Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

GRAPHICAL MODELS AND IMAGE PROCESSING

Vol. 58, No. 4, July, pp. 345–359, 1996


ARTICLE NO. 0028

Jordan Graphs
RON AHARONI,*,1 GABOR T. HERMAN,* AND MARTIN LOEBL†,2
*Medical Image Processing Group, Department of Radiology, University of Pennsylvania, Philadelphia, Pennsylvania 19104; and
†Department of Combinatorics and Optimisation, University of Waterloo, Waterloo, Ontario, Canada

Received January 9, 1995; revised November 20, 1995; accepted March 7, 1996

erties which make their use advantageous whenever the under-


Early development of digital topology concentrated on tessel- lying situation allows us to use them.  1996 Academic Press, Inc.
lations of the plane into square-shaped pixels, each one of which
had a 1 or a 0 assigned to it. It became quickly apparent that
in order to avoid some ‘‘paradoxes’’ one needs to consider 1. MOTIVATION AND BACKGROUND
adjacencies in addition to that provided by the edge-adjacency.
The customary choice became the use of a pair of adjacencies; Jordan curves have an interior and an exterior which
one for the 1-pixels and another for the 0-pixels. While this between them contain all the points not on the curve and
approach can be treated in a mathematically rigorous fashion, both of which are connected (for any two interior points
it nevertheless remains attractive to consider those digital there is a path from one to the other which lies entirely
spaces in which a single adjacency sufficies to usefully define in the interior and for any two exterior points there is a
connectivity. This not only makes the resulting mathematics path from one to the other which lies entirely in the exte-
more elegant, but it also brings the subject nearer to classical
rior), but are disconnected from each other (every path
graph theory and its enormous wealth of results. This is what
motivated us to search for an appropriate new concept. Jordan
from an interior point to an exterior point must contain a
curves have an interior and an exterior which between them point on the curve). Generalization to surfaces in digital
contain all the points not on the curve and both of which are spaces is useful when displaying (only the exterior of a
connected, but are disconnected from each other. Generaliza- surface is visible from any direction) or analyzing (the
tion to surfaces in digital spaces is useful when displaying (only interior volume is well defined). An example of a Jordan
the exterior of a surface is visible from any direction) or analyz- surface in the medical imaging application area is shown
ing (the interior volume is well defined). Boundaries in digital in Fig. 1. Boundaries in digital spaces may or may not be
spaces may or may not be Jordan, but are ‘‘guaranteed’’ to be Jordan (we will show examples), but are ‘‘guaranteed’’ to
Jordan in spaces we call strong Jordan graphs. In this paper be Jordan in those digital spaces which we will call (strong)
we define such a notion of a strong Jordan graph. We show Jordan graphs.
that some previously studied classes (such as those of 1-simply
We now give the necessary background definitions and
connected digital spaces and of bridged graphs) are subclasses
results from [1–3]. For some of the terms used, the primary
of strong Jordan graphs. We also define Jordan graphs, as
those digital spaces in which finite boundaries are guaranteed definitions given in these papers are not identical. How-
to be Jordan and show that there are Jordan graphs which are ever, it is proved in [3] that the defined concepts themselves
not strong Jordan graphs. (Strong) Jordan graphs are charac- are the same and so it is valid to quote any of the results
terized by the existence of ‘‘cuts’’ in associated graphs, by the from these papers.
acyclic nature of the adjacency graphs of binary pictures, and A digital space is a pair (V, f), where V is a nonempty
also by the connectedness of the immediate interiors (or exteri- set (of spels) and f is a symmetric binary relation (called
ors) of certain types of surfaces. The surfaces of the last category the proto-adjacency) on V, under which V is connected;
include the so-called minimally near-Jordan surfaces, which that is, f is a set of ordered pairs of elements of V such that
are shown to be of some interest. Finally, we tie some of these
notions to standard notions of graph theory, such as the notion (i) if (c, d ) [ f, then (d, c) [ f and
of a separating set. Our overall conclusion is that those digital (ii) for any c and d in V, there exists a finite sequence
spaces which are Jordan graphs have some very desirable prop- c (0), . . . , c (m) of elements of V such that m $ 0, c (0) 5 c,
c (m) 5 d, and (c (k), c (k11)) [ f, for 0 # k , m.
1
Visiting from the Department of Mathematics, Technion, Haifa,
Israel. This definition of a digital space is much more general
2
On leave from the Department of Applied Mathematics, Charles than is necessary to do ‘‘digital topology,’’ since there is
University, Prague, Czech Republic. no geometrical restriction on the nature of the set V. In
345
1077-3169/96 $18.00
Copyright  1996 by Academic Press, Inc.
All rights of reproduction in any form reserved.
346 AHARONI, HERMAN, AND LOEBL

sical topology (such as those based on cell complexes [5]


and on Khalimsky spaces [6]) at the end of our paper.
For the sake of simplicity in illustrating our ideas, we
use V 5 Z 2, the set of ordered pairs of integers, in all our
figures. One intuitive geometrical interpretation of this is
that the underlying space is the Euclidean plane and the
grid points are the points with integer rectangular coordi-
nates. Thus the corresponding Voronoi neighborhoods are
the closed unit squares (called pixels) whose centers are
at the grid points. An ordered pair of pixels which share
an edge is interpreted as that edge with an orientation
across it from the first pixel to the second. Formally, the
digital space is (Z 2, g), where

g 5 h((i, j ), (i9, j9)) u [ui 2 i9u 1 u j 2 j9u 5 1]j. (1)

This is the classical ‘‘4-adjacency’’ [7, 8]. An alternative


intuitive geometrical interpretation is provided by defining
the grid points using a slanted coordinate system, as in
Fig. 2. The corresponding Voronoi neighborhoods are the
closed hexagons whose centers are at the grid points. An
ordered pair of hexagons which share an edge is interpre-
ted as that edge with an orientation across it from the first
to the second hexagon. Formally, this digital space is (Z 2,
FIG. 1. Computer graphic display of a Jordan surface in three-dimen- b), where
sional space. (Data provided by Dr. J. K. Udupa.)

b 5 h((i, j ), (i9, j9)) u [ui 2 i9u 1 u j 2 j9u 5 1]


(2)
or (i 2 i9) 3 ( j 2 j9) 5 1]j.
graph-theoretical terms, (V, f) is a connected directed
graph in which the adjacency is symmetric; note that we
allow V to be infinite and do not rule out self-adjacency.
(Even (V, V 2) is a legitimate example.) Although this will
not enter the mathematical discussion at all, the way we
imagine these spaces is much more restrictive. We think
of an n-dimensional Euclidean space and some arbitrary
but fixed collection of ‘‘grid points’’ in this space. Then we
think of V as the set of so-called Voronoi neighborhoods of
the grid points. (The Voronoi neighborhood of a grid point
is the set of all points in the n-dimensional Euclidean space
which are at least as close to that grid point as to any other
grid point.) These are closed convex polyhedral sets. We
then allow two elements of such a set V to be in the relation
f only if their intersection is (n 2 1)-dimensional. In the
two-dimensional case, this would mean that two Voronoi
neighborhoods can be in f only if they share an edge.
We emphasize that a digital space is not a topological
space in the sense of classical topology [4]. In fact, the
most fundamental notion of topological spaces (open sets)
do not even enter into our discussion. Nevertheless, as we
will see, we are able to define other notions corresponding
to those of classical topology, such as a ‘‘simply connected
set.’’ We will return to the relationship between our ap-
proach and alternative approaches more embedded in clas- FIG. 2. Hexagonal tessellation of the plane.
JORDAN GRAPHS 347

between the spels c and d, with an orientation from the


first to the second. This is illustrated in Fig. 2. The surfel
((0, 0), (1, 1)) is not to be interpreted as the vector from
(0, 0) to (1, 1), but rather as the edge between the hexagons
whose centers are at (0, 0) and at (1, 1), respectively (indi-
cated by the heavy line in Fig. 2), with the orientation of
the edge indicated by the vector. Also note that in this
interpretation spels are n-dimensional objects and surfels
are (oriented) (n 2 1)-dimensional objects and so surfaces
are not subsets of V.
The boundary between two subsets O and Q of V is
defined as

­(O, Q) 5 h(c, d ) u c [ O, d [ Q, (c, d ) [ fj. (4)


