Crescco CRESCCO IST-2001-33135 Critical Resource Sharing for Cooperation in Complex Systems
Workpackage 1: Efficient Resource Assignment and Communication Protocols in Wireless Networks
Deliverable D1.2 Algorithmic Solutions and Technical Recommendations for Wireless Networks I
Responsible Partner: Computer Technology Institute (GR)
Report Version: 1.0 Report Preparation Date: 15/01/03 Classification: PUB Contract Start Date: 01/01/02 Duration: 36 months Project Co-ordinator: University of Patras
Partners: Computer Technology Institute (GR) University of Geneva (CH) Centre National de la Recherche Scientific (F) Universite de Nice-Sophia Antipolis (F) Christian-Albrechts-Universitaet zu Kiel (D) Universita degli studi di Salerno (IT) Universita degli studi di Roma “Tor Vergata” (IT)
Project funded by the European Com- munity under the “Information Society Technologies” Programme (1998-2002) Contents
1 Frequency Assignment in Radio Networks 3 1.1 The Radiocoloring Problem ...... 3 1.2 Radiocoloring general and planar graphs: Complexity and approximations 5 1.3 Efficient coloring of Squares of Planar Graphs ...... 10 1.4 Radiocoloring Periodic Planar Graphs: PSPACE-completeness and ap- poximations ...... 13 1.5 Radiocoloring Hierarchically Specified Planar Graphs: PSPACE- completeness and appoximations ...... 17
2 Fault Tolerant Broadcasting, Energy Consumption and Connectivity of WNs 25 2.1 Fault Tolerant Broadcasting in WN ...... 25 2.2 Energy consumption and connectivity of WNs ...... 26 2.3 Optimal F-Reliable Protocols for the Do-All problem on Single-Hop Wire- less Networks ...... 28 2.4 Distributed Broadcast in Radio Networks for Unknown Topology . . . . . 28 2.5 The Minimum Range Assignment Problem on Linear Radio Networks . . . 29 2.6 The Minimum Broadcast Range Assignment Problem on Linear Multi-Hop Wireless Networks ...... 29 2.7 Energy Efficient Broadcasting in General Wireless Networks ...... 29 2.8 Some Recent Theoretical Advances and Open Questions on Energy Con- sumption in Ad-Hoc Wireless Networks ...... 30 2.9 On the Approximation Ratio of the MST-based Heuristic for the Energy- Efficient Broadcast Problem in Static Ad-Hoc Radio Networks ...... 30 2.10 Minimum-Energy Parameterized Problems in (Wireless Network) Graphs . 31
1 2 Chapter 1
Frequency Assignment in Radio Networks
In this chapter we present the results achieved in the project regarding the Frequency Assignment Problem (FAP) in Wireless Networks and in particular the Radiocoloring problem (RCP) as described in Deliverable D1.1. We study the computational complex- ity and provide efficient approximation algorithms for the optimizations versions of the radiocoloring problem for some important network topologies (i.e. planar) and kinds of networks (i.e. flat, periodic, hierarchical) as specified in Deliverable D1.1. These results highlight the difficulty of efficiently assigning frequencies in networks that often appear in practice and also provide approximating algorithms that may be applied to get good and simple bandwidth management solutions in reasonable time.
1.1 The Radiocoloring Problem
The Problem of Frequency Assignment in radio networks (FAP) is a well-studied, inter- esting problem. It is usually modelled by variations of graph coloring. The interference between transmitters are usually modelled by the interference graph G(V,E), where the set V corresponds to the set of transmitters and the set E represents distance constraints. The set of colors represents the available frequencies. In addition, the color of each vertex in a particular assignment gets an integer value which has to satisfy certain inequalities compared to the values of colors of nearby nodes in the interference graph G (frequency- distance constraints). We here study a variation of FAP, called the Radiocoloring Problem (RCP), that models co-channel and adjacent interference constraints.
Definition 1.1.1 (Radiocoloring) Given a graph G(V,E) consider a function Φ: V → N ∗ such that |Φ(u) − Φ(v)| ≥ 2 if D(u, v) = 1 and |Φ(u) − Φ(v)| ≥ 1 if D(u, v) = 2. The least possible number of colors (order) that can be used to radiocolor G is denoted by Xorder(G). The number ν = maxv∈V Φ(v) − minu∈V Φ(u) + 1 is called span of the radiocoloring of G and the least such number is denoted as Xspan(G).
Real networks usually reserve bandwidth (range of frequencies) rather than distinct frequencies. The objective of an assignment here is to minimize the bandwidth (span) used. The optimization version of RCP related to this objective, called min span RCP,
3 tries to find a radiocoloring for G of minimum span, Xspan(G). However, there are cases where the objective is to minimize the distinct number of frequencies used so that unused frequencies are available for other use by the system. The related optimization version of RCP here, called min order RCP, tries to find a radiocoloring that uses the minimum number of distinct frequencies, Xorder(G). The min span order RCP tries to find one from all minimum span assignments that uses the minimum number of colors. Similarly, the min order span RCP tries to find one from all minimum order assignments that uses the minimum span. Another variation of FAP related to the radiocoloring problem uses the square of a graph G:
Definition 1.1.2 Given a graph G(V,E), G2 is the graph having the same vertex set V and an edge set E0 : {u, v} ∈ E0 iff D(u, v) ≤ 2 in G.
The related problem tries to color the square of a graph G, G2, with the minimum number of colors, denoted as χ(G2). so that neighbour vertices have different colors.
