See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/237306987

An Introduction to Graph Theoretical Methods in Geography

Book · January 1977

CITATIONS READS

27 916

1 author:

Keith Tinkler
Brock University
84 PUBLICATIONS 993 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

I retired long before I ran into Research Gate Nowadays I doodle along as suits my fancy View project

none at all View project

All content following this page was uploaded by Keith Tinkler on 25 June 2015.

The user has requested enhancement of the downloaded file.

 AN INTRODUCTION TO GRAPH THEORETICAL

METHODS IN GEOGRAPHY

K.J.Tinkler

ISSN 0306 - 6142

ISBN 0 90224668 2

© K.J. Tinkler 1977

 CONCEPTS AND TECHNIQUES IN MODERN GEOGRAPHY
CATMOG(Concepts and Techniques in Modern Geography)
AN INTRODUCTION TO GRAPH THEORETICAL METHODS IN GEOGRAPHY
1 An introduction to Markov chain analysis - L. Collins by
2 Distance decay in spatial interactions - P.J. Taylor Keith J. Tinkler
3 Understanding canonical correlation analysis - D. Clark (Brock University, Ontario, Canada).
4 Some theoretical and applied aspects of spatial interaction
shopping models - S. Openshaw
5 An introduction to trend surface analysis - D. Unwin
CONTENTS
Page
6 Classification in geography - R.J. Johnston
7 An introduction to factor analytical techniques - J.B. Goddard & A. Kirby
I INTRODUCTION
(i) What is a graph? 3
8 Principal components analysis - S. Daultrey
(ii) Prerequisites and Literature 4
9 Causal inferences from dichotomous variables - N. Davidson (iii) Some definitions: Types of graphs 5
(iv) Deriving graphs in actual examples 12
10 Introduction to the use of logit models in geography - N. Wrigley
(v) Matrix representations of a graph 12
11 Linear programming: elementary geographical applications of the
transportation problem - A. Hay II MATRIX OPERATIONS AS PROCESSES IN GRAPHS
12 An introduction to quadrat analysis - R.W. Thomas (i) Matrix addition and subtraction 14
(ii) Multiplying or dividing a matrix by a scalar 17
i3 An introduction to time-geography - N.J. Thrift
(iii) The Null matrix, 0, and the Identity matrix, I 17
14 An introduction to graph theoretical methods in geography - K.J. Tinkler (iv) Matrix multiplication 18
(v) Multiplying a matrix by a vector 22
15 Linear regression in geography - R. Ferguson
(vi) O's or is in the main diagonal of A 23
16 Probability surface mapping. An introduction with examples and (vii) Matrix multiplication involving digraphs - asymmetric 24
Fortran programs - N. Wrigley matrices
(viii) Matrix multiplication of two different matrices or 25
17 Sampling methods for geographical research - C. Dixon & B. Leach
graphs
18 Questionnaires and interviews in geographical research - 27
C. Dixon & B. Leach 27
28
19 Analysis of frequency distributions - V. Gardiner & G. Gardiner
29
20 Analysis of covariance and comparison of regression lines - J. Silk
21 An introduction to the use of simultaneous-equation regression analysis
III 1st ORDER PROPERTIES OF GRAPHS
in geography - D. Todd (i) Nodality 30
(ii) The degree vector of a graph and the fixed vector of 31
22 Transfer function modelling: relationship between time series
the transition matrix of the same graph
variables - Pong-wai Lai (in preparation)
31
23 Stochastic processes in one-dimensional series: an introduction - 32
K.S. Richards (in preparation )
(v) Simple graphs 32
24 Linear programming: the Simplex method with geographical applications -
James E. Killen (in preparation) IV HIGHER ORDER PROPERTIES OF GRAPHS
Price each: 75p or S1,50 except Catmogs 16 s 19 which are £1.25 (S2.50) (i) Routes, paths, chains and cycles 35
(ii) The cyclomatic number, p, and the x index 36
The series is produced by the Study Group in Quantitative Methods, of (iii) Connectivity and distance 37
the Institute of British Geographers. For details of membership of the (iv) Components of a graph, articulation points and 41
Study Group, write to the Institute of British Geographers, 1 Kensington bridges
Gore, London SW7. The series is published by GEO ABSTRACTS, University
of East Anglia, Norwich NR4 7JT , to whom ail other enquiries should be
addressed.
1
 V EXTENSIONS
(i) Nodal Analysis 42
(ii) Traversability 44 I INTRODUCTION
(iii) Covering a graph 45
(iv) An Algorithm for minimal independent point covers 46 (i) What is a graph?
(v) Colouring a graph 47
A graph consists of a set of points and a set of relationships between
VI AN APPLICATION : KISII DISTRICT, KENYA 48 pairs of points. A pair of such points may be termed a line, link, connect-
ion or relation from the first point to the second one. In geographical
VII PROBLEMS AND LIMITATIONS 52 applications the points may be places, towns, road junctions, telephone
exchanges, regions, etc. and the relationships between them can be defined
BIBLIOGRAPHY 53 as is appropriate to the problem. Relationships fall in three main cate-
gories: (1) equal and reciprocal : town i is joined to town j (and vice-
versa) by a road, (2) unequal and reciprocal : telephone exchange i calls
exchange j with 20 calls a day whereas exchange j calls i with only 10
calls, (3) non-reciprocal : in one-way systems cars can go from junction i
to junction j but not reciprocally (except perhaps by a circuitous route).
Category (1) gives rise to graphs, and categories (2) and (3) give rise to
directed graphs or digra hs. Because geographical problems frequently deal
with a finite number of p aces, and with the inter-relationships between
them, it follows that graph theory has much to offer in elucidating the
form and structure of geographic space.

Acknowledgements Graph theory is primarily a theoretical framework for looking at a

problem and in this respect differs from most volumes in this series which
I am indebted to Neil Wilson, formerly Teaching Assistant in Geography are primarily concerned with a single well-defined technique. The liter-
at Brock University, and Thomas A. Jenkyns of the Mathematics Department ature using graph theoretical methods naturally reflects this characteristic
for their careful review of an early version of the manuscript. An anonymous and consequently there is no single article that exemplifies all the tech-
referee made valuable comments on a later draft. niques discussed here, and there is no reason why there should be. Most
of the basic ideas of graph theory can be exemplified by small diagrams,
Phyllis Riesberry has withstood a siege of drafts and corrections and but most of the useful applications involve applying these ideas to very
I am grateful to her for her patience. I thank David Bell for drawing the large and complex systems for which a visual display and solution are im-
diagrams. possible for all practical purposes. The reader should not assume that
because simple examples are used here for obvious reasons of space that
therefore graph theory has nothing to offer. For most of these examples
we have a totally unfair advantage : the ability, visually, to comprehend
the totality of graph. However, given a graph in its abstract form :
as a set of elements together with relationships between the elements we
do not have this advantage; and this is the form in which a large empirical
example is usually found. For example, given the graph: Fig 1 we can
Acknowledgements to Publishers and Authors immediately comprehend that it is in one piece so to speak but given the
following list of lines:
We thank - R.J. Wilson for permission to reproduce Table 1 from Table 1 in
An Introduction to Graph Theory, 1972, Longman Group Ltd. 2 - 3
L.J. WOOD and the editor of Geografiska Annaler, Series B, for permission 4 - 5
to reproduce Figures 3 and 4 and Tables 1 and 2 from "Functional 1 - 3
Structure of a rural market system", 1975, Geografiska Annaler, 4 - 3
p. 112 et seq.. R. SYMANSKI and M.J. WEBBER and the Executive 3 - 5
Director of the Association of American Geographers, for permission
to reproduce Figures 1 and 2, p. 205 from "Complex Periodic Market it takes much longer to discover this fact. Given a long list it is often
Cycles" in Annals, Association of American Geographers, 1974. difficult to draw the graph corresponding to Figure 1 and even more trouble-
some to analyse it by eye. Consequently graphs are usually represented as
matrices using the data in the list and the algebra of matrices is used to
define and discover the various properties of graphs. Throughout this

2 3
 (iii) Some definitions : Types of graphs

In Wilson's terminology a graph with n points and q links (pairs of

connected points) is termed a simple graph whenever every distinct pair of
points is joined either exactly once, or not at all, and where no point is
linked directly to itself, (Wilson 1972).

A graph is defined when either or both of the above conditions do not

hold with respect to single connections and self-connections. Graphs turn up
frequently in geographical applications because it is often necessary to
represent two places as being joined not just by one link but by several.
For example, two towns may be connected by road, rail and air links and so a
proper representation would show three links between the places; or two towns
may make several telephone calls to one another. In later sections, in con-
sidering the graph of a transportation structure as a framework for possible
movements as well as a mere spatial form, a place may be defined as being
connected to itself. Consequently a simple graph is just a special, and useful,
kind of graph. The term graph will be used to refer to either type, while
the adjective simple will be used when simple graph is specifically needed.