FIG. 3. Illustration of an g-component (heavy shading) of one subset
of Z 2 (the spels labeled by 1) and a d-component (light shading) of If ­(O, Q) is nonempty, then it is a surface. (Furthermore,
another subset of Z 2 (the spels labeled by 0). The *’s in two of the spels
are in there for ease of identification of them in a discussion following
this surface is uniquely specified by O and Q.) If in Fig. 3
Theorem 2.6. we let O be represented by the heavily shaded region and
Q be represented by the lightly shaded region, then the
unique ­(O, Q) is represented by the heavy edges, with
orientation from the spel labeled by a 1 to the spel labeled
A spel-adjacency r in the digital space (V, f) is defined
by a 0. (To avoid confusion, it is worth pointing out that
as a symmetric binary relation on V such that f , r. (We
while in classical topology [4] the boundary of a subset of
use the notation , for the subset relation; whenever we
a topological space X is always itself a subset of X, the
need to specify a proper subset, we do so in words.) Two
boundary between two sets of spels is a set of surfels, rather
spels c and d are referred to as r-adjacent (or, in case r 5
than a set of spels.)
f, as proto-adjacent), if (c, d ) [ r. A r-path kc (0), . . . , c (m)j
For any surface, S, its immediate interior is
is a sequence of spels such that, for 0 # k , m, (c (k),
c (k11)) [ r. We refer to a f-path as a proto-path. The
notions of a r-connected subset of V and of a r-component II(S) 5 hc [ V u (c, d ) [ S for some d [ V j (5)
of a subset of V are defined in the customary way.
To illustrate these ideas consider the digital space (Z 2, and its immediate exterior is
g). Examples of spel-adjacencies are g itself, the b defined
by (2), and IE(S) 5 hd [ V u (c, d ) [ S for some c [ V j. (6)

d 5 h((i, j ), (i9, j9)) u [ui 2 i9u 1 u j 2 j9u 5 1] A proto-path kc (0), . . . , c (m)l is said to cross a surface S if,
(3)
or [ui 2 i9u 5 2 and u j 2 j9u 5 1]j. for some 0 # k , m, either (c (k), c (k11)) [ S or (c (k11), c (k))
[ S. For a surface S, its interior is defined as
In Fig. 3 we illustrate an g-components of one subset of
Z 2 and a d-component of another subset of Z 2. I(S) 5 hd [ V u there exists a f-path
An element of f is called a surfel (short for ‘‘surface connecting d to an element of II(S) (7)
element’’). A nonempty set of surfels is called a surface in which does not cross S j,
(V, f). (Note that (c, d ) ? (d, c) and so if (c, d) is in a
surface, then (d, c) may or may not also be in the surface.
This is an essential feature of the theory that we are trying and its exterior is defined as
to develop; this theory deals with surfels with an orienta-
tion in which we can distinguish between their ‘‘insides’’ E(S) 5 hc [ V u there exists a f-path
and their ‘‘outsides.’’) While it is again not used in our connecting c to an element of IE(S) (8)
abstract development, it is worthwhile to give a geometrical which does not cross S j.
interpretation of the idea of a surfel in the case when V
consists of n-dimensional convex polyhedral sets in an n- In Fig. 3, both the interior and the exterior of the surface
dimensional space with adjacency interpreted as having an represented by the heavy edges (with orientation from the
(n 2 1)-dimensional intersection. The interpretation of the spel labeled by a 1 to the spel labeled by a 0) is the whole
surfel (c, d ) is exactly this (n 2 1)-dimensional intersection of Z 2.
348 AHARONI, HERMAN, AND LOEBL

LEMMA 1.1. For any surface S, A kl-boundary in a binary picture is a nonempty bound-
ary between a k-component of the set of 1-spels and a l-
I(S) < E(S) 5 V. (9) component of the set of 0-spels. Note that a kl-boundary
is always a surface. In Fig. 3, the surface represented by
Proof. This is Proposition 2.1 of [1]. n the heavy edges (with orientation from the spel labeled
by a 1 to the spel labeled by a 0) is an gd-boundary.
A surface S is said to be near-Jordan if every proto-path
For any binary relation r on V, a sequence kc (0), . . . ,
from an element of II(S) to an element of IE(S) crosses
c l of spels is said to be r-tight if, for 0 # k # m, either
(m)
S. Clearly, the surface represented in Fig. 3 is not near-
(c (0), c (k)) [ r or (c (m), c (k)) [ r, or both. A spel-adjacency
Jordan.
r in the digital space (V, f) is said to be tight if
LEMMA 1.2. For a near-Jordan surface S,
(i) (c, d ) [ r ⇒ (there exists a r-tight proto-path from c to d ).
(12)
I(S) > E(S) 5 B, (10)
Specific spel-adjacencies that have been proposed in the
(ii) every proto-path from a spel in I(S) to a spel in literature [7, 8] with some application in mind are all tight.
E(S) crosses S, and In particular, g, b, and d are all tight in (Z 2, g) and b and
(iii) d are tight in (Z 2, b).
THEOREM 1.3. If k and l are tight spel-adjacencies in a
S 5 ­(I(S), E(S)). (11)
digital space (V, f), then any kl-boundary in any binary
picture over (V, f) has a k-connected interior and a l-
Proof. This is Proposition 2.2 of [1]. n
connected exterior.
Near-Jordanness of a surface does not guarantee the
connectedness either of its interior or of its exterior. (As Proof. This is Theorem 4.6 of [1]. n
an example, consider again Fig. 3. The surface which is A consequence of this theorem is that (for tight k and
represented by all the oriented edges from a spel labeled l) in order to check whether a kl-boundary is kl-Jordan,
by a 1 to a spel labeled by a 0 is clearly near-Jordan. Its ‘‘all’’ we have to do is to check whether it is near-Jordan.
interior consists of all the spels labeled by a 1 and it is not Unfortunately, near-Jordanness is a global concept. In this
g-connected. Its exterior consists of all the spels labeled paper, we seek a property of digital spaces which guaran-
by a 0 and it is not d-connected.) For spel-adjacencies k tees near-Jordanness.
and l, a kl-Jordan surface S is a near-Jordan surface such An unordered pair hk, lj of tight spel-adjacencies is
that I(S) is k-connected and E(S) is l-connected. The no- called a strong Jordan pair (respectively, a Jordan pair)
tion of a kl-Jordan surface shares important properties for (V, f), if every (respectively, every finite) kl-boundary
with the notion of a Jordan curve: its interior and exterior in every binary picture over (V, f) is near-Jordan. It follows
partition the whole space; both are connected in some from Fig. 3 that hg, d j is not a Jordan pair (and hence
sense, but one cannot get from the interior to the exterior definitely not a strong Jordan pair) for (Z 2, g). On the
without crossing the surface which is, in fact, the boundary other hand, it is known ([2, Corollary 5.13]) that h b, bj is
between the interior and the exterior. (Mathematically a strong Jordan pair (and hence a Jordan pair) for (Z 2, b).
speaking there is also a difference: a simple closed curve
is a subset of the plane and the Jordan curve theorem [4] THEOREM 1.4. If hk, lj is a strong Jordan pair (respec-
says that the complement of the curve, rather than the tively, a Jordan pair) for (V, f) and k9 and l9 are tight spel-
whole plane, has precisely two components. Since our sur- adjacencies in (V, f) such that k , k9 and l , l9, then hk9,
faces are subsets of f, rather than of V, it is possible to l9j is also a strong Jordan pair (respectively, a Jordan pair)
demand that the interior and exterior partition the whole for (V, f).
of V; if anything, a more attractive-sounding aim than that Proof. This is Theorem 5.1 of [1]. n
of the classical theorem.)
A binary picture is a triple (V, f, f ), where f is a function It follows from this theorem and the discussion just
whose domain is V and which takes its values from the set above it that hg, gj is not a Jordan pair (and hence not a
h0, 1j. For a binary picture (V, f, f ), a spel c is called a 1- strong Jordan pair) for (Z 2, g) but every pair hk, lj of tight
spel if f (c) 5 1 or a 0-spel if f (c) 5 0. A bindary picture spel-adjacencies in (Z 2, b) is a strong Jordan pair (and
is said to be finite if its set of 1-spels is finite. If all nonindi- hence a Jordan pair) for (Z 2, b). This is clearly a very
cated spels in Fig. 3 are considered labeled by a 0, then desirable property of the digital space (Z 2, b); in this space
the figure represents a finite binary picture. every kl-boundary is kl-Jordan, as long as both k and l
JORDAN GRAPHS 349

are tight in the space. In this paper we investigate in general arcs rather than to points and lines and also we allow
the nature of those digital spaces which have such a graphs to be infinite.
property. For any surface S in a digital space we define the arc set
associated with S as
2. JORDAN GRAPHS AND
STRONG JORDAN GRAPHS A(S) 5 hhc, d j u (c, d ) [ S or (d, c) [ S j. (14)