Observation 1.1.1 The min order RCP of G, is equivalent to the problem of vertex 2 2 coloring of G , i.e. Xorder(G) = χ(G ).
This is so because from any coloring of G2 we can get a radiocoloring of G of the same order by doubling the color (integer) assigned to each vertex. On the other hand, every radiocoloring of G is a valid coloring of its square.
1.1.1 Related work
The problem of coloring the square of a graph G has been studied in [45] named as Distance-2-Coloring (D2C). Actually, the D2C problem tries to color the vertices of the graph G so that any two vertices of distance at most two get different colors, with the minimum number of colors. Obviously, this problem is equivalent to the problem of coloring the square of the graph G. So, in the following we use the terms coloring the square of the graph, D2C and min order RCP equivalently. In [32] it is proved that the problem of coloring the square of a general graph is NP- complete. In [45] it is proved that the problem remains NP-complete even for the class of planar graphs. In [27] and [22] it has been proved that the problem of min span radiocoloring is NP-complete, even for graphs of diameter 2. Several researchers ([6, 31, 50, 44]) studied the approximability of optimization versions of the radiocoloring problem giving approximations and upper bounds for some interest- ing families of graphs such outerplanar graphs, k-outerplanar, planar graphs, graphs of bounded treewidth, permutation split, (r,s)-civilized graphs.
4 1.2 Radiocoloring general and planar graphs: Com- plexity and approximations
In [23] we have shown the following: 0 (a) The number of colors Xorder(G) used in the min span order RCP of graph G is different from the chromatic number of the square of the graph, X(G2). (b) The radiocoloring problem for general graphs can not be approximated within a factor of n1/2−² by any polynomial algorithm, where n is the network size. However, when restricted to some special cases of graphs, the problem becomes easier. (c) The problems of min span order RCP, min order span and min span RCP are NP-complete for planar graphs. (d) We then present an O(n∆) algorithm that approximates the minimum order of
RCP, Xorder, of a planar graph G by a constant ratio which tends to 2 as the maximum degree ∆ of G increases. (e) Finally, we study the problem of estimating the number of different radio colorings of a planar graph G. This is a #P-complete problem (as can be easily seen from our completeness reduction that can be done parsimonious). We employ here techniques of rapidly mixing Markov Chains and the method of coupling for purposes of proving rapid convergence (see e.g. [30]) and we present a fully polynomial randomised approximation scheme (fpras) for estimating the number of radio colorings with λ colors for a planar graph G, when λ ≥ 4∆ + 50. Our algorithm is motivated by a constructive coloring theorem of Heuvel and McGui- ness ([29]). Their construction can lead (as we show) to an O(n2) technique assuming that a planar embedding of G is given. We improve the time complexity of the approximation, and we present a much more simple algorithm to verify and implement. Our algorithm does not need any planar embedding as input. Recently, Agnarsson and Halld´orsson ([3]) presented upper bounds for the chromatic number of square and power graphs (Gk). For planar graphs of large degree (∆ > 749) they obtain an 1.8-approximation while for planar graphs of general degree they get a 2 approximation. Bodlaender et al ([6]) proved recently that the problem of min span radiocoloring, they call it λ-labeling, is NP-complete for bipartite planar graphs, using a reduction which is very similar to our reduction. In the same work the authors presented approximations for the best λ for some interesting families of graphs: outerplanar graphs, graphs of treewidth k, permutation and split graphs. Another relevant work is that of Formann et al ([21]), where the authors proved the 13∆ 2/3 chromatic number of the square of any planar graphs is at most ( 7 ) + Θ(∆ ). Our work has motivated further relevant research. Very recently Moloy and Salavatipour in [39] improved our approximation by providing a constructive upper bound 5 of 3 ∆(G) + constant for a more general problem called L (p,q)-Labeling, which is equiv- alent to RCP for p = 2 and q = 1.
5 1.2.1 The difference between min span order radiocoloring and distance-2-coloring a) Planar Graphs
We first present (Figure 1.1) an instance of a planar graph G for which the order of min span order RCP of G is different from the minimum order of coloring the square of graph G (distance-2-coloring).
Theorem 1.2.1 There is an instance of a planar graph G for which the order of min span order RCP of G is different from the minimum order of coloring the square of graph G (distance-2-coloring).
In [23] we prove that for the graph G of Figure 1.1 the order of min span order radiocoloring is 8 while the order of distance-2-coloring is 6.
radio coloring with min span AND order: coloring the square of G, with min order:
clique vertex 6 clique vertex 2 4 1
central vertex 8 4 2 clique vertex 3
new1 5 new2 5 vertex 6 7 vertex 3 6
7 1 1 1 3 5 1 8 4 2 3 3 5 2
radial radial vertices vertices order=6 (we minimize the order) span=8 (minimum span) order=8 (minimize order)
Figure 1.1: An instance where the problem of min span order radiocoloring and the problem of distance-2-coloring have different orders.
b) Arbitrary large difference between min span order and distance-2-coloring in general graphs
Theorem 1.2.2 For any integer k, there exists an instance of (general) graph G where the order of min span order Radiocoloring is 4k while the minimum order of coloring the square of graph G (distance-2-coloring) is 3k.
In [23] we prove that for the graph G of Figure 1.2 the order of a min order radiocoloring assignment is 3k, while the minimum order distance-2-coloring 4k.
