and fault diameters [8,19,29] were studied on various faulty networks. Among them, two fault models were adopted; one is the random fault model [5,8,10,12,16,17,24,25,32], and the other is the conditional fault model [4,6,19,29,30].
The random fault model assumed that the faults might occur anywhere without any restriction, whereas the conditional
fault model assumed that the fault distribution must be subject to some constraint, e.g., that two or more fault-free links
are incident to each node. As a consequence of the constraint, it is in general more difficult to solve problems under the conditional fault model than the random fault model.
∗Corresponding author. Fax: +886 2 23628167.
E-mail address:ghchen@csie.ntu.edu.tw(G.-H. Chen).
0304-3975/$ – see front matter©2008 Elsevier B.V. All rights reserved.
Fig. 1. S4and four embedded S3’s.
In this paper, under the conditional fault model and with the assumption of at least two fault-free links incident to each node, we show that an n-dimensional star network can tolerate up to 2n
−
7 link faults, while retaining strongly (fault-free) Hamiltonian laceability, where n≥
4. The result is optimal with respect to the number of link faults tolerated. For the same problem, at most n−
3 link faults can be tolerated if the random fault model is adopted [24]. With our results, all parallel algorithms developed on a linear array of length n! −
1 or n! −
2 can be executed as well on an n-dimensional star network with up to 2n−
7 link faults.Previous results under the conditional fault model are described as follows. With the same assumption as ours, an n-dimensional hypercube (n-cube for short) with 2n
−
5 link faults is strongly (fault-free) Hamiltonian laceable [30], and an m-ary n-cube with 4n−
5 link faults has a fault-free Hamiltonian cycle [4]. On the other hand, with the assumption of each node having at least k fault-free neighbors, the minimum number of node faults whose removal may disconnect an n-cube is(
n−
k)
2k, where 1≤
k≤ b
n/
2c
[20]. Such a minimum number was named the restricted-node-connectivityand denoted by Rk-node-connectivity [20]. The node-connectivity of an n-cube is known to be n. There is a lower bound of
md
((
n−
d−
1)(
m−
1)(
s+
1) + (
m−
s−
1))
on the Rk-node-connectivity of an m-ary n-cube [33], where d
= b
k/(
m−
1)c
and s
=
k mod(
m−
1)
.When k
=
1, the R1-node-connectivity of an n-cube (an m-ary n-cube) is 2n−
2 (4n−
2 if m≥
4, and 4n−
3 if m=
3)[11] ([7]), and the R1-node-connectivities of cube-connected cycles, undirected binary de Bruijn networks and Kautz graphs
are all greater by one at most than their node-connectivities [23]. Besides, the maximal diameters of an n-cube with 2n
−
3 node faults and an n-dimensional star network with 2n−
5 node faults are n+
2 [19] andb
3(
n−
1)/
2c +
2 [29], respectively. When they are fault-free, their diameters are n andb
3(
n−
1)/
2c
, respectively.In the next section, the structure of the star network is reviewed. Some necessary definitions, notations and previous results are also introduced. In Section3, some new properties of the star network are derived in order to prove the main result. The proof of the main result is shown in Section4. Finally, this paper concludes with some remarks in Section5.
2. Preliminaries
It is convenient to represent a network as a graph G, where each vertex (edge) of G uniquely represents a node (link) of the network. We use V
(
G)
and E(
G)
to denote the vertex set and edge set of G, respectively. Throughout this paper, we use network and graph, node and vertex, link and edge, interchangeably. The following is a definition of star networks, in terms of graph theory.Definition 1 ([1]). An n-dimensional star network, denoted by Sn, has the node set V
(
Sn) = {a1a2· · ·
an|
a1a2· · ·
anis apermutation of 1
,
2, . . . ,
n}
and the link set E(
Sn) = {(a1a2· · ·
an,aia2· · ·
ai−1a1ai+1· · ·
an)|a1a2· · ·
an∈
V(
Sn)and 2≤
i
≤
n}
.Snhas n
!
nodes, each of degree n−
1. S1is a node, S2is a link, and S3is a cycle of length six. S4is shown inFig. 1. Thelink (a1a2
· · ·
an, aia2· · ·
ai−1a1ai+1· · ·
an) is referred to as an i-dimensional link. We use e(i)(v)
to denote the i-dimensionallink that is incident to node
v
, and let E(i)(
Sn) ={e(i)(v)|v ∈
V(
Sn)} be the set of all i-dimensional links in Sn. Snis both node
768 P.-Y. Tsai et al. / Theoretical Computer Science 410 (2009) 766–775
It can be observed fromFig. 1that S4contains four embedded S3’s, denoted by
h∗ ∗ ∗
1i
3,h∗ ∗ ∗
2i
3,h∗ ∗ ∗
3i
3andh∗ ∗ ∗
4i
3,respectively (
∗ ∗ ∗
1, for example, represents any permutation of 1, 2, 3, 4 ending with 1). In general, Sncontains n embeddedSn−1’s.
For 1
≤
r≤
n−
1, an embedded Srin Snis denoted byh
s1s2· · ·
sni
r, where s1= ∗
and there are exactly r−
1 occurrencesof
∗
in s2s3· · ·
sn. For example,h∗
4∗
2i
2 denotes an embedded S2 in S4. When Sn is partitioned intoh∗
d−11∗
n−di
n−1,h∗
d−12∗
n−di
n−1,
. . .
,h∗
d−1n∗
n−di
n−1, Sn is said to be partitioned along dimension d, where 1<
d≤
n and∗
d−1 (∗
n−d)represents d
−
1 (n−
d) consecutive∗
’s. These n embedded Sn−1’s are connected by d-dimensional links. When d=
n,we useE
˜
p(n,q() Sn)to represent the set of those n-dimensional links in Snthat connecth∗
n−1pi
n−1andh∗
n−1qi
n−1, where p6=
q.Clearly, we have
| ˜
E(pn,q() Sn)| = (n−
2)!
