The Topology of Evolutionary Biology

Bärbel M. R. Stadler1 and Peter F. Stadler2,3,4

1
Max Planck Institute for Mathematics in the Sciences. Inselstrasse 22-26,
D-04103 Leipzig, Germany
[email protected], http://www.tbi.univie.ac.at/∼baer/
2
Lehrstuhl für Bioinformatik, Institut für Informatik, Universität Leipzig,
Kreuzstrasse 7b, D-04103 Leipzig, Germany.
[email protected],
http://www.bioinf.uni-leipzig.de/∼studla/
3
Insitut für Theoretische Chemie und Molekulare Strukturbiologie, Universität
Wien, Währingerstrasse 17, A-1090 Wien, Austria
4
Santa Fe Institute, Santa Fe, New Mexico

Summary. Central notions in evolutionary biology are intrinsically topological.

This claim is maybe most obvious for the discontinuities associated with punctuated
equilibria. Recently, a mathematical framework has been developed that derives the
concepts of phenotypic characters and homology from the topological structure of
the phenotype space. This structure in turn is determined by the genetic operators
and their interplay with the properties of the genotype-phenotype map.

1 Introduction
Evolutionary change is the result of the spontaneous generation of genetic
variation and the subsequent fixation of variants in the population through
natural selection and/or genetic drift.
This is the basic assumption of the Neo-Darwinian model. Population ge-
netics appears therefore as a natural framework for studying the evolution of
phenotypic adaptation, the evolution of gene sequences, and the process of
speciation, see e.g. [1, 2]. Patterns of phenotypic evolution [3], however, such
as the punctuated mode (the partially discontinuous nature) of evolutionary
change [4], developmental constraints or constraints to variation [5,6], innova-
tion [7], directionality in evolution, and phenotypic stability or homology [8,9]
are not adequately described by population genetics models.
The reason for this apparent discrepancy is that selection can determine
the fate of a new phenotype only after it has been produced or “accessed” by
means of variational mechanisms [?, 10]. Phenotypes are not varied directly
in a heritable fashion, but through genetic mutation and its consequences on
development. The accessibility of a phenotype is therefore determined by the
genotype-phenotype map (GP-map) which determines how phenotypes vary
with genotypes [11–13].
The motivation for emphasizing the central role of the GP-map originates
from studies in which RNA folding from sequences to secondary structures
2 B. M. R. Stadler & P. F. Stadler

is used as a biophysically realistic, yet extremely simplified, toy-model of a

genotype-phenotype map. Simulated populations of replicating and mutating
sequences under selection exhibit many phenomena known from organismal
evolution: neutral drift, punctuated change, plasticity, environmental and
genetic canalization, and the emergence of modularity, see e.g. [13–18]. Lab-
oratory experiments have also generated phenomena consistent with these
patterns [19–21].
In this contribution we discuss in some detail the surprisingly far-reaching
consequences of formalizing evolutionary theory starting from accessibility.
This approach is to a large extent the consequence of the realization that there
is a significant discrepancy between the mathematical setup of population
genetics (the current implementation of the Neo-Darwinian model of evolu-
tion) and computational case studies of simple evolution processes mentioned
above. Population genetics theory typically assumes that the set of possible
phenotypes is organized into a highly symmetric and regular space equipped
with a notion of distance; most conveniently, a Euclidean vector space [1].
Computational studies using an explicit genotype-phenotype model based on
the RNA folding, however, suggest a quite different picture [13,17,22]: If phe-
notypes are organized according to genetic accessibility, the resulting space
lacks a metric and is conveniently formalized by an unfamiliar structure that
generalizes topological spaces [23–25].