6 clique of size k formed by vertices level 1 of level 1 clique of size 2k formed by vertices of levels 1,2 clique of size k formed by vertices level 2 of level 2 clique of size 2k formed by vertices of levels 2, 3 level 3 clique of size k formed by vertices of level 3 clique of size 2k formed by vertices of levels 3, 4 level 4 clique of size k formed by vertices of level 4
Figure 1.2: An instance of a graph of Theorem 1.2.2 where the order of min span radio- coloring and the order of distance-2-coloringlevel 1 differs by1, 2, 3k, 4,, 5 for, ..., k any given k. clique of size 2k formed by vertices
of levels 1,2
+1, +2, +3, +4, ..., 2 level 2 1.2.2 The inapproximabilityclique of size 2k formed of by verti Radiocoloringces in general net-
of levels 2, 3
level 3 2 +1, 2 +2, 2 +3, 2 +4, ..., 3 works clique of size 2k formed by vertices of levels 3, 4 The problem of min orderleve Radiocoloringl 4 for general1, 2, graphs 3, 4, 5, ..., k is proved to be hard to be approximated within the ratio of n1/2−² (where n is the number of vertices of the graph), by any polynomial time approximation algorithm. min order Theorem 1.2.3 The minimum order Radiocoloring Problem for general graphs cannot be approximated within the ratio of n1/2−² unless P=NP.
The reduction is performed by reducing from the ordinary vertex coloring problem which is known to be non-inapproximable within a factor of n1−², see [23].
1.2.3 The NP-Completeness of the RCP for Planar Graphs In [23], we have proved that the decision version of min span radiocoloring remains NP- complete for planar graphs.
Theorem 1.2.4 The min span radiocoloring problem is NP-complete for planar graphs.
The proof is performed by reducing from ordinary 3-coloring of planar graphs. The new graph G0 with the desired properties is obtained by any planar graph G applying local transformation and the component design technique ([26]). We remark that this result is not implied by the NP-completeness result for the distance-2-coloring (which is equivalent to min order RCP), since by Theorem 1.2.1 the two problems are different.
1.2.4 A Constant Ratio Approximation Algorithm for Planar Networks The Algorithm
We provide here an approximation algorithm for radiocoloring of planar graphs by mod- ifying the constructive proof of the theorem presented by Heuvel and McGuiness in [29]. The algorithm presented here is of better time complexity (i.e. O(n∆)) compared to the (implicit) algorithm in [29] which needs O(n2). The improvement is achieved by per- forming the heavy part of the computation of the algorithm only in some instances of G
7 instead of all in [29]. This enables less checking and computations in our algorithm. Also, the behavior of our algorithm is very simple and more time efficient for graphs of small maximum degree. Finally, the algorithm provided here needs no planar embedding of G, as opposed to the algorithm implied in [29]. The main theorem of [29] provides an upper bound for planar graphs for a more general problem called L (p,q)-Labeling, which is equivalent to RCP for p = 2 and q = 1. By setting p = q = 1 and using the observation λ(G; 1, 1) = χ(G2), where χ(G2) is the chromatic number of the graph G2, we get immediately, that:
Theorem 1.2.5 ([29]) If G is a planar graph with maximum degree ∆ then χ(G2) ≤ 2∆ + 25.
The theorem is proved using two Lemmas. Let d(v) the degree of vertex v. The first of the two Lemmas is the following:
Lemma 1.2.6 ([29]) Let G be a simple planar graph. Then there exists a vertex v with k neighbors v1, v2, . . . , vk with d(v1) ≤ · · · ≤ d(vk) such that one of the following is true:
(i) k ≤ 2; (ii) k = 3 with d(v1) ≤ 11;
(iii) k = 4 with d(v1) ≤ 7 and d(v2) ≤ 11;
(iv) k = 5 with d(v1) ≤ 6, d(v2) ≤ 7, and d(v3) ≤ 11.
The second Lemma, is quite similar. We apply the following operation to G: For an edge e ∈ E let G/e denote the graph obtained from G by contracting e. We provide below a high level description of our algorithm. Algorithm Radiocoloring(G) [I] Sort the vertices of G by their degree.
[II] If ∆ ≤ 12 then follow Procedure (1) below: Procedure (1): Every planar graph G has at least one vertex of degree ≤ 5. Now, inductively assume that any proper (in vertices) subgraph of G can be radiocolored by 66 colors. Consider a vertex v in G with degree(v) ≤ 5. Delete v from G to get G0. Now recursively radiocolor G0 with 66 colors. The number of colors that v has to avoid is at most 5∆ + 5 ≤ 65. Thus, there is one free color for v.
[III] If ∆ > 12 then
1. Find a vertex v and a neighbor v1 of it, as described in Lemma 1.2.6, and set e = vv1. 2. Form G0 = G/e (G0 = (V 0,E0) with |V 0| = n − 1, while |V | = n) and denote the new vertex in G0 obtained by the contraction of edge e as v0. 0 Modify the sorted degrees of G by deleting v, v1, and inserting v at the appropriate place, and also modify the possible affected degrees of the neighbors of both v and v1. 3. φ(G0) =Radiocoloring(G0)
8 4. Extend φ(G0) to a valid radiocoloring of G: 0 0 (a) Set v1 = v and give to v1 the color of v . (b) Color v with one of the colors used in the radiocoloring φ of G0.
Notice first that Procedure (1) produces a proper coloring of G2 with X = 66 colors. Then, assign the frequencies 1, 3,..., 2X − 1 to the obtained color classes of G2. This is a proper radiocoloring of G with the same number of colors.