A digital space (V, f) is called a (strong) Jordan graph For any digital space (V, f), we define the graph associated
if hf, fj is a (strong) Jordan pair for (V, f). The following with (V, f) as
is an immediate consequence of Theorem 1.4.
COROLLARY 2.1. In a (strong) Jordan graph every pair G(V, f) 5 (V, A(f)). (15)
of tight spel-adjacencies is a (strong) Jordan pair. In particu-
lar, if r is a tight spel-adjacency in a (strong) Jordan graph A nonempty subset F of A(f) is called a cut in the graph
(V, f), then (V, r) is also a (strong) Jordan graph. G(V, f), if there exists a subset O of V such that
Our previous discussion shows that (Z 2, b) is a strong
Jordan graph and that (Z 2, g) is not even a Jordan graph. F 5 A(­(O, V 2 O)). (16)
We show below that there are Jordan graphs which are
not strong Jordan graphs. To see that the condition on (We use a minus symbol to denote set theoretical subtrac-
tightness is necessary in Corollary 2.1, consider the follow- tion; i.e., V 2 O is the set of elements in V which are not
ing one-dimensional example. For any positive integer, n, in O.)
we define a binary relation nn on Z by LEMMA 2.2. If a surface S in a digital space (V, f) is
near-Jordan, then A(S) is a cut in the graph G(V, f). If a
nn 5 h(i, i9) u ui 2 i9u 5 1 or nj. (13) nonempty subset F of A(f) is a cut in the graph G(V, f),
then there is a near-Jordan surface S in (V, f) such that
It is trivial to prove that (Z, n1) is a strong Jordan graph A(S) 5 F.
(in fact every n1 n1-boundary has exactly one element in
this digital space) and also that n2 and n3 are tight spel- Proof. Suppose first that S is near-Jordan. Let O 5
adjacencies in (Z, n1). It follows from Corollary 2.1 that I(S). By Lemmas 1.1 and 1.2, E(S) 5 V 2 O. Substituting
(Z, n2) and (Z, n3) are also strong Jordan graphs. On the into (11) and applying A to both sides, we see using (16)
other hand, n4 is not a tight spel-adjacency in (Z, n1) and that A(S) is a cut in G(V, f).
(Z, n4) is not even a Jordan graph. This can be seen from Now suppose that F is a cut in G(V, f). By definition,
Fig. 4, in which the heavily shaded region is a n4-component there exists a subset O of V such that F satisfies (16).
of 1-spels, the lightly shaded region is a n4-component of Letting S 5 ­(O, V 2 O), we see that A(S) 5 F. S must
0-spels (consisting of one spel), and the heavy line indicates be near-Jordan, since every proto-path from an element
the n4 n4-boundary between them; it consists of the single of its immediate interior (which must be in O) to an ele-
surfel (9, 10). The n4-path (9, 13, 12, 11, 10) connects the ment of its immediate exterior (which cannot be in O)
immediate interior of this n4 n4-boundary to its immediate must at some point move from O to a spel not in O and
exterior without crossing it. hence cross S. n
Prior to continuing our discussion of (strong) Jordan THEOREM 2.3. A digital space (V, f) is a strong Jordan
graphs using the terminology of digital topology, we will graph (respectively, a Jordan graph) if, and only if, for
explain their nature in terminology that is more familiar every binary picture (V, f, f ) and for every pair of f-
to graph theorists. Even though we have been emphasizing components O of the 1-spels and Q of the 0-spels such that
the importance of orientation in the previous section, it ­(O, Q) is nonempty (respectively, nonempty and finite),
turns out that being Jordan can be introduced as a property A(­(O, Q)) is a cut in the graph G(V, f).
of graphs, rather than of directed graphs. When discussing
ordinary (i.e., not directed) graphs, we generally follow Proof. Suppose first that (V, f) is a strong Jordan graph
the terminology of [9]; however, we refer to nodes and (respectively, a Jordan graph). Let (V, f, f ) be a binary
picture with O a f-component of the 1-spels and Q a f-
component of the 0-spels, such that ­(O, Q) is nonempty.
By definition, this means that ­(O, Q) is a ff-boundary.
Since hf, fj is a strong Jordan pair (respectively, a Jordan
pair) for (V, f), ­(O, Q) is near-Jordan (respectively, near-
FIG. 4. Illustration that (Z, n 4) is not a Jordan graph. Jordan if finite). It now follows from Lemma 2.2 that if
350 AHARONI, HERMAN, AND LOEBL

­(O, Q) is nonempty (respectively, nonempty and finite), Consider, for example, Fig. 4. There are three n4-compo-
then A(­(O, Q)) is a cut in the graph G(V, f). nents of 0-spels (O1 5 h. . . , 1, 2, 3, 4, 5, 7j, O2 5 h10j,
Suppose now that for every binary picture (V, f, f ) and and O3 5 h12, 15, 16, 17, 18, . . .j) and three n4-components
for every pair of f-components O of the 1-spels and Q of of 1-spels (I1 5 h6j, I2 5 h8, 9, 13, 14j, and I3 5 h11j). The
the 0-spels such that ­(O, Q) is nonempty (respectively, corresponding n4 n4-adjacency graphs in (Z, n1) and in (Z,
nonempty and finite), A(­(O, Q)) is cut in the graph G(V, n4) are shown in Figs. 5a and 5b, respectively. We note
f). In order to show that (V, f) is a strong Jordan graph that kl-adjacency graphs are necessarily bigraphs ([9, p.
(respectively, a Jordan graph), we need to show that for 17]; other authors quite commonly refer to these as ‘‘bipar-
every pair of f-components O of the 1-spels and Q of 0- tite graphs’’).
spels such that ­(O, Q) is nonempty (respectively, non- Our next theorem characterizes strong Jordan pairs in
empty and finite), ­(O, Q) is near-Jordan. Since, by the terms of adjacency graphs. Its proof makes use of the
above, A(­(O, Q)) is a cut in the graph G(V, f), we know following technical lemmas.
from Lemma 2.2 that there is a near-Jordan surface S in
LEMMA 2.4. Let (V, f, f ) be a binary picture, O be a
(V, f) such that
k-component of the 1-spels, Q be a l-component of the 0-
spels, such that ­(O, Q) is nonempty. If k is a tight spel-
A(S) 5 A(­(O, Q)). (17) adjacency in (V, f), then O , I(­(O, Q)) and if l is a tight
spel-adjacency in (V, f), then Q , E(­(O, Q)).
Now we show that ­(O, Q) itself is near-Jordan. In order
Proof. This is Lemma 4.5 of [1]. n
to do that, we need to show that every proto-path from
an element c in the immediate interior of ­(O, Q) to an LEMMA 2.5. Let (V, f, f ) be a binary picture.
element d in the immediate exterior of ­(O, Q) crosses
(i) If k is a tight spel-adjacency in (V, f) and O is a
­(O, Q). Due to (17), both c and d are in II(S) < IE(S).
k-component of the 1-spels, then for every pair of elements
If one is in II(S) and the other in IE(S), then a proto-path
of O, there is a proto-path from one to the other such that
from one to the other crosses S and, equivalently, ­(O,
all 1-spels occurring in this proto-path are also in O.
Q). Otherwise, suppose without loss of generality that they
(ii) If l is a tight spel-adjacency in (V, f) and Q is a
are both in II(S); we now show that this leads to a contra-
l-component of the 0-spels, then for every pair of elements
diction. Let e be that element of V such that (d, e) [ S.
of Q, there is a proto-path from one to the other such that
Due to (17), we know that either (e, d ) [ ­(O, Q) or (d,
all 0-spels occurring in this proto-path are also in Q.
e) [ ­(O, Q). The second possibility cannot occur, since
d [ Q and hence cannot also be in O. Hence both c and Proof. We prove only (i), since the proof of (ii) is
e are in O and so there is a proto-path entirely in O (and strictly analogous. Let (V, f, f ) be a binary picture, k be
hence not crossing ­(O, Q) or equivalently, S) from c to a tight spel-adjacency in (V, f), and O be a k-component
e, i.e., from an element of II(S) to an element of IE(S), of the 1-spels. Since O is k-connected, we can get from
contradicting the fact that S is near-Jordan. n any element in it to any other element in it by a k-path
entirely in O. This k-path can be used to create the desired
In view of this theorem, we can give the following purely proto-path, provided only that we can show that, for any
graph-theoretical definition. A (finite or infinite) con- k-adjacent c and d in O, there is a proto-path connecting
nected graph G is strong Jordan (respectively, Jordan), if them such that all 1-spels occurring in this proto-path are
for every partition of its node set into two sets X and Y also in O. Since k is tight, there exists a proto-path kc 5
and for every pair of connected components A of kX l and c (0), . . . , c (m) 5 d l from c to d such that, for 0 # k # m,
B of kY l (the subgraphs induced by X and Y, respectively either (c, c (k)) [ k or (d, c (k)) [ k, or both. Since O is a
[9]), if the set F of arcs between A and B is nonempty k-component of the 1-spels, the required result follows. n
(respectively, nonempty and finite), then F is a cut in G.
Thus, all that we say can be considered to be about certain
entities of graph theory; nevertheless from now on we will
stick to the terminology and notation of digital topology,
which is the field that motivated our work. In spite of this,
graph theory enters in an essential way into our develop-
ment, as can be seen already from the following definition.
Let k and l be tight spel-adjacencies in the digital space
(V, f). The kl-adjacency graph of a binary picture (V, f,
f ) has as its nodes the k-components of the 1-spels and
the l-components of the 0-spels, with an arc between two FIG. 5. n 4n 4-adjacency graphs associated with Fig. 4: (a) in (Z, n 1)
such nodes O and Q if, and only if, ­(O, Q) is nonempty. and (b) in (Z, n 4).
JORDAN GRAPHS 351