2 Genotype space
The structure of genotype space is uniquely determined by the genetic oper-
ators at work: mutation, recombination, genome rearrangements, etc. In the
case of point mutations and constant length genomes the situation is straight
forward. Naturally, sequences that differ by a single mutation are neighbors
in “sequence space” [26, 27]. The sequence space can thus be represented as
a graph, Fig. 1(a). The same is true for rearrangements.
In the RNA example of Fig. 2 below, a genotype is a sequence of nu-
cleotides encoded by A, C, G, and U of a given length n. The genotype space
consists of all 4n sequences. Two sequences x and y are neighbors if they are
different in just one position, i.e., if their Hamming distance is dH (x, y) = 1.
The resulting graph, a generalized hypercube (Fig. 1) is highly symmetric.
The situation becomes more complicated, however, when recombination
(crossover) is considered [28]. The analogue of the adjacency relation of the
graph is the recombination set R(x, y), which is defined as the set of all
(possible) recombinants of two parents x and y. Recombination sets satisfy
at least two axioms:
(X1) {x, y} ∈ R(x, y),
(X2) R(x, y) = R(y, x).
Condition (X1) states that replication may occur without recombination, and
(X2) means that the role of the two parents is interchangeable. Often a third
 The Topology of Evolutionary Biology 3

condition
(X3) R(x, x) = {x}
is assumed, which is, however, not satisfied by models of unequal crossover
[24, 29]. Functions R : X × X → P(X) satisfying (X1), (X2), and (X3) were
considered recently as transit functions [30] and as P-structures [31, 32].
On the other hand, classical models of population genetics and quantita-
tive genetics (tacitly) assume a euclidean vector space as the natural frame-
work for studying the evolution of phenotypic adaptation, the evolution of
gene sequences, and the process of speciation, see e.g. [1, 2]. This begs the
question whether there is a mathematical framework that contains graphs,
recombination sets, and euclidean vector spaces as special cases. After all,
accessibility in terms of the genetic operators is the common mechanism that
“creates” the structure of genotype space.
We base our discussion on the notion of accessibility [13, 17, 23]. Let us
write x xU y to mean that x is accessible from y “at level U ”. Whether x can
be obtained from y in practice depends on how long we are willing to wait.
Alternatively, we might be interested whether we obtain x from y within a
fixed time with a certain probability. The symbol xU emphasizes that we
consider accessibility w.r.t. a user-defined criterion.
The relation x xU y is represented in an equivalent way by the set
U = {(x, y)|x xU y}. Then U [x] = {y|x xU y} can be interpreted as
the U -neighborhood of x: it is the set of all y that have access to x at level U .
Suppose we have two accessibility relations xU 0 and xU 00 . Then their com-
position, defined by x(xU 0 ◦ xU 00 )y if there is a z ∈ X such that x xU 0 z
and z xU 0 y, is again an accessibility relation at a weaker level. Note that we
do not require that two relations xU 0 and xU 00 are comparable. There is a
natural partial order, however, which is determined by the inclusion relation
of the associated sets U 0 and U 00 respectively: If U 0 ⊂ U 00 then xU 0 is a more
stringently defined accessibility relation than xU 00
From the mathematical point of view it is natural to consider the collec-
tion U of all accessibility relations on a given genotype space. This construc-
tion follows the spirit of [23] with small technical differences. What are the
natural properties U? We propose:
(U0) X × X ∈ U (ergodicity).
(U1’) U 0 , U 00 ∈ U implies U 0 ◦ U 00 ∈ U.
(U1) U ∈ U and U ⊆ U 0 implies U 0 ∈ U.
(U2) x xU x for all x ∈ X and all U ∈ U.
The ergodicity hypothesis says that at some level everything is accessible
from everywhere, if we just wait long enough or if we are content with suf-
ficiently small probabilities. Axiom (U1’) simply allows us to “combine” ac-
cessibility relations to multi-step processes that define weaker accessibility
relations, and (U2) states the trivial observation that every point is acces-
sible from itself. Axiom (U1) allows us to construct a (weaker) accessibility
relations from a given one by adding arbitrary pairs — after all, by axiom
4 B. M. R. Stadler & P. F. Stadler