Theorem 1.2.7 (Performance) The algorithm Radiocoloring(G) outputs a radiocolor- ing of G using no more than max{66, 2∆ + 25} colors and approximates Xorder(G) by a 25 66 constant factor of at most max{2 + ∆ , ∆ }. It runs in O(n∆) sequential time.
A key observation for the proof of the Theorem is that the graph G0 obtained (G0 = G/e with e is chosen so that Lemma 1.2.6 holds for this edge) has ∆(G0) ≤ ∆(G). Now, the proof is performed by induction. Assume a radiocoloring assignment of G0 using the above number of colors. By the way edge e is chosen (so that to fulfills the requirements of Lemma 1.2.6), we can prove that we can extend this assignment to a radiocoloring assignment of G. See [23] for the proof of the Theorem.
1.2.5 An FPRAS for the Number of Radiocolorings of a Planar Graph Sampling and Counting
Let G be a planar graph of maximum degree ∆ = ∆(G) on vertex set V = {0, 1, . . . , n−1} and C be a set of λ colors. Let φ : V → C be a (proper) radiocoloring of the vertices of G. Such a radiocoloring always exists if λ ≥ 2∆ + 25 and can be found by our O(n∆) time algorithm of the previous section.
Consider the Markov Chain (Xt) whose state space Ω = Ωλ(G) is the set of all radio- colorings of G with λ colors and whose transition probabilities from radiocoloring Xt are modelled by: 1. choose a vertex v ∈ V and a color c ∈ C uniformly at random (u.a.r.) 2. recolor vertex v with color c. If the resulting coloring X0 is a proper radiocoloring 0 then let Xt+1 = X else Xt+1 = Xt. The procedure above is similar to the “Glauber Dynamics” of an antiferromagnetic Potts model at zero temperature, and was used by [30] to estimate the number of proper colorings of any low degree graph with k colors.
The Markov Chain (Xt), which we refer to in the sequel as M(G, λ), is ergodic, provided λ ≥ 2∆ + 26, in which case its stationary distribution is uniform over Ω. We show here that M(G, λ) is rapidly mixing i.e. converges, in time polynomial in n, to a close approximation of the stationary distribution, provided that λ ≥ 2(2∆ + 25). This can be used to get a fully polynomial randomised approximation scheme (fpras) for the number of radiocolorings of a planar graph G with λ colors, in the case where λ ≥ 4∆ + 50. For some definitions and measures used below see [23].
Theorem 1.2.8 The above method leads to a fully polynomial randomized approximation scheme for the number of radiocolorings of a planar graph G with λ colors, provided that λ > 2(2∆ + 25), where ∆ is the maximum degree of G.
9 1.3 Efficient coloring of Squares of Planar Graphs
In [2] we have considered the min order RCP problem on planar graphs. The motivation for this work is important since most real networks are planar. We provide a new coloring algorithm, called MDsatur, and we analyze its performance on the square of any planar graph. As we have already said, in section 1.1, this problem is equivalent to the min order RCP. We prove that MDsatur colors the square of any planar graph G with at most 1.5∆(G) + c colors, where c is a constant, when ∆(G) ≥ 8. When ∆(G) < 8 then as concluded by the work of Heuvel and McGuiness [29] a constant number of colors are enough to color the squares of such planar graphs. Hence, we do not analyze the performance of our algorithm on such graphs. MDsatur is inspired by algorithm Dsatur of Br´elaz [7], which is a greedy coloring algorithm with very good behaviour on classes of graphs with similar characteristics as planar graphs and their squares. Dsatur colors the vertices of a given graph, say G, one at a time based on a dynamic order, O, of them. Dsatur finds at each point, i.e. i, which of the currently uncolored vertices of G will be the ith element of O (denoted by Oi). Then, it correctly colors Oi with the smallest available color (i.e. it is not assigned to any of its colored neighbors). If Oi = v, then v is the current uncolored vertex of G which has the biggest value on its degree of saturation.
Definition 1.3.1 The degree of saturation of vertex v is denoted by ds(v) and is the number of distinct colors that have been assigned to its neighbors.
Our algorithm, MDsatur, also colors the vertices of a given graph, i.e. G2, in the same way (namely, greedily choosing the smallest available color for each vertex) but following a different dynamic order on its vertices. Our algorithm differs from the original algorithm Dsatur in the way of growing the dynamic order O. More precisely, at each of its points, i.e. i, it orders the current uncolored vertices of G by taking into account not only their degree of saturation, but also by considering how ‘close’ is each of the potential vertices for the position i of order O to the current last colored vertex, let us call it CurrLast, following a procedure similar to Breadth First Search. The term ‘close’ is related to the closeness of the behaviour of the neighbors of a vertex, i.e. u, to its behaviour. The most ‘close’ neighbor of u is the one whose coloring is determined at most by the colors of the other neighbors of u. In most of the cases the location of this neighbor of u, in an embedding of G in the plane, is very near to the location of u.
1.3.1 Presentation of the coloring Algorithm M Dsatur
In this paragraph, we present the coloring algorithm MDsatur in pseudo-code.
10 Algorithm: MDsatur Input : a graph G = (V,E). Output: a coloring of G. Begin
1. (a) Set each vertex of G as uncolored, and set a dynamic order on them, called O, as empty. (b) Select any vertex of G.
Add it in the beginning of O (i.e. in O1), and color it with color 1.