Mathematicians conceive of graphs as relationships (lines, edges, arcs,

Fig. 1 links) between pairs of points (nodes or vertices). In this form a graph is
treatment it will be understood that a graph, G, and its matrix form A abstract. In many geographical applications, however, the points are places
(the adjacency matrix A of the graph G, defined as A = A(G)) are inter- and the links are transportation or communication links between them. When
changeable. It is simplest to regard both the visual graph and the numer- we 'draw' the graph on paper we are making a 'map' of some reality. A
ical matrix as different representations of an abstract entity, an abstract mathematician calls this an embedding of a graph in another space, in this
graph, comprising a set of points and the set of relationships between the case a Euclidean plane surface, a small region on the earth's surface approx-
points. imately, or a sheet of paper). This particular embedding is very common and
gives rise to many fundamental problems in graph theory exemplified by the
(ii) Prerequisites and Literature next class of graphs.

The only prerequisites for this volume are a willingness to work A plane graph is a graph which is embedded in the plane in such a way
through all the examples, to read a representative selection of the liter- that no lines in the graph intersect. The ends of lines may only join at
ature for each section and to get to grips with the matrix algebra required points of the graph, (Fig. 2).
to operate graph theory. Additional familiarity with matrix algebra will
be useful but is not strictly necessary.

Of the basically mathematical introductions to graph theory Wilson

(1972) is to be recommended but Harary (1969), Ore (1963), Berge (1962),
Flament (1963), Busacker and Saaty (1965) are all alternatives. An extens-
ive treatment of digraphs is given to Harary, Norman and Cartwright (1965).
Within Geography, Kansky (1963) is frequently cited, although only Chapter
II deals with strictly structural notions and there is no treatment of
matrix methods.

A similar introduction but including the powers of an adjacency matrix

is found in Garrison and Marble (1965) which has only recently become
widely available through reprinting in Eliot Hurst (1974). More recently,
Campbell (1971) has provided a brief introduction to digraph analysis as
a basis for examining inter-industry relationships. Other than these, most
articles have taken from graph theory just those notions needed for a part-
icular problem. In this respect Garrison (1960), Nystuen and Dacey (1961)
and Pitts (1965) are early classics in the field. A comprehensive review
of the uses of graph theory in geography can be found in Tinkler (1979).
Fig. 2 Fig. 3
4 5
 A planar graph is a graph whose lines may be represented in the plane (a) a ring (polygon or simple
without intersections, i.e. as a plane graph. Hence Fig. 3(a) is planar but circuit) on n points and
not plane. Any planar graph may be drawn as a plane graph and any plane graph n links
is planar. There is no need to labour the point but it is important to realise
that Fig. 3(a) is not non-planar. Some authors (e.g. Golledge 1968) fail to for n >= 3
realise that some simple graphs drawn with an intersection could
in fact be drawn in the plane without such an intersection, e.g. Fig 3(a) be-
comes Fig. 3(b), and thus despite first appearances are planar. (b) line (simple path) on n
points and (n-1) links,
A non-planar graph is any graph which cannot be drawn as a plane graph
without crossings, which is why Fig. 3 is planar but Fig. 4 is non-planar. for n >= 2

(c) a star on n points, a hub

and n-1 satellites and n-1
links,

for n >= 3

(d) a wheel on n points - a

Fig. 4 union of a ring and a star,

A simple graph with five points and ten links can never be drawn in the for n >= 5,
plane without at least one crossing. If a G is planar and simple it is useful for n = 5, see f(ii).
to know that it has a plane representation where all the edges are represented
by straight line segments. In practice for all matrix operations concerning
graphs the distinction between planar and non-planar graphs is immaterial,
and for a given large graph it is not always an easy problem to determine trees on
whether or not it is planar. (i) 8 points
(ii) 2 points
Many simple structures are found so frequently that they have been re- (iii) 7 points
cognised and named. While they are not always very interesting on their own
they often occur as parts of larger structures or graphs and it is convenient Note a star is a particular
to have a name to reference them. In general it will be understood that the kind of tree, and so is a line.
structure defined can be drawn for as many points as we like. The degree of a
point is the number of links that connect to it, as exemplified in Fig. 5(a),
(b) and (c).

The diagrams show examples of rings, stars, wheels, trees, complete graphs,
nullgraphs and regular graphs. The last three categories f, g and h may re-
quire some explanation. A complete graph (f) is a graph in which every pair Complete graphs (cliques - see
of points is connected by a link, and is of course defined for any number of text) on
points. A complete graph which is a sub-graph (see below) is termed a clique.
A null graph (g) on n points has n isolated nodes and no links. While trivial (i) 3 points
in itself, the idea is needed in describing the structures which may arise (ii) 4 points
when certain operations are performed on graphs, and in ensuring that two (iii) 5 points, etc.

countered in actual problems are like these ideal examples but certain parts
Fig. 5 Types of Graph
of them may be. James, Cliff, Haggett and Ord (1970) have devised an index that
can discriminate amongst groups similar in type to these. Tinkler (1972(b)) 7
has characterised a broad class of graphs with a systematic radial structure.
6
 A sub-graph of a graph, G, is defined as any collection of points and
lines taken from the graph G. Fig. 6(a) is a graph, G, and Fig. 6(b)
is a sub-graph of G based on the points (1,2,5,6,8,9,10) together with the
links which inter-connect those points. This particular sub-graph is a
wheel centred on point 8. Fig. 6(c) is another sub-graph based on points
(9,13,10,4,5,6,12) and their interconnecting links. This sub-graph consists
of two disconnected pieces, and both parts are both lines and trees (see
Fig. 5(b) and (e)). In a real-world example G might be the graph of a

Fig. 6 A Transportation system and its subgraphs

transportation system and sub-graph Fig. 6 (b) might be the bus system
whereas the sub-graph 6 (c) might represent the graph of the street trolleys
The remaining links, and the points at each end of them may represent links
on which there is only private transportation. A sub-graph may be derived
just as well by taking a collection of links from the graph together with
the points to which they are connected.

For example the links on which there is only private transportation

define the following sub-graph of G, Fig. 7.

A super graph is the reverse concept of a sub-graph. Fig. 7 may be

Fig. 5 (continued) called a graph G one of whose super-graphs is Fig. 6 (a). A graph of the

8 9
 In a stream obviously all links are uni-directionally downstream but
in a road system this need not be so; for example Fig. 9 contains one-way
and two-way links and the whole includes circuits so that following certain
links will bring a traveller back to his starting point.

Fig. 7 Private transportation sub-graph of Figure 6

railway system of any country has as a super-graph the graph of all trans-
port links in the country. The general idea of a super-graph is that a
graph may be in fact a sub-graph of some larger graph.
Fig. 9 One-way road system as a graph
So far all links in graph have been considered as two-directional but
in many cases we want to define one-way relationships, for example stream
networks and one-way road systems. In the former case a stream such as It is salutary to reflect briefly on the number of graphs which are
Fig. 8(a) is drawn as a digraph in Fig. 8(b). structurally distinct from each other for each n. Table 1, simplified, is
taken from Wilson, p.162.

Table 1

n 1 2 3 4 5 6 7 8

graphs 1 2 4 11 34 156 1044 12346

trees 1 1 1 2 3 6 11 23

connected* planar graphs 1 1 2 6 20 105 ? ?

6 12
digraphs 1 3 16 218 9608 2x10 9x10 2x10 8

(Wilson, 1972) *For definition see Chapter IV of this monograph.