For an illustration of Lemma 2.5(ii), consider Fig. 3. from d9 to e9 such that all 0-spels occurring in this proto-
The lightly shaded d-component of the 0-spels is not g- path are also in Q9. Since Q and Q9 are different l-compo-
connected and so there are pairs of elements in this d- nents of the 0-spels, this proto-path does not cross ­(O,
component which cannot be connected by a proto-path of Q). Similarly, we can extend the proto-path to an element
(Z 2, g) which does not leave the d-component. However, of the k-component of the 1-spels which follows Q9 in the
for any pair of elements of the d-component, there is a cycle, without crossing ­(O, Q). This way we have moved
connecting proto-path of (Z 2, g) all of whose 0-spels are from Q9 in the cycle to the node of the kl-adjacency graph
in the d-component. which follows it in the cycle. By repeating this process,
using alternatively Lemma 2.5(i) and Lemma 2.5(ii) (and
THEOREM 2.6. For tight spel-adjacencies k and l in the
noting that until and including the point when we have
digital space (V, f), hk, lj is a strong Jordan pair for (V,
taken care of O9, the nodes that we are dealing with are
f) if, and only if, the kl-adjacency graph of any binary
different from O and from Q), we can generate the re-
picture is acyclic.
quired proto-path from c to d. n
Proof. Suppose first that hk, lj is not a strong Jordan
It is easy to illustrate the existence of a cycle in the gd-
pair for (V, f). Then there is a binary picture (V, f, f )
adjacency graph associated with the binary picture of Fig.
with a k-component O of the 1-spels and a l-component
3. Let O be the heavily shaded g-component of the 1-spels
Q of the 0-spels such that there is a proto-path kc (0), . . . ,
and Q be the lightly shaded d-component of the 0-spels.
c (m)l not crossing ­(O, Q) from an element of II(­(O, Q))
Let Q9 consist of the 0-spel marked by an asterisk (it is
to an element of IE(­(O, Q)). Clearly, there is an arc
not d-adjacent to any other 0-spel) and let O9 consist of
between O and Q in the kl-adjacency graph of (V, f, f ).
the 1-spel marked by an asterisk (it is not g-adjacent to
This arc can be shown to be an arc in a cycle in this graph
any other 1-spel). Then O, Q9, O9, Q, O is a cycle in the
by following the proto-path kc (0), . . . , c (m)l. Since this proto-
gd-adjacency graph associated with the binary picture of
path begins in O and ends in Q but does not cross ­(O,
Fig. 3.
Q), the 0-spel in it which follows the last spel in it which
While this theorem is of some general interest (since it
is also in O must be in a l-component Q9 of the 0-spels
gives a new characterization of the concept of a strong
other than Q and the 1-spel in it which precedes the first
Jordan pair), from the point of view of the current paper
spel in it which is also in Q must be in a k-component O9
its relevance lies in the fact that we can in particular con-
of the 1-spels other than O. It follows that there are arcs
clude the following:
between O and Q9 and between O9 and Q in the kl-
adjacency graph of (V, f, f ). We get a cycle O, Q9, . . . , COROLLARY 2.7. A digital space (V, f) is a strong Jor-
O9, Q, O in the kl-adjacency graph of (V, f, f ) by putting dan graph if, and only if, the ff-adjacency graph of any
in place of the ‘‘. . .’’ all the k-components of the 1-spels binary picture is acyclic.
and l-components of the 0-spels visited by the proto-path
between Q9 and O9 (eliminating any repeated visits that As an illustration consider Fig. 6. The spel e is adjacent
may occur). to the other four spels. Hence, in any binary picture in
Suppose now that there is a binary picture (V, f, f ) such which e is a 1-spel there is at the most one f-component
that its kl-adjacency graph contains the cycle O, Q9, . . . , of the 1-spels, and in any binary picture in which e is a 0-
O9, Q, O, where O is a k-component of the 1-spels and Q spel there is at most one f-component of the 0-spels. In
is a l-component of the 0-spels. To show that hk, lj is not either case, the ff-adjacency graph is acyclic and so, by
a strong Jordan pair for (V, f), it is sufficient to show that Corollary 2.7, the digital space is a strong Jordan graph.
­(O, Q) is not near-Jordan. By Lemma 1.2(ii) and Lemma The same corollary can be used to show that all elements
2.4, it is sufficient to show that there is a proto-path from of some previously studied classes of digital spaces are
a spel in O to a spel in Q which does not cross ­(O, Q).
Let (c, d9) [ ­(O, Q9) and (c9, d ) [ ­(O9, Q). Since O
and O9 are different k-components of the 1-spels and Q
and Q9 are different l-components of the 0-spels, neither
of these surfels is in ­(O, Q). We now discuss how to fill
in the ‘‘. . .’’ in the proto-path kc, d9, . . . , c9, d l from an
element of O to an element of Q so that it does not cross
­(O, Q). Let e9 be an element of Q9 which is proto-adjacent
to some element of the k-component of the 1-spels which
follows Q9 in the cycle whose existence we are assuming.
(This k-component of the 1-spels could be, but does not FIG. 6. Illustration of a finite strong Jordan graph. Two spels are
have to be, O9.) By Lemma 2.5(ii), there is a proto-path considered proto-adjacent if, and only if, they share an edge.
352 AHARONI, HERMAN, AND LOEBL

Jordan. In this paper we introduce, based on [2, 10], two which is N-equivalent to a trivial loop by the induction hy-
such classes and prove that they have the stated property. pothesis.
Let N be a nonnegative integer. If (iii) The very sequence of elementary N-equivalences
which transform kd (0), . . . , d (u)l into kc (i)l can be used to
P 5 kc (1), . . . , c (m), d (0), . . . , d (n), e (1), . . . , e (l)l (18) transform kc (0), . . . , c (i 21), d (0), . . . , d (u), c (i 11), . . . , c (m)l
into kc (0), . . . , c (i), . . . , c (m)l. n
and In order to get to our next result we need a further
definition from [2]. A ff-boundary S is said to be 1-locally
P9 5 kc (1), . . . , c (m), f (0), . . . , f (k), e (1), . . . , e (l)l (19) Jordan if for any proto-path kc (0), c (1), c (2)l such that (c (0),
c (2)) [ S, exactly one of (c (0), c (1)) and (c (1), c (2)) is also in S.
are proto-paths in the same digital space, such that THEOREM 2.9. Every 1-simply connected digital space
is a strong Jordan graph.
f (0) 5 d (0), f (k) 5 d (n), and 1 # k 1 n # N 1 2, (20)
Proof. As a special case of Corollary 4.8 of [2], we
have that a 1-simply connected digital space (V, f) is a
then P and P9 are said to be elementarily N-equivalent. strong Jordan graph if every ff-boundary in every binary
Two proto-paths, P and P9, in a digital space are said to picture over (V, f) is 1-locally Jordan. On the other hand,
be N-equivalent, if there is a sequence of proto-paths P0 , it is a special case of Proposition 5.1 of [2] that indeed
. . . , Pp ( p $ 0) in the digital space, such that P0 5 P, every ff-boundary in every binary picture over (V, f) is
Pp 5 P9, and, for 0 # q , p, Pq and Pq11 are elementarily 1-locally-Jordan. n
N-equivalent.
A loop (of length m) in a digital space is a proto-path This theorem is as strong as it can possibly be, since we
kc (0), . . . , c (m)l such that c (m) 5 c (0). In particular, for any cannot replace in its statement 1-simply connected by 2-
spel c, kcl is a loop and is called a trivial loop. A digital simply connected. This is so, since (Z 2, g) is 2-simply con-
space is said to be N-simply connected if every loop in the nected [2, Lemma 5.5] and we have already seen that it is
digital space is N-equivalent to a trivial loop. The digital not a Jordan graph. Also, the converse of Theorem 2.9 is
space of Fig. 6 is 1-simply connected (and hence N-simply not valid; we now give an example, due to [11], of a finite
connected for any positive N ). The 1-equivalence of the strong Jordan graph which is not a 1-simply connected
loop ka, b, c, d, al to a trivial loop can be shown by the digital space. The set of spels is h0, 1, 2, 3, 4, 5, 6j and
following sequence of loops, each one of which is elemen- two spels are proto-adjacent if, and only if, the modulo 7
tarily 1-equivalent to its neighbors in the sequence: ka, b, difference between them is 61 or 62. The digital space is
c, d, al, ka, e, b, c, d, al, ka, e, c, d, al, ka, e, d, al, kal. We now the graph called C 27 in [11]. It is easy to see that it is strong
prove some technical results which will be needed later. Jordan. Consider the surface T 5 h(0, 5), (5, 0), (0, 6), (6,
0), (1, 6), (6, 1)j and define the T-parity of a proto-path
LEMMA 2.8. Let (V, f) be a digital space and N be a kc (0), c (1), . . . , c (m21), c (m) 5 c (0)l to be even if the number
nonnegative integer. of occurrences of (c (0), c (1)), (c (1), c (2)), . . . , (c (m21), c (m))
(i) If a loop kc (0), . . . , c (m)l is N-equivalent to a trivial in T is even and to be odd otherwise. It is easy to check
loop, then that trivial loop is kc (0) 5 c (m)l. that if two proto-paths are elementarily 1-equivalent, then
(ii) For any proto-path kc (0), . . . , c (m)l, the loop kc (0), they have the same T-parity. Hence the loop k0, 1, 2, 3, 4,
. . . , c (m), c (m), . . . , c (0)l is N-equivalent to a trivial loop. 5, 0l, which has an odd T-parity cannot possibly be 1-
(iii) For any proto-path kc (0), . . . , c (i), . . . , c (m)l and equivalent to a trivial loop (which has an even T-parity).
for any loop kd (0), . . . , d (u)l which is N-equivalent to the We now consider the relationship of the concepts intro-
trivial loop kc (i)l, kc (0), . . . , c (i), . . . , c (m)l is N-equivalent to duced so far to a class of commonly studied graphs, the
kc (0), . . . , c (i 21), d (0), . . . , d (u), c (i 11), . . . , c (m)l. so-called bridged graphs. A loop kc (0), c (1), . . . , c (m21),
c (m) 5 c (0)l in a digital space (V, f) is called a cycle if its
Proof. (i) If two proto-paths are elementarily N- length m is at least 3 and c (i) ? c ( j), for 0 # i , j , m. A
equivalent, then they have the same first spel and the same proto-path kd (0), . . . , d (u)l is said to be a bridge of this
last spel; see (18)–(20). cycle, if for some i and j, 0 # i , j , m, d (0) 5 c (i),
(ii) This condition is clearly valid if m 5 0. Now suppose d (u) 5 c ( j), and both (c (0), . . ., c (i21), c (i) 5 d (0), d (1), . . . ,
that it is valid for some nonnegative integer m. kc (0), . . . , d (u21), d (u) 5 c ( j), c ( j11), . . . , c (m) 5 c (0)l and kc ( j) 5 d (u),
c (m), c (m11), c (m11), c (m), . . . , c (0)l is elementarily N-equiva- d (u21), . . . , d (1), d (0) 5 c (i), c (i11), . . . , c ( j21), c ( j)l are cycles
lent to kc (0), . . . , c (m), c (m11), c (m), . . . , c (0)l, which is elemen- in (V, f) of length less than m. A digital space (V, f) is
tarily N-equivalent to kc (0), . . . , c (m), . . . , c (0)l, which is said to be bridged (see, e.g., [10]) if every cycle of length
elementarily N-equivalent to kc (0), . . . , c (m), c (m), . . . , c (0)l, at least 4 has a bridge. The digital space represented in
JORDAN GRAPHS 353