2. For i = 2 to n = |V | do
(a) Set X to be the set of the current uncolored vertices which have the maximum value on their degree of saturation. (b) For each vertex v ∈ X do Find which colored neighbor of v is colored first (i.e. its oldest colored neighbor) (see req. A2). (c) Select from X the vertex v which has the oldest colored neighbor, denoted by ON, and also has the maximum number of common neighbors with the last colored vertex (namely Oi−1) (see req. A3).
Set Oi = v. If more than one of the vertices of X satisfies these requirements, then select at random one of them with equal probability. If none of the uncolored neighbors of ON, which are in X, has common
neighbors with vertex Oi−1, then select one of them at random. (d) color v with the smallest of its allowable colors. If none of the used colors is allowable for this vertex, insert a new color.
3. Return the color of each vertex.
End
This algorithm consists a novel approach in the coloring of squares of planar graphs. The inspiration for the design of MDsatur originates from the behaviour of the coloring algorithms Dsatur and NoChoice of Br´elaz and Turner, respectively, on random [47] and k-colorable random graphs [48], respectively. These algorithms have very good and optimal performance on the above families of graphs. This happens, because each vertex of these graphs belongs to a clique of size almost maximum and maximum, respectively, and the chromatic number of these graphs is strongly related to the size of their maximum clique. For our case, it is easily concluded from the known results, that the chromatic number of the square of any planar graph G is strongly related to the size of its maximum clique.
11 More precisely it is known by Molloy and Salavatipour that χ(G2) ≤ 1.66∆ + 24 see [39]. Hence, χ(G2) ≤ 1.66ω(G2) + 24. An experimental evaluation of the coloring algorithm Dsatur on squares of planar graphs reveals some interesting properties for the performance of this algorithm on this class of graphs. Focusing our attention on the structure of the planar graphs where Dsatur has its worst performance and studying carefully its behaviour we conclude, that we could guide it to have better performance. Based on these experimental results we modified Dsatur, in a not straightforward way, by taking into account additional requirements. These limit the number of the potential uncolored vertices of G2 which claim to be the 2 ith element of order O (Oi). Suppose that we apply MDsatur to graph G . Let i be the current point of this algorithm in its application on G2. Then, the current uncolored vertex which manages to become the vertex in position i of O, let it be v, satisfies the following requirements in turn: A1) has the maximum value on its degree of saturation in the current point of this algorithm (as Dsatur), A2) has the oldest colored neighbor (i.e. sits in the leftmost position in O among the neighbors of v) and A3) is the ‘closer’ neighbor of CurrLast. Our algorithm has better performance than Dsatur, because it locally colors a given graph and the colored subgraph obtained in each point of its application on this graph is more dense, than the corresponding graph obtained by Dsatur.
1.3.2 Our results Our main result is presented in the following Theorem.
Theorem 1.3.1 Let G be a planar graph. Algorithm MDsatur colors G2 using at most 1.5∆(G) + c colors, where c is a constant, when ∆(G) ≥ 8.
In order to prove the assertion of this Theorem we have characterized the worst case graph instances of squares of planar graphs for our algorithm according to the next The- orem. Then, by a case analysis we prove that MDsatur colours any such graph with at most 1.5∆(G) + c colors.
Theorem 1.3.2 The minimal worst case squares of planar graphs for the coloring algo- rithm MDsatur are the squares of the graphs obtained by any planar r−regular graph, replacing each of its edges, i.e. e, with a set of vertices and joining them with e0s end- points.
The proof of Theorem 1.3.2 is based on the fact, that MDsatur is a greedy coloring algorithm (it colors the vertices of a given graph based on an ordering of its vertices with an available already used color. If any such color does not exist, then it uses a new one.) without backtracking. Suppose that MDsatur is applied on a graph G and let H be an uncolored subgraph of G in an intermediate point of the algorithm. If the colored neighbors of each vertex of H have all the currently used colors (used in the colouring of H0), then MDsatur has to increase the number of colors used at least by χ(H). This process can continue recursively on the current uncolouled subgraph of G. We denote the graph H in the first case by H1. After the coloring of H1, the value of XMDsatur (that denotes the number of colors already used) is increased and it continues to increase whenever there is another uncolored subgraph of G such that all the vertices of this graph have colored neighbors with XMDsatur different colors. Seeing abstractly this procedure we can notice that there is a sequence of graphs having the properties of graph H1. We
12 denote the graphs of this sequence as Hi, for i ≥ 1. Observe that the coloring of graph Hi is based on the coloring of each of the graphs Hj, for j < i. The index of each graph Hi shows its level in the above sequence. In the case of squares of planar graphs, we prove that the maximum number of levels of graph H is three. Thus, we focus our attention on the structure of graphs H0,H1,H2. To specify the possible structures of these graphs we have firstly determined the structure of each possible square root of any clique of size more than eight.
1.3.3 Conclusions Our algorithm colors the square of any planar graph G with at most 1.5∆(G) + c colors. We conjecture that this is not a tight bound in the case where χ(G2) = ∆(G) + 1. So, this is an open problem for this case. We also conjecture that MDsatur colors any planar graph with at most 5 colors.