(U0) arbitrary pairs are accessible at some level! If we assume that U is non-
empty and axioms (U1) and (U2) are satisfied, then (U0) and (U1’) hold as
well. The set-system U on X × X is a generalized version of a uniformity.
In the theory of uniform space [33] one usually assumes additional axioms,
which, however, do not seem to be satisfied naturally in the case of genetic
accessibility relations. These are:
(U3) U, U 0 ∈ U implies U ∩ U 0 ∈ U.
(U4) For each U ∈ U there is a V ∈ U with V ◦ V ⊆ U .
(U5) U ∈ U implies U −1 = {(x, y)|(y, x) ∈ U } ∈ U.
One speaks of pre-uniformities if (U0-U3) are satisfied, if (U4) also holds
we have a quasi-uniformity, while for a uniformity (U0-U5) must hold. The
(generalized) uniform structure U is associated with a generalized topology
on X. In the following section we will briefly review this connection and the
underlying mathematical framework.

3 Generalized Topological and Uniform Spaces

Textbooks on topology such as [34, 35] usually start by defining a topology
on a set X by means of a collection O ⊆ P(X) of “open sets” (the power-set
P(X) is the set of all subsets of X). For our purposes it is more convenient
to use the (equivalent) collection C = {A|(X \ A) ∈ O} of closed sets as
the primitive concept. A pair (X, C) is a topological space if C satisfies the
following four axioms1 :
(I0) ∅ ∈ C.
(I1) X ∈ C. T
(I2) If Ai ∈ C for all i ∈ I, then {Ai |i ∈ I} ∈ C.
(I3) If A, B ∈ C then A ∪ B ∈ C.
Here, I is an arbitrary, possibly infinite, index set. In lattice theory more
general so-called intersection structures are considered that fulfill only (I2),
see e.g., [36]. The closure function cl : P(X) → P(X) defined by
\
cl(A) = {B ∈ C|A ⊆ B} (1)

associates with each set A ⊆ X its “closure” cl(A). The closure function cl has
two important properties: (i) it is isotone, i.e., A0 ⊆ A implies cl(A0 ) ⊆ cl(A)
and (ii) it is idempotent, i.e., cl(cl(A)) = cl(A). Given an isotone and idem-
potent closure function cl, one recovers the associated intersection structure
by setting C = {C = cl(A)|A ∈ P(X)}. One can therefore just as well use
the closure function cl to define a topology by means of the Kuratowski’s
axioms [37]:
(K0) cl(∅) = ∅.
1
The corresponding axioms for open sets are obtained by exchanging unions and
intersections.
 The Topology of Evolutionary Biology 5

1111
1111
1101
1110 1101 1011 0111 1110 1011 0111

1001
1010 0101 0110 1001 0110
1010 0101
0011 1100 0011
1100

1000 0001
0100 0010 0001 1000 0100 0010

0000 0000

(a) (b)

Fig. 1. Closure in Sequence Space.

A population A = {1111, 1110, 1101} (red) and its closure cl(A) (in green) is shown
for (a) point mutations, which give rise to the usual Hamming graph structure,
and (b) 1-point recombination. In the latter case we show the embedding of A and
cl(A) in the Hamming graph since the resulting neighborhood space itself does not
have a convenient representation. The homeomorphism of recombination spaces and
Hamming graphs is discussed in detail in [28].

(K1) A ⊆ B implies cl(A) ⊆ cl(B).

(K2) A ⊆ cl(A).
(K3) cl(A ∪ B) ⊆ cl(A) ∪ cl(B) holds for all A, B ⊆ X.
(K4) cl(cl(A)) = cl(A) holds for all A ⊆ X.
The dual of the closure function is the interior function int defined by
int(A) = X \ cl(X \ A) and hence cl(A) = X \ int(X \ A). A set N is a
neighborhood of a point x ∈ X if and only if x ∈ int(N ). We write N (x) for
the collection of all neighborhoods of x.
Kuratowski’s closure axioms can be translated into the language of neigh-
borhoods where (Ni) is equivalent to (Ki) above:
(N0) X ∈ N (x) for all x ∈ X.
(N1) N ∈ N (x) and N ⊆ N 0 implies N ∈ N (x).
(N2) N ∈ N (x) implies x ∈ N .
(N3) N 0 , N 00 ∈ N (x) implies N 0 ∩ N 00 ∈ N (x)
(N4) N ∈ N (x) if and only if int(N ) ∈ N (x).
Generalized topological spaces have been used in various applied domains
of computer science, such as digital image processing, information represen-
tation, the semantics of modal logic,handwriting recognition, and artificial
chemistry, see e.g. [?, 38–44] for a few examples.
A generalized neighborhood function N can be constructed from a gener-
alized uniform structure U in a natural way. For each x ∈ X and each U ∈ U
we define the sets
U [x] = {y ∈ X|(x, y) ∈ U } for each U ∈ U. (2)
and consider the collection
U[x] = {N |∃U ∈ U : U [x] ⊆ N } (3)
It is not hard to verify that NU : X → P(X), x 7→ U[x] is a neighborhood
function on X. Conversely, given a neighborhood function N we may con-
6 B. M. R. Stadler & P. F. Stadler

