The Jordan Brouwer Theorem for Graphs
Uploaded by
epidendrum2The Jordan Brouwer Theorem for Graphs
Uploaded by
epidendrum2THE JORDAN-BROUWER THEOREM FOR GRAPHS
OLIVER KNILL
Abstract. We prove a discrete Jordan-Brouwer-Schoenflies sep-
aration theorem telling that a (d − 1)-sphere H embedded in a
d-sphere G defines two different connected graphs A, B in G such
a way that A ∩ B = H and A ∪ B = G and such that the comple-
mentary graphs A, B are both d-balls. The graph theoretic defini-
tions are due to Evako: the unit sphere of a vertex x of a graph
G = (V, E) is the graph generated by {y |; (x, y) ∈ E}. Inductively,
a finite simple graph is called contractible if there is a vertex x such
that both its unit sphere S(x) as well as the graph generated by
V \ {x} are contractible. Inductively, still following Evako, a d-
sphere is a finite simple graph for which every unit sphere is a
(d − 1)-sphere and such that removing a single vertex renders the
graph contractible. A d-ball B is a contractible graph for which
each unit sphere S(x) is either a (d − 1)-sphere in which case x is
called an interior point, or S(x) is a (d − 1)-ball in which case x
is called a boundary point and such that the set δB of boundary
point vertices generates a (d − 1)-sphere. These inductive defini-
tions are based on the assumption that the empty graph is the
unique (−1)-sphere and that the one-point graph K1 is the unique
0-ball and that K1 is contractible. The theorem needs the follow-
ing notion of embedding: a sphere H is embedded in a graph G
if it is a subgraph of G and if any intersection with any finite set
of mutually neighboring unit spheres centered in H is a sphere. A
knot of co-dimension k in G is a (d − k)-sphere H embedded in a
d-sphere G.
1. Introduction
The Jordan-Brouwer separation theorem [21, 4] assures that the image
of an injective continuous map H → G from a (d − 1)-sphere H to a
d-sphere G divides G into two compact connected regions A, B such
that A ∪ B = G and A ∩ B = H. Under some regularity assumptions,
the Schoenflies theorem assures that A and B are d-balls. Hypersphere
Date: June 21, 2015.
1991 Mathematics Subject Classification. Primary: 05C15, 57M15 .
Key words and phrases. Topological graph theory, Knot theory, Sphere embed-
dings, Jordan, Brouwer, Schoenflies.
1
2 OLIVER KNILL
embeddings belong to knot theory, the theory of embedding spheres in
other spheres, and more generally to manifold embedding theory [9].
While H is compact and homeomorphic to the standard sphere in Rd ,
already a 1-dimensional Jordan curve H ⊂ R2 can be complicated, as
artwork in [41] or Osgood’s construction of a Jordan curve of positive
area [40] illustrate. The topology and regularity of the spheres as well
as the dimension assumptions matter: the result obviously does not
hold for surfaces G of positive genus. For codimension 2 knots H in
a 3-sphere G, the complement is connected but not simply connected.
Alexander [2] gave the first example of a topological embedding of S 2
into S 3 for which one domain A is simply connected while the other B
is not. With more regularity of H, the Mazur-Morse-Brown theorem
[36, 38, 5] assures that the complementary domains A, B are homeo-
morphic to Euclidean unit balls if the embedding of H is locally flat,
a case which holds if H is a smooth submanifold of G diffeomorphic
to a sphere. In the smooth case, all dimensions except d = 4 are set-
tled: one does not know whether there are smooth embeddings of S 3
into S 4 such that one of the domains is a 4-ball homeomorphic but not
diffeomorphic to the Euclidean unit ball. Related to this open Schoen-
flies problem is the open smooth Poincaré problem, which asks whether
there are is a smooth 4-sphere homeomorphic but not diffeomorphic to
the standard 4-sphere. If the smooth Poincaré conjecture turns out to
be true and no exotic smooth 4-spheres exist, then also the Schoenflies
conjecture would hold (a remark attributed in [6] to Friedman) as a
Schoenflies counter example with an exotic 4-ball would lead to an ex-
otic 4-sphere, a counter example to smooth Poincaré.
Even in the particular case of Jordan, various proof techniques are
known. Jordan’s proof in [21] was unjustly discredited at first [24] but
rehabilitated in [17]. The Schoenflies theme is introduced in [42, 43,
44, 45]. Brouwer [4] proves the higher dimensional theorem using p-
dimensional “nets” defined in Euclidean space. His argument is similar
to Jordan’s proof for d = 2 using an intersection number is what we will
follow here. The theorem was used by Veblen [48] to illustrate geometry
he developed while writing his thesis advised by Eliakim Moore. The
Jordan curve case d = 2 has become a test case for fully automated
proof verifications. Its deepness in the case d = 2 can be measured by
the fact that ”4000 instructions to the computer generate the proof of
the Jordan curve theorem” [16]. There are various proofs known of the
Jordan-Brouwer theorem: it has been reduced to the Brouwer fixed
point theorem [35], proven using nonstandard analysis [39] or dealt
THE JORDAN-BROUWER THEOREM FOR GRAPHS 3
with using tools from complex analysis [10]. Alexander [1] already
used tools from algebraic topology and studied the cohomology of the
complementary domains when dealing with embeddings of with finite
cellular chains. In some sense, we follow here Alexander’s take on
the theorem, but in the language of graph theory, language formed by
A.V. Evako in [19, 11] in the context of molecular spaces and digital
topology. It is also influenced by discrete Morse theory [13, 14].
When translating the theorem to the discrete, one has to specify what
a “sphere” and what an “embedding” of a sphere in an other sphere
is in graph theory. We also need notions of “intersection numbers”
of complementary spheres as well as workable notions of “homotopy
deformations” of spheres within an other sphere. Once the definitions
are in place, the proof can be done by induction with respect to the di-
mension d. Intersection numbers and the triviality of the fundamental
group allow to show that the two components in the (d−1)-dimensional
unit sphere of a vertex x ∈ H ⊂ G lifts to two components in the d-
dimensional case: to prove that there are two complementary compo-
nents one has to verify that the intersection number of a closed curve
with the (d − 1)-sphere H is even. This implies that if a curve from
a point in A to B in the smaller dimensional case with intersection
number 1 is complemented to become a closed curve in G, also the new
connection from A to B has an odd intersection number, preventing
the two regions to be the same. The discrete notions are close to the
“parity functions” used by Jordan explained in [17]. Having estab-
lished that the complement of H has exactly two components A, B, a
homotopy deformation of H to a simplex in A or to a simplex B will
establish the Schoenflies statement that A, B are d-balls. We will have
to describe the homotopy on the simplex level to regularizes things.
The Jordan-Schoenflies theme is here used as a test bed for definitions
in graph theory. Indeed, we rely on ideas from [31, 33].
Lets look at the Jordan case, an embedding of a circle in a 2-sphere.
A naive version of a discrete Jordan theorem is the statement that a
“simple closed curve in a discrete sphere divides the complement into
two regions”. As illustrated in Figure (1), this only holds with a grain
of salt. Take the octahedron graph G and a closed Hamiltonian path
which visits all vertices, leaving no complement. It is even possible
for any m ≥ 0 to construct a discrete sphere G and a curve H in G
such that the complement has m components as the curve can bubble
off regions by coming close to itself without intersecting itself but still
dividing up a disk from the rest of the sphere. As usual with failures
of discrete versions of continuum results, the culprit is the definition,
4 OLIVER KNILL
Figure 1. We see a homotopy deformation of an em-
bedded C4 graph H1 in an octahedron G to a Hamilton-
ian path H2 of length 6 in the 2-sphere G. The initial
curve H1 divides the octahedron into two complemen-
tary wheel graph domains A, B. After two steps, the
deformed H2 is no more an embedding as its closure is
G. The deformations are obtained by taking a trian-
gle t containing an edge of H and forming H → H∆t.