1.4 Radiocoloring Periodic Planar Graphs: PSPACE-completeness and appoximations
In [25] we investigate the min span RCP for an interesting family of infinite planar graphs, called periodic planar graphs. A periodic graph G is defined by an infinite sequence of repetitions of the same finite graph Gi(Vi,Ei). The edge set of G is derived by connecting some of the vertices of each iteration Gi to some of the vertices of the next iteration Gi+1, the same for all iterations. Infinite periodic graphs usually represent finite networks that accept periodic (in time, e.g. daily) requests for frequency assignment. We note that periodic interference graphs usually represent networks of great practical interest, since in many networks the requests for frequency assignment exhibit some periodic behavior. That is, the network accepts periodic (e.g. daily) requests for frequency assignment. Each request has a starting and ending time and a node where it is applied. Two requests interfere if they apply for nearby nodes and their time intervals overlap. The assignment should be such that there is no time overlap between any two nearby requests of the same or the preceding and following periods of requests. Alternatively, infinite periodic graphs can model very large networks produced by the repetition of a small graph. Note in this context that many real networks consist of the repetition of the same component. We focus here on planar periodic graphs, because in many cases real networks are planar and because of the independent mathematical interest of this family of graphs.
Definition 1.4.1 Linear Periodic Planar Graph G: A linear periodic planar graph is defined as follows: Let Ge be an arbitrary finite connected planar graph. Let V the vertex set of Ge. Let e also E0 be the edge set of G. Let E+ be a specific set of ordered pairs (u, v) of the nodes e of G. Note that E+ must be a set of ordered pairs of vertices whose connection according to the rule (c2) below leads to planarity preservation. Consider the two-way infinite sequence of graphs ...,Gi,Gi+1,..., where each Gi is isomorphic to Ge. The infinite graph G is obtained from this sequence as follows:
(a) We assume a line (in fact, any 1-dimensional infinite simple curve) on which we select discrete points . . . , i, i + 1, i + 2,..., such that:
13 (a1) Each point in the line is replaced by Ge.
(a2) Each edge (i, i + 1) in the line is replaced by E+. (a3) For any finite subset of consecutive points in the line, replacing the points e of the line by graphs G end the edges between them by E+, the resulting graph is planar.
(b) The vertex set of G is the union of the vertex sets of the sequence ...,Gi,Gi+1,....
(c) The edges of G are (c1) The union of edge sets of the sequence of Gis (i.e., the e e edge set E0 of G) (c2) For each pair of adjacent copies of G, call them Gi, Gi+1, we use the E+ specification of G to connect the nodes of Gi corresponding to the first elements of the pairs in E+ to the nodes of Gi+1, corresponding to the second elements of the pairs in E+. e We denote a linear periodic planar graph by G = (G(V,E0),E+). This pair is called the finite specification of G.
Summary of Results In [25] we have given the following results: 1. We first prove that the min span RCP is PSPACE-complete for periodic planar graphs. (The space is polynomial with respect to the size of the finite specification of G.)
2. We provide an O(n(∆(G)) time algorithm, which obtains a radiocoloring of a periodic planar graph G that approximates the minimum span within a ratio which tends to R as ∆(G) tends to infinity. The term R is the approximation ratio obtained by any approximation algorithm for the min span RCP of a finite planar graph G achieving a span of at most R · ∆(G) + constant. The best known radiocoloring algorithm for planar graphs has an approximation ratio which tends to 1, 66 [39], so this is the approximation ratio obtained by our algorithm too.
Related Work A model for periodic graphs (called l-dimensional periodic graphs) was first presented by Orlin in [41]. The model of periodic graphs considered in this work is similar to that of Orlin for the 1-dimensional case, l = 1 (also called 1-dimensional periodically specified graphs or simply periodically specified graphs), when restricted to planar instances. The complexity of various basic problems of periodically specified graphs was studied by Orlin [41] and Wanke [49]. In [41, 36, 43] it is proved that the problems of Maximum Independent Set (MIS), Hamiltonial Path, Partition into Triangles, SAT, 3-coloring for periodically specified graphs are PSPACE-complete. The appoximability of basic prob- lems on infinite periodic graphs was studied by several researchers ([19, 28, 42]) giving efficient algorithms for solving problems such as determining strongly connected compo- nents, testing the existence of cycles, bipartiteness, planarity and minimum cost spanning forests for periodically specified graphs. Marathe et al [35] presented several PSPACE-hardness results and also efficient ap- proximation schemes for partitioning problems including MIS, min vertex cover and max- SAT for periodically specified graphs when restricted to planar instances. However, their approximation technique for periodically specified graphs (illustrated for the MIS prob- lem) can not directly apply for coloring problems, considered here, because it takes the
14 union of partial solutions-subsets of the infinite graph and thus it does not consider all the vertices; something not allowed in coloring problems.
1.4.1 Embeddings of Periodic Planar Graphs Studying periodic planar graphs, we use the notions of embedding and discover some interesting properties on the embeddings of a linear periodic planar graph. We use the notion of an embedding of a planar graph. Definition 1.4.2 Planar Embedding (of a periodic graph G)([40]): For each node v of G, there is an adjacency list, such that all neighbours of v appear in clockwise order with respect to an actual drawing of G. The following Lemma reveals important information about the structure of a linear peri- odic planar graph. Lemma 1.4.1 Any linear periodic planar graph G can be embedded in the plane by inter- changing at most two different planar embeddings of the graph obtained by an iteration i e of G, Gi (which is isomorphic to G) and the set of edges connecting Gi with the previous e and next iterations, sets Ei−, Ei+ (each of which is equal to E+), called Extended G. The Lemma is proved basically via exhaustive check of all possible edge sets that can lead to a planar graph.