It is clear that one soon encounters an astronomical number of

structurally distinct graphs. The numbers remain large even when, for a
particular n, the numbers of graphs with a fixed number of lines is con-
sidered. For example, Fig. 24 (a) which has n=6 points and q=7 lines has
24 possible, different, structures and a graph with n=9 and q=14 would
have 15615 different representations: Obviously in these circumstances it
is natural to ask what the 'average' structure of an n=9, q=14, graph would
look like. Section III (iv) addresses itself to this question in terms
of the frequency distribution of the number of lines incident at any one
Fig. 8 A stream network and its digraph point; a limited introduction to a difficult topic.
10 11
(iv) Deriving graphs in actual examples

There is really no end to the possible applications of graphs to geo-

graphical problems - the main requirement is a willingness to try to see
a problem in graph theory terms and to apply a little ingenuity. The read-
er is urged to see applications of graph theory in the literature where
apart from the 'obvious' examples of transportation systems they have also
been applied to trying to simplify the definition of atmospheric pressure
surfaces, (Warntz and Waters, 1975), to industrial structure (Schmidt, 1975,
Campbell, 1974), and to patterns of through valleys in heavily glaciated
terrain, (Haynes, 1975). Rural periodic market systems, Jackson (1972),
Tinkler (1973), Symanski and Webber (1974), and Wood (1975) have yielded
a particularly rich variety of examples. By representing the routes of
travelling vendors between rural markets Symanski and Webber (1974)
define digraphs (Fig. 17), which may then be analysed by the methods of
this booklet to reveal patterns of possible interactions between markets.
Similarly Good (1975) has examined the efficiency of vendor routes in
Buganda using a modification of a travelling salesman algorithm. This
springs from the ideas of traversability mentioned in Section V. Wood
(1975) in a paper summarised briefly in Section VI has examined the art-
iculation of a market system by domestic market visits using connectivity
analysis, (Section IV). In a theoretical analysis Tinkler (1973) conceives
of each market as lying at the centre of its own market area, and being
surrounded by a variable number of other market areas with which it has a
boundary in common. The assignment of days to the markets (and market Fig. 10 Deriving the matrix of a simple graph
areas) so that no two mutually adjacent markets operate on the same day
is the same problem as colouring a planar graph - a topic outlined in are discussed in Section II.
Section V. In a final example Jackson (1972) derives valued digraphs to
represent the expected flow of crops between rural markets with different Any graph may be represented as a matrix and conversely any symmetric
price levels. Indeed, the power of graph theory lies not so much in simply matrix whose entries are non-negative integers can be interpreted as a
drawing a representation on a piece of paper, as from the application of the graph. However, for simple graphs the matrices will always have entries
powerful algebraic methods discussed in later sections. of either 0 or 1, so called (0,1) matrices. A graph, however, will lead
to matrices with (0, integer) entries because two places may be linked by
(v) Matrix representation of a graph more than one link. Fig. 11 (a) and (b) illustrates this.

Drawing a graph on a sheet of paper to represent a situation does The adjacency matrix of a simple digraph is constructed in the same
little more than clarify and simplify the pattern. In addition, as men- fashion but the matrix is not symmetric. For example for a stream system:
tioned in the introduction, a visual analysis is only possible for small Figure 12 (a) to (d) and the matrix is not symmetric. If there are one-way
examples and we need reliable methods for obtaining the same results with and two-way links then parts of the matrix will be symmetric and parts will
very large examples. To achieve this a graph is represented as a matrix. not. Therefore, the matrix is asymmetric, for example Fig. 13 (a), (b).
The points of the graph are labelled arbitrarily and these labels are used As shown in Figs. 10(c), 12(d) and 13(b) the usual representation of the
matrix is with curved brackets, the labelled rows and columns are implicitly
understood, and of course the matrix is square.
links between the two places. A simple graph has 1 or 0 whereas a graph In some cases we may wish to place some numerical weight on a link in
may have any non-negative integers. For simple graphs and graphs if i is a structure. So far these weights have been 0,1 or integer, but the princ-
joined to j then j is joined to i and the same entry appears in two places. iple is easily extended to real-valued weights. A weight with a fractional
Fig. 10 (a) shows a simple graph and 10 (b) shows its matrix, A, usually

problem reverts to that of a graph or digraph. Each link in the graph

or digraph can then be thought of as representing a smaller unit than the
original one. In practice most weights will be integer.
ric, or is not, irrespective of the values along the main diagonal. Diag-
onal entries are used to represent self-connections. Their use and meaning

12 13
 Fig. 12 The asymmetric matrix of the digraph of a stream network

Fig. 11 A graph and its matrix

It makes no difference in matrix calculations how the labels are

assigned to points of the graph but for a large empirical example it helps
if some order is imposed on the labelling process. Often labels are set
up by working systematically through the system, as a diffusion might
spread throughout it. Another method that may be of help in structural
studies is to label points in descending order according to the number
links connected to them. For example Figs. 14 (a), (b). It is easy to
see that for this example the places with many connections are themselves
linked together (States 1,2,3) while the places with few connections, nodes
5 to 10, are not connected amongst themselves.
Fig. 13 An asymmetric matrix
II MATRIX OPERATIONS AS PROCESSES IN GRAPHS

Once a graph has been represented by a matrix then operations on the

matrix can be used to reveal properties of the graph, some of them obvious
and others much less obvious. The important issue is that every matrix
operation performed on the graph must be interpreted properly in terms of
what it 'does' to the graph. It is convenient to learn the operations of
matrix algebra in terms of what they mean for graphs.

(i) Matrix addition and subtraction

Two matrices may be added by simply adding their corresponding elem-

ents. To be conformable for addition two matrices must have an identical
number of rows and columns. Suppose we have two different graphs with the
same number of nodes, and we are able to identify the nodes in each graph
as identical, as for example in Fig. 15, then by taking the adjacency mat-
rix for each graph and adding we obtain an adjacency matrix for the add-

(1969)).

Fig. 14 Canonical ordering of a matrix of a graph

14 15
 As with normal subtraction if two matrices are chosen at random on the right
then the result may have negative elements.

The obvious principle is that we can 'add' graphs (and their matrices)
whenever the set of nodes involved is the same - which makes obvious sense -

Symanski and Webber, (1974, p 205-6) provide an example of graph addi-

tion through matrix addition representing the different patterns of movement
of trading vendors in periodic market systems. The node sets defined are the
Fig. 15 Adding graphs and matrices

(ii) Multiplying or dividing a matrix by a scalar

A scalar is merely an ordinary common or garden number but is so called

no need for the result to be just (0,1) in form and if in some transport to distinguish it from matrices which are collections of numbers. Multiplying
mode a city has no links then it appears as an isolated node; in the adjacency a matrix by a scalar is the process of multiplying each element of the
matrix it has a row and column of zeros. For example Figure 16. matrix by that scalar. If the scalar is an integer n then multiplying an
adjacency matrix A by n is equivalent to adding A to itself n times. If the
scalar is a real number then it is equivalent to scaling every element of
the matrix by the same weight. If the scalar is a fraction smaller that 1
the same argument applies - it is again a weight on each element of the
matrix, obviously if the weight is zero then the whole matrix reduces to
zero entries - a system of n nodes with no links. At this stage these applic-
ations would appear to be trivial or nonsensical but they do in fact arise
in more sophisticated studies arising from matrix operations dealt with in
later sections (see Section III). Some articles relevant to this are
Garrison (1960), Nystuen and Dacey (1961), Harvey (1972), Stutz (1973) and
Tinkler (1974).

(iii)The Null matrix, 0, and the Identity matrix, I

These matrices fulfill a similar role in matrix algebra to the numbers

0 and 1 in ordinary algebra. They are square matrices defined on whatever n
is appropriate to the problem in hand. As graphs they are trivial, but they
are essential to matrix operations and interpretations. The null matrix is an
n by n matrix full of zeros. It is equivalent to a graph of n nodes with no
links. The identity matrix is an n by n matrix with l's down the principal
diagonal and zeros elsewhere. It is equivalent to a graph of isolated nodes
but with a self-connection defined for each node. There are no links between
the nodes. When written with other matrices the null and the identity matrices
Fig. 16 Adding graphs and matrices are assumed to take on appropriate dimensions to ensure conformability.

16
17
(iv) Matrix multiplication

Matrix multiplication is the basis for many methods for manipulating the two inner indices (m,m) are identical in size. In the example they are
graphs. Most of these methods involve multiplying a matrix by itself but a and the reverse in this example is also true, (n,n), viz:
few involve multiplying a matrix by a vector or by another matrix. In order
to multiply two matrices the matrices must be conformable but the rules for
conformability are different to those for addition. To make this clear let

then the outer indices indicate the size of the product. Matrices for
graphs are always square and hence we only need to check that the matrices

i.e. another row vector following

the rules given above whereas:

is not conformable for multiplication. On the

other hand: another column vector and again:

is not conformable for multiplication.

For convenience vectors are usually written in lower case boldface,

(underlined in hand-written work), to make them easy to distinguish in
reading equations, from matrices, usually written in capitals boldface,
(underlined when hand written).