In the lower row, we see the situation on the simplex
level, where the octahedron G has become the Catalan
solid G1 and the deformed spheres remain sphere em-
beddings. While our theorem will be formulated for em-
bedded spheres in G, the proof of the Schoenflies case
requires the ability to perform homotopy deformations
and have a picture in which spheres remain embedded.
The complementary domains A, B in the lower case are 2-
balls: 9 interior vertices representing 4 triangles, 4 edges
and the vertex of each pyramid at first. After the de-
formation there are 7 interior vertices representing the
4 triangles and 3 edges of the domain. On the simplex
level the deformation steps are done using unit balls at
vertices belonging to original triangles.
THE JORDAN-BROUWER THEOREM FOR GRAPHS 5
in this case, it is the definition of an “embedding”. The example of a
Hamiltonian path is more like a discrete Peano curve in the continuum
as it visits all vertices but without hitting all directions or area forms of
the plane. We need to make sure that also the closure of the embedding
is the same curve. In the graph theoretical case, we ask that the graph
generated in G by the vertex set of H is still a sphere. For curves in
a 2-sphere for example we have to ask that the embedded curve fea-
tures no triples of vertices forming a triangle in G. There is an other
reinterpretation to make the theorem true for simple closed curves as
we will see in the proof: there is a regularized picture on the simplex
level, where two complementary domains always exist. We need the
simplex regularization because embeddings do not play well with de-
formations: When making homotopy deformation steps, we in general
lose the property of having an embedded sphere. Already in discrete
planar geometry, where we work in the flat 2-dimensional hexagonal
lattice [27] one has to invoke rather subtle definitions to get conditions
which make things work. Discrete topological properties very much
depend on the definitions used. It would be possible to build a ho-
motopy deformation process which honors the embeddability, but it
could be complicated. The construction of a graph product [33] pro-
vided us with an elegant resolution of the problem: we can watch the
deformation of an “enhanced embedding” H1 in G1 , where H1 is the
graph obtained from H by taking all the complete subgraphs of H as
vertices and connecting two of them if one is a subgraph of the other.
It turns out that even if H is only a subgraph of G, the graph H1 is
an embedding of G1 . This holds in particular if H is a Hamiltonian
path in G, the closure of H1 , the graph generated by the vertex set of
H1 remains geometric in G1 . The Cartesian product [33] allows also to
look at homotopy groups geometrically: a deformation of a curve in a
2-sphere G for example is now described as a geometric surface in the
3-dimensional solid cylinder G × Ln , where Ln is the 1-dimensional line
graph with n vertices and (n − 1) edges. This is now close to the defi-
nition of homotopy of a curve using a function F (t, s) in two variables
so that F (t, 0) is the first curve and F (t, 1) the second.
This paper is not the first take on a discrete Jordan theorem as var-
ious translations of the Jordan theorem have been constructed to the
discrete. They are all different from what we do here: [8] uses notions
of discrete geometry, [18] looks at Jordan surfaces in the geometry of
digital spaces, [12] proves a Jordan-Brouwer result in the discrete lat-
tice Z d , [47] extends a result of Steinhaus on a m × n checkerboard
G, a minimal king-path H connecting two not rook-adjacent elements
6 OLIVER KNILL
of the boundary divides G into two components. A variant of this on
a hexagonal board [15] uses such a result to prove the Brouwer fixed
point theorem in two dimensions. [49] deals more generally with graphs
which can have multiple connections. The result essentially establishes
what we do in the special case d = 2. [46] looks at two theorems in
L × L or L ⋆ L, where × is the usual Cartesian product and ⋆ the tight
product. In both cases, the complement of C has exactly two path
components. The paper [34] deals with graphs, for which every unit
sphere S(x) is a Hamiltonian graph. Also here, a closed simple path
C in a connected planar graph G divides the complement into exactly
two components. [49] deals with graphs which can have multiple con-
nections. On 2-spheres, it resembles the Jordan case d = 2 covered
here as a special case.
The main theorems given here could readily be derived from the con-
tinuum by building a smooth manifold from a graph, and then use the
Jordan-Brouwer-Schoenflies rsp. Mazur-Morse-Brown theorem. The
approach however is different as no continuum is involved: all defi-
nitions and steps are combinatorial and self-contained and could be
accepted by a mathematician avoiding axioms invoking infinity. Some-
times, constructabily can be a goal [3]. New is that we can prove
Jordan-Brouwer-Schoenflies entirely within graph theory using a gen-
eral inductive graph theoretical notion of “sphere” [11]. Papers of Jor-
dan, Brouwer or Alexander show that the proofs in the continuum often
deal with a combinatorial part only and then use an approximation ar-
gument to get the general case. As the Alexander horned sphere, the
open Schoenfliess conjecture or questions about triangulations related
to the Hauptvermutung show, the approximation part can be difficult
within topology and we don’t go into it. Our proof remains discrete
but essentially follows the arguments from the continuum by defin-
ing intersection numbers and use induction with respect to dimension.
The induction proof is possible because of the recursive definition of
spheres and seems not have been used in the continuum, nor in discrete
geometry or graph theory. But what really makes the theorem go, is to
watch the story on the simplex level, where geometric graphs G remain
geometric and where sub-spheres H of G can be watched as embedded
spheres H1 in G1 . The step G → G1 can be used in the theory of trian-
gulations because it has a “regularizing effect”. Given a triangulation
G described as a graph, the new triangulation G1 has nicer properties
like that the unit sphere of a point is now a graph theoretically defined
sphere.
THE JORDAN-BROUWER THEOREM FOR GRAPHS 7
2. Definitions
Definition 1. A subset W of the vertex set V = V (G) of a graph G
generates a subgraph (W, F ) of G, where F is defined as the subset
{(a, b) ∈ E | a ∈ W, b ∈ W } of the edge set of E. The unit sphere
S(x) of a vertex x in a graph is the subgraph of G generated by all
neighboring vertices of x. The unit ball B(x) is the subgraph of G
generated by the union of {x} and the vertex set of the unit sphere
S(x).
If H is a subgraph of G, one can think of the graph generated by H
within G as a “closure” of H within G. It is in general larger than H.
For example, the closure of the line graph H = ((a, b, c), (a, b), b, c))
within the complete graph G with vertex set V = {a, b, c } is equal to
G.
Definition 2. Starting with the assumption that the one point graph
K1 is contractible, recursively define a finite simple graph G = (G, V )
to be contractible if it contains a vertex x such that its unit sphere
S(x) as well as the graph generated by V \ {x} are both contractible.
Definition 3. A complete subgraph Kk+1 of G will also be denoted
k-dimensional simplex. If vk (G) is the set of k-dimensional sim-
plices in G, then the Euler characteristic of G is defined as χ(G) =
k
P
k=0 (−1) vk (G).
Examples.
1) The Euler characteristic of a contractible graph is always 1 as re-
moving one vertex does not change it. One can use that χ(A ∩ B) =
χ(A) + χ(B) − χ(A ∩ B) and use inductively the assumption that unit
balls as well as the spheres S(x) in a reduction are both contractible.