Table 1. Correspondence between generalized uniformities and generalized topolo-

gies.

(U0-U1) ↔ (N0-N1) isotone space

(U0-U2) ↔ (N0-N2) neighborhood space
preuniformity (U0-U3) ↔ (N0-N3) pretopology
quasiuniformity (U0-U4) ↔ (N0-N4) topology
semiuniformity (U0-U3,U5) ↔ symmetric pretopology [45]
uniformity (U0-U5) ↔ completely regular topology

struct a corresponding generalized uniformity as the collection UN of all sets

U that contain a set of the form

U = (x, y)|x ∈ X and y ∈ Nx for a fixed Nx ∈ N (x) , (4)

UN is the generalized uniformization of the neighborhood system N . The ax-

ioms for the generalized uniformities translate into properties of the resulting
closure spaces as listed in Tab. 1. The correspondence of pre-uniformities and
pretopologies is described in [23], for a proof that every topology is quasiuni-
formizable and for the characterization of uniformizable spaces see e.g. [33].

4 Phenotype Space

Accessibility at the phenotypic level is the crucial determinant for evolution

since it is the phenotype that is subject to selection. We have seen in the pre-
vious section that the structure of genotype space is determined by physical
processes, namely mutation and recombination, acting on genes. The phe-
notype, on the other hand, is therefore not modified directly, but indirectly
through the modification of the genome from which it arises (We simplify
here by disregarding e.g. epigenetic inheritance [46]). The accessibility rela-
tion at phenotypic level therefore has to be understood as a consequence of
the interplay of genotypic accessibility and the GP-map.
The motivation for emphasizing the central role of the GPmap arose from
a series of computer simulations in which fitness was modeled as a function
of RNA secondary structures [13–18].
The RNA model, see Fig. 2, is a simplified, yet biophysically realistic,
model for a GP-map that has the advantage that its predictions are amenable
to experimental tests, see e.g. [47]. In nature, RNA molecules can act both
as genotype (e.g. as genome of certain viruses) and as phenotype (the 3D
structure can bind specifically to other molecules or even perform catalytic
functions). These functions can then be “evaluated” by the environment e.g.
in a replication experiment [48] or in a SELEX process [49]. RNA secondary
structures are a convenient computational model because they can also be
 The Topology of Evolutionary Biology 7

GCGGGAGCUUCGCGGCGAUAUAGCGGAGAGAGCGCGAUUUUAGAGGCGCGCCCGAGAGCG
CGAGAUUUCUUCGAGAAAGAGCGCGCGAUUUCGCGCGGCGCGAGAGAGCGCGCGCGAAAA
GC
A G
U G
A C
U A G GC
C
G U
G U GA
A C G
G
A G
G G AGA
A G C CG G G
G
G C GC C GC C A
A G G G
U
CG GC
C G C C GGC
U
U
U A GA
G CG GG A A
G CC G
C
G A
UU
AG
A A UCG
C G
42
A G
GA G C
CG
GC
CG
GCG
A A
G G
AUA
U A
UA
> Sequence