.A path from vertex u (
=
v
0) to vertexv
(=
vk
) in a graph G, represented ash
v
0,v
1,. . .
,vk
i
, is referred to as a u–v path. Apath (cycle) in G is a Hamiltonian path (Hamiltonian cycle) if it contains every vertex of G exactly once. G is bipartite if there is a partition of V
(
G)
into V0(
G)
and V1(
G)
such that every(
u, v) ∈
E(
G)
has either u∈
V0(
G)
andv ∈
V1(
G)
or u∈
V1(
G)
andv ∈
V0(
G)
. The two subsets V0(
G)
and V1(
G)
are referred to as partite sets of G. Star networks are known to be bipartite [1].A bipartite graph G with
|
V0(
G)| = |
V1(
G)|
is Hamiltonian laceable if it has a u–v
Hamiltonian path for every u∈
V0(
G)
andevery
v ∈
V1(
G)
[31], and strongly Hamiltonian laceable if it additionally has a longest u–v
path of length|
V0(
G)|+|
V1(
G)|−
2for all u,
v ∈
V0(
G)
or u,v ∈
V1(
G)
[15]. A Hamiltonian laceable graph G is hyper Hamiltonian laceable if for everyw ∈
V0(
G)(w ∈
V1(
G))
, G–w
contains a u–v
Hamiltonian path for all u,v ∈
V1(
G)
(u,v ∈
V0(
G)
) [22]. Clearly, a hyperHamiltonian laceable graph is strongly Hamiltonian laceable.
Given a vertex u in G, we define N
(
u) = {v|(
u, v) ∈
E(
G)}
to be the neighborhood of u, which is the set of vertices that are adjacent to u in G. The size of N(
u)
, i.e.,|
N(
u)|
, is the degree of u. The minimum vertex degree of G is denoted byδ(
G) =
min{|
N(
u)||
u∈
V(
G)}
. Let V0be a vertex subset of G. We define N(
V0) = Su
∈V 0N
(
u) −
V0to be the neighborhood of V0. Besides, we use G
[
V0]
to denote the subgraph of G induced by V0. Throughout this paper, we use F (⊆
E(
Sn)) to denote the set of link faults in Sn.Lemma 1 ([24]). Sn
−
F is strongly Hamiltonian laceable if|
F| ≤
n−
3, and hyper Hamiltonian laceable if|
F| ≤
n−
4, wheren
≥
4.Lemma 2 ([14]). If
|
F| ≤
2n−
7 andδ(
Sn−
F) ≥
2, then there exists 1<
d≤
n such that|
E(d)(
Sn) ∩F| ≥
1 andδ(h∗
d−1q∗
n−di
n−1
−
F) ≥
2 for all 1≤
q≤
n, where n≥
4. 3. Properties and main resultIn this section, we first introduce some properties of Sn. Then we present our main result.
Lemma 3. Suppose u
=
u1u2· · ·
un∈
V(h∗
n−1qi
n−1)
, where un=
q and n≥
3. For every r∈ {
1,
2, . . . ,
n} − {
u1,
q}
, thereexists
w ∈
V(h∗
n−1ri
n−1
)
andv ∈
N(
u) ∩
V(h∗
n−1qi
n−1)
such that(w, v) =
e(n)(v)
.Proof. We assume uc
=
r, where 1<
c<
n. Selectv =
ucu2· · ·
uc−1u1uc+1· · ·
un(∈ N(
u) ∩
V(h∗
n−1qi
n−1))
andw =
unu2· · ·
uc−1u1uc+1· · ·
uc(∈
V(h∗
n−1ri
n−1))
. Clearly,(w, v) =
e(n)(v)
.Lemma 4. Suppose that P is a path in
h∗
n−1qi
n−1
−
F , where n≥
4 and 1≤
q≤
n. If|
F| ≤
2n−
7 and P is Hamiltonian or oflength
(
n−
1)! −
2, then P has a link(
u, v)
with e(n)(
u),
e(n)(v) /∈
F .Proof. Suppose conversely that no such (u,
v
) can be found in P. Then,|
E(n)(
Sn) ∩F| ≥
((
n−
1)! −
2)/
2, which is greaterthan 2n
−
7 as n≥
4, a contradiction.In subsequent discussion, we let VA
=
S
r∈AV
(h∗
n−1r
i
n−1
)
, where A⊆ {
1,
2, . . . ,
n}
. Lemma 5. Suppose that A⊆ {
1,
2, . . . ,
n}
and n≥
5. For any s∈
V(h∗
n−1pi
n−1
)
and t∈
V(h∗
n−1qi
n−1)
, there exists a longests–t path of length
|
A| ×
(
n−
1)! −
1 or|
A| ×
(
n−
1)! −
2 in Sn[
VA] −
F , where p, q∈
A and p6=
q, provided the following twoconditions hold:
(1)
| ˜
Ei(,nj)(
Sn) ∩F|
< (
n−
2)!/
2 for all i,
j∈
A and i6=
j;(2)
h∗
n−1ri
n−1
−
F is strongly Hamiltonian laceable for every r∈
A. Proof. Since the distance between node ja2∗
n−4an−1i and node jan−1
∗
n−4a2i is three, they belong to different partitesets of
h∗
n−1ii
n−1. For each link (ja2∗
n−4a
n−1i, ia2
∗
n−4an−1j) connectingh∗
n−1ii
n−1 withh∗
n−1ji
n−1, there exists anotherlink (jan−1
∗
n−4a2i, ian−1∗
n−4a2j)
connectingh∗
n−1ii
n−1withh∗
n−1ji
n−1. It is implied that there are an equal number (i.e.,(
n−
2)!/
2) of nodes in V0(h∗
n−1ii
n−1)
and V1(h∗
n−1ii
n−1)
, respectively, that are connected toh∗
n−1ji
n−1.Suppose A
= {
a1,
a2, . . . ,
a|A|}
. A longest s–t path in Sn[
VA] −
F is shown inFig. 2, where a1=
p and a|A|=
q are assumed. Since| ˜
Ea(n1),a2(
Sn)| = (n−
2)!