2) Also by induction, using that a unit sphere of a d-sphere is (d − 1)-
sphere, one verified that χ(G) = 1 + (−1)d for a d-sphere. This holds
also in the case d = −1 as the Euler characteristic of the empty
graph is 0. The Euler characteristic of an octahedron for example
is 6 − 12 + 8 = 2 as there are 6 vertices, 12 edges and 8 triangles. The
cube graph G is not a sphere as the unit sphere at each vertex is P3 . Its
Euler characteristic is 8 − 12 = −4. G is a sphere with 6 holes punched
in, leaving only a 1-dimensional skeleton. A 2-dimensional cube can be
constructed as the boundary δB of the solid cube B = L2 × L2 × L2
defined in [33].
Definition 4. Removing a vertex x from G for which S(x) is con-
tractible is called a homotopy reduction step. The inverse opera-
tion of performing a suspension over a contractible subgraph H of G by
8 OLIVER KNILL
adding a new vertex x and connecting x to all the vertices in H is called
a homotopy extension step. A finite composition of reduction or
extension steps is called a homotopy deformation of the graph G.
Remarks.
1) Since simple homotopy steps removing or adding vertices with con-
tractible S(x) do not change the Euler characteristic, it is a function
on the homotopy classes [19]. If we add a vertex for which S(x) is
not contractible, we add a vertex with index 1 − χ(S(x)) which is a
Poincaré-Hopf index [28]. Given a function on the vertex set giving
an ordering on the build up of the graph one gets the Poincaré-Hopf
theorem.
2) Examples like the Bing house or the Dunce hat show that homo-
topic to a one-point graph K1 is not equivalent to contractible: some
graphs might have to be expanded first before being contractible. This
is relevant in Lusternik-Schnirelmann category [22].
3) The discrete notion of homotopy builds an equivalence relation on
graphs in the same way homotopy does in the continuum. The problem
of classifying homotopy types can not be refined as one can ask how
many types there are on graphs with n vertices.
4) The discrete deformation steps were put forward by Whitehead [50]
in the context of cellular complexes. The graph version is due to [19]
and was simplified in [7].
5) The definition of d-spheres and d-balls is inductive. Introduced in
[30, 32] we were puzzled then why this natural setup has not appeared
before. But it actually has, in the context of digital topology [12] going
back to [11] and we should call such spheres Evako spheres. Alexander
Evako is a name shortcut for Alexander Ivashchenko who also intro-
duced homotopy to graph theory and also as I only learned now while
reviewing his work found in [20] a similar higher dimensional Gauss-
Bonnet-Chern theorem [26] in graph theory.
Definition 5. The induction starts with the assumption that the empty
graph is the only (−1)-sphere and that the graph K1 is the only 0-ball.
A graph is called a d-sphere, if all its unit spheres S(x) are (d − 1)
spheres and if there exists a vertex x ∈ V such that the graph generated
by V \ {x} is contractible. A contractible graph G is called a d-ball, if
one can partition its vertex set V into two sets int(V ) = {x ∈ V | S(x)
is a (d − 1) sphere } and δ(V ) = {x ∈ V | S(x) is a (d − 1)-ball } such
that δV generates a (d − 1)-sphere called δG, the boundary of G.
By induction, χ(G) = 1 + (−1)d if G is a d-sphere and χ(G) = 1 if G
is a d-ball.
THE JORDAN-BROUWER THEOREM FOR GRAPHS 9
Examples.
1) The boundary sphere of the 0-ball K1 is the (−1) sphere ∅, the
empty graph.
2) The boundary sphere of the line graph Ln with n > 2 vertices, is
the 0-sphere P2 . Line graphs are 1-balls.
3) The boundary sphere of the wheel graph Wn with n ≥ 4 is the
circular graph Cn . The wheel graph is an example of a 2-ball and Cn
is an example a 1-sphere.
4) The boundary sphere of the 3-ball obtained by making a suspension
of a point with the octahedron is the octahedron itself.
5) We defined in [32] Platonic spheres as d-spheres for which all unit
spheres are Platonic (d−1)-spheres. The discrete Gauss-Bonnet-Chern
theorem [26] easily allows a classification: all 1-dimensional spheres
Cn , n > 3 are Platonic for d = 1, the Octahedron and Icosahedron
are the two Platonic 2-spheres, the sixteen and six-hundred cells are
the Platonic 3-spheres. As we only now realize while looking over the
work of Evako, we noticed that the Gauss-Bonnet theorem [26] appears
in [20]. The d-cross polytop P2 ⋆ P2 ⋆ · · · ⋆ P2 obtained by repeating
suspension operations from the 0-sphere P2 is the unique Platonic d-
sphere for d > 3.
Definition 6. P The dimension of a graph is inductively defined as
dim(G) = 1 + x∈V dim(S(x))/v0 , where v0 = |V | is the cardinality
of the vertex set. The induction foundation is that the empty graph ∅
has dimension 0. The dimension of a finite simple graph is a rational
number.
Remarks.
1) This inductive dimension for graphs has appeared first in [27, 25]. It
is motivated by the Menger-Uryson dimension in the continuum but it
is different because with respect to the metric on a graph, the Menger-
Uryson dimension is 0.
2) Much of graph theory literature ignores the Whitney simplex struc-
ture and treat graphs as one dimensional simplicial complexes. The
inductive dimension behaves very much like the Hausdorff dimension
in the continuum, the product [33] is super additive dim(H × K) ≥
dim(H) + dim(K) like Hausdorff dimension of sets in Euclidean space.
3) There are related notions of dimension like [37], who look at the
largest dimension of a complete graph and then extend the dimension
using the usual Cartesian product. This is not equivalent to the di-
mension given above.
10 OLIVER KNILL
Examples.
1) The complete graph Kn has dimension n − 1. 2) The dimension of
the house graph obtained by gluing C4 to K3 along an edge is 22/15:
there are two unit spheres of dimension 0 which are the base points,
two unit spheres of dimension 2/3 corresponding to the two lower roof
points and one unit sphere of dimension 1 which is the tip of the roof.
3) The expectation dn (p) of dimension on Erdoes-Renyi probability
spaces G(n, p) of all subgraphs in Kn for which edges are turned on with
probability p can be computedP explicitly. It is an explicit polynomial
in p given by dn+1 (p) = 1 + nk=0 nk pk (1 − p)n−k dk (p) [25].
Definition 7. A finite simple graph G = (V, E) is called a geometric
graph of dimension d if every unit sphere S(x) is a (d − 1)-sphere. A
finite simple graph G is a geometric graph with boundary if every
unit sphere S(x) is either a (d − 1) sphere or a (d − 1)-ball. The subset
of the vertex set V , in which S(x) is a ball generates the boundary
graph of G. We denote it by δG and assume it to be geometric of
dimension (d − 1).
Examples.
1) By definition, d-balls are geometric graphs of dimension d and d-
spheres are geometric graphs of dimension d.
2) For every smooth d-manifold one can look at triangulations which
are geometric d-graphs. The class of triangulations is much larger.
Remarks.
1) Geometric graphs play the role of manifolds. By embedding each
discrete unit ball B(x) in an Euclidean space and patching these charts
together one can from every geometric graph G generate a smooth com-
pact manifold M . Similarly, if G is a geometric graph with boundary,
one can “fill it up” to generate from it a compact manifold with bound-
ary.
2) The just mentioned obvious functor from geometric graphs to man-
ifolds is analogue to the construction of manifolds from simplicial com-
plexes. We don’t want to use this functor for proofs and remain in the
category of graphs. One reason is that many computer algebra systems
have the category of graphs built in as a fundamental data structure.