We now come to matrix multiplication in detail; it is described first

Specifically, and in this case letting A, the adjacency matrix of a

graph, be both the first and the second matrix we write:

Fig. 17 Routes of travelling vendors in rural periodic markets

19
(after Symanski and Webber, 1974, with permission)

18
To calculate the 1,1 element of A2 take the first row of the first matrix (0 x 0) + (1 x 1) + (1 x 1) + (0 x 0) = 2
A and the first column of the second matrix, again A; ignore their nature
as rows or columns now and place side by side viz: (0 x 1) + (1 x 0) + (1 x 2) + (0 x 1) = 2
1st row 1st column (0 x 1) + (1 x 2) + (1 x 0) + (0 x 1) = 2
0 x 0 = 0 (0 x 0) + (1 x 1) + (1 x 1) + (0 x 0) = 2
1 x 1 =1
1 x 1 = 1
and therefore:
0 x 0 = 0

Add 2

multiply the corresponding elements and add. We could also represent this
process as:

(0 x 0) + (1 x 1) + (1 x 1) + (0 x 0) = 2

(0 x 1) + (1 x 0) + (1 x 2) + (0 x 1) =2 row 2 col. 3 through routes

row 2 column 3
interpretation interpretation possible
(0 x 1) + (1 x 2) + (1 x 0) + (0 x 1) =2

1 1 from 2 to 1 from 1 to 3 yes 1

(0 x 0) + (1 x 1) + (1 x 1) + (0 x 0) = 2. from 2 to 3 no 0
0 2 from 2 to 2
2 0 from 2 to 3 from 3 to 3 no 0
1 1 from 2 to 4 from 4 to 3 yes 1

total 2

and interpreting each element then since the row is 2 then the column
positions in this row, as it lies in the matrix, indicate the number of
The remaining elements are:
routes available from node 2 to all nodes in the system. For column 3
the row positions indicate routes from all nodes in the system to node 3.
(1 x 0) + (0 x 1) + (2 x 1) + (1 x0) =2
Placed side by side we see that they can be used to show the availability
of routes between nodes 2 and 3 (the values of i and j being considered)
(1 x 1) + (0 x 0) + (2 x 2) + (1 x 1) = 6
via all other nodes in the system. The row vector element 1 shows that
one route exists from node 2 to node 1 and the column vector element 1
(1 x 1) + (0 x 2) + (2 x 0) + (1 x 1) = 2
shows that one route is possible from node 1 to 3, hence there is a total
of one route, (1 x 1), available. 1i the case of node 2 to node 2 there
(1 x 0) + (0 x 1) + (2 x 1) + (1 x 0) = 2
is no route available, 0, hence no onward route to node 3 is possible.
For routes from node 2 to 3 via node 3 it is again seen that none are
(1 x 0) + (2 x 1) + (0 x 1) + (1 x 0) = 2 possible because there are none from 3 to 3. The remaining possibility
is from 2 to 3 via node 4 and in this case a through route is possible.
(1 x 1) +(2 x 0) + (0 x 2) + (1 x 1) = 2
Adding all these terms gives the total number of possible routes which is
2. When the two terms being multiplied are larger than 0 or 1 the inter-
(1 x 1) + (2 x 2) + (0 x 0) + (1 x 1) = 6 pretation is that if there are say 3 routes from i to k and then 2 routes
from k to j then the 3 routes on the first part combine with the 2 in the
(1 x 0) + (2 x 1) + (0 x 1) + (1 x 0) = 2
second part giving a total of 2 x 3 = 6 possible routes.

20 21
 The discussion has been long and tedious, but understanding it is crit- The procedure for pre-multiplying a matrix A by a row vector v,
—
v.A = c,
ical to an appreciation of graph theoretical methods. is simply this procedure. The procedure for post-multiplying A by a column
vector v' is analogous: combine each row in A in turn with the column
vector v'to produce each element in the resultant column vector r

2
2nd row of A 3rd. col. of A row col. possible routes
2 1 Total 2 step routes 2- 1 1- 3 yes 2
6 2 Total 2 step routes 2- 2 2- 3 yes 12
2 0 Total 2 step routes 2- 3 3- 3 no 0
2 1 Total 2 step routes 2- 4 4- 3 yes 2
-
Tr

The reader should check that A 3 is:

is the vector, c, of in-degrees. The vectors r and c may be different

when the matrix is the adjacency matrix of a digraph, and will be ident-
ical for the adjacency matrix of a graph.
While this information can be obtained visually in simple graphs,
the number of nodes and the number of steps required do not have to be (vi) O's or l's in the main diagonal of A
very large for the process to become impossibly tedious and very prone to ,

error. In most practical examples use is made of computer routines to do

the matrix multiplication. In structural studies it makes little sense to connect a node to it-
self with a loop but in the previous section it is clear that the existence
(v) Multiplying a matrix by a vector of a route implies movement along it, that is to say movement in the graph.
In these circumstances connecting nodes to themselves, placing l's in the
main diagonal, allows a 'movement' to remain at a node as a move equivalent
to moving one step - a sort of graph-theoretical stopover: In this case
the 'routes' are counted with the understanding that a route may remain
at a node for one or several moves. One might wish to examine the results
of allowing certain nodes to be retentive in this fashion, while others
are not. A notation to distinguish between 0's and l's in the main diagonal

23
22
 A problem that can occur with 0's in the main diagonal can be exempli-
fied by the following example: (from Fig. 12(a)) -

Inspection of these examples shows that while node 3 can be reached in two
steps from node 1 it cannot be reached in 3 steps, although it is again

and in this case node 3 is reachable from node 1 in 6 different ways. This
(viii) Matrix multiplication of two different matrices or graphs
is achieved by combining holding moves at nodes 1 and 3 with the two ways to
get to 3 from 1, plus holding moves on either route to 3 at nodes 2 and 4 In powering a single matrix, A, we examined possible routings within
respectively; one such sequence is 1-)2-42-43 where the 2-42 is the 'holding'
move. the A system. In applying the same basic technique of multiplication to
two or more matrices, A, B, C etc., we can examine routings available
(vii) Matrix multiplication involving digraphs - asymmetric matrices in a multi-mode system whe e routes are followed first in one system, A,
r

and then in another, B, and so on.

All the results obtained above hold just as well for digraphs as for
To be conformable two matrices of graphs must be defined over the
graphs except that the matrices involved, and their products, are no longer
same node set (p.17). This is no problem for powers of an adjacency matrix
symmetric matrices. To illustrate briefly consider the digraph, D, Fig. 18,
with adjacency matrix A = A(D), since the nodes remain the same. Conformability is ensured by augmenting
the node sets where necessary as in matrix addition, see p.14. Let the
first matrix, A, be the adjacency matrix of one transport mode over a set
of nodes, perhaps an inner city one-way rapid transit bus system. Let
the second, B, be the adjacency matrix of a second transport mode; perhaps
a suburban railway system. Note that to ensure conformability A is aug-
mented with the isolated nodes 4, 5, 6, 7, whereas B involves a11 the
nodes. If now we write :

AB=C

then we can still resort to the previous interpretations.

Fig. 18 A digraph

24 25
 but the interpretation is that the first move is made in graph A and the
second in graph B. The only two-step route from A to B starting in node
1 is to node 6 as the reader may verify. Note that the 2-step route from
1 to 3 is not recorded since this requires 2 steps in the A network. In

give the results of one move in A followed by two moves in the graph con-
sisting of A and B together followed by a move in B alone. A full dis-
cussion may be found in Tinkler (1975).

(ix) Powers of A as a non-conserved diffusion system

The availability of routes, and their numerical computation in

applications of these ideas would seem to imply their utilisation. A
technical argument (Tinker, 1973, 1974) shows that in fact any diffusion
interpretation placed on powers of A is isomorphic to a non-conserved diff-
usion where total inflow to a node does not necessarily balance with total
outflow. This behaviour characterizes the diffusion of information as
opposed to physical entities and the powers of A are analagous to a gen-
Fig. 19 A two mode transportation system eralised Hagerstrand model in which the telling rate varies from node to
node instead of being constant. A classical conserved diffusion requires
that A be converted to a transition matrix before the powering commences;
In full we have : (see Section III (ii) p. 32).

ments of the matrix.) The problem is to determine whether a matrix has a

solution matrix and if so to determine the smallest m. Alao (1970) gives

ed in Section VI but intuitively corresponding to the idea that the graph

is in 'one piece'), and (2) the graph contains an odd cycle. For details

26 27
and definitions of cycles see Section VI, but the principle is that if it thirdly we can usually substitute a smaller, more exact number, d, called
is possible, for some node, to return to the same node in an odd number of
steps then there is an odd cycle. Most empirical networks contain odd

The earliest study of accessibility in the geographical literature

is that of Garrison (1960) who analysed a portion of the Interstate High-
way using the methods outlined in this section. For each node he computed
an accessibility index by summing the numbers in each row of the final

matrix to obtain the accessibility index for each node. The results were
then expressed as proportions. The higher the row sum is, the more out-
going routes are available from that node, and the higher the 'centrality'
or 'accessibility' is judged to be for that node. Another historical
study, modelled on Pitts, is that of Carter (1969) for the Serbian Oecumene.

(xii) Weighted sums of powers of adjacency matrices

To counteract the influence of the excessive weight of the higher powers

in a direct sum such as:

(xi) Sums of powers of adjacency matrices

One way to take into account all possible routings of all possible a modified procedure is to use a weight on each matrix. The most commonly
lengths in a graph is to add all the matrices obtained e.g. A, A2, A3 etc. used is of the form:
so that:

where r is suitably chosen. The suitable choice of r requires some care.