, two links inE˜
(n)
a1,a2
(
Sn) −F , one incident to V0(h∗
n−1a1
i
n−1)
and the other incidentto V1
(h∗
n−1a1i
n−1)
, can be found, as a consequence of (1). Hence, a link(
x1,
y2) ∈ ˜
E( n)a1,a2
(
Sn) −F can be determined suchthat x1and s belong to different partite sets of
h∗
n−1a1i
n−1and y2∈
V(h∗
n−1a2i
n−1)
. As a consequence of (2), there exists aFig. 2. A longest s–t path in Sn[VA] −F .
Fig. 3. A longest s–t path in Snthat contains (u,v), where s, t∈V(h∗kqi k). Similarly, by the aid of (1) and (2), links
(
xk,yk+1) ∈ ˜
E(n)
ak,ak+1
(
Sn) −F and Hamiltonian yk–xkpaths inh∗
n−1ak
i
n−1−
F canbe obtained, where 2
≤
k≤ |
A| −
1. All these Hamiltonian paths together with a longest y|A|–t path inh∗
n−1a|A|i
n−1−
Fconstitute a longest s–t path in Sn
[
VA] −
F . If the longest y|A|–t path has length(
n−
1)! −
1, the length of the longest s–t path is computed as(|
A| −
1) × ((
n−
1)! −
1) + (|
A| −
1) + ((
n−
1)! −
1) = |
A| ×
(
n−
1)! −
1. Similarly, if the longesty|A|–t path has length
(
n−
1)! −
2, the longest s–t path has length|
A| ×
(
n−
1)! −
2.Lemma 6. Suppose that s, t are two distinct nodes of Snand
(
u, v) ∈
E(
Sn), where n≥
4. If{
s,
t} 6= {
u, v}
, then there exists aHamiltonian s-t path or an s-t path of length
(
n−
1)! −
2 in Snthat contains(
u, v)
.Proof. We prove this lemma by induction on n. This lemma holds for S4, which can be verified by exhaustive search (see
[34]). So, we assume that this lemma holds for Sk, and then consider Sk+1below, where k
≥
4.Without loss of generality, suppose u,
v ∈
V(h∗
kqi
k)for some 1≤
q≤
k+
1. We also assume(
u, v) ∈
E(l)(
Sk+1
)
, where2
≤
l≤
k. With the following three cases, we show a longest s–t path of length(
k+
1)! −
1 or(
k+
1)! −
2 in Sk+1thatcontains (u,
v
).Case 1. s
∈
V(h∗
kqi
k)and t/∈
V(h∗
kqi
k)or s/∈
V(h∗
kqi
k)and t∈
V(h∗
kqi
k). We consider the situation of s∈
V(h∗
kqi
k)and t
/∈
V(h∗
kqi
k). The discussion for the situation of s/∈
V(h∗
kqi
k)and t∈
V(h∗
kqi
k)is very similar. A desired s–t path canbe obtained using the construction method ofFig. 2. We only need to change n
−
1 to k and|
A|
to k+
1, and set a1=
q.Besides, the node x1is selected with
{
s,
x1} 6= {
u,v}
. The induction hypothesis assures a Hamiltonian s–x1path inh∗
ka1i
kthat contains (u,
v
). The desired s–t path has length(
k+
1) ×
k! −
1=
(
k+
1)! −
1 or(
k+
1) ×
k! −
2=
(
k+
1)! −
2. Case 2. s, t∈
V(h∗
kqi
k). A desired s–t path can be obtained as shown inFig. 3, where a1=
q is assumed. The inductionhypothesis assures a longest s–t path in
h∗
ka1
i
kthat contains (u,v
). A link (w
,w
0) 6= (
u, v)
can be selected from the path.Let z
=
e(k+1)(w)
and z0=
e(k+1)(w
0)
. A Hamiltonian z–z0path in Sk+1
− h∗
ka1i
kcan be obtained using the constructionmethod ofFig. 2(changing s to z, t to z0, n
−
1 to k,|
A|
to k, andh∗
n−1ari
n−1toh∗
kar+1i
kfor all 1≤
r≤ |
A|
)
. The Hamiltonianz–z0path, combining with (
w
, z), (w
0, z0), and the s–w
andw
0–t paths inh∗
ka1
i
k, forms a desired s–t path in Sk+1. The desireds–t path has length
(
k×
k! −
1) +
2+
((
k! −
1) −
1) = (
k+
1)! −
1 or(
k×
k! −
1) +
2+
((
k! −
2) −
1) = (
k+
1)! −
2. Case 3. s, t/∈
V(h∗
kqi
k). Suppose s∈
V(h∗
kgi
k)and t∈
V(h∗
khi
k), where g, h∈ {
1,
2, . . . ,
k+
1} − {
q}
. First we assumeg
6=
h. A desired s–t path can be obtained using the construction method ofFig. 2(changing n−
1 to k and|
A|
to k+
1, and letting a1=
g, a2=
q, and ak+1=
h). The node x2is selected with{
x2,
y2} 6= {
u, v}
. The induction hypothesis assures aHamiltonian y2–x2path in
h∗
ka2i
kthat contains (u,v
). The desired s–t path has length(
k+
1) ×
k! −
1=
(
k+
1)! −
1 or(
k+
1) ×
k! −
2=
(
k+
1)! −
2.Next we assume g
=
h. A desired s–t path can be obtained by slightly modifying the construction method ofFig. 3. We only need to set a1=
g=
h and a2=
q. The node x2is selected with{
z,
x2} 6= {
u, v}
. The induction hypothesis assures aHamiltonian z–x2path in
h∗
ka2i
kthat contains (u,v
). A Hamiltonian y3–z0path in Sk+1− h∗
ka1i
k− h∗
ka2i
kcan be obtained,similarly, using the construction method ofFig. 2. The desired s–t path, which consists of the s–
w
andw
0–t paths inh∗
ka 1i
k,the Hamiltonian z–x2path, the Hamiltonian y3–z0path, and three links (
w
, z), (w
0, z0), (x2, y3), has length(
k+
1)! −
1 or(
k+
1)! −
2.Lemma 7. For any two distinct links (s, t), (u,
v
) in Sn, there exists a Hamiltonian cycle in Snthat contains both of them, wheren
≥
3.Proof. Since S3is a cycle of length 6, this lemma holds for S3. When n
≥
4, since s and t belong to different partite sets of770 P.-Y. Tsai et al. / Theoretical Computer Science 410 (2009) 766–775
Fig. 4. A distribution of 2n−6 link faults over Sn.