An other reason is that we want to explore notions in graph theory and
stay combinatorial.
3) Graph theory avoids also the rather difficult notion of triangular-
ization. Many triangularizations are not geometric. In topology, one
would for example consider the tetrahedron graph K4 as a triangula-
tion of the 2-sphere. But K4 is not a sphere because unit spheres are
THE JORDAN-BROUWER THEOREM FOR GRAPHS 11
K3 which are not spheres etc. And K4 is also not a ball. While it is
contractible, it coincides with its boundary as it does not have interior
points. Free after Euclid one could say that Kd+1 is a d-dimensional
point, as it has no d-dimensional parts.
4) Every graph defines a simplicial complex, which is sometimes called
the Whitney complex, but graphs are a different category than simpli-
cial complexes. Algebraically, x + y + z + xy + yz + xz is a simplicial
complex which is not a graph as it does not contain the triangle simplex.
The graph completion K3 described by x + y + z + xy + yz + zx + xyz
however, is a graph.
Definition 8. A geometric graph of dimension d is called orientable
if one can assign a permutation to each of its d-dimensional simplices
in such a way that one has compatibility of the induced permutations
on the intersections of neighboring simplices. For an orientable graph,
there is a constant non-zero d-form f , called volume form. It satisfies
df = 0 for the exterior derivative d but which can not be written as dh.
Remarks.
1) A connected orientable d-dimensional geometric graph has a 1-
dimensional cohomology group H d (G). This is a special case of Poincaré
duality, assuring an isomorphism of H n−d (G) with H n (G), which holds
for all geometric graphs.
2) For geometric graphs, an orientation induces an orientation on
theR boundary.
R Stokes theorem for geometric graphs with boundary
is G df = δG f [29], as it is the definition on each simplex.
Examples.
1) All d-spheres with d ≥ 1 are examples of orientable graphs.
2) If a d-sphere G has the property that antipodal points have dis-
tance at least 4, then the antipodal identification map T factors out a
geometric graph G/T , we get a discrete projective space P d . For even
dimensions d, this geometric graph is not orientable.
3) The cylinder Cn × Lm , with n ≥ 4, m ≥ 2 is orientable. One can get
a sphere, a projective plane, a Klein bottle or a torus from identifica-
tions of the boundary of Ln × Lm in the same way as in the continuum.
For example, the graph L3 ×L3 is obtained by taking the 25 polynomial
monoid entries of (a + ab + b + bc + c)(u + uv + v + vw + w) as vertices
and connecting two if one divides the other.
The following definition of the graph product has been given in [33]:
Definition 9. A graph G = (V, E) withP vertex set V = {x1 , . . . , xn }
k1 kn
defines a polynomial fG (x1 , . . . , xn ) = x x, where x = x1 · · · xn
12 OLIVER KNILL
with ki ∈ {0, 1} represents a complete subgraph of G. The polynomial
defines the graph G1 with vertex set V = {x | simplex } and edge set
E = {(x, y) | x|y or y|x}, where x|y means x divides y, geometrically
meaning that x is a sub-simplex of y. Given two graphs H, K, define
its graph product H × K = G(fH · fK ). The graph G1 = G × K1 is
called the enhanced graph obtained from G.
Examples.
1) For G = C4 we have fG = x + xy + y + yz + z + zw + w + wx and
G1 = C 8 .
2) For G = K3 , we have fG = x + y + z + xy + yz + zx + xyz and
G1 = W6 .
3) For a graph without triangles, G1 is homeomorphic to G in the clas-
sical sense.
Remarks.
1) The graph G1 has as vertices the complete subgraphs of G. Two
simplices are connected if and only one is contained in the other. If G
is geometric, then G1 is geometric. For example, if G is the octahedron
graph with v0 = 6 vertices, v1 = 12 edges and v2 = 8 triangles, then
G1 is the graph belonging to the Catalan solid with v0 + v1 + v2 = 26
vertices and which has triangular faces. Also G1 is a 2-sphere.
2) In full generality, the graph G1 is homotopic to G and has therefore
the same cohomology. The unit balls of G1 form a weak Čech cover
in the sense that the nerve graph of the cover is the old graph G and
two elements in the cover are linked, if their intersection is a d − 1
dimensional graph. To get from G1 to G, successively shrink each unit
ball of original vertices analogue to a Vietoris-Begle theorem. If G is
geometric, it is possible to modify the cover to have it homeomorphic
in the sense of [31] so that if G is geometric then G and G1 are homeo-
morphic. As the dimension of G1 can be slightly larger in general, the
property that G1 and G are homeomorphic for all general finite simple
graphs does not hold.
Examples.
1) A Hamiltonian path H in the icosahedron G, a 2-sphere, does not
leave any room for complementary domains. However, the graph H1
in G1 divides G1 into two regions. The graph G1 by the way is the
disdyakis triacontahedron, a Catalan solid with 62 vertices.
2) Let G be the octahedron, a 2-sphere with 6 vertices. Assume a, b, c, d
are the vertices of the equator sphere C4 and that n, p are the north and
south pole. Define the finite simple curve a, b, c, d, s, a of length 5. It is
THE JORDAN-BROUWER THEOREM FOR GRAPHS 13
a circle H in G, but it is not embedded. In this case, the complement
of H has only one region A = {n}. The Jordan-Brouwer theorem is
false. But its only in this picture. If we look at the embedding of H1
in G1 , then this is an embedding which divides the Catalan solid G1
into two regions A1 , B1 .
Definition 10. A graph H = (W, F ) is a subgraph of G = (V, E) if
W ⊂ W and F ⊂ E. Let H, G be geometric graphs. A graph H is
embedded in an other graph G if H is a subgraph such that for any
collection of unit spheres S(xj ) in G, where xj is in the vertex set of
some Kk , the intersection H ∩ kj=1 S(xj ) is a sphere.
T
Examples.
1) A 0-sphere H is embedded in a geometric graph if the two vertices
are not adjacent. A 1-sphere H is embedded if two vertices of H are
connected in G if and only if they are connected in H.
2) A graph Ck can only be immersed naturally in Cn if n divides k
and n ≥ 4. It is a curve winding k/n times around Cn . For example,
C15 can be immersed in C5 and described by the homomorphism ale-
braically described by x ∈ Z15 → x ∈ Z5 if Cn is identified with the
additive group Zn . In this algebraic setting dealing with the funda-
mental group it is better to look at the graph homomorphism rather
than the physical image of the homomorphism.
3) If H is embedded in G, then also H1 is an embedding of G1 . But H1
is an embedding in G1 even if H is only a subgraph of G. See Figure (1).
The following definition places the sphere embedding problem into the
larger context of knot theory:
Definition 11. A knot of co-dimension k is an embedding of a
(d − k)-sphere H in a d-sphere G.
Remarks.
1) A knot can be called trivial if it is homeomorphic to the (d − k)-
cross polytop embedded in the d-cross polytop in the sense of [31].
2) As we don’t yet know whether there are graphs homeomorphic to
spheres which are not spheres or whether there are graphs homeomor-
phic to balls which are not balls.
3) The Jordan-Brouwer-Schoenflies theorem can not be stated in the
form that a (d − 1)-sphere in a d sphere is trivial.
Definition 12. A closed curve in a graph is a sequence of vertices
xj with (xj , xj+1 ) ∈ E and xn = x0 . A simple curve in the graph is
the image of an injective graph homomorphism Ln → G, where Ln is
14 OLIVER KNILL
the line graph. A simple closed curve is the image of an injective
homomorphism Cn → G with n ≥ 3. It is an embedding of the circle
if the image generates a circle. In general, a simple closed curve is not
an embedding of a circular graph.