CGA
AGGGCGCG

U G
U C

Genotype Phenotype Fitness

Fig. 2. An RNA molecule folds by first establishing the secondary structure, i.e.,
an outerplanar graph in which every nucleotide (letter) is connected to its sequence
neighbors and to at most one other nucleotide with which it forms hydrogen bonds.
The three-dimensional structure of the molecule is formed only in a second step, [52].
The main part of the energy of structure formation can be explained in terms of the
base pair stacking at the secondary structure level [53]. The secondary structure of
an RNA molecule is therefore a useful model for a biophysically realistic GP-map
which serves as the basis for fitness landscapes that are obtained by assigning a
fitness value to each secondary structure graph.

computed efficiently given only the sequence information by means of a dy-

namic programming algorithm [50, 51].
A generic feature of the GP-map at least of biopolymers is redundancy,
i.e., there are many more sequences than structures (at a resolution of practi-
cal interest). Computational studies for both RNA [15] and protein [54] show
quite convincingly that the set ϕ−1 (α) of sequences that fold into a common
phenotype α forms a connected network in sequence space [15].
In practice, any useful notion of nearness will have to correlate with the
likelihood of a transition from one phenotype β to a “neighboring” one α.
For concreteness, let us consider a finite pretopological space, i.e., a directed
graph, that might arise from a particular mutation operator. The possible mu-
tants of a phenotype α then form the graph-theoretical boundary bdϕ−1 (α)
of the neutral network of α. The set ϕ−1 (β) ∩ bdϕ−1 (α) is the subset of all
these mutants that have phenotype β. Since the neutral networks are — at
least in the RNA and protein case — approximately homogeneous we can
estimate the probability of reaching β from α as
8 B. M. R. Stadler & P. F. Stadler

neutral network
neutral network

neutral network

sequence space shape space

Fig. 3. Schematic representation of the neutral networks arising in RNA and pro-
tein folding.

|ϕ−1 (β) ∩ bdϕ−1 (α)|

χ(β ← α) = (5)
|bdϕ−1 (α)|

where |A| denotes the cardinality of the set A. Equ.(5) in turn defines a
collection of accessibility relations xp by setting

β xp α ⇐⇒ χ(β ← α) ≥ p (6)

An approach to formulation accessibility in terms of so-called probabilistic

closure spaces or, alternatively, fuzzy closure space is briefly discussed in [55].
Depending on the size of the underlying neutral nets, nearness between
two phenotypes does not have to be symmetric. Consider a shape α with
a very large, and a shape β with a very small neutral network. It might
be possible that almost all sequences in bdϕ−1 (β) fold into shape α, i.e.,
χ(α ← β) ≈ 1, and α xp β for a large fixed value of p. On the other hand, we
have χ(β ← α) ≈ 0 because |ϕ−1 (β)|  |ϕ−1 (bdα)|, i.e, β 6xp α. This effect
can indeed be observed in the case of RNA secondary structures [13, 17]. As
a consequence, such asymmetries may introduce a directionality in evolution
that is purely structural and hence independent of fitness.

5 Continuity
The notion of continuity lies at the heart of topological theory. Its importance
is emphasized by a large number of equivalent definitions, see e.g. [56, 57].
Let (X, cl) and (Y, cl) be two isotone spaces. Then f : X → Y is continuous
if one (and hence all) of the following equivalent conditions holds:
 The Topology of Evolutionary Biology 9

(i) cl(f −1 (B)) ⊆ f −1 (cl(B)) for all B ∈ P(Y ).

(ii) f −1 (int(B)) ⊆ int(f −1 (B)) for all B ∈ P(Y ).
(iii) B ∈ N (f (x)) implies f −1 (B) ∈ N (x) for all x ∈ X.
(iv) f (cl(A)) ⊆ cl(f (A)) for all A ∈ P(X).
We say that f : X → Y is continuous in x if B ∈ N (f (x)) implies f −1 (B) ∈
N (x). Obviously, f : X → Y is continuous if it is continuous in each x ∈
X. Continuity is preserved under function composition. If f : X → Y and
g : Y → Z are continuous (in x and y = f (x)), then g ◦ f : X → Z is also
continuous (in x).
The least stringent form of accessibility in phenotype space is defined by
setting α x β iff there is a sequence y ∈ ϕ−1 (β) that has a potential offspring
with phenotype α, i.e.,

cl(α) = ϕ(cl(ϕ−1 )(α)) (7)

In fact, this closure function on phenotype space is the most restrictive (finest)
one with the property that ϕ is continuous. We have argued in the previous
section that in most cases a finer closure structure will reflect the practical
accessibilities in phenotype space. It follows that the GP-map will in general
not be continuous everywhere. We shall return to this topic in section 7.