The main result of this paper is presented in the following theorem whose proof is shown in the next section.
Theorem 1. With the assumption of two or more fault-free links incident to each node, an n-dimensional star network can tolerate
up to 2n
−
7 link faults, and be strongly (fault-free) Hamiltonian laceable, where n≥
4.Theorem 1is optimal with respect to the number of link faults tolerated.Fig. 4shows a distribution of 2n
−
6 link faults over Sn, whereh
s, u,v
, ti
is an s–t path and (s, u), (u,v
) ((u,v
), (v
, t)) are the only two fault-free links incident to u (v
). It iseasy to see that no fault-free Hamiltonian s–t path exists in the faulty Sn.
4. Proof of Theorem 1
With
|
F| ≤
2n−
7 andδ(
Sn−
F) ≥
2, we show by induction that there exists a longest s–t path of length n! −
1 or n! −
2between every two distinct nodes s, t of Sn
−
F . ByLemma 1(n−
3=
2n−
7 as n=
4), the theorem holds for S4. So, weassume that this theorem holds for Sk, and then consider Sk+1in the rest of this section, where k
≥
4.ByLemma 2, we can partition Sk+1along some dimension d such that
|
E(d)(
Sk+1)∩
F| ≥
1 andδ(h∗
d−1q∗
k+1−di
k−
F) ≥
2for all 1
≤
q≤
k+
1, where 1<
d≤
k+
1. Without loss of generality, we assume d=
k+
1. Now that|
E(k+1)(
Sk+1)∩
F| ≥
1,we have
|
E(h∗
kqi
k) ∩F| ≤
2k−
6 for all 1≤
q≤
k+
1. A desired s–t path in Sk+1
−
F is constructed in Section4.1if|
E(h∗
kqi
k) ∩ F| ≤
2k−
7 for all 1≤
q≤
k+
1, and constructed in Section4.2else.4.1.
|
E(h∗
kqi
k) ∩F| ≤
2k−
7 for all 1≤
q≤
k+
1Suppose that
|
E(h∗
kqi
k) ∩ F|
≤
2k−
7 for all 1≤
q≤
k+
1. Since|
F|
≤
2k−
5, we haveP
i,j∈{1,2,...,k+1} and i6=j
| ˜
E(k+1)
i,j
(
Sk+1) ∩
F| ≤
2k−
5, where k≥
4. Notice that 2k−
5=
((
k+
1) −
2)!/
2 when k=
4,and 2k
−
5< ((
k+
1) −
2)!/
2 when k>
4. Two cases are discussed below.Case 1. k
>
4. We have| ˜
Ei(,kj+1)(
Sk+1)∩
F|
< ((
k+
1)−
2)!/
2 for all i, j∈ {
1,
2, . . . ,
k+
1}
and i6=
j. The induction hypothesisassures that
h∗
kqi
k
−
F is strongly Hamiltonian laceable for all 1≤
q≤
k+
1. If s∈
V(h∗
kgi
k)and t∈
V(h∗
khi
k)for someg, h
∈ {
1,
2, . . . ,
k+
1}
and g6=
h, then byLemma 5, Sk+1−
F is strongly Hamiltonian laceable. If s, t∈
V(h∗
kgi
k)for some1
≤
g≤
k+
1, then a desired s–t path in Sk+1−
F can be obtained by slightly modifying the construction method ofFig. 3.We only need to set a1
=
g. The induction hypothesis assures a longest s–t path inh∗
ka1i
k−
F . ByLemma 4, a link (w
,w
0)with
(w,
z) ∈ ˜
Ea(k1+,a12)−
F and(w
0
,
z0
) ∈ ˜
Ea(k1+,a1k+1)−
F can be selected from the path. ByLemma 5, a Hamiltonian z–z0 path in
Sk+1
− h∗
ka1i
k−
F can be obtained. The resulting longest s–t path in Sk+1−
F has length(
k+
1)! −
1 or(
k+
1)! −
2.Case 2. k
=
4. If| ˜
Ei(,5j)(
S5) ∩
F|
< ((
4+
1) −
2)!/
2=
3 for all i, j∈ {
1,
2, . . . ,
5}
and i6=
j, then the discussion is the sameas Case 1. So, we consider
| ˜
Ei0(5,j0() S5) ∩
F| =
3 for i0, j0∈ {
1,
2, . . . ,
5}
and i06=
j0. Since|
F| ≤
3 as k=
4, all link faults areinE
˜
i0(5,j0() S5)
. Assume that s∈
V(h∗
4gi
4)
and t∈
V(h∗
4hi
4)
, where g, h∈ {
1,
2, . . . ,
5}
. When g6=
h, a desired s–t path inS5
−
F can be obtained using the construction method ofFig. 2. We only need to change n−
1 to 4 and|
A|
to 5, set a1=
gand a5
=
h, and set a2, a3,
a4with{
i0,
j0} 6= {
ar,
ar+1}
for all 1≤
r≤
4. The induction hypothesis assures a Hamiltonians–x1path in
h∗
4a1i
4−
F , a Hamiltonian yr–xr path inh∗
4ari
4−
F for all 2≤
r≤
4, and a longest y5–t path inh∗
4a5i
4−
F .The desired s–t path has length 5
! −
1 or 5! −
2.When g
=
h, a desired s–t path in S5−
F can be obtained using the construction method ofFig. 3. We only need to changek
+
1 to 5, set a1=
g=
h, and set a2, a3, a4, a5with{
i0,
j0} 6= {
ar,
a(rmod5)+1}
for all 1≤
r≤
5. The induction hypothesisassures a longest s–t path in
h∗
4a1
i
4−
F . ByLemma 4, a link (w
,w
0) with(w,
z) ∈ ˜
Ea(51),a2−
F and(w
0
,
z0) ∈ ˜
E(5)a1,a5
−
F can beselected from the path. The induction hypothesis assures a Hamiltonian yr–xrpath in
h∗
4ari
4−
F for all 2≤
r≤
5, wherey2
=
z and x5=
z0. The desired s–t path has length 5! −
1 or 5! −
2.4.2.