Example.
1) A Hamiltonian path is a simple closed curve in a graph G which
visits all vertices exactly once. Such a path is not an embedding if G
has dimension larger than 1 as illustrated in Figure (1).
2) While we mainly deal with geometric graphs, graphs for which all
unit spheres are spheres, the notion of a simple closed curve or an
embedding can be generalized for any pair of finite simple graphs H, G:
if H is a subgraph of G, then there is an injective graph homomorphism
from H to G. If the intersection of an intersection of finitely many
neighboring unit spheres with H is a sphere, we speak of an embedding.
Definition 13. An embedding of a graph H in G separates G into
two graphs A, B if A ∩ B = H, A ∪ B = G and A \ H, B \ H are disjoint
nonempty graphs. The two graphs A, B are called complementary
subgraphs of the embedding H in G.
Examples.
1) The empty graph separates any two connectivity components of a
graph.
2) By definition, if a graph G is k-connected but (k + 1)-disconnected,
there is a graph H consisting of k vertices such that H separates G.
3) For G = Kn , there is no subgraph H which separates G.
4) The join of the 1-sphere C4 with the 0-sphere P2 is a 2-sphere, the
disdyakis dodecahedron, a Catalan solid which by the way is G × K1 ,
where G is the octahedron. See Figure (1).
The C4 subgraph embedded as the equator in the octahedron G sepa-
rates G into two wheel graphs A, B.
Remarks.
1) A knot H of co-dimension 2 in a 3-sphere G is a closed simple curve
embedded in G. Classical knots in R3 can be realized in graph theory
as knots in d-spheres, so that the later embedding is the same as in
the discrete version. The combinatorial problem is not quite equiva-
lent however as it allows refined questions like how many different knot
types there are in a given 3-sphere G and how many topological invari-
ants are needed in a given 3-sphere to characterize any homotopy type
or a knot of a given co-dimension. .
2) An embedded curve has some “smoothness”. In a 2-sphere for ex-
ample, it intersects every triangle in maximally 2 edges. An extreme
THE JORDAN-BROUWER THEOREM FOR GRAPHS 15
case is a Hamiltonian graph H inside G which by definition generates
the entire graph. A Hamiltonian path H which is a subgraph of a
higher dimensional graph plays the role of a space-filling Peano curve
in the continuum, a continuous surjective map from [0, 1] to the 2-
manifold M . For a simple curve H which is not an embedding in G,
the complement of H can therefore be empty.
Definition 14. A simple homotopy deformation of a (d−1)-sphere
H in a d-sphere G is obtained by taking a d-simplex x in G which
contains a non-empty set Y of (d − 1)-simplices of H and replacing
these simplices with Y ′ , the set of (d − 1)-simplices in x which are in
the complement Y .
Examples.
1) If H is a simple curve in a 2-sphere G and x is a triangle containing
a single edge e of H, replace e with the two other edges of the trian-
gle. This stretches the curve a bit. The reverse operation produces a
“shortcut” between two vertices (a, b) visited by the curve initially as
(a, c, b).
2) If H is a 2-dimensional graph, then a homotopy step is done by
replacing a triangle in the tetrahedron with the 3 other triangles of a
tetrahedron.
Remarks.
1) It is allowed to replace an entire d-simplex with the empty graph to
allow a smaller dimensional sphere to be deformed to the empty graph.
This is not different in the continuum, where we deform curves to a
point.
2) A homotopy step does not honor embeddings in general. However,
it preserves the class of simple curves with the empty curve included.
3) We have in the past included a second homotopy deformation which
removes or adds backtracking parts (a, b, a). Since is only needed if
one looks at homotopy deformations of general curves, we don’t use it.
Homotopy groups must be dealt with using graph morphisms, rather
than graphs. The backtracking deformation steps would throw us from
the class of simple curves.
Definition 15. We say that a k-sphere H is trivial in a d-sphere G
if there is a sequence of simple homotopy deformations of H which
deforms H to the empty graph. If every 1-sphere is trivial in G, then
G is called simply connected. If every k-sphere is trivial in G, we
say the k’th homotopy class is trivial.
16 OLIVER KNILL
Remarks.
1) The set of simple closed curves is not a group, as adding a curve
to itself would cross the same point twice. Similarly, the set of simple
k-spheres is not a group. In order to define the fundamental group, one
has to look at graph homomorphism and not at the images. This is
completely analogue to the continuum, where one looks at continuous
maps from T 1 to G.
2) Unlike in the continuum, where the zero’th homotopy set π0 (G) is
usually not provided with a group structure, but π0 (G) can has a group
structure. It is defined as the commutative group of subsets of V with
the symmetric difference ∆ as addition, modulo the subgroup generated
by sets {{a, b} | (a, b) ∈ E}. It is of course Z0b0 where b0 = dim(H 0 (G))
is the number of connectivity components.
3) The Hurewicz homomorphism π0 (G) → H 0 (G) maps a subset A of
V to a locally constant function obtained by applying the heat flow
e−L0 t on the characteristic function 1A (x) which is 1 on A and 0 else,
playing the role of a 0-current = generalized function in the contin-
uum.
4) Also the Hurewicz homomorphism π1 (G) → H 1 (G) is explicit by
applying the heat flow e−L1 t on the function on edges telling how many
times the curve has passed in a positive way through the edge.
Examples.
1) Every simple curve in a 2-sphere is trivial if it can be deformed to
the empty graph. This general fact for d-spheres is easy to prove in the
discrete setup because by definition, a d-sphere becomes contractible
after removing one vertex. The contraction of this punctured sphere to
a point allows a rather explicit deformation of the curve to the empty
graph.
2) The deformation works also for 0-spheres. In a connected graph,
any embedded 0 sphere can be homotopically deformed to the empty
graph. So, a graph is connected if and only if every 0-sphere in G is
trivial.
Definition 16. Fix a geometric d-dimensional graph G. Let πk (G)
denote the union of all graph homomorphisms from a graph in the set
{Ck , k ≥ 3} to G. Any such homomorphism φ : H → G defines the
homomorphism graph, for which the vertices are the union of the
vertices of H and G and for which the edge set is the union of the
edges in C and G together with all pairs (a, φ(a)). Two such homo-
morphisms are called homotopic, if the corresponding homomorphism
THE JORDAN-BROUWER THEOREM FOR GRAPHS 17
graphs are homotopic. The homotopy classes π1 (G) define the fun-
damental group of G. The 0-element in the group is the homotopy
class of a map from the empty graph to G. The addition of two maps
Ck → G, Cl → G is a map from Ck+l → G obtained in the usual way by
first deforming each map so that φi (0) = x0 is a fixed vertex x0 , then
define φ(t) = φ1 (t) for t ≤ k and then φ(t) = φ2 (t−k) for k ≤ t ≤ k +l.
Remarks.
1) If one would realize the graph in an Euclidean space and see it as
a triangularization of a manifold, then the fundamental groups of G
and M were the same. The groups work also in higher dimensions.
As we have to cut up a sphere at the equator to build the addition in
the higher homotopy groups, it would actually be better to define the
addition in the enhanced picture and look at maps H1 → G1 where H1
is the enhanced graph of the k-sphere H and G1 the enhanced version
of the graph G. A deformation of a graph H to a graph K is then
geometrically traced as a surface.