The problem with this approach is two-fold. On the one hand it is Obviously to provide attenuation r must be initially smaller than 1, but
not obvious at what power level, k, the summation should stop, and on
the other because the higher powers contain much larger numbers than the
smaller powers they consequently dominate the total sum. (In fact the
entries in successive powers increase in an approximately geometric
fashion - and for matrices of graphs the common ratio is larger than 1 and
smaller than the largest marginal sum of A, the node with the largest num-
ber of links into or out of it. For a full discussion see Tinkler (1972 by the largest marginal sum of A. For example if A is:
(a) and Tinkler (1974) ). The excessive weight of the higher power sums
in S is contrary to our geographical expectations, where we expect a de-
creasing influence with increasing distance, not an increasing one. Be-
cause of this problem direct sums of this sort are not usually made and
accessibility, if it is studied in these terms, is analysed with respect
to a single matrix Ak. The problem still remains, which m or k to choose?
There are three rules of thumb. The first is to stop at the solution num-
ber (see the last section). The second one is to choose k n - 1 where
n is the number of nodes in the graph. This ensures that, if the graph then 4 is the largest marginal sum. Successive powers of A will not
is strongly connected (see Section IV), the two most separated points grow by a factor larger than 4 and so can be counteracted BY a factor of
are connected by at least one route. In fact this is a fail-safe and so 1/4. Genuine attenuation can then be achieved by choosing the attenuation

28 29
maximum sum of the matrix A then the proper range for r is: (ii) The degree vector of a graph and the fixed vector of the transition
matrix of the same graph.

(A full discussion can be found in Tinkler (1974) which discusses the lit-
erature on this issue.)

The structure of the graph can then be studied using random processes
and the full and powerful methods involved in Markov Chain Analysis for
which the reader is referred to CATMOG No. 1 (L. Collins). Fortunately
it is possible to state a theorem which relates, under certain circum-
stances, the degree vector of a graph, the row sums of A, to the unique
fixed point probability vector of T = T(A).

Theorem (Tinkler (1973))

Of course if in some study step lengths larger than some index p are imp-
ossible or inappropriate then the series sum may be arbitrarily terminated
at index p :

This theorem is used to justify many procedures in which the degree

of a node is used as a measure of its importance in the graph and is a

III 1ST ORDER PROPERTIES OF GRAPHS divided by the sum of the degrees for the whole graph. Any graph, by
virtue of its symmetry, satisfies the theorem and its degree vector, norm-
(i) Nodality alised to sum to 1, may be taken as the fixed vector of its transition
matrix. A digraph may, or may not, have an adjacency matrix with equal
The nodality (degree, local degree, valency are alternative terms) in- and out- degree vectors.
of a point is simply the number of links in the graph incident at that
node. Given the adjacency matrix A of a graph the ith node has a nodality
given by the sum of the ith row of A. If 1 is a column vector of ones
then (A 1) r is the degree vector ofthe graph, it contains the nodality
of every point or node. When A is the adjacency matrix of a digraph then
A 1 r defines the out-degrees of the nodes, outgoing links, and l' A = c
where T' is a row vector of l's defines the in-degrees of the nodes, in-
-

coming links. In itself nodality looks too simple to be interesting des-

pite a hallowed geographical tradition embodied in the map-reading ritual
at school where the canonical invocation is to a 'town important because section. The numerator in this expression is exactly twice the number of
it is a focus of routeways'. However, for a recent appeal to simple nodal- links, q, in the graph since for nodality the two ends of each link are
ity, see Burghart, (1969, p. 436). counted. Hence the mean nodality is 2q, which is just twice the B index
n
In fact because nodality can be linked precisely to random processes described by Kansky (1963). As more links are added to the system for
in a graph its use can be justified in operations where it might otherwise
seem trivial.

work structures, a criticism made, with examples, by James, Cliff, Haggett

and Ord (1970).

30
31
 Werner and Smart (1973).

IV HIGHER ORDER PROPERTIES OF GRAPHS

The section on matrix operations, particularly matrix multiplication,

has introduced the idea of higher than 1st order properties in graphs, i.e.
those properties which only become apparent when information on n-step
adjacencies is available. In this section extensive use is made of n-step
adjacencies and the reader is assumed to be familiar with them and their
calculation, but, first some definitions.

(i) Routes, paths, chains and cycles

A path in a graph is a sequence of edges or links, defined by success-

ive pairs of nodes, which are all distinct and any successive pair in the
sequence have exactly one node in common. For example in Fig. 21 a path

Fig. 21 A graph to illustrate routes, etc.

The routes discussed earlier are less restrictive. They are edge or link
sequences in which links may be repeated so that a route from 1 to 3 might

Fig. 22 A graph with self-loops to illustrate routes, etc.

34 35
 Symanski & Webber, (1974), Haynes, (1975). Modifications of the index are
2, 3, 4 and 6. A path is closed if the first and the last nodes are the given by Kansky (1963) in which it is related to the maximum number of
same. A chain is a path in hich no link or node repeats itself and a
w
cycles in planar and non-planar simple graphs respectively.
closed chain is termed a circuit or cycle. (These are definitions from
Wilson, but different authors have widely different terminology.)

where q and n are number of links and nodes in the graph and p = 1 if the
graph is connected, or equals the number of components in the graph if it
In this form the index is a little more discriminating although it
is not. (Connectivity and components are discussed below.)
has not attracted a wide practical use. Examples of its use can be found
in the pavers just mentioned.
If we consider a tree, Fig. 23, with n nodes and n - 1 links it is
clear that the addition of any new link between two nodes not currently
Marble (1965) found a correlation coefficient of +0.998 between them for
a sample of 22 empirical graphs. This would indicate one can serve for the
other. For completeness we define them :

where N = (1/2) (n2 - n), and q and n are as defined above. The gamma
index is seen to be a simple ratio of the observed number of links to the
total possible in either a planar or a non-planar graph.

(iii) Connectivity and distance

The connectivity of a graph is easy to determine at a glance for a
small graph drawn on paper. It may be defined in the following way. A
linked directly, will produce a cycle, for example between nodes 3 and 7. graph is strongly connected or connected if there exists a path between
For the tree, every pair of nodes in the graph. For a graph the symmetry ensures that
if node i is connected to node j then j is connected to i. For digraphs
this need not be the case. When it is true the digraph is strongly
connected. The question of whether any particular pair of nodes is
connected or not can be settled by calculating the powers of the adjacency
matrix of the graph. In a simple example connectivity can be determined
The addition of a new link either splits the existing cycle into two at a glance. For a complex empirical graph, perhaps a graph of eighty
if it has more than three nodes in it, e.g. linking 2 and 7, or creates a rural markets in which connections are determined by whether two markets
new cycle outside it. In either case a new cycle is created. Consequent - are visited by a householder in the sample period, such an approach is
impossible, (L.J. Wood, 1975), see below VI, Fig. 29.
graph theory as a mathematical study it is not especially useful in geo-
graphic work since it is unrelated to the size of the graph. It has been
used or mentioned by Garrison & Marble, (1965), Warntz & Waters, (1975),

36
37
Fig. 24 Types of connectivity in graphs and digraphs

reference to a line graph with n nodes we see that the two end nodes can-
not be more than n - 1 lines apart. In practice we will not usually
have to power as far as that. At the same time as the powering process
we can incidentally determine the distance between every pair of nodes.
The distance (geodesic, shortest path) between any pair of nodes is the
length of the shortest path between them. Suppose two nodes i and j

apart. By keeping track of when each ij position first becomes positive

in some power of A a distance matrix D can be compiled. The computations
will be illustrated for digraph Fig. 24 (c).
An equivalent way to obtain the distance matrix D which is especially
useful when A has cost or distance values on the links is to use 'modified
arithmetic' in the matrix power computations. The modified operations are:
for a x b substitute a+ b
for a + b substitute min (a,b),
and the latter operation may apply to a term with several elements:
38 39
 a + b + c + + h = min (a, b, c, .......... , h).