|
E(h∗
kαik) ∩
F| =
2k−
6 for some 1≤
α ≤
k+
1Suppose that
|
E(h∗
kαik) ∩
F| =
2k−
6, where 1≤
α ≤
k+
1. Now that|
F| ≤
2k−
5 and|
E(k+1)(
Sk+1
) ∩
F| ≥
1,we have
|
E(k+1)(
Sk+1
) ∩
F| =
1. Besides, we have|
E(h∗
kq0i
k) ∩F| =
0 for all q0∈ {
1,
2, . . . ,
k+
1} − {
α}
, and| ˜
Ei(,kj+1)(
Sk+1) ∩
F| ≤
1(< ((
k+
1) −
2)!/
2)
for any i, j∈ {
1,
2, . . . ,
k+
1}
and i6=
j, where k≥
4.If s, t
∈
V(h∗
kαik)
, then a link fault, say (h, h0), is chosen from E(h∗
kαik) ∩
F such that{
s,
t} 6= {
h,
h0}
and e(k+1)(
h)
,Fig. 5. A longest s–t path in Sk+1−F when h, h0/∈N(V(h∗kak+1ik)).
of length k
! −
1 or k! −
2 inh∗
kαik
−
(
F− {
(
h,
h0)})
. A desired s–t path in Sk+1−
F can be obtained using the constructionmethod ofFig. 3. We only need to set a1
=
α
, and(w, w
0) = (
h,
h0)
if P contains (h, h0). The desired s–t path has length(
k+
1)! −
1 or(
k+
1)! −
2.If s
/∈
V(h∗
kαik)
or t/∈
V(h∗
kαik)
, then a desired s–t path in Sk+1−
F is constructed in Sections4.2.1and4.2.2, wherewe use a1, a2, . . . , ak+1to denote the k
+
1 distinct integers from 1 to k+
1 (i.e.,{
a1,
a2, . . . ,
ak+1} = {
1,
2, . . . ,
k+
1}
).4.2.1. s
∈
V(h∗
kαik)
and t/∈
V(h∗
kαik)
or s/∈
V(h∗
kαik)
and t∈
V(h∗
kαik)
Suppose that s
∈
V(h∗
kαik)
and t/∈
V(h∗
kαik)
. The discussion for s/∈
V(h∗
kαik)
and t∈
V(h∗
kαik)
is similar. Weassume t
∈
V(h∗
kβik)
, whereβ 6= α
. A link fault (h, h0) is chosen from E(h∗
kαik) ∩
F such that s/∈ {
h,
h0}
and e(k+1)(
h)
,e(k+1)
(
h0) /∈
F . Set a1
=
α
and ak+1=
β
.We first consider the situation of h, h0
/∈
N(
V(h∗
kak+1
i
k)). Imagine that (h, h0) is fault-free. A vertexv
1∈
V(h∗
ka1i
k)isdetermined such that
v
1and s belong to different partite sets of Sk+1, e(k+1)(v
1) /∈
F , andv
1/∈
N(
V(h∗
kak+1i
k). The inductionhypothesis assures a longest s–
v
1path of length k! −
1 inh∗
ka1i
k−
(
F− {
(
h,
h0)})
. If the longest s–v
1path does not contain(h, h0), then a desired s–t path of length
(
k+
1)! −
1 or(
k+
1)! −
2 in Sk+1
−
F can be obtained using the constructionmethod ofFig. 2. We only need to set x1
=
v
1, and change n−
1 and|
A|
to k and k+
1, respectively.If the longest s–
v
1path contains (h, h0), then a desired s–t path in Sk+1−
F can be obtained as shown inFig. 5. We set a2and a3such that h
∈
N(
V(h∗
ka2i
k))and h0∈
N(
V(h∗
ka3i
k)). There is an additional restriction tov
1:
v
1/∈
N(
V(h∗
ka2i
k))and
v
1/∈
N(
V(h∗
ka3i
k)). Also we set a4such thatv
1∈
N(
V(h∗
ka4i
k)). Then, three vertices u2∈
V(h∗
ka2i
k),v
3∈
V(h∗
ka3i
k),and u4
∈
V(h∗
ka4i
k)are determined such that(
u2,
h) =
e(k+1)(
h)
,(v
3,
h0) =
e(k+1)(
h0)
, and(
u4, v
1) =
e(k+1)(v
1)
.ByLemma 5, there is a longest u4–t path of length
(
k−
2) ×
k! −
1 or(
k−
2) ×
k! −
2 in Sk+1[
VA] −
F , whereA
= {
1,
2, . . . ,
k+
1} − {
a1,
a2,
a3}
. Again, byLemma 5, there is a longest u2–v
3path of length 2×
k! −
1 in Sk+1[
VA] −
F ,where A
= {
a2,
a3}
. The desired s–t path has length((
k! −
1) −
1) + (
2×
k! −
1) + ((
k−
2) ×
k! −
1) +
3=
(
k+
1)! −
1or
((
k! −
1) −
1) + (
2×
k! −
1) + ((
k−
2) ×
k! −
2) +
3=
(
k+
1)! −
2.Then we consider the situation of h
∈
N(
V(h∗
kak+1i
k)), without loss of generality. Notice that at most one of h and h0belongs to N
(
V(h∗
kak+1
i
k)). Imagine that (h, h0) is fault-free. A vertex u1∈
V(h∗
ka1i