Fig. 6 is not bridged, since the cycle ka, b, c, d, al does not LEMMA 2.11. Let k and l be finitary spel-adjacencies in
have a bridge. On the other hand the digital space (Z 2, b) a digital space (V, f) such that at least one of them is very
is bridged; in Fig. 2 the cycle k(0, 1), (1, 1), (1, 0), (0, 0), tight. If S is a finite kl-boundary in a binary picture (V, f,
(0, 1)l has the bridge k(1, 1), (0, 0)l and the cycle k(0, 1), g), then it is a kl-boundary in some finite binary picture
(1, 1), (1, 0), (0, 21), (21, 21), (21, 0), (0, 1)l has the (V, f, f ).
bridge k(0, 1), (0, 0), (0, 21)l. Thus some (even infinite)
Proof. This is Lemma 5.2 of [1]. n
strong Jordan graphs are bridged and some (even finite
and 1-simply connected) ones are not. On the other hand, The proof given in [1] for this lemma is constructive. In
we have the following result. fact, the finite binary picture of Fig. 3 is the result of
such a construction applied to a nonfinite binary picture
THEOREM 2.10. Every bridged digital space is 1-simply
containing the same gd-boundary as is indicated in Fig. 3
connected and so is a strong Jordan graph.
by the heavily drawn eges.
Proof. Suppose that a digital space (V, f) is not 1-
simply connected. We show that assuming that it is bridged THEOREM 2.12. If k and l are finitary spel-adjacencies
leads to a contradiction. By the assumption that (V, f) is in the digital space (V, f) such that at least one of them is
not 1-simply connected, there exists a loop which is very tight, then hk, lj is a Jordan pair if, and only if, the
not 1-equivalent to a trivial loop. Let kc (0), kl-adjacency graph of any finite binary picture is acyclic.
c (1), . . . , c (m21), c (m) 5 c (0)l be a such a loop of minimal Proof. Suppose first that hk, lj is not a Jordan pair for
length m. Clearly m . 3; otherwise, the loop is elementarily (V, f). Then there is a finite kl-boundary S in a binary
1-equivalent to the trivial loop kc (0)l. Furthermore, there picture (V, f, g), such that S is not near-Jordan. By Lemma
can be no i and j such that 0 # i , j , m and c (i) 5 c ( j) ; 2.11, S is also a kl-boundary in some finite binary picture
otherwise, kc (i), c (i11), . . . , c ( j) 5 c (i)l would be a loop of (V, f, f ). Now we can show that the kl-adjacency graph
length less than m (and hence 1-equivalent to a trivial of (V, f, f ) is cyclic, just as we have done in the first half
loop) and so, by Lemma 2.8(iii), kc (0), c (1), . . . , c (m21), of the proof of Theorem 2.6.
c (m) 5 c (0)l would be 1-equivalent to the shorter loop Conversely, suppose that the kl-adjacency graph of a
kc (0), . . . , c (i21), c (i) 5 c ( j), c ( j11), . . . , c (m21), c (m) 5 c (0)l and, finite binary picture is cyclic. We can then generate a kl-
hence, to a trivial loop. This shows that kc (0), c (1), . . . , boundary in this finite binary picture which is not near-
c (m21), c (m) 5 c (0)l is a cycle. By the assumption that Jordan, just as we have done in the second half of the proof
(V, f) is bridged, there exist an i and a j such that 0 # of Theorem 2.6. In addition, we observe that finiteness of
i , j , m, d (0) 5 c (i), d (u) 5 c ( j), and both kc (0), . . . , c (i21), the binary picture together with the finitariness of f (which
c (i) 5 d (0), d (1), . . . , d (u21), d (u) 5 c ( j), c ( j11), . . . , c (m) 5 has to be finitary, since it is a subset of both k and l)
c (0)l and kc ( j) 5 d (u), d (u21), . . . , d (1), d (0) 5 c (i), c (i11), . . . , implies this not near-Jordan kl-boundary is finite. This
c ( j21), c ( j)l are cycles in (V, f) of length less than m and shows that hk, lj is not a Jordan pair for (V, f). n
hence are 1-equivalent to trivial loops. By Lemma 2.8, kc (0),
COROLLARY 2.13. If f is finitary, then (V, f) is a Jordan
c (1), . . . , c (m21), c (m) 5 c (0)l is 1-equivalent to kc (0), c (1), . . . ,
graph if, and only if, the ff-adjacency graph of any finite
c (i) 5 d (0), d (1), . . . , d (u) 5 c ( j), d (u), . . . , d (0) 5 c (i), . . . ,
binary picture is acyclic.
c (m21), c (m) 5 c (0)l, which in its turn is 1-equivalent to kc (0),
c (1), . . . , c (i) 5 d (0), d (1), . . . , d (u) 5 c ( j), c ( j11), . . . , c (m21), We now discuss further characterizations of (strong) Jor-
c (m) 5 c (0)l and, hence, to a trivial loop. n dan graphs in terms of contractibility [9, p. 113]. This discus-
sion will be informal, without giving details of the proofs
Somewhat oddly, the characterization of Jordan pairs (as
of the statements in it; its purpose is to strengthen the links
opposed to strong Jordan pairs) requires some additional
between our study and traditional graph theory. We note
machinery. A spel-adjacency r in a digital space (V, f) is
that if G(V, f) is contractible to a bigraph H, then we can
said to be finitary if, for every spel c, hd u (c, d ) [ rj is
use H to define an f so that the ff-adjacency graph of (V,
finite and is said to be very tight if whenever (c, d ) [ r,
f, f ) is H. Hence, a necessary condition for the strong
there exists a proto-path kc (0), . . . , c (m)l from c to d such
Jordanness of (V, f) is that if G(V, f) is contractible to a
that, for 0 , k # m, (c (0), c (k)) [ r. Clearly, all spel-
bigraph H, then H is acyclic. In the special case when (V,
adjacencies that we have dealt with so far are finitary and
f) is a digital space with a finite V, then (V, f) is contractible
all specific spel-adjacencies studied in the digital topology
to the ff-adjacency graph of any binary picture (V, f, f )
literature [7, 8] are finitary. Being very tight is somewhat
and the following are equivalent:
more restrictive from the point of view of eliminating po-
tentially useful spel-adjacencies; in (Z 2, g) b is very tight, (i) (V, f) is a strong Jordan graph.
but d is not. The need for these concepts comes from the (ii) (V, f) is a Jordan graph.
fact that they appear in the conditions of the following (iii) If G(V, f) is contractible to a bigraph H, then H
result. is acyclic.
354 AHARONI, HERMAN, AND LOEBL