6 Fitness and Fitness Landscapes

Fitness landscapes were introduced in the 1930s by Sewall Wright [58, 59] as
a means of visualizing evolutionary adaptation. In this picture a population
moves uphill on a kind of “potential function” due to the combined effects
of mutation and selection, Fig. 4. We have seen in the previous sections,
however, that the horizontal — phenotypic — axis, in many cases, is not
a real line but a neighborhood space. Alternatively, we may want to regard
fitness landscapes as the graphs of functions from the genotype space with its
neighborhood structure into the real numbers R. This latter case is usually
studied in simulations of molecular evolution.
Fitness and energy landscapes have become a unifying theme in fields
as diverse as drug design, spin glass physics, molecular structure, protein
folding, combinatorial optimization, and evolutionary theory, see e.g. [60] for
a recent review. In each case, there is a function f , e.g., a molecular index, a
Hamiltonian, a cost-function, or a fitness, that evaluates each member x ∈ X
of a (usually very large) configuration set X. These configurations can be
(organic) molecules, spin configurations, conformations of polypeptide chain,
tours of a TSP, or genotypes of living organisms.
Characteristic properties of fitness landscapes such as local minima and
saddle points are inherently topological notions as well [61, 62]. For instance,
a point x ∈ X is a local minimum of f : X → R if there is a neighborhood
N ∈ N (x) such that f (y) ≥ f (x) for all y ∈ N . It is instructive to consider
10 B. M. R. Stadler & P. F. Stadler

Fitness

Phenotype

Fig. 4. Fitness landscapes assign fitness values to each (in this case) phenotypic
state. The action of selection drives a population to local maxima of the fitness
landscape.

also the following alternative definition: For each neighborhood U ∈ N (x)

there is a neighborhood N 0 ⊆ U such that f (x) ≤ f (y) for all y ∈ N 0 .
Of course, in pretopological spaces both versions are equivalent: If f (y) ≥
f (x) for all y in a neighborhood N then N 0 = U ∩ N is another, smaller,
neighborhood of x which obviously satisfies f (x) ≤ f (y) for all x ∈ N 0 .
In general neighborhood spaces, however, the second definition is strictly
stronger because the intersection U ∩ N of two neighborhoods is in general
not a neighborhood of x any more. This example highlights the pitfalls that
appear because of the generality of the framework: “obvious” equivalences
of properties that are familiar from topology textbooks might no longer be
valid.
Related approaches to defining saddle points are discussed in [25]. The
topological approach to fitness landscapes is largely unexplored. It appears
promising, however, because it provides a common framework for the discrete
landscapes of combinatorial optimization and for the continuum models that
are used in evolutionary computation [63] and the analysis of chemical po-
tential energy surfaces [64, 65].

7 Evolutionary Trajectories

An evolutionary trajectory can be regarded as a function f from the time

axis into phenotype space, where f (t) represents the phenotype (e.g., the
dominating phenotype in a population) at time t. Computer simulations [15,
16] using the RNA model described in Fig. 2 reveal a pattern of periods
of stasis with intermittent bursts of adaptive evolution that is reminiscent
 notion of continuous versus discontinuous phenotypic transitions [13, 17, 67].

time in the form of discrete steps, it is more natural to use the pretopology
simulations and samples from the fossil record, which intrinsically represent
of the structure of the fossil record and consistent with in vitro evolution
experiments [19, 20], see also [66]. The analysis of the patterns of changes