k)is determined such that u1and sbelong to different partite sets of Sk+1, e(k+1)
(
u1) /∈
F , and u1/∈
N(
V(h∗
kak+1i
k). The induction hypothesis assures a longests–u1path of length k
! −
1 inh∗
ka1i
k−
(
F− {
(
h,
h0)})
. If the longest s–u1path does not contain (h, h0), a desired s–t path inSk+1
−
F can be obtained using the construction method ofFig. 2. We only need to set x1=
u1, and change n−
1 and|
A|
tok and k
+
1, respectively. If the longest s–u1path contains (h, h0), two cases: h∈
N(
t)
or h/∈
N(
t)
, are discussed below.Case 1. h
∈
N(
t)
. A desired s–t path in Sk+1−
F can be obtained as shown inFig. 6(a). We set a2with h0∈
N(
V(h∗
ka2i
k)). Thereis an additional restriction to u1
:
u1/∈
N(
V(h∗
ka2i
k))and u1/∈
N(
V(h∗
kak+1i
k)). Also we set akwith u1∈
N(
V(h∗
kaki
k)).Then, six vertices u2,
v
2∈
V(h∗
ka2i
k), u3∈
V(h∗
ka3i
k),vk
∈
V(h∗
kaki
k), and uk+1,vk
+1∈
V(h∗
kak+1i
k)are determinedsuch that
(
u2,
h0) =
e(k+1)(
h0)
,v
2and u2belong to different partite sets of Sk+1, e(k+1)(v
2) /∈
F ,(
uk+1, v
2) =
e(k+1)(v
2)
,vk
+1and uk+1belong to the same partite set of Sk+1, e(k+1)(vk
+1) /∈
F ,(
u3, vk
+1) =
e(k+1)(vk
+1)
, and(vk,
u1) =
e(k+1)(
u1)
.ByLemma 1, there are a longest u2–
v
2path of length k! −
1 inh∗
ka2i
kand a longest uk+1–vk
+1path of length k! −
2 inh∗
kak+1
i
k− {
t}
. ByLemma 5, there is a longest u3–vk
path of length(
k−
2)×
k! −
1 or(
k−
2)×
k! −
2 in Sk+1[
VA] −
F , whereA
= {
1,
2, . . . ,
k+
1}−{
a1,
a2,
ak+1}
. The desired s–t path has length((
k!−
1)−
1)+(
k!−
1)+(
k!−
2)+((
k−
2)×
k!−
1)+
4=
(
k+
1)! −
1 or((
k! −
1) −
1) + (
k! −
1) + (
k! −
2) + ((
k−
2) ×
k! −
2) +
4=
(
k+
1)! −
2.Case 2. h
/∈
N(
t)
. A desired s–t path in Sk+1−
F is shown inFig. 6(b) whenw
06=
t and(v
2, w
0) /∈
F , and shown inFig. 6(c) whenw
0=
t or
(v
2, w
0) ∈
F . We set a2such that h0∈
N(
V(h∗
ka2i
k)). There is an additional restriction to u1:
u1/∈
N(
V(h∗
ka2i
k))and u1
/∈
N(
V(h∗
kak+1i
k)). Also we set akwith u1∈
N(
V(h∗
kaki
k)). Then, three vertices u2∈
V(h∗
ka2i
k),vk
∈
V(h∗
kaki
k),and
w ∈
V(h∗
kak+1
i
k)are first determined such that(
u2,
h0) =
e(k+1)(
h0)
,(w,
h) =
e(k+1)(
h)
, and(vk,
u1) =
e(k+1)(
u1)
. By Lemma 3, there existv
2∈
V(h∗
ka2i
k)andw
0∈
N(w) ∩
V(h∗
kak+1i
k)such that(v
2, w
0) =
e(k+1)(w
0)
(seeFig. 6(b)).If
w
06=
t and(v
2
, w
0) /∈
F , then, again byLemma 3, there exist u3∈
V(h∗
ka3i
k)andvk
+1∈
N(
t)∩
V(h∗
kak+1i
k)such that(
u3, vk
+1) =
e(k+1)(vk
+1)
. Besides,(
u3, vk
+1) /∈
F can be satisfied, because|
E(k+1)(
Sk+1) ∩
F| =
1. ByLemma 7, there exists aHamiltonian cycle in
h∗
kak+1i
kthat contains (t,vk
+1) and (w
,w
0). ByLemma 1, there is a longest u2–v
2path of length k! −
1in
h∗
ka2
i
k. ByLemma 5, there is a longest u3–vk
path of length(
k−
2) ×
k! −
1 or(
k−
2) ×
k! −
2 in Sk+1[
VA] −
F , whereA
= {
1,
2, . . . ,
k+
1}−{
a1,
a2,
ak+1}
. The desired s–t path has length((
k!−
1)−
1)+(
k!−
1)+(
k!−
2)+((
k−
2)×
k!−
1)+
4=
(
k+
1)! −
1 or((
k! −
1) −
1) + (
k! −
1) + (
k! −
2) + ((
k−
2) ×
k! −
2) +
4=
(
k+
1)! −
2.772 P.-Y. Tsai et al. / Theoretical Computer Science 410 (2009) 766–775
Fig. 6. A longest s–t path in Sk+1−F when h∈N(V(h∗kak+1ik)). (a) h∈N(t). (b) h/∈N(t)and (w06=t and(v2, w0) /∈F ). (c) h/∈N(t)and (w0=t or(v2, w0) ∈
F ).