2) As a single basic homotopy extension step Ck → G to Ck+1 → G
keeps the map in the same group element of π1 (G), the verification
that the group operation is well defined is immediate.
There are various generalized notions of “geometric graphs”, mirror-
ing the definitions from the continuum. We mention them in the next
definition, as we still explore discrete versions of questions related to
Schoenflies problem in the continuum. The main question is whether
there are discrete versions of exotic spheres, spheres which are homeo-
morphic to a d-sphere but for which unit spheres are not spheres.
Definition 17. A homology d-sphere is a geometric graph of dimen-
sion d which has the same homology than a d-sphere. It is a geometric
graph of dimension d with Poincaré polynomial pG (x) = 1 + xd . A ho-
mology graph of dimension d is a graph for which every unit sphere
is a homology sphere. A pseudo geometric graph of dimension d is
a graph for which every unit sphere is a finite union of (d − 1) spheres.
A discrete d-variety is defined inductively as a graph for which every
unit sphere is a (d − 1)-variety with the induction assumption that a
(−1)-variety is the empty graph.
Examples.
1) An example of a homology sphere can be obtained by triangulating
the dodecahedron and doing identifications as in the continuum. A
suspension of a homology sphere is an example of a homology graph.
2) A figure eight graph is an example of a pseudo geometric graph of
dimension 1.
18 OLIVER KNILL
3) The cube graph or dodecahedron graph are examples of discrete
1-varieties; their unit spheres are the 0-dimensional graphs P3 which
are not 0-spheres but 0-varieties.
Definition 18. Two k-spheres H, K in a d-sphere G are called geo-
metric homotopic within G if there is a geometric (k+1)-dimensional
graph with boundary M in the (d + 1)-dimensional graph G × Ln such
that M ∩ (G × {0}) = H × {0} and M ∩ (G × {n}) = K × {0} and
such that the boundary of M is included in the boundary of G × Ln .
Remarks.
1) Given a (d − 1)-sphere H embedded in a d-sphere G. The deforma-
tion H → H ′ = H∆S(x) is equivalent to a homotopy deformation of
the complement.
2) The above definition can can also be done for more general k-spheres
(where k is not necessarily d−1) by taking intersections of unit spheres
with (k + 1)-sphere and performing the deformation within such a
sphere.
3) Any homotopy deformation of H within G defines a deformation
of the embedding of H1 in G1 and can be seen as a geometric ho-
motopy deformation, a surface in G × L2 . We will explore this else-
where. Let f be the polynomial in the variables x1 , . . . , xn represent-
ing vertices in G. The function f (x1 , . . . , xn )(a + ab + b) describes the
(d + 1)-dimensional space G × L2 in which we want to build a sur-
face. Let g(y1 , . . . , ym ) be the function describing the surface H in
G, where yi are the vertices in H. Make the deformation at x: De-
fine ag(x1 , . . . , xn ) + ab[go (S(x))x] + b[x + ge (S(x)) + g(x1 , . . . , xn )].
For example if fH = a + c + ac, fK = a + b + c + ab + bc, f =
u(a + b + c + ab + bc) + uv(abc + a + c) + v(a + c + ac).
4) In [31] we wondered what the role of 1 in the ring describing graphs
could be. It could enter in reduced cohomology which is used in the
Alexander duality theorem bk (H) = bd−k−1 (G−H) (going back to [23]).
Here bk = bk if k > 0 and bk = 1 + bk if k = 0. In the Jordan case
for example, where b(H) = (1, 1, 0) and b(G − H) = (2, 0, 0) one gets
b(H) = (0, 1) and b(G − H) = (1, 0). When embedding a 2 sphere H
in a 3-sphere G, then b(H) = (1, 0, 1) and b(H) = (0, 0, 1) as well as
b(G − H) = (2, 0, 0) and b(G − H) = (1, 0, 0). This works for any d as
b(H) = (1, 0, . . . , 0, 1) and b(G − H) = (2, 0, . . . , 0) by Schoenflies.
THE JORDAN-BROUWER THEOREM FOR GRAPHS 19
3. Tools
In this section, we put together three results which will be essential
in the proof. The first is the triviality of the fundamental group in a
sphere:
Lemma 1 (trivial fundamental group). Every embedding of a 1-sphere
H in a d-sphere G for d > 1 is homotopic to a point.
Proof. Look at the sphere H1 in G1 . Remove a vertex x disjoint from
H1 . By definition of a d-sphere, this produces a d-ball. The curve H1 is
contained in this ball B. We can now produce a homotopy deformations
at the boundary of B until H is at the boundary. Note that B does
not remain a ball in general during this deformation as the boundary
might not generate itself but a larger set. But B1 remains a ball. Once
the H is at the boundary switch an make homotopy deformations of
H until H again in the interior of B. Continuing like that, perform
alternating homotopy deformations of B and H. Because the ball B
can be deformed to a point, we can deform H to the empty graph.
We now show that if sphere H is a subgraph of G, then the enhanced
sphere H1 is embedded. It is an important point but readily follows
from the definitions:
Proposition 2. If H is a (d − 1)-sphere which is a subgraph of a d-
sphere G, then H1 is embedded in G1 .
Proof. We have to show that H1 ∩ S(x1 ) ∩ · · · ∩ S(xk ) is a (d − k − 1)-
sphere for every k and every simplex x with vertices x1 , . . . , xk in H1 .
Any intersections are the same whether we see H as part of G or
whether H is taken alone. The reason is that any of the unit spheres
S(xk ) consists of simplices which either contain xk or are contained in
xk . None of these simplices invoke anything from G. So, the state-
ment reduces to the fact that the intersection of spheres S(xk ) with xk
belonging to a simplex form a sphere. But this is true by induction.
For one sphere it is the definition of a sphere. If we add an addi-
tional sphere, we drill down to a unit sphere in a lower dimensional
sphere.
Finally, we have to look at an intersection number. At appears at first
that we need a transversality condition when describing spheres K, H
of complementary dimension 1, d − 1 in a d-sphere G. While we will
not require transversality, the notion helps to visualize the situation.
Definition 19. Given an embedding of a (d−1)-sphere H in a d-sphere
G and a simple curve C in G. We say it C crosses H transversely
20 OLIVER KNILL
if for every t such that C(t) ∈ H, both C(t − 1) and C(t + 1) are not
in H.
More generally:
Definition 20. Given a d-sphere G, let K be an embedded k-sphere
and let H be an embedded (d − k) sphere. We say K, H intersect
transversely if H1 , K1 intersect in a 0-dimensional geometric graph.
Remarks
1) Given two complementary spheres H, K in a sphere G. Look at the
spheres H1 , K1 in G1 . There is always a modification of the spheres
so that they are transversal. Consider for example the extreme exam-
ple of two identical 1-spheres H, K in the equator of the octahedron
G. The graphs H1 , K1 are closed curves of length 8 inside the Catalan
solid G1 . Now modify the closed curves by forcing H1 to visit the ver-
tices in G1 corresponding to the original 4 upper triangles and K1 visit
the vertices in G1 corresponding to the original 4 lower triangles. The
modified curves now intersect in 4 vertices.
2) Given a (d − 1)-sphere H embedded in a d-sphere G and given a
closed curve C which is transverse to H let xj be the finite intersection
points. We need to count these intersection points. For example, if a
curve is just tangent to a sphere, we have only one intersection point
even so we should count it with multiplicity 2. When doing homotopy
deformations, we will have such situations most of the time. As we
can not avoid losing transversality when doing deformations, it is bet-
ter to assign intersection numbers in full generality, also if we have no
transversality.