Since one can also prove that a bigraph which has a cycle Q)) to a spel d [ IE(­(O, Q)). We will give a construction
is contractible to a complete bigraph Km,n (for a definition which with any such picture will associate a binary picture
of these see, e.g., [9, p. 17]) for some integers m and n, both (Z 2, b, g) such that if O9 is the b-component of the 1-spels
greater than 1, we have the further equivalent condition: containing c and Q9 is the b-component of the 0-spels
containing d, then there is a b-path not crossing ­(O9, Q9)
(iv) For any integers m and n, both greater than 1, G(V,
from c to d. This would indicate that (Z 2, b) is not a
f) is not contractible to the complete bigraph Km,n .
Jordan graph.
This last result, combined with Theorem 1.2 of [11], implies The construction is done as follows. Since (T, t, f ) is
that—in the special case of a finite V—(V, f) is a (strong) finite, there exists an integer u such that f (i, j) 5 0 if j $
Jordan graph if, and only if, G(V, f) is a well-connected u, for 0 # i # 3. We define
graph in the sense defined in [11]. As a result of this, special
instances (for the case of finite V) of some of our results D 5 h(i, j) u 0 # i # 3, j # uj (25)
have been foreshadowed in [11]; we discuss the relationship
between our results and those of [11] in greater detail near and we set g equal to f for spels in D (which is a subset
the end of our paper. of both Z 2 and of T ). Outside D we set g equal to 0 with
We conclude this section by giving an example of a the following exception. For any k such that f (3, k) 5 1,
Jordan graph which is not a strong Jordan graph. Let we assign the value 1 to g for any spel (i, j) such that either
4 # i # 3 1 u 2 k and j 5 k, or i 5 3 1 u 2 k and k #
T 5 h0, 1, 2, 3j 3 Z (21) j # 2u 2 k, or 2u 1 k # i # 3 1 u 2 k and j 5 2u 2 k,
or i 5 2u 1 k and k # j # 2u 2 k, or 2u 1 k # i # 21
and and j 5 k. This is done to simulate the wraparound in the
digital space (T, t) by ‘‘corridors’’ (of all 1-spels or all 0-
t 5 h((i, j), (i9, j9)) u [ui 2 i9u 1 u j 2 j9u 5 1] spels) outside D in (Z 2, b). We illustrate the process in
or [ui 2 i9u 5 3 and u j 2 j9u 5 0] Fig. 7. We leave it to the reader to check that the construc-
(22)
or [(i 2 i9) 3 ( j 2 j9) 5 1] tion has the following properties. If O9 is the b-component
or [ui 2 i9u 5 3 and ( j 2 j9) 5 (i9 2 i)/3]j. of the 1-spels containing c and Q9 is the b-component of
the 0-spels containing d, then O > D 5 O9 > D and Q >
This is a wrapped around version of the hexagonal tessella- D 5 Q9 > D. We can simulate the t-path from c to d not
tion of the plane (see Fig. 2): hexagons with coordinates crossing ­(O, Q) by a b-path not crossing ­(O9, Q9), by
(3, j) are proto-adjacent to those two hexagons with first following its moves inside D and using the corridors when-
coordinate 0 which are proto-adjacent to (21, j) in (Z 2, ever the t-path wraps around.
b). Consider now the binary picture (T, t, t), where Some of the results discussed in this section can be sum-

H
marized in the following single statement.
0, if i is even,
t(i, j) 5 (23) COROLLARY 2.14. There are strong Jordan graphs, there
1, if i is odd. are Jordan graphs which are not strong Jordan graphs, and
there are digital spaces which are not Jordan graphs.
There are four nodes in the tt-adjacency graph of this
binary picture; they are 3. MINIMALLY NEAR-JORDAN SURFACES

Ni 5 h(i, j) u j [ Z j, for 0 # i # 3. (24) Near-Jordan surfaces have been introduced and studied
in [1–3] since the concept captures some of the properties
There is an arc between nodes Ni and Nj if, and only if, of Jordan curves (simple closed curves). There is a property
the modulo 4 difference between the nodes is 61. Hence of Jordan curves which is not reflected by near-Jordan
N0 , N1 , N2 , N3 , N0 is a cycle in the tt-adjacency graph of surfaces; namely that they are minimal in the sense that a
(T, t, t) and so, by Corollary 2.7, (T, t) is not a strong proper subset of a Jordan curve is not a Jordan curve.
Jordan graph. Correspondingly, we define a minimally near-Jordan sur-
To show that it is a Jordan graph, we show that assuming face as a near-Jordan surface such that any surface which
that it is not a Jordan graph implies that (Z 2, b) is also is a proper subset of it is not near-Jordan.
not a Jordan graph, which we know to be false [2, Corollary We will see that minimally near-Jordan surfaces help us
5.13]. If (T, t) is not a Jordan graph, then there is a finite to give an alternative characterization of (strong) Jordan
binary picture (T, t, f ) with a t-component O of the 1- graphs. Before getting into that, we recall some relevant
spels and a t-component Q of the 0-spels such that there results and discuss some basic properties of minimally near-
is a t-path not crossing ­(O, Q) from a spel c [ II(­(O, Jordan surfaces.
JORDAN GRAPHS 355

FIG. 7. Illustration of the construction for showing that (T, t) is a Jordan graph. T is the shaded area, with D the more heavily shaded part of
it. The unbroken heavy lines (with orientation from 1-spels to 0-spels) form a tt-boundary in (T, t) which is extended by the construction into a
bb-boundary in (Z 2, b) as indicated by the broken heavy lines (with orientation from 1-spels to 0-spels).

THEOREM 3.1. Let k and l be spel-adjacencies in a digi- Recall that we said earlier that near-Jordanness of a
tal space (V, f). Let O be a nonempty k-connected proper surface does not guarantee the connectedness either of its
subset of V and let Q be one of the l-components of the interior or of its exterior. We see from Theorem 3.2 that
complement of O. Then ­(O, Q) is a kl-Jordan surface. this situation changes essentially when we replace the con-
cept of ‘‘near-Jordan’’ by that of ‘‘minimally near-Jordan.’’
Proof. This is Theorem 3.2 of [1]. n
As an immediate consequence, we have the following.
THEOREM 3.2. If S is a minimally near-Jordan surface
in a digital space (V, f), then both I(S) and E(S) are f-con- COROLLARY 3.3. If S is a minimally near-Jordan surface
nected. and k and l are spel-adjacencies in a digital space (V, f),
then S is kl-Jordan.
Proof. We prove the result only for E(S), the proof
for I(S) is strictly analogous. Since S is near-Jordan, it A further immediate consequence follows with the help
follows from Lemmas 1.1 and 1.2 that E(S) is the comple- of another previously proven result.
ment of I(S) and S 5 ­(I(S), E(S)). We can always find a LEMMA 3.4. If k and l are spel-adjacencies in a digital
spel-adjacency k in (V, f) such that I(S) is k-connected. space (V, f) and S is a kl-Jordan surface in that space,
(Recall that even V 2 is a spel-adjacency.) Also, since S is then there exists a binary picture (V, f, f ) in which S is the
a surface, neither I(S) nor E(S) is empty and so I(S) is a unique kl-boundary.
nonempty proper subset of V. We assume that E(S) is not
f-connected and show that this leads to a contradiction. Proof. This is Proposition 4.2 of [1]. n
Let Q9 and Q0 be two distinct (and hence disjoint) f-
COROLLARY 3.5. If k and l are spel-adjacencies in a
components of E(S). Then, by Theorem 3.1 both ­(I(S),
digital space (V, f) and S is a minimally near-Jordan surface
Q9) and ­(I(S), Q0) are kf-Jordan surfaces, which in partic-
in that space, then there exists a binary picture (V, f, f ) in
ular means that they are nonempty and near-Jordan. They
which S is the unique kl-boundary.
are also disjoint, implying that they are both near-Jordan
surfaces which are proper subsets of S, contradicting the We may ask whether the converse of either Corollary
minimal near-Jordanness of S. n 3.3 or Corollary 3.5 holds. This turns out not to be the
356 AHARONI, HERMAN, AND LOEBL