1.4.2 The PSPACE-Completeness of min span RCP for Periodic Planar Graphs We have proved that min span RCP is PSPACE-complete for periodic planar graphs. In order to show this, we need to prove that a number of problems are PSPACE-complete. Our PSPACE-completeness proofs utilize known constructions for the NP-completeness of corresponding problems for ordinary graphs. However, note that applying those con- structions on the infinite periodic graph G we need infinite time to get the transformed graph G0 with the desired properties. We manage to apply the transformation only to a part of the infinite graph (an iteration) and from the obtained graph to get the transfor- mation of the whole infinite periodic graph, thus the new graph G0, in time polynomial to the size of the finite specification of G. Moreover, the new graph fulfills the desired properties. These are achieved by exploiting some ‘locality characteristics’ that the con- structions utilized here exhibit, i.e. the construction applies locally on a part (vertex or edge) of this part involving only information of the neighborhood of the part and it affects only to this neighborhood. This, combined with the repetitive structure of an infinite pe- riodic graph enables us to get constructions of polynomial time in the size of the finite specification of the graph. The main complexity Theorem of [25] is the following: Theorem 1.4.2 Let r ≥ 8 be an even integer. The problem of deciding whether a periodic e planar graph G = (G(V,E0),E+) of maximal degree r −2 can be radiocolored using a span of size at most r is PSPACE-complete. The reduction adapts the ‘local’ construction by Bodlaender et al in [6] to show NP- completeness of min span RCP of ordinary planar graphs reducing it from the 3-coloring of planar graphs with a given 4-edge coloring. Here we reduce from 3-coloring of periodic planar graphs with a given 4-edge coloring which we also prove to be PSPACE-complete using the same methodology. For the proof of the Theorem see [25].
15 1.4.3 An Efficient, Constant Ratio Approximation Algorithm for min span RCP for Periodic Planar Graphs We present an efficient time, constant ratio approximation algorithm that approximates the min span radiocoloring problem for periodic planar graphs with the same ratio as the ratio obtained by the best known approximation algorithm for ordinary planar graphs
(which we use as a subroutine for the finite specification), for the same problem.
The modified graph 3=2
1=1
The algorithm groups together every four successive iterations of the infinite graph. Call 2=2 the j − th such group as Ggroup j. Denote the first graph of such a group as G(j)1 or G1, the second as G(j)2 or G2 and so on until the fourth. Consider any such group Ggroup j. The algorithm modifies the graph to obtain a new 0 graph, Ggroup j as follows: The new graph has the same vertex and edge set as the graph Ggroup j except from the following modifications on the first and the fourth graphs of group Ggroup j: The edges of the first graph of the group connecting it to the previous iteration are removed. Also, the edges of the last iteration, the fourth, connecting a vertex of this iteration to the next iteration are removed. Then for each removed edge uv u ∈ G(j)1 and v ∈ G , add edge vu0 where u0 is the corresponding to u vertex in G . An example (j−1)4 G (j)4 group j G of the graph obtained by a periodic graph is illustrated in Figure 1.3.group (j+1)
added
u''≠ G j*3-1 u d u ele ' ted deleted u ≠G G G j*3+1 1 G2 3 G4
E1 E2 E1 E2
0 Figure 1.3: The graph Ggroup j produced by the Group j of the periodic graph G
0 The graph Ggroup j has two critical properties compared to the initial periodic planar 0 graph G: (i) it has the same maximum degree as the initial graph G, i.e. ∆(Ggroup j) = ∆(G) and (ii) using Lemma 1.4.1 it is proved that it is a planar graph too.
The Periodic Radiocoloring Partitioning Algorithm (PRPA) The following definition is needed by the Algorithm. The definition uses the observation that the optimal span, S∗, of a radiocoloring of a graph G with maximum degree ∆(G) is clearly S∗ ≥ ∆(G).
Definition 1.4.3 RC Algorithm: Let an RC Algorithm be any known min span radiocoloring polynomial time approximation algorithm for finite planar graphs with per- formance ratio R (when ∆(G) is used as a lower bound), i.e. if S∗ is the optimal span and SRC is the span obtained by the algorithm then there are constants R > 1 and b such that ∗ ∆(G) ≤ S ≤ SRC ≤ R · ∆(G) + b For example the algorithm of [29] is an RC algorithm with R = 2 and b = 35.
16 Algorithm PRPA
0 1. Run an RC algorithm, on graph Ggroup j. Let SRC be the span obtained 0 by RC on Ggroup j.
2. For all j = 1, 2,... color the four graphs G(j−1)∗4+1, G(j−1)∗4+2, G(j−1)∗4+3, , G(j−1)∗4+4 of the group Ggroup j as follows: Set the color of each vertex of graph G(j−1)∗4+k, k = 1, 2, 3, 4 to the color of 0 its corresponding vertex, in Vk of V (Ggroup j). Step 2 produces a radiocoloring of the whole periodic graph G with span SRC .
Theorem 1.4.3 The Algorithm PRPA produces a radiocoloring on the infinite linear periodic graph G, runs in time O(T (RC)) and approximates the span within an asymptotic 0 ratio of R, where T (RC) is the time needed for the RC Algorithm to run on Ggroup j.
Since the modified graph is planar and has the same maximum degree as the infinity graph, we get:
Corollary 1.4.4 If the RC algorithm is that of [39] for planar graphs, then algorithm PRPA has R = 1, 66 and b = 24 and runs in O(n(∆(G)) time, where n = |Vi|.