In these operations of course a, b, c, .... etc. are the elements of the

matrix in hand. In addition all zero elements in the initial matrix are

A number of indices can be derived from the distance matrix. Taking

each ith row the maximum number in each row, indicating the distance to
the most distant node from i, is the Konig number or Associated Number.
Taking the set of all Konig numbers for a graph the largest is the diameter
of the graph, agreeing with the previous definition and the smallest Konig
number is the radius of the graph. The row sums of D, specifically D.1
have been used as measures of access since the smallest sum indicates the
place closest to all other places in simple topological distance terms.
Dividing each row sum by (n-1) yields the average distance from that node
to all others. Remembering that (D.1) is a column vector, then summing
all these elements, or in matrix notation l' (D.1) where l' is a row vector
of ones, gives a measure that has been termed the dispersion of the graph.
Computing these various indices for the distance matrix of Fig. 24(a) we
get the following values:

nodal access mean

node Konig row sums distance
Fig. 25 Graphs to illustrate cutpoints and bridges

1 4 11 2.2
2 3 8 1.6 (iv) Components of a graph, articulation points and bridges
3 3 8 1.8
4 2 7 1.4 When the distance matrix indicates that certain nodes are not connect-
5 3 9 1.8 ed then graph has two or more strong components. A strong component is a
6 4 13 2.6 sub-graph which is strongly connected. Fig. 24 (b) (c) and (e) all have
two strong components. Fig. 24 (a) and (d) have one strong component and
the graphs are said to be strongly connected. Werner (1971) compared the
56 number of components in a developing transport system to the numbers ex-
pected in a comparable series of random graphs. Connectivity analysis
diameter = 4 = maximum Konig number dispersion = 56 may be combined with two additional notions in the analysis of graph struct-
radius = 2 = minimum Konig number ure. An articulation node or cutpoint is a point whose removal together
with the incident links will result in a graph becoming disconnected. A
The following authors illustrate the use of the short path matrix cutpoint set is a collection of points that give the same result. Fig.
and a variety of the derived indices; Garrison (1960), Pitts (1965) and 25 illustrates both concepts. A bridge is a single link whose removal dis-
Carter (1969). Reed (1970) investigates changes in the average node dis- connects the graph, (the nodes at either end are not removed), and a bridge
tances when certain nodes and links are deleted from the network. Kissling set is a collection of links with the same property. There are no bridges
(1969) uses the measure of nodal access applied to a least cost-distance in he examples above but in (a) bridge sets are {l-2, 2-3} and {3-4, 3-5},
t

matrix. James, Cliff, Haggett and Ord (1970) have used the frequency and in (b) {4-5, 4-6, 3-6} is a bridge set. Removing a single point is equival-
distribution of distances in the short path matrix to construct a very ent to removing the bridge set consisting of all links incident to that
sensitive index (S-I) which is able to discriminate between broad struct- point.
ural classes of graphs similar to those identified in Section II above.
The same approach may be applied to the frequency distribution of the In the empirical case one might want to either identify cutpoint or
short path distances for each node to discriminate the structural positions bridge sets or test whether specific elements, or sets of elements have
of nodes within a network. these properties. In either case the selected set is deleted and then the
distance matrix is computed to identify the components of the graph. In
this way critical elements in the system can be identified, or the critical-
ity of elements can be tested. Reed (1970) used this approach by removing
critical nodes in the Indian National Airline system and examining the
effect on the structure. A systematic method for investigating an empirical

41
40
would be to remove nodes in descending order of nodality, or sets of nodes TELEPHONE TRUNK CENSUS JULY 1967 (5 Days Total 30 hrs.)
based on the same principle. Wood (1975) provides an example of just such KENYA
a procedure for investigating the structure of a market visiting network
for both nodes and links, (see Section VI, Fig. 29).

Bridge sets arise in an apparently different connection in dealing

with flows in a network. If values attached to the directed links of a
digraph are used to describe the maximum capacity of links with respect to
flows going from a node i to a node j then the Ford-Fulkerson max-flow
min-cut theorem (1956) states that the maximal possible flow from i to j
through the network is equal to the total capacity in that bridge set, out
of all possible bridge sets separating i and j, whose total capacity is
the smallest. It is a characteristic of graph theory that an apparently
simple idea leads on to very powerful results. A lucid treatment of this
topic may be found in Wilson (1972). The Ford-Fulkerson algorithm has been
applied by Gauthier (1968), Kissling (1969) and Muraco (1972) following
on from simpler accessibility analyses.

V Extensions

Given the basic elements of graph theory reviewed so far it is possible

to outline a number of directions which yield applications in broadly
geographical fields. The first extension involves nodal analysis in which
a particular type of accessibility matrix from Section II is used as the
basis for an investigation of dominant telephone flows amongst the towns
of an urban hierarchy. The relationship that this methodology bears to
Mean First Passage Times in Markov Chain Analysis (see CATMOG No. 1) is
discussed in Tinkler (1972(a), 1976). A second direction concerns the
existence of certain routes and optimal routes within networks - the issue
of traversability, and the third concerns coverings and colourings of
graphs in graphs - notions fundamental to optimising location and alloc-
ation problems for graphs. For the latter two directions Scott (1971) pro-
vides a detailed review.

(i) Nodal Analysis The highest entry in each row, the potential nodal flow, is underlined.
Note that Nairobi to Mombosa is not a nodal flow, nor is Eldoret to
Some mention should be made of Nodal Analysis since it has become Kitale.
widely referenced as a graph theoretical technique, (Nystuen and Dacey,
1961; Clayton, 1974; Soja, 1968, p. 46, 47; Taafe and Gauthier, 1973, We get a nodal flow matrix :
p. 149). The method may be applied to any interaction matrix to identify
nodal flows. A flow is nodal if a given centre's largest flow is to a
centre higher in the city hierarchy than itself. Usually a city's hier-
archical position is measured by the total flows incoming to the city - the
appropriate column sum in the interaction matrix. A city is a terminal
node in the nodal flow system if its largest flow is to a centre lower in
the hierarchy than itself. The system is constructed from the interaction
matrix by identifying the largest value in each row. If the column sum for
that value is larger than the column sum indexed by the row number then a
1 is entered in the nodal graph matrix, and a zero otherwise. Using a
matrix of telephone flows in Kenya, 1967, between exchange regions, the
analysis is performed below. (Nystuen and Dacey (1961) also give a worked
example). which is the adjacency matrix of the digraph, Fig. 26.

43
42
 Fig. 27 Graph to illustrate types of cover

required to traverse the graph with a closed path which visits each node
exactly once. There is no general solution to this problem, the simplest
result being that if the graph has n nodes and every node has degree equal
Fig. 26 or larger than (1/2)n then a Hamiltonian path exists. A related problem
that is commonly encountered in scheduling and allocation problems is to
select, out of a large number of Hamiltonian paths, where each link in the
The method was originally applied by Nystuen and Dacey to the sum of a system has a fixed cost value attached, that which has the lowest total
power series expansion of a weighted adjacency matrix, as discussed on cost associated with it. This is the travelling salesman problem and it
p. 31. It may, however, be applied to a raw interaction matrix, as in the is usually posed for complete, valued-graphs in which there are (n-1): poss-
example just given. Clearly, in the example, the polarising effect of ible Hamiltonian paths. An extension of the problem with an application to
Nairobi is evident, but Eldoret at the northern end of the former White periodic market rings is given by Harvey, Hocking and Brown, (1974). Good
Highlands is able to maintain a local nodal position. (1975) has attempted to delineate optimal paths for travelling traders in
Buganda using ideas based on travelling salesman algorithms.
(ii) Traversability
(iii) Covering a graph
An Eulerian path is a closed path (i.e. starting and ending at the
same node) which includes every link in the graph just once. Euler's A point cover of a graph G is a set of nodes such that all other nodes
eponymous solution to this problem is credited with creating graph theory. in the graph are adjacent to at least one of the nodes in the set. Fig.
The necessary and sufficient condition is that the degree or nodality of 27 (a), (b), (c) illustrates three different independent point covers. The
every node in the graph be even. covers are termed independent since no pair of nodes in the cover are them-
selves adjacent in the graph. Figure 27 (d) shows a point cover which is
An apparently similar problem, but one that offers many unsolved pro- not independent. Often but not always, dropping the independence require-
blems in graph theory, is the Hamiltonian path problem in which it is ment reduces the number of points in the cover; in this example it is not

44 45
 Step 3 - Choose the point with the largest row sum, (in case of ties make
an arbitrary choice). Enter into the cover.
Step 4 - Delete from A the point chosen, i.e. set that row and column to
crosses, and alsodelete from A the rows and columns of all points
to which it was connected, similarly setting these to crosses.
Step 5 - If A is now a null matrix enter any currently isolated points
(zero row sums), and the point cover is complete. Otherwise go
to Step 2 and use the current A matrix, i.e. with the various
rows and columns deleted.

algorithm is not foolproof and pathological graphs can be devised that

have point covers that the algorithm does not find, much smaller than those
it does find.