case in general, as is illustrated by the following example We complete this section with some alternative charac-
in (Z 2, g). In Fig. 8, there is only one g-component of terizations of strong Jordan graphs and of Jordan graphs,
the 1-spels and only one d-component of the 0-spels. The which we can now give due to our results on minimally
unique gd-boundary between them (composed of all sur- near-Jordan surfaces.
fels between a 1-spel; and a 0-spel) is gd-Jordan. However, THEOREM 3.9. Let (V, f) be a digital space. The follow-
it is not minimally near-Jordan, since the proper subset ing are equivalent:
indicated by the heavy lines in Fig. 8 also forms a near-
Jordan surface. Nevertheless, we have the following partial (i) (V, f) is a strong Jordan graph.
converse to these corollaries. (ii) The immediate interior of every ff-boundary is
f-connected.
LEMMA 3.6. In a digital space (V, f), every near-Jordan (iii) The immediate interior of every minimally near-
ff-boundary is minimally near-Jordan. Jordan surface is f-connected.
Proof. Let S be a near-Jordan ff-boundary. In order (iv) The immediate exterior of every ff-boundary is
to show that it is minimally near-Jordan, we have to demon- f-connected.
strate for any surface S9 which is a proper subset of S that (v) The immediate exterior of every minimally near-
it is not near-Jordan. Consider a surfel (c, d ) [ S 2 S9. Jordan surface is f-connected.
Let e be a spel in the immediate interior of S9 and f be a Proof. We prove only the equivalence of (i), (ii), and
spel in the immediate exterior of S9. We now show that (iii). The rest of the proof is clearly strictly analogous.
there is a proto-path from e to f which does not cross S9. To show that (i) implies (ii), we assume that (V, f) is a
This proto-path first goes from e to c. It can do this without strong Jordan graph. Let S be a ff-boundary. By Theorem
crossing S (and hence S9), since the interior of the ff- 1.3 both I(S) and E(S) are f-connected. Define a binary
boundary S is f-connected (see Theorem 1.3). It then picture (V, f, f ) so that the set of 1-spels is exactly the
moves from c to d (and we know that this step does not immediate interior of S. In this binary picture, E(S) is a
cross S9). Finally, it moves from d to f ; which again can f-connected set (as discussed above) and is in fact a f-
be done without crossing S in view of Theorem 1.3. n component of the 0-spels of (V, f, f ). This follows from
Lemmas 1.1 and 1.2, according to which the complement
THEOREM 3.7. Let S be a surface in a digital space (V, of E(S) is I(S), but any element of I(S) which is proto-
f). The following are equivalent: adjacent to an element of E(S) is in II(S) and is therefore
(i) S is a near-Jordan ff-boundary. a 1-spel in (V, f, f ). Also, clearly, if O is a f-component
(ii) S is minimally near-Jordan. of the 1-spels of (V, f, f ), then ­(O, E(S)) is not empty.
(iii) S is ff-Jordan. We suppose now that II(S) is not f-connected and show
that this leads to a contradiction. Let c and d be two spels in
Proof. That (i) implies (ii) is Lemma 3.6. That (ii) different f-components of II(S). Since I(S) is f-connected,
implies (iii) is a special case of Corollary 3.3. That (iii) there is a proto-path from c to d entirely in I(S). This path
implies (i) follows from Lemma 3.4. n must begin in one f-component J of II(S) (which is also
COROLLARY 3.8. Let S be a surface (respectively, a finite a f-component of the 1-spels of (V, f, f )) and must eventu-
ally arrive at another f-component of L of II(S) (which is
surface) in a strong Jordan graph (respectively, a Jordan
also a f-component of the 1-spels of (V, f, f )), having gone
graph) (V, f). The following are equivalent:
through a f-component K of the 0-spels of (V, f, f ) which
(i) S is a ff-boundary. is other than E(S). Thus, in the ff-adjacency graph of (V,
(ii) S is minimally near-Jordan. f, f ), there is a cycle E(S), J, K, . . . , L, E(S) which, by
(iii) S is ff-Jordan. Corollary 2.7 contradicts that (V, f) is a strong Jordan
graph.
That (ii) implies (iii) follows immediately from Theorem
3.7, according to which a minimally near-Jordan surface is
a ff-boundary.
To show that (iii) implies (i), we assume that the ff-
adjacency graph of some binary picture (V, f, f ) has cycle
Q, O, . . . , O9, Q (where Q is a f-component of the 0-spels
and O and O9 are two different f-components of the 1-
spels) and produce from this a minimally near-Jordan sur-
face whose immediate interior is not f-connected. Let
FIG. 8. Illustration that the converses of Corollary 3.3 and of Corol-
lary 3.5 do not hold. 2 5 hO u ­(O, Q) is not emptyj. (26)
JORDAN GRAPHS 357

We define a binary relation p on elements of 2 by O p have given an equivalent definition of when a digital space
O9 if, and only if, either O 5 O9 or there is a cycle Q, is a Jordan graph without ever referring to binary pictures.
O, . . . , O9, Q in the ff-adjacency graph of (V, f, f ). It is So, for example, the surface indicated by heavy lines in
easy to show that p is an equivalence relation. According Fig. 8 is by itself sufficient to tell us that (Z 2, g) is not a
to our assumption, there is at least one equivalence class Jordan graph (even if there were no 1’s and 0’s in that
which contains more than one element; let us refer to the figure). This is because the surface oriented from the inside
union of all the elements of this class as P. We next show to the outside of this rectangle is clearly a finite minimally
that ­(P, Q) is near-Jordan. Let c (0), . . . , c (m) be a proto- near-Jordan surface, but its immediate exterior is clearly
path from II(­(P, Q)) to IE(­(P, Q)). Let c (0) be in O and not g-connected. Hence condition (v) of Theorem 3.10 is
let c (k) be the last element in the proto-path which is not not satisfied, implying that (Z 2, g) is not a Jordan graph.
in Q. Then c (k) [ O9 for some O9 in 2. If O9 ? O, then There are approaches to digital topology which do not
the proto-path can be used to create a cycle Q, O, . . . , make use of surfaces as we defined them (i.e., as a set of
O9, Q in the ff-adjacency graph of (V, f, f ). This shows surfels), but rather use subsets of spels to play a similar
that O p O9 and so (c (k), c (k11)) [ ­(P, Q). Now we show role. (Such approaches are also quite natural when we take
that ­(P, Q) is minimally near-Jordan. Let S be any surface a more pure graph-theoretical point of view.) We complete
which is a proper subset of ­(P, Q) and suppose that (c, this paper with results corresponding to the last two theo-
d ) [ ­(P, Q) 2 S and (c9, d9) [ S. Since c9 [ O9 and rems phrased in terms of this alternative approach. For any
c [ O for some O and O9 such that O p O9, there is a digital space (V, f), for any spel c, and for any nonempty set
proto-path from c9 to c not crossing ­(P, Q) and, hence, D of spels not containing c, a set U of spels is called cD-
S. (This is obvious if O 5 O9, which is a f-component of separating if c is not in U and, for every spel d in D, every
the 1-spels. In the alternative case, we can make use of proto-path from c to d contains an element of U. A cD-
the ‘‘inside’’ of the cycle Q, O9, . . . , O, Q to get from separating set is called minimally cD-separating, if no
c9 to c without crossing ­(P, Q).) By our choice that proper subset of it is cD-separating. It turns out that this
(c, d ) [ ­(P, Q) 2 S, we can extend the proto-path to d new concept is not so different from those already intro-
without crossing S. Finally, we can extend it to d9 without duced, as can be seen from the following.
crossing ­(P, Q), since both d and d9 are in Q, which is a
f-component of the 0-spels. This shows that S is not near- THEOREM 3.11. Let (V, f) be a digital space and U be
Jordan. To complete the proof, let O and O9 be two distinct a subset of V. The following are equivalent:
elements of 2 which are subsets of P. Since II(­(P, Q)) (i) U is a minimally cD-separating set for some c and
contains at least one element from both O and O9, it cannot a f-connected set D.
be f-connected (otherwise O and O9 would be the same (ii) U is the immediate exterior of a minimally near-
f-component of the 1-spels). n Jordan surface.
THEOREM 3.10. Let (V, f) be a digital space such that
Proof. Assume first that U is a minimally cD-separat-
f is finitary. The following are equivalent:
ing set for some spel c and a f-connected set D. We define
(i) (V, f) is a Jordan graph. a binary picture (V, f, f ) so that a spel is given the value
(ii) The immediate interior of every finite ff-bound- 1 if, and only if, there is a f-path from c to it which does
ary is f-connected. not contain an element of U. Clearly, c itself is a 1-spel,
(iii) The immediate interior of every finite minimally while all elements of D are 0-spels. Also, the set O of 1-
near-Jordan surface is f-connected. spels is f-connected and all f-components of the 0-spels
(iv) The immediate exterior of every finite ff-bound- are f-components of the complement of O. Let Q be the
ary is f-connected. f-component of the 0-spels which contains D (since D is
(v) The immediate exterior of every finite minimally f-connected, there is such a f-component) and let S be the
near-Jordan surface is f-connected. ff-boundary ­(O, Q) (necessarily nonempty). By Theorem
3.1, S is ff-Jordan and hence, by Theorem 3.7, is minimally
Proof. The proof of Theorem 3.9 can be adopted with
near-Jordan. Clearly, the immediate exterior of S is a sub-
a very minor change: we need to use Corollary 2.13 instead
set of U. In fact, it is the whole of U, since it is a cD-
of Corollary 2.7. n
separating set and U is minimally cD-separating.
A significant aspect of these theorems is worth emphasiz- Now assume that U is the immediate exterior of mini-
ing: in both theorems, (i), (ii), and (iv) deal with concepts mally near-Jordan surface S. By Theorem 3.2, both I(S)
that were defined in terms of binary pictures; however, the and E(S) are f-connected. Let c be any element in I(S)
equivalent conditions (iii) and (v) do not in any way refer and let D 5 E(S). By Lemma 1.2 IE(S) is a cD-separating
to assignment of 0’s and 1’s to spels. They are just proper- set. Let d be any element of IE(S). Then d is also in D.
ties of certain types of surfaces. In other words, we could Let e be an element of I(S) which is f-adjacent to d. There
358 AHARONI, HERMAN, AND LOEBL