For the following definition, we fix an orientation of the 1-sphere C
which we only require to be a simple closed curve and not an embedding
and we fix also an orientation on the (d − 1)-sphere H. Since both are
spheres and so orientable, this is possible.
Definition 21. If the vertex set of C is contained in the vertex set
of H, we define the intersection number to be zero. Otherwise,
let {a = C(t1 ), . . . , C(tk ) = b} ⊂ H1 be a connected time interval
in H1 . Let y = C(t0 ) be the vertex in G \ H just before hitting H1
and z = C(tk+1 ) the vertex just after leaving H1 . The vertex a is
contained in a d-dimensional simplex generated by the edge (y, a) in C
and a simplex σ in H1 containing a. This d-simplex has an orientation
from the orientation of the edge in the curve and the simplex σ in H.
If the orientation of the simplex generated by C and H at the point
agrees with the orientation of the d-simplex coming from G1 , define the
THE JORDAN-BROUWER THEOREM FOR GRAPHS 21
incoming intersection number to be 1. Otherwise define it to be
0. Similarly, we have an outgoing intersection number which is again
either 0 or 1. The sum of all the incoming and outgoing intersection
numbers is called the intersection number of the curve.
In other words, if the curve C touches H and bounces back after pos-
sibly staying in H for some time then the intersection number is 2 or
−2, depending on the match of orientation of this “touch down”. If the
curve C passes through H, entering on and leaving on different sides,
then the intersection number belonging to this “crossing” is assumed
to be 1. If a curve is contained in H, then the intersection number is 0.
The intersection number depends on the choice of orientations chosen
on C and H, but these orientations only affect the sign. If the orien-
tation of the curve or the (d − 1)-sphere is changed, the sign changes,
but only by an even number.
Lemma 3. Two homotopic curves in a sphere G with embedded (d − 1)
sphere have the same intersection number modulo 2.
Proof. Just check it for a single homotopy deformation step of the
curve. As these steps are local, a change of the intersection number can
only happen if a curve makes a touch down at H before the deformation
and afterwards does no more intersect. This changes the intersection
number by 2.
The following lemma will be used in Jordan-Brouwer and is essentially
equivalent to Jordan-Brouwer:
Lemma 4. The intersection number of a closed curve K with an em-
bedded (d − 1)-sphere H in a d-sphere G is always even.
Proof. The sphere G is simply connected, so that every curve C can be
deformed to a vertex not in H. For the later, the intersection number
is zero.
In the case of transversal intersections, the result can be illustrated
geometrically. Look at the intersection of the 2-dimensional deforma-
tion surface of the curve with the d-dimensional deformation cylinder
H × Ln . As long as the curve never intersects H in more than two
adjacent vertices, this is a 1-dimensional geometric graph, which must
consist of a finite union of circular graphs.
4. The theorems
Theorem 5 (Discrete Jordan-Brouwer). A (d − 1)-sphere H embedded
in a d-sphere G separates G into two complementary components A, B.
22 OLIVER KNILL
Proof. We use induction with respect to the dimension d of the graph
G. For d = 0, the graph G is the 0-sphere and H is the empty graph S−1
and the complementary components A, B are both K1 graphs which
by definition are 0-balls.
To prove the theorem, we prove a stronger statement: if H is a (d − 1)-
sphere which is a subgraph of H, then H1 separates G1 into two com-
ponents A, B.
Take a vertex x ∈ H ⊂ G. When removing it, by definition of spheres,
it produces the (d − 1)-ball H ′ = \{x} and the d-ball G′ = G \ {x}. By
definition of embedding, the boundary of H ′ is (d − 2)-sphere which is
a subgraph of the boundary S(x) of G′ .
By induction, H1′ divides G′1 into two connected parts A′ and B ′ .
Let A′′ be the path connected component in G1 \ H1 containing A′ and
let B ′′ be the path connected component in G1 \ H1 containing B ′ . The
union of A′′ and B ′′ is G if there was a third component it would have
H in its boundary and so intersect S(x), where by induction only two
components exist.
The graphs A = A′′ ∪ H1 and B = B ′′ ∪ H1 are d-dimensional graphs
which cover G1 and have the (d − 1) sphere H1 as a common boundary.
It remains to show that the two components A′′ , B ′′ are disconnected.
Assume they are not. Then there is a path C in G1 connecting a vertex
x ∈ A′ ⊂ S(x)1 with a vertex y ∈ B ′ ⊂ S(x)1 such that C does not
pass through H. Define an other path C ′ from x to y but within S(x).
The path C ′ passes through H and the sum of the two paths C, C ′ form
a closed path whose intersection number with H is 1. This contradicts
the intersection lemma (4).
Remarks.
1) As we know χ(G) = 1 + (−1)d and χ(H) = 1 − (−1)d we have
2 = χ(G) + χ(H) = χ(A) + χ(B) so that χ(A) + χ(B) = 2. We will of
course know that χ(A) = χ(B) = 1 but χ(A) + χ(B) = 2 is for free.
2) The proof used only that G is simply connected. We don’t yet know
that A, B are both simply connected. At this stage, there would still
be some possibility that one of them is not similarly as the Alexander
horned sphere.
Examples.
1) The empty graph separates the two point graph P2 without edges
into two one point graphs. 2) The two point graph P2 without edges
THE JORDAN-BROUWER THEOREM FOR GRAPHS 23
separates C4 into two 1-balls K2 .
3) The equator C4 in an octahedron O separates O into two wheel
graphs W4 .
4) The unit circle S(x) = C5 of a vertex x in an icosahedron separates
it into a wheel graph W5 and a 2-ball with 11 vertices.
5) The (d − 1)-dimensional cross polytop embedded in a d-dimensional
cross polytop separates it into two d-balls.
6) More generally, a d-sphere H divides the suspension G = S0 ⋆ H
into two balls, which are both suspensions of H with a single point.
Theorem 6 (Jordan-Schoenflies). A (d − 1)-sphere H embedded in a
d-sphere G separates G into two complementary components A, B such
that A, B are both d-balls.
Proof. Jordan-Brouwer gives two complementary domains A, B in G.
They define complementary domains A1 , B1 in G1 , where A1 , B1 have a
nonzero number of vertices. The proof goes in two steps. We first show
that A1 , B1 are balls: to show that A1 is a d-ball, (the case of B1 is
analog), we only have to show that A1 is contractible, as the boundary
is by definition the sphere H. It is enough to verify that we can make
a homotopy deformation step of H so that the number of vertices in
A1 gets smaller. This can be done even if the complement of H has
no vertices left, like for example if H would be a Hamiltonian path
where A is empty. A simple homotopy deformation of H will reduce
the number of vertices in A1 and since A1 only has a finite number of
vertices, this lead to a situation, where H1 has been reduced to a unit
sphere meaning that H has reduced to a simplex verifying that A1 is
a d-ball. The deformation step is done by taking any vertex x in A1
which belongs to a d-simplex x in A. The step H → H∆x produces
the step H1 → H1 ∆S(x) in G1 . This finishes the verification that
A1 , B1 are d-balls. But now we go back to the original situation and
note that A and A1 originally are homeomorphic. There is a concrete
cover of A1 with d-balls such that the nerve graph is A. So, also A is
contractible.
Remarks.
1) Without the enhanced picture of the embedding H1 in G1 , there
would be a difficulty as we would have to require the embedded graph
H to remain an embedding. The enhanced picture allows to watch the
progress of the deformation.