If
w
0=
t or(v
2
, w
0) ∈
F , then byLemma 3, there existv
3∈
V(h∗
ka3i
k) andw
00∈
N(w) ∩
V(h∗
kak+1i
k)suchthat
(v
3, w
00) =
e(k+1)(w
00)
. Again byLemma 3, there exist u4∈
V(h∗
ka4i
k) andvk
+1∈
N(
t) ∩
V(h∗
kak+1i
k)with(
u4, vk
+1) =
e(k+1)(vk
+1)
. Besides, (v
3,w
00),(
u4, vk
+1) /∈
F andw
006=
t can be satisfied. ByLemma 7, there exists aHamiltonian cycle in
h∗
kak+1
i
kthat contains (t,vk
+1) and (w
,w
00). ByLemma 5, there is a longest u2–v
3path of length2
×
k! −
1 in Sk+1[
VA] −
F , where A= {
a2,
a3}
. If k=
4,Lemma 1assures a longest u4–v
4path of length 4! −
1 or 4! −
2 inh∗
4a4
i
4. If k>
4,Lemma 5assures a longest u4–vk
path of length(
k−
3) ×
k! −
1 or(
k−
3) ×
k! −
2 in Sk+1[
VA] −
F , whereA
= {
1,
2, . . . ,
k+
1} − {
a1,
a2,
a3,
ak+1}
. The desired s–t path has length(
k+
1)! −
1 or(
k+
1)! −
2.4.2.2. s
,
t/∈
V(h∗
kαik)
Suppose that s
∈
V(h∗
kβik)
and t∈
V(h∗
kγ ik)
, whereβ, γ ∈ {
1,
2, . . . ,
k+
1} − {
α}
. We first consider the situation ofβ = γ
. A link fault (h, h0) is chosen from E(h∗
kαik) ∩
F such that e(k+1)(
h),
e(k+1)(
h0) /∈
F . Set a1
=
β = γ
, and set a3=
α
ifh, h0
/∈
N(
V(h∗
ka1
i
k)), and a2=
α
else. A desired s–t path in Sk+1−
F can be constructed in a way similar toFig. 3. We onlyexplain below the construction for
α =
a3. The construction forα =
a2is similar.Refer toFig. 3again. For the purpose of our construction, we let A
= {
1,
2, . . . ,
k+
1}
,(
x3,
y3) = (
h,
h0)(∈
E(h∗
ka3i
k)),and select
(w, w
0)(∈
E(h∗
ka1i
k))such that{
w, w
0} 6= {
s,
t}
,w
,w
0/∈
N(
V(h∗
ka3i
k)),w
and y3belong to different partitesets of Sk+1, and e(k+1)
(w)
, e(k+1)(w
0) /∈
F . Four vertices z, x2∈
V(h∗
ka2i
k), y4∈
V(h∗
ka4i
k), and z0∈
V(h∗
kak+1i
k)aredetermined such that
(
z, w) =
e(k+1)(w)
,(
x2
,
y3) =
e(k+1)(
y3)
,(
y4,
x3) =
e(k+1)(
x3)
, and(
z0, w
0) =
e(k+1)(w
0)
.There is a longest z–x2path of length k
! −
1 inh∗
ka2i
k. ByLemma 6, there exists a longest s–t path of length k! −
1 ork
! −
2 inh∗
ka1
i
kthat contains (w
,w
0). Imagine that (x3, y3) is fault-free. The induction hypothesis assures a longest y3–x3path of length k
! −
1 inh∗
ka3i
k−
(
F− {
(
x3,
y3)})
. ByLemma 5, there is a longest y4–z0path of length(
k−
2) ×
k! −
1 inSk+1
[
VA] −
F , where A= {
1,
2, . . . ,
k+
1} − {
a1,
a2,
a3}
. The desired s–t path has length(
k! −
1) + ((
k! −
1) −
1) + (
k! −
1
) + ((
k−
2) ×
k! −
1) +
4=
(
k+
1)! −
1 or(
k! −
1) + ((
k! −
1) −
2) + (
k! −
1) + ((
k−
2) ×
k! −
1) +
4=
(
k+
1)! −
2. Then we consider the situation ofβ 6= γ
. Set a1=
β
and ak+1=
γ
. A link fault (h, h0) is chosen from E(h∗
kαik) ∩
F suchthat e(k+1)
(
h),
e(k+1)(
h0) /∈
F . Three cases are further discussed below. Case 1. Neither of h and h0belongs to N(
V(h∗
ka1
i
k)) ∪N(
V(h∗
kak+1i
k)). Set a3=
α
. A desired s–t path in Sk+1−
F can beconstructed in a way similar toFig. 2. Refer toFig. 2again. For the purpose of our construction, we let A
= {
1,
2, . . . ,
k+
1}
and
(
x3,
y3) = (
h,
h0)(∈
E(h∗
ka3i
k)). We assume that s and y3belong to the same partite set of Sk+1. In case s and y3belongto different partite sets of Sk+1, we have s and x3in the same partite set of Sk+1and exchange x3and y3inFig. 2.