The topological language outlined here is the result of attempting to recasting

The time axis is usually described as well-known topological space, namely

the real line R endowed with its standard topology. In the case of computer
an adaptive walk with neutrality on an RNA landscape with a fitness function
of the form F (α) = F0 − d(α, α0 ), where α0 is a randomly chosen target struc-

along evolutionary trajectories obtained from RNA simulations lead to a

used by the Vienna RNA Package [51]. Discontinuous transitions w.r.t. the Fontana-
ture and d( . , .) denotes the number of basepairs by which two structures differ.
11

Each attempted move in genotype space corresponds to one step on the time axis.
Fig. 5. Evolutionary trajectory. Here we show the time-dependence of fitness for

Novel phenotypes (secondary structures) are indicated below in the string notation
time

these findings in a more traditional mathematical framework [22, 23].

The Topology of Evolutionary Biology

...........(((((.(((((...((....))....).)))))))))..............(((((((...))))))).

...........(((((.(((((...((....))....).))))))))).............((((((((...))))))))
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))
...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))

...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).

...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))
...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))
...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))

... → • → • → • → • → • → ...
...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))

...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))

corresponding to the directed infinite path graph

...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))
...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))
...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))

...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((..(((....)))...).))))))))).............((((((((...))))))))

...........(((((.(((((..(((....)))...).)))))))))..............(((((((...))))))).
...........(((((.(((((...............).)))))))))..............(((((((...))))))).

...........(((((.(((((...............).))))))))).............((((((((...))))))))
...........(((((.(((((...............).)))))))))..............(((((((...))))))).
...........(((((.(((((...............).))))))))).............((((((((...))))))))
...........(((((.(((((...............).))))))))).............(((((((.....)))))))
...........(((((.(((((...............).)))))))))..............((((((.....)))))).

Schuster topology are marked by a + sign.

...........(((((.(((((...............).))))))))).............(((((((.....)))))))
...........(((((.((((..................))))))))).............(((((((.....)))))))
...........(((((.((((..................))))))))).............((((((.......))))))
+

...........(((((.((((..................))))))))).............(((((.........)))))

...........(((((.((((..................)))))))))................................
+

...........(((((.((((..................)))))))))......................(((...))).

((.........(((((.((((..................)))))))))....((....))............))......
((.........(((((.((((..((((....))))....)))))))))....((....))............))......

((.........(((((.(((...((((....)))).....))))))))....((....))............))......

.......((..(((((.(((...((((....)))).....))))))))....((....))............))......
.......((..(((((.(((.(.((((....))))...).))))))))....((....))............))......
.......((..(((((.(((...(((........)))...))))))))....((....))............))......
.......((..((((((.((...(((........)))...))))))))....((....))............))......
.......((.(((((((.((...(((........)))...)))))))))...((....))............))......
.......((...(((((.((...(((........)))...))))))).....((....))............))......
.......(((..(((((.((...(((........)))...))))))).....((....))...........)))......
.......(((..(((((((....((((....)))).....))))))).....((....))...........)))......
+

.......(((..(((((((....((((....)))).....)))))))........((.......)).....)))......
.......((...(((((((....((((....)))).....)))))))..........((((......)))).))......
......((....(((((((....((((....)))).....)))))))......))..((((......)))).........
......((....(((((((..(.((((....)))).)...)))))))......))..((((......)))).........
......(((...(((((((..(.((((....)))).)...))))))).....)))..((((......)))).........
+
+
+

......(((...(((((((..(.((........)).)...))))))).....)))..((((......)))).........
......(((...(((((((..(.((........)).)...))))))).....)))..(((((....))))).........
......((((..(((((((..(.((........)).)...)))))))....))))..(((((....))))).........
......((((..(((((((((....))..((....))...)))))))....))))..(((((....))))).........
......((((..(((((((..((((....)))).......)))))))....))))..(((((....))))).........
.(((..((((..(((((((..((((....)))).......)))))))....))))..(((((....)))))...)))...
.(((..((((..((((((((.((((....))))..))....))))))....))))..(((((....)))))...)))...
...........(((((.(((((...((....))....).)))))))))..............(((((((...))))))).
0.0
80.0