is a f-path from c to e entirely in I(S). It follows that any there is what they call a null-homotopic graph and they
subset of IE(S) which does not contain d would not be show (in their Theorem 2.2) that every (connected) null-
cD-separating. n homotopic graph is well connected. This can also be de-
rived as an immediate corollary of our more general Theo-
This combined with Theorems 3.9 and 3.10 immediately
rem 2.9.
yields the following new characterizations of strong Jordan
Finally, we give our promised discussion of the relation-
graphs and Jordan graphs. Note that these characteriza-
ship of our approach to parallel works nearer in spirit to
tions do not make use of the concept of a surface; they
classical topology [4]. A basic notion which might be
use only the concepts of f-connected and of cD-separating
thought of as topological in our approach is the notion of
sets of spels.
f-connectedness. A very natural question is: can we impose
COROLLARY 3.12. Let (V, f) be a digital space. The a topology on V such that a subset of V is f-connected if,
following are equivalent: and only if, it is connected (in the classical sense) in the
resulting topological space? The answer to this is some-
(i) (V, f) is a strong Jordan graph.
times yes; this has been shown for the digital space (Z 2,
(ii) For every spel c and for every f-connected non-
g) (and for its n-dimensional generalizations) in [12]. Un-
empty set D of spels not containing c, every minimally cD-
fortunately, for other interesting spaces the answer is no;
separating set is f-connected.
for example, the proof given in [13] can easily be adopted
COROLLARY 3.13. Let (V, f) be a digital space such that to show that no such topology can be found for the 1-
f is finitary. The following are equivalent: simply connected digital space (Z 2, b) of Fig. 2. Thus
connectedness in Jordan graphs cannot be made in this
(i) (V, f) is a Jordan graph.
fashion to be a special case of classical topological connect-
(ii) For every spel c and for every f-connected non-
edness. One approach that has been advocated to over-
empty set D of spels not containing c, every finite minimally
come this difficulty is to extend the set of the proposed
cD-separating set is f-connected.
topological space to include elements in addition to those
For readers interested in graph theory, we mention the of V. This is the idea underlying the use of cellular com-
consequence of these corollaries to ordinary graphs, with plexes as advocated by Kovalevsky (see, e.g., [5] and, for
a finite number of nodes. For such graphs there is no a critical discussion, [14]) and of Khalimsky spaces [6].
difference between Jordanness and strong Jordanness (see These certainly make good geometrical sense: when our
the discussions following Theorem 2.3 and Corollary 2.13). spels are interpreted as n-dimensional Voronoi neighbor-
Hence, we may conclude that such a graph is Jordan (equiv- hoods, then they are in fact n-cells and the unoriented
alently, strong Jordan) if, and only if, every set of nodes versions of the surfels can be interpreted as (n 2 1)-spels.
which is a minimal separator of a node and a connected We can also carry over the Khalimsky idea to our digital
set of nodes is connected. spaces: consider the set V < f and define the topology on
This result brings us around again to [11]. We now repro- it as that generated by the basis [4]
duce the basic definitions of that paper. A minimal relative
cutset in a (finite ordinary connected) graph G 5 G(V, f) hhcj u c [ V j < hhc, (c, d ), d j u (c, d ) [ fj. (27)
is a cutset C of G (i.e., a separating subset of V ) for which
there exist nodes c and d such that C is inclusion-minimal We certainly could have attempted to work with such a
with the property of separating c and d. G is said to be topology; for example, it is easy to see that if O and Q are
well connected if every minimal relative cutset of G is sets of spels which have no elements in common, then
connected. It does not immediately follow from this defini- A(­(O, Q)) 5 O < Q, where the overbar indicates closure
tion and the result stated at the end of the previous para- in the topological space. The choice for our more graph-
graph that being well connected is the same as being theoretical approach was made mainly based on our previ-
(strong) Jordan in case of ordinary graphs. That is because ous success in obtaining results in that framework [1–3];
there are minimally cD-separating sets which are not mini- the further results presented in here do somewhat validate
mal relative cutsets; e.g., any set of the form hd j which is this choice.
not a cutset is nevertheless a minimally chd j-separating
set. However, as we have already discussed, well connect- 4. DISCUSSION
edness is in fact the same as Jordanness for ordinary con-
nected graphs and so the definition of [11] can be consid- Early development of digital topology concentrated on
ered to be yet another characterization of a (strong) Jordan the space we call in this paper (Z 2, g). It became quickly
graph in this special case. Similarly, the notion of a 1- apparent that in order to avoid some ‘‘paradoxes’’ (unnatu-
simply connected digital space has also been considered ral behavior of the interior and exterior of boundaries;
for ordinary graphs in [11]; the corresponding notion in such as the failure of the Jordan curve theorem) one needs
JORDAN GRAPHS 359

to introduce adjacencies in addition to that provided by rected) graphs, by the acyclic nature of the adjacency
the proto-adjacency. The customary choice became to use a graphs of binary pictures, and also by the connectedness
pair of adjacencies; one for the 1-spels and another (usually of the immediate interiors (or exteriors) of certain types
different) for the 0-spels [7, 8]. While this approach can of surfaces. The surfaces of the last category include the
be treated in a mathematically rigorous fashion (our own so-called minimally near-Jordan surfaces, which have been
treatment of it can be found in [1–3] and is somewhat shown to be of some interest (e.g., they are kl-Jordan for
further developed in this paper; see Theorems 2.6 and any spel-adjacencies k and l). Finally, we have tied some
2.12), it nevertheless remains attractive to consider those of these notions to standard notions of graph theory, such
digital spaces in which the proto-adjacency alone can serve as the notion of a separating set.
to usefully define connectivity. This not only makes for Our overall conclusion is that those digital spaces which
more elegant mathematics of digital topology (and possibly are Jordan graphs have some very desirable properties
leads to more efficient computer algorithms when dealing which make their use advantageous (as compared to arbi-
with discrete geometrical objects), but it also brings the trary digital spaces in general) whenever the underlying
subject nearer to classical graph theory together with its situation allows us to use them.
enormous wealth of results. This is what motivated us to
search for the concept that we have defined as a ‘‘(strong) ACKNOWLEDGMENTS
Jordan graph.’’
The research of the first two authors was supported in part by NIH
In the body of the paper we have given only one specific
Grant HL28438 and NSF Grant IRI-9013341. The third author was sup-
example of an infinite strong Jordan graph (that which posed by Canada International Fellowship and CACR 2167, GAUK 261.
arises from the hexagonal tessellation of the plane) and
one example of a Jordan graph that is not a strong Jordan REFERENCES
graph. However, there are other interesting examples of
strong Jordan graphs: in fact, arbitrary dimensional gener- 1. G. T. Herman, Oriented surfaces in digital spaces. CVGIP: Graphical
alizations of the hexagonal tessellation of the plane (such Models Image Process. 55, 1993, 381–396.
as the hexahedral tessellations in three dimensions and the 2. G. T. Herman and E. Zhao, Jordan surfaces in simply connected
digital spaces. J. Math. Image Vision 6, 1996, 121–138.
octahedral tessellations in four dimensions) result in strong
3. G. T. Herman, Boundaries in digital spaces: Basic theory, in Topologi-
Jordan graphs [2].
cal Algorithms in Digital Image Processing (A. Rosenfeld and T. Y.
In addition to the new characterizations of (strong) Jor- Kong, Eds.), Elsevier, Amsterdam, 1996.
dan pairs in terms of acyclic adjacency graphs (Theorems 4. J. R. Munkres, Topology: A First Course, Prentice Hall, Englewood
2.6 and 2.12), the original work that we have presented in Cliffs, 1975.
this paper consists of the following. We have defined the 5. V. A. Kovalevsky, Finite topology as applied to image analysis, Com-
notion of a (strong) Jordan graph so that if (V, f) is a put. Vision Graphics Image Process. 46, 1989, 141–161.
strong Jordan graph (respectively, a Jordan graph), then 6. E. Khalimsky, Pattern analysis of n-dimensional digital images, in
(for arbitrary tight spel-adjacencies k and l and for an IEEE International Conference on Systems, Man, and Cybernetics,
Atlanta, GA, 1983, pp. 1559–1562.
arbitrary assignment of 1’s and 0’s to spels) a nonempty
7. T. Y. Kong and A. Rosenfeld, Digital topology: Introduction and
(respectively, nonempty and finite) boundary between a survey, Comput. Vision Graphics Image Process. 48, 1989, 357–393.
k-component of the 1-spels and a l-component of the 0- 8. T. Y. Kong, A. W. Roscoe, and A. Rosenfeld, Concepts of digital
spels will be a kl-Jordan surface, in the sense that its k- topology. Topology Appl. 46, 1992, 219–262.
connected interior (which contains the original k-compo- 9. F. Harary, Graph Theory, Addison–Wesley, Reading, MA, 1969.
nent of the 1-spels) and its l-connected exterior (which 10. M. Farber, Bridged graphs and geodesic convexity, Discrete Math.
contains the original l-component of the 0-spels) partition 66, 1987, 249–257.
V in such a way that every proto-path from the interior to 11. P. Duchet, M. Las Vergnas, and H. Meyniel, Connected subsets of
the exterior must cross the boundary. We have shown that a graph and triangle basis of the cycle space, Discrete Math. 62,
1986, 145–154.
every 1-simply connected digital space (and that so, in
12. D. Marcus, A special topology for integers, Am. Math. Monthly 77,
particular, every bridged graph) is a strong Jordan graph, 1970, 1119.
that there are Jordan graphs which are not strong Jordan 13. L. Latecki, Topological connectedness and 8-connectedness in digital
graphs, and that there are digital spaces which are not pictures, CVGIP: Image Understanding 57, 1993, 261–262.
Jordan graphs. (Strong) Jordan graphs have been charac- 14. G. T. Herman, On topology as applied to image analysis, Comput.
terized by the existence of ‘‘cuts’’ in associated (not di- Vision Graphics Image Process. 52, 1990, 409–415.

You might also like