Four vertices x1
∈
V(h∗
ka1i
k) − {s}
,
y2,
x2∈
V(h∗
ka2i
k), and y4∈
V(h∗
ka4i
k)are determined such that(
x2,
y3) =
e(k+1)
(
y3
)
, y2and x2belong to different partite sets of Sk+1,
e(k+1)(
y2) /∈
F ,(
x1,
y2) =
e(k+1)(
y2)
, and(
y4,
x3) =
e(k+1)(
x3)
.There are a longest s–x1path in
h∗
ka1i
kand a longest y2–x2path inh∗
ka2i
k, each of length k! −
1. Imagine that (x3, y3) isFig. 7. A longest s–t path in Sk+1−F whenβ 6= γ, h∈N(V(h∗ka1ik)), and h0 /∈N(V(h∗ka1ik)) ∪N(V(h∗kak+1ik)).
Fig. 8. A longest s–t path in Sk+1−F whenβ 6= γ, h∈N(V(h∗ka1ik)), and h0∈N(V(h∗kak+1ik)). (a) z6=t. (b) z=t.
a longest y4
−
t path of length (k−
2)×
k! −
1 or(
k−
2)×
k! −
2 in Sk+1[
VA] −
F , where A= {
1,
2, . . . ,
k+
1} − {
a1,
a2,
a3}
. Thedesired s–t path has length 3
×
(
k!−
1)+((
k−
2)×
k!−
1)+
3=
(
k+
1)!−
1 or 3×
(
k!−
1)+((
k−
2)×
k!−
2)+
3=
(
k+
1)!−
2. Case 2. One of h and h0belongs to N(
V(h∗
ka1
i
k)) ∪ N(
V(h∗
kak+1i
k)). Without loss of generality, we assume that h∈
N
(
V(h∗
ka1i
k))and h0/∈
N(
V(h∗
ka1i
k))∪N(
V(h∗
kak+1i
k)). A desired s–t path in Sk+1−
F can be constructed as shown inFig. 7.Set a2
=
α
and(
u2, v
2) = (
h,
h0) ∈
E(h∗
ka2i
k). Six verticesv
1∈
V(h∗
ka1i
k),u1∈
N(v
1) ∩
V(h∗
ka1i
k) − {s}
,w ∈
V(h∗
ka1i
k),u3
∈
V(h∗
ka3i
k),vk
∈
V(h∗
kaki
k), and uk+1∈
V(h∗
kak+1i
k) − {t}
are determined such that(v
1,
u2) =
e(k+1)(
u2)
,(
u3, v
2) =
e(k+1)(v
2)
, e(k+1)(
u1) /∈
F ,(vk,
u1) =
e(k+1)(
u1)
,w
and s belong to different partite sets of Sk+1,
e(k+1)(w) /∈
F ,and
(
uk+1, w) =
e(k+1)(w)
.Imagine that (u2,
v
2) is fault-free. The induction hypothesis assures a longest u2–v
2path of length k! −
1 inh∗
ka2i
k−
(
F−
{
(
u2, v
2)})
. ByLemma 6, there exists a longest s–w
path of length k! −
1 inh∗
ka1i
kthat contains (u1,v
1). ByLemma 1,there is a longest uk+1–t path of length k
! −
1 or k! −
2 inh∗
kak+1i
k. ByLemma 5, there is a longest u3–vk
path oflength
(
k−
2) ×
k! −
1 in Sk+1[
VA] −
F , where A= {
1,
2, . . . ,
k+
1} − {
a1,
a2,
ak+1}
. The desired s–t path has length(
k!−
1)+((
k!−
1)−
1)+(
k!−
1)+((
k−
2)×
k!−
1)+
4=
(
k+
1)!−
1 or(
k!−
1)+((
k!−
1)−
1)+(
k!−
2)+((
k−
2)×
k!−
1)+
4=
(
k+
1)! −
2.Case 3. Both of h0and h belong to N
(
V(h∗
ka1i
k)) ∪N(
V(h∗
kak+1i
k)). If h∈
N(
V(h∗
ka1i
k))(∈ N(
V(h∗
kak+1i
k))), then h0∈
N
(
V(h∗
kak+1i
k))(∈N(
V(h∗
ka1i
k))). Without loss of generality, we assume that h∈
N(
V(h∗
ka1i
k))and h0∈
N(
V(h∗
kak+1i
k)).Set a2
=
α
. Two verticesw ∈
V(h∗
ka1i
k) and z∈
V(h∗
kak+1i
k)are determined such that(w,
h) =
e(k+1)(
h)
and(
z,
h0) =
e(k+1)(
h0)
. A desired s–t path in Sk+1
−
F can be constructed as shown inFig. 8(a) if z6=
t, and as shown in Fig. 8(b) if z=
t.If z
6=
t, then ten vertices u1∈
N(w) ∩
V(h∗
ka1i
k) − {s}
,v
1∈
V(h∗
ka1i
k) − {s}
, u2,v
2∈
V(h∗
ka2i
k), u3,v
3∈
V(h∗
ka3i
k),u4
∈
V(h∗
ka4i
k),vk
∈
V(h∗
kaki
k), uk+1∈
V(h∗
kak+1i
k), andvk
+1∈
N(
z) ∩
V(h∗
kak+1i
k) − {t}
are further determinedsuch that e(k+1)
(
u1) /∈
F , (v
3, u1) =
e(k+1)(
u1)
, u3 andv
3 belong to different partite sets of Sk+1, e(k+1)(
u3) /∈
F ,(v
2,
u3) =
e(k+1)(
u3)
, u2andv
2belong to different partite sets of Sk+1, e(k+1)(
u2) /∈
F ,(vk,
u2) =
e(k+1)(
u2)
,v
1and t belongto the same partite set of Sk+1, e(k+1)
(v
1) /∈
F ,(
uk+1, v
1) =
e(k+1)(v
1)
, e(k+1)(vk
+1) /∈
F , and(
u4, vk
+1) =
e(k+1)(vk
+1)
.There is a longest u3–
v
3path of length k!−
1 inh∗
ka3i
k. Imagine that (h, h0) is fault-free. The induction hypothesis assures alongest u2–