60.0

40.0

20.0

fitness
12 B. M. R. Stadler & P. F. Stadler

Let us write T for the generalized topological space that represents the time
axis, and let ξ : T → (X, cl) be a continuous function from the time axis
into genotype space. Continuity means here simply that we assume that sub-
sequently sampled genotypes are (easily) accessible from their predecessors.
The composition ϕ ◦ ξ : T → (Y, cl) the describes the sequence of phenotypes
as a function of time, i.e., ϕ(ξ(t)) is the phenotype at time t. Note that ϕ ◦ ξ
cannot be guaranteed to be continuous everywhere, since the GP-map ϕ will
in general not be continuous.
The simplest approximation of an evolutionary process is the adaptive
walk [68]. A population is represented as a single genotype which could
be interpreted as the the consensus genotype. In each time step a mutant
is explored and accepted if the fitness does not decrease. Thus the fitness
f = F ◦ ϕ ◦ ξ is monotonically increasing function of time. A typical trajec-
tory is shown in Fig. 5. It does not differ qualitatively from the trajectories
obtained in a more complex, population-based setting. In the RNA case, at
least, there is a rather clear-cut distinction between “continuous transitions”
(where ϕ ◦ ξ is locally continuous at a time t0 where the phenotype changes)
and discontinuous transitions. In particular, there is a close correspondence
between continuity of a transition and the kind of structural changes that the
RNA structure undergoes at the transition point. For more details we refer
to [13, 17].

8 Characters as Factors of Phenotype Space

The correspondence of structural changes and transition type suggests that
the topological language might be suitable to deal with a much more gen-
eral question: Can one use the information about the evolutionary process
that is represented by the topological structure of phenotype space to de-
fine a biologically meaningful character concept? In [23, 69] we proposed
that the most promising avenue is to start with Lewontin’s notion of “quasi-
independence” [70] which was introduced to clarify the mechanistic assump-
tions underlying the adaptationist research program.
Explaining a character state as an adaptation caused by natural selection
requires the assumption that the character state can be produced by muta-
tion without significantly affecting the functionality and/or structure of the
rest of the body. This notion does not assume that genetic and mutational
variation among characters is stochastically independent (it may in fact be
strongly correlated). All that is assumed is that genetic variation can be
produced at not too low rate so that natural selection can adjust one charac-
ter without permanently altering other attributes of the phenotype. We can
therefore interpret the notion of quasi-independence as a statement about the
accessibility relations in the phenotypic configuration spaces, namely that the
phenotype space can be represented as a product of generalized topological
space, Fig. 6.
 The Topology of Evolutionary Biology 13

Fig. 6. Characters as factors of a graph. In this case we have two factors w.r.t.
the Cartesian graph product, the “faces” and the “bodies” of the “animals”. These
are identified as the unique prime factors of their phenotype space graph. Note
that while the representations of the “bodies” suggests “heads” and “tails” as finer
subdivisions, these are not characters in our setting, because states of “head” and
“tails” do not vary independently here.

In the following few paragraphs we outline very briefly how this approach
works at the formal level. First we recall the definition the usual topological
product and one of its interesting variants:
Let (X, N ) and (Y, M) be two isotone space defined by their neighborhood
functions. Then the (canonical) product space (V, V) = (X, N ) × (Y, M) is
defined on the set V = X × Y and has the following neighborhood function
V: U ∈ V iff there is N ∈ N and M ∈ M such that N × M ⊆ U . The space
(W, W) is the inductive product of X and Y if W = X × Y and U ∈ W
iff there is N ∈ N and M ∈ M such that (N × {y}) ∪ ({x} × M ) ⊆ U .
When the neighborhood function associated with each space is clear we write
simply V = X × Y and W = XY , respectively. The closure function of
14 B. M. R. Stadler & P. F. Stadler

the inductive product can be characterized as follows. Consider two isotone

spaces (X, cX ) and (Y, cY ) and a subset S ⊆ X × Y . The (p, q) ∈ cl