(v) Colouring a graph

Fig. 28
A proper colouring of a graph is an assignment of distinct labels -
colours - to the points of a graph with the property that no two points
so but the next one shows the type of structure in which this can occur, adjacent in the graph have the same label (colour). The minimum number
Fig. 28 (a), (b). of colours that can be used to achieve this requirement is termed the

In practice it is often required to find a minimum point cover - in

allocation problems the need is to use as few facilities as possible and
to locate them as close as possible to all remaining nodes. We need cover. These points are labelled colour 1. Return now to the original
therefore a minimum point cover. A set X is minimal with property P if and A for the graph and delete simply those rows and columns corresponding to
only if X has property P and no proper subset of X has property P. Hence points in the cover obtained. Using this reduced matrix, A', re-use
Fig. 27 (a), (b) and (c) are all minimal whereas Figure 27 (d) is not the algorithm to obtain a point cover amongst the remaining nodes. Label
minimal since node z can be deleted without affecting the property P, these nodes colour 2. Reduce the A' matrix to A" by removing the rows
i.e. the covering of the remaining nodes. A minimum set with property P and columns for points coloured with 2 and continue in like fashion until
is a set with the smallest possible number of elements in it; i.e. the no nodes remain to be coloured. The number of colours used will be a good
minimum or minima of all the minimal sets. Hence Figure 27 (b) and (c)
are both different minimum sets whereas Figure 27 (a) is simply minimal.
of colouring in the geographical literature is that by Tinkler (1973),
where rural periodic markets and market areas surrounding them are con-
ceived of as a planar graph. Two rural markets are adjacent when they
share a common boundary. The four-day market week typical of West Africa
matrix of the graph we are interested in and construct point covers in which is interpreted as a natural response to the four-colour conjecture for
points of the graph are guaranteed to be within n-steps of a node in the planar graphs. The conjecture states that any planar graph can be colour-
point cover. (The probable multiplicity of many of the n-step adjacencies ed with only four colours. It was long ago proved that any planar graph
is, of course, irrelevant.) is five-colourable, and has been proved for planar graphs with up to 51
points (Stromquist, 1975). A computer proof for an arbitrary number of
(iv) An algorithm for minimal independent point covers points has recently been given by Appel and Haken (1976). For graphs with

Most algorithms to obtain minimum independent point covers are ex-

tremely laborious. It is easy to find a minimal independent point cover length are two-colourable. A graph with an odd cycle necessitates three
but hard to find a minimum one, i.e. that which is smallest of all the colours and one containing as a subgraph a wheel on n = even points necess-
minimal covers. The procedure is to use a reasonably 'good' method to itates four colours. (See Section VIII, p.54, for an application of the
find a minimal cover and to hope that it is close to being a minimum cover.

Step 1 - Compute A for the graph, this is the current A matrix.

Step 2 - Sum the rows of the current A matrix, enter points with row sums
of zero into the cover, provided they have not already been ent-
ered. Deleted rows do not count as zero row sums.

46 47
 VI AN APPLICATION : KISII DISTRICT, KENYA.
Wood (1975) has used graph theoretical concepts to examine the funct-
ional structure and integration of rural markets in Kisii District, Kenya.
In total 83 rural markets were identified, but two isolated ones, and one
with only one connection, were excluded from the analysis. For a period
of 14 days, market visits by the mothers in 103 households were recorded.
Two markets were deemed to be connected if they were visited in the sample
period by a member of the household. In this way the pattern of connect-
ions between markets was established and gives rise to Fig. 29, which
simply records connections. In fact many links represent multiple connect-
ions. In either case the system is obviously highly polarised. An analy-
sis may be based on the binomial method in Section V, p. 34, where n=80

giving a standard deviation of 3.55. Table 3 shows the observed freq-

uency distribution of nodality and obviously it contains several nodes
lying well outside a three standard deviations limit. Kisii, for example,
has a nodality in terms of visits of 78 and is directly connected to 38
of the 79 other places. The behaviour patterns displayed are adjudged
very significantly polarised.

Table III

Frequency Distribution of Nodality in Kisii District

Market Visiting Network.

Nodality Class

0-4 5-9 10-14 15-19 20-24 25-29 30-34 35-39 40+

Frequency 15 23 13 15 4 2 3 2 3

Maximum = 78
Minimum = 3
Mean = 12.93

A short path analysis forms the basis for a connectivity analysis.

Table 4 tabulates the Konig number frequency distribution and for the net-
work of both domestic marketers and traders* the diameter of the system

*Wood was able to determine from the questionnaires that some mothers were
trading items part-time for profit - traders - as distinct from purely dom-
estic buying and selling - marketers. Full time itinerant traders were not Fig. 29 Inter-market connections in Kisii District, Kenya (from Wood, 1975,
included in the analysis, see Wood (1975), footnote 11. with permission)

48 49
 is only 6, (despite the 80 markets), which indicates a high degree of
integration. As expected, Kisii is closest to all points; only three
links at most from any place. When the traders are removed from the net-
work a slight shift occurs : the diameter moves to 7 and Kisii has a
Konig number of 4. This indicates that the traders do knit the network
together more closely, but apart from 3 markets which become isolated when
the trader links are removed, their removal from the system does not split
it into major regional components.

Table 4 - The distribution of Konig numbers (from Wood, 1975, with permiss-
ion).

Konig marketer and

number trader network marketer network alone

3 1 0
4 41 13
5 32 33
6 6 25
7 0 6

Total 80 77

An analysis of the subsystems comprising the network was made using

the procedures described in Section VI, p. 37. In the absence of useful
official data to establish a hierarchy of markets in terms of market use,
recourse was had to the hierarchy established by the total numbers of
visits received as recorded by the survey. Kisii received 78 visits,
8 markets received between 28 and 56 visits and 71 markets had less than
25 visits. Wood tested the structural integrity of the system as success-
ively more nodes were removed. First Kisii was removed, then Kisii and
the next largest, and so on for the first nine nodes. The significant
results are recorded in Table 5 and in Figure 30.

Table _5 - The decomposition of the Kisii market system

Fig. 30 Disintegration of the market visiting structure with the removal

of various market nodes (from Wood, 1975, with permission)

a The disintegration of the networks is cumulative, i.e. the isolation of market

51 in extraction 3 for the marketer network means that this market is isolated 51
in all subsequent extractions. (From Wood 1975, with permission)

50
 The analysis demonstrated the highly connected nature of the marketer optimal solutions to most of the meaningful problems. A tradeoff must be made
and trader network and its independence from an integration caused just by between the speed of the algorithm and the quality of the solution and a
part-time traders or one or two critically placed, frequently visited, markets. good recent review of algorithms arising from graph theoretical problems is
Interestingly, while emphasising the tightly woven structure caused by indivi- Christofides (1975). Nevertheless, it is evident that graph theory and re-
dual trading, the one major regional decomposition that was revealed, see lated combinatorial theory in transversals, matchings and matriods has an
Fig. 30, can be interpreted as a split between the subsistence agriculture enormous potential for finite state problems, in geography as in other sub-
practiced in the south and the more intensive cash cropping in the north. Four jects. As the application discussed in Section VI illustrates a good deal of
major markets lying along zonal boundaries appear to have an integrating structural clarity is achieved by fairly simple methods.
effect. A similar analysis for the marketer network alone, (i.e. with trader
links removed), had similar results, Table 5, column 4, although in this case BIBLIOGRAPHY
the northern (cash crop) component splits in two when the first five nodes
were extracted. The integrating role of the trader is therefore seen to be A. Graph Theory Texts
relatively limited. An application of the graph colouring algorithm, Section
V, yields a chromatic number of 8. (This analysis is not reported in Wood Berge, C., 1962, The theory of graphs and its applications. John
(1975), but was made with the market matrix kindly made available by Wood). Wiley.
Therefore, despite its complexity, 8 colours (days) would be sufficient to
ensure that no householder, on visiting one market in his market set would Busacker, R.G. and Saaty, T., 1965, Finite graphs and networks: An in-
miss a market on the same day in the remaining markets in his set. Since the troduction with applications. McGraw-Hill.
8th. colour is used to resolve only one conflict, and the 7th. to resolve only Christofides, N., 1975, Graph theory: an algorithmic approach.
four conflicts it may be concluded that the system can work with almost total Academic Press.
efficiency on a six day week. (Kisii is a stronghold of Seventh Day Adventists
and Saturday is hardly ever used as a market day, (Wood, 1975)). Flament, C., 1963, Applications of graph theory to group structure.
Prentice-Hall.
Wood's analysis reveals the utility of a straightforward graph theoreti- Harary, F., 1969, Graph theory. Addison-Wesley.
cal approach and yields useful answers to problems of spatial patterning and
interactions which would not be achieved easily by other methods. Harary, F., Norman, R.Z., and Cartwright, D., 1965, Structural models:
an introduction to the theory of directed graphs.
VII PROBLEMS AND LIMITATIONS John Wiley.
Ore, Oystein, 1963, Graphs and their uses. Random House.
The possible limitations of graph theory were pointed out by some of the
earliest users - Garrison (1960), Nystuen and Dacey (1961) and Garrison and Wilson, R.J., 1972, Introduction to graph theory. Oliver and Boyd.
Marble (1965). The latter concluded that "the present investigations, while
illuminating certain points, have failed to provide an entirely satisfactory B. Texts or Papers in Geography with considerable instructional component
set of results. Part of the problem appears to lie in the line of attack which in graph theory.
was chosen; the graph theoretical model of network structure has proved to
be strong in some areas, but disappointingly weak in others ....". Burton Alao, N., 1973, Some aspects of network theory. Geographica Polonica.
(1961) was more hopeful when he stated: "It is conceivable that a body of 25, pp 107-137.
useful theory could be built up around the application of graph theory to Campbell, J., 1971, Growth pole theory, digraph analysis and inter-industry
geographical problems". If the field has not moved as fast as Burton antici- relationships. Tidjschrift V. Econ. en Soc. Geografie,
pated this may be because a rather limited interpretation was put upon graph 63, pp 79-87.
theory and what it had to offer Geography. This restricted view within Geo-
graphy may have stemmed from an over-reliance on the summaries provided in Haggett, P., and Chorley, R.J., 1969, Network analysis in geography.
Garrison (1960) and Kansky (1963); both angled at a transportation audience. Edward Arnold.
The results obtained by Garrison and Marble (1965) undoubtedly derive from the Garrison, W., 1960, Connectivity of the Interstate highway system. Papers,
elementary form of analysis that they undertook. No such limitation is evident Reg. Sci. Ass., 6, pp 121-37.
in the work of Gauthier (1968) who applied the Ford-Fulkerson theorem to net-
Garrison, W., and Marble, D.F., 1965, A prolegomenon to the forecasting of
work flows in Argentina, yet this same theorem may be proved directly from transportation development. Technical Report 65-35. Prepared for the
the basic idea of a bridge set outlined in Section V. As the dates of many U.S. Army. (Reprinted as 'Graph theoretic concepts', in: Trans-
of the references indicate, the more fruitful and imaginative applications portation geography: comments and readings, ed M. Eliot
have come in graph theory's second decade in geography.
Hurst, (McGraw-Hill) 1974, p 58).
Of course there are limitations. Graph theory is a theoretical system Hay, A., 1973, Transport for the space economy. Macmillan, see Chapter 4.
for finite state problems and it is not well adapted to dealing with problems
Kansky, K.J., 1963, The structure of transport networks. university of
in the continuous plane. A problem that arises in dealing with combinatorial Chicago, Dept. of Geography, Res. Paper, 84.
structures in general is the complexity of the algorithms needed to find the