2) As shown in the proof, one could reformulate the result by saying
24 OLIVER KNILL
that if a (d − 1)-sphere H is a subgraph of a d-sphere G (not neces-
sarily an embedding) then H1 separates G1 into two complementary
components A1 , B1 which are both d-balls.
References
[1] J.W. Alexander. A proof and extension of the Jordan-Brouwer separation the-
orem. Trans. Amer. Math. Soc., 23(4):333–349, 1922.
[2] J.W. Alexander. An example of a simply connected surface bounding a region
which is not simply connected. Proceedings of the National Academy of Sciences
of the United States of America, 10:8–10, 1924.
[3] G.O. Berg, W. Julian, R. Mines, and F. Richman. The constructive Jordan
curve theorem. Rocky Mountain J. Math., 5:225–236, 1975.
[4] L. E. J. Brouwer. Beweis des Jordanschen Satzes für den n-dimensionlen raum.
Math. Ann., 71(4):314–319, 1912.
[5] M. Brown. Locally flat embeddings of topological manifolds. Annals of Math-
ematics, Second series, 75:331–341, 1962.
[6] D. Calegari. Scharlemann on Schoenflies.
https://lamington.wordpress.com/2013/10/18/scharlemann-on-schoenflies,
Oct 18, 2013.
[7] B. Chen, S-T. Yau, and Y-N. Yeh. Graph homotopy and Graham homotopy.
Discrete Math., 241(1-3):153–170, 2001. Selected papers in honor of Helge
Tverberg.
[8] Li M. Chen. Digital and Discrete Geometry. Springer, 2014.
[9] R.J. Daverman and G.A. Venema. Embeddings of Manifolds, volume 106 of
Graduate Studies in Mathematics. American Mathematical Society, Provi-
dence, 2009.
[10] M. Dostál and R. Tindell. The Jordan curve theorem revisited. Jahresber.
Deutsch. Math.-Verein., 80(3):111–128, 1978.
[11] A.V. Evako. Dimension on discrete spaces. Internat. J. Theoret. Phys.,
33(7):1553–1568, 1994.
[12] A.V. Evako. The Jordan-Brouwer theorem for the digital normal n-space space
Z n . http://arxiv.org/abs/1302.5342, 2013.
[13] R. Forman. Morse theory for cell complexes. Adv. Math., page 90, 1998.
[14] R. Forman. Combinatorial differential topology and geometry. New Perspec-
tives in Geometric Combinatorics, 38, 1999.
[15] D. Gale. The game of Hex and the Brouwer fixed point theorem. Amer. Math.
Monthly, 86:818–827, 1979.
[16] T.C. Hales. The Jordan curve theorem, formally and informally. Amer. Math.
Monthly, 114(10):882–894, 2007.
[17] T.C. Hales. Jordan’s proof of the jordan curve theorem. Studies in logic, gram-
mar and rhetorik, 10, 2007.
[18] G.T. Herman. Geometry of digital spaces. Birkhäuser, Boston, Basel, Berlin,
1998.
[19] A. Ivashchenko. Contractible transformations do not change the homology
groups of graphs. Discrete Math., 126(1-3):159–170, 1994.
[20] A.V. Ivashchenko. Graphs of spheres and tori. Discrete Math., 128(1-3):247–
255, 1994.
THE JORDAN-BROUWER THEOREM FOR GRAPHS 25
[21] M.C. Jordan. Cours d’Analyse, volume Tome Troisieme. Gauthier-
Villards,Imprimeur-Libraire, 1887.
[22] F. Josellis and O. Knill. A Lusternik-Schnirelmann theorem for graphs.
http://arxiv.org/abs/1211.0750, 2012.
[23] II J.W. Alexander. A proof of the invariance of certain constants of analysis
situs. Trans. Amer. Math. Soc., 16(2):148–154, 1915.
[24] J.R. Kline. What is the jordan curve theorem? American Mathematical
Monthly, 49:281–286, 1942.
[25] O. Knill. The dimension and Euler characteristic of random graphs.
http://arxiv.org/abs/1112.5749, 2011.
[26] O. Knill. A graph theoretical Gauss-Bonnet-Chern theorem.
http://arxiv.org/abs/1111.5395, 2011.
[27] O. Knill. A discrete Gauss-Bonnet type theorem. Elemente der Mathematik,
67:1–17, 2012.
[28] O. Knill. A graph theoretical Poincaré-Hopf theorem.
http://arxiv.org/abs/1201.1162, 2012.
[29] O. Knill. The theorems of Green-Stokes,Gauss-Bonnet and Poincare-Hopf in
Graph Theory.
http://arxiv.org/abs/1201.6049, 2012.
[30] O. Knill. Coloring graphs using topology. http://arxiv.org/abs/1410.3173,
2014.
[31] O. Knill. A notion of graph homeomorphism.
http://arxiv.org/abs/1401.2819, 2014.
[32] O. Knill. Graphs with Eulerian unit spheres. http://arxiv.org/abs/1501.03116,
2015.
[33] O. Knill. The Kuenneth formula for graphs. http://arxiv.org/abs/1505.07518,
2015.
[34] V. Neumann Lara and R.G. Wilson. Digital Jordan curves—a graph-theoretical
approach to a topological theorem. Topology Appl., 46(3):263–268, 1992. Spe-
cial issue on digital topology.
[35] R. Maehara. The Jordan curve theorem via the Brouwer fixed point theorem.
The American Mathematical Monthly, 91:641–643, 1984.
[36] B. Mazur. On embeddings of spheres. Bulletin of the American Mathematical
Society, 65:59–65, 1959.
[37] R. Tsaur M.B. Smyth and I. Stewart. Topological graph dimension. Discrete
Math., 310(2):325–329, 2010.
[38] M. Morse. Differentiable mappings in the Schoenflies theorem. Compositio
Math., 14:83–151, 1959.
[39] L. Narens. A nonstandard proof of the jordan curve theorem. Pacific Journal
of Mathematics, 36, 1971.
[40] W.F. Osgood. A Jordan curve of positive area. Trans. Amer. Math. Soc.,
4(1):107–112, 1903.
[41] F. Ross and W.T. Ross. The jordan curve theorem is non-trivial. Journal of
Mathematics and the Arts, 00:1–4, 2009.
[42] A. Schoenflies. Beiträge zur Theorie der Punktmengen. I. Math. Ann., 58(1-
2):195–234, 1903.
[43] A. Schoenflies. Beiträge zur Theorie der Punktmengen. II. Math. Ann., 59(1-
2):129–160, 1904.
26 OLIVER KNILL
[44] A. Schoenflies. Beiträge zur Theorie der Punktmengen. III. Math. Ann.,
62(2):286–328, 1906.
[45] A. Schoenflies. Bemerkung zu meinem zweiten Beitrag zur Theorie der Punk-
tmengen. Math. Ann., 65(3):431–432, 1908.
[46] L.N. Stout. Two discrete forms of the Jordan curve theorem. Amer. Math.
Monthly, 95(4):332–336, 1988.
[47] W. Surówka. A discrete form of Jordan curve theorem. Ann. Math. Sil., (7):57–
61, 1993.
[48] O. Veblen. Theory on plane curves in non-metrical analysis situs. Transactions
of the AMS, 6:83–90, 1905.
[49] A. Vince and C.H.C. Little. Discrete Jordan curve theorems. J. Combin. The-
ory Ser. B, 47(3):251–261, 1989.
[50] J.H.C. Whitehead. Simplicial spaces, nuclei and m-groups. Proc. London Math.
Soc., 45(1):243–327, 1939.