1.5 Radiocoloring Hierarchically Specified Planar Graphs: PSPACE-completeness and appoxima- tions
Many practical applications of graph theory and combinatorial optimization in CAD sys- tems, VLSI design, parallel programming and software engineering involve the processing of large (but regular) objects constructed in a systematic manner from smaller and more manageable components. As a result, the graphs that abstract such circuits (designs) also have a regular structure and are defined in a systematic manner using smaller graphs. The methods for specifying such large but regular objects by small specifications are referred to as succinct specifications. One way to succinctly represent objects is to specify the graph hierarchically. Hierarchical specifications are more concise in describing objects than or- dinary graph representations. A well known hierarchical specification model, considered in this work, is that of Lengauer, introduced in [33, 34], referred to as L-specifications. Real communication networks, especially wireless and large ones, that may be struc- tured in a hierarchical way and are usually planar. In [1] we study min order RCP on L-specified hierarchical graphs.
1.5.1 Summary of Results In [1] we investigate the computational complexity and provide efficient approximation algorithms for the RCP on a class of L-specified hierarchical planar graphs which we call Well-Separated (WS) graphs. In such graphs, levels in the hierarchy are allowed to directly connect only to their immediate descendants. In particular:
1. We prove that the decision version of the RCP for Well-Separated L-specified hier- archical planar graphs is PSPACE-complete.
17 2. We present two approximation algorithms for RCP for this class of graphs. These algorithms offer alternative trade-offs between the quality and the efficiency of the solution achieved. The first one is a simple and very efficient 3,33-approximation algorithm, while the second one achieves a better solution; it is a 2,66-approximation algorithm, but is less efficient, although of polynomial time.
We note that the class of WS L-specified hierarchical graphs considered here can lead to graphs that are exponentially large in the size of their specification. The WS class is a subclass of the class of L-specified hierarchical graphs considered in [35], called k-level-restricted graphs.
1.5.2 Related Work In a fundamental work, Lengauer and Wagner [34] proved that the following problems are PSPACE-complete for L-specified hierarchical graphs: 3-coloring, hamiltonian circuit and path, monotone circuit value, network flow and independent set. For L-specified graphs, Lengauer ([33]) have given efficient algorithms to solve several important graph theoretic problems including 2-coloring, min spanning forest and planarity testing. Marathe et al in [36, 35] studied the complexity and provided approximation schemes for several graph theoretic problems for L-specified hierarchical planar graphs including maximum independent set, minimum vertex cover, minimum edge dominating set, max 3SAT and max cut. We remark that the PSPACE-completeness proof of planar 3-COLORING of WS L-specified hierarchical graphs provided in this work, is not implied by known PSPACE- completeness results of the same problem for similar (but different) classes of planar graphs. This is so because our PSPACE-completeness proof for planar 3-COLORING concerns a subclass studied in [36] of the L-specified hierarchical planar graphs for the same problem. Moreover, the PSPACE-completeness proof of planar 3-COLORING of [37] for L-specified hierarchical graphs which are simultaneously planar and unit disks concerns a different class of hierarchical planar graphs than the class of WS L-specified hierarchical planar graphs considered here. Note, also that most of the work done so far on approximations of PSPACE-complete problems, has basically addressed such ‘finding a subset’ problems and not coloring prob- lems. The methodologies applied for such problems, such as maximum independent set (MIS) in [35], do not directly apply to coloring problems, since they exclude from a solu- tion some vertices of each graph of the L-specification of the hierarchical graph, something not allowed in graph coloring problems. To our knowledge, the only work studying ap- proximations to coloring problems on hierarchical graphs is the work of [37]. We remark that the best currently known approximation ratio for the RCP on ordinary (non-hierarchical) planar graphs (which are much simpler to color than the hierarchical ones) is 1,66 ([39]). Also, the only known results on any kind of coloring problems have been shown for the vertex coloring for a special kind of hierarchical graphs (k-level re- stricted unit disk graphs) achieving a 6-approximation solution ([37]). We do not see an easy way of using their algorithm for the radiocoloring problem studied here.
1.5.3 Preliminaries We study the RCP on hierarchical graphs as specified by Lengauer [33].
18 Definition 1.5.1 (L-specifications, [33]) An L-specification Γ = (G1, ··· ,Gi, ··· ,Gn), where n is the number of levels in the specification, of a graph G is a sequence of labeled undirected simple graphs Gi called cells. The graph Gi has mi edges and ni vertices. The pi of the vertices are called pins. The other (ni −pi) vertices are called inner vertices. The ri of the inner vertices are called nonterminals. The (ni −ri) vertices are called terminals. The remaining ni − pi − ri vertices of Gi that are neither pins nor nonterminals are called explicit vertices. Each pin of Gi has a unique label, its name. The pins are assumed to be numbered from 1 to pi. Each nonterminal in Gi has two labels (v, t), a name and a type. The type t of a nonterminal in Gi is a symbol from G1, ··· ,Gi−1. The neighbours of a nonterminal vertex must be terminals. If a nonterminal vertex v is of type Gj in Gi, 1 ≤ j ≤ i − 1, then v has degree pj and each terminal vertex that is a neighbor of v has a distinct label (v, l) such that 1 ≤ l ≤ pj. We say that the neighbor of v labeled (v, l) matches the l-th pin of Gj.
: explicit vertices : pins : nonterminals G3
G G 1 1 G 2 G G1 G2 1 Hierarchy Tree