52 53
Nystuen, J.D., and Dacey, M.F., 1961, A graph theory interpretation of nodal James, G.A., Cliff, A.D., Haggett, P., and Ord, J.K., 1970, Some discrete
regions. Papers Reg. Sci. Ass., 7, pp 29-42. distributions with applications to regional transport networks.
Taafe, E.J., and Gauthier, H.L., 1973, Geography of transportation. Geografiska Annaler, 52B, pp 14-21.
Prentice-Hall, especially chapters 4, 5 and 6. Kissling, C.C., 1969, Linkage importance in a regional highway network.
Canadian Geographer, 13(2), pp 113-129.
C. Examples of 'rash theoretic applications in geography and further theoret-
ics developments. Muraco, W.A., 1972, Intraurban accessibility. Economic Geography, 48(4),
pp 388-405.
Alao, N., 1970, A note on the solution matrix of a network. Geographical
Analysis, 2, pp 83-88. Pitts, F.R., 1965, A graph theoretic approach to historical geography, Pro-
fessional Geographer, 17(5), pp 15-20.
Appel, K. and Haken, W., 1976, Every planar map is four colorable. Bulletin
of the American Mathematical Society, 82(5), pp 711-2. Reed, W.E., 1970, Indirect connectivity and hierarchies of urban dominance.
Annals, Association of American Geographers, 60, pp 770-85.
Burghart, A., 1969, The origin and development of the road network of the
Niagara Peninsula, Ontario, 1770-1851. Annals, Assoc. Amer. Saaty, T., 1972, Thirteen colorful variations on Guthries conjecture. Amer.
Geogr., 59(3), pp 417-440. Math. Monthly, 79(1), pp 2-43.
Burton, Ian, 1963, The quantitative revolution and theoretical geography. Scheidegger, A.E., 1967, Random graph patterns of drainage basins. Inter-
Canadian Geographer, 7, pp 151-162. national Association for Scientific Hydrology, Public-
ation No. 76, pp 415-25.
Carter, F.W., 1969, An analysis of the medieval Serbian Oecumene: a theoret-
ical approach. Geografiska Annaler, 51B, 1, p 39. Schmidt, C.G., 1975, Firm linkage structure and structural change - a graph
theoretical analysis. Economic Geography, 51(1), pp 27-36.
Clayton, C., 1974, Communication and spatial structure. Tijdschrift voor
Econ en Soc. Geografie, 65(3), pp 271-7. Scott, A.J., 1971, Combinatorial programming, spatial analysis and
planning. Methuen.
Collins, L., 1974, An introduction to Markov Chain Analysis. Catmog No. 1,
Geo Abstracts, Norwich. Smart, J.S., 1969, Topological properties of channel networks. Bull. Geol.
Soc. Am., 80, pp 1757-74.
Cummings, L.P., Manly, B.J., Weinand, H.C., 1974, Measuring association in
link-node problems. Geoforum, 13, pp 43-51. Soja, E.W., 1968, The Geography of Modernization in Kenya, Syracuse
Geographical Series, No.2
Cummings, L.P., Manly, B.J., Weinand, H.C., 1974, Measuring association in
link-node problems: a reply. Geoforum, 17, p 92. Stutz, F.P., 1973, Accessibility and the effect of scalar variations on the
powered transportation connection matrix. Geographical Analysis,
Ford, L.R. Jr. & Fulkerson, D.R., 1956, Maximal flow through a network. 5, pp 61-66.
Canadian Journal of Mathematics, 8, pp 395-404.
Symanski, R., and Webber, M.J., 1974, Complex periodic market cycles.
Gauthier, H.L., 1968, Least-cost flows in a capacitated network: a Brazilian Annals, Association of Amer. Geogr., 64(2), pp 208-213.
example. in: Geographic Studies of Urban Transportation and
Network Analysis, ed F. Horton, Northwestern University Studies Stromquist, W., 1975, The four-color theorem for small maps. J. of Com-
in Geography, No. 16. binatorial Theory, Series B, 19, pp 256-268.
Golledge, R.G., 1968, Methods in Australian trade and transportation studies. Tapiero, C.S. and Boots, B.N., 1974, Structural inference in transportation
in: Studies in Australian Geography, ed Dury, G.H., and Logan, networks, Environment and Planning A, 6, pp 411-18.
M.I., (Heinemann) pp 303-327. Tinkler, K.J., 1972(a), Physical interpretation of eigen functions of dicho-
Good, C.M., 1975, Periodic markets and travelling traders in Uganda. Geo- tomous matrices. Trans. Inst. Brit. Geog., 55, A-46.
graphical Review, 65(1), pp 45-72. Tinkler, K.J., 1972(b), Bounded planar networks: a theory of radial structures.
Hay, A.M., 1974, Measuring association in link-node problems: a comment. Geographical Analysis, 4(1), pp 5-33.
Geoforum, 17, pp 90-92. Tinkler, K.J., 1973, The topology of rural periodic market systems. Geograf-
Haynes, V., 1975, in: Cirque morphometry. Area, 7(1), p 69. iska Annaler, 55, Ser. B., pp 121-133.
Harvey, M.E., 1972, The identification of development regions in developing Tinkler, K.J., 1974, On summing power series expansions of accessibility
countries. Economic Geography, 48, pp 229-43. matrices by the inverse method. Geographical Analysis, 6,
pp 175-178.
Harvey, M.E., Hocking, R.T., and Brown, J.R., 1974, The chromatic travelling-
Tinkler, K.J., 1975, Access in several networks used successively. Area,
salesman problem and its application to planning and structuring
geographic space. Geographical Analysis, 6, pp 33-52. 7(4), pp 296-299.
Jackson, R.T., 1972, Economic Rationality amongst African rural traders.
Geografiska Annaler, 54B, pp 85-94.
55
54
Tinkler, K.J., 1976, On functional regions and indirect flows. Geographical
Review, 65(4), pp 476-492.
Tinkler, K.J., 1979, Graph Theory in Progress in Human Geography, 3,
pp 85-116.
Warntz, W., and Waters, N., 1975, Network representation of critical
elements of pressure surfaces. Geographical Review, 65(4),
pp 476-492.
Werner, C., 1971, Formal study of the evolution of the railway network in
Schleswig-Holstein. in: Studies in Geography, ed F.E. Ian
Hamilton, (Northwestern University), 18.
Werner, C., and Smart, J.S., 1973, Some new methods of topologic
classification of channel networks, Geographical Analysis,
5, pp 271-295.
Wood, L.J., 1975, The functional structure of a rural market system.
Geografiska Annaler, 57B, pp 109-18.

56

View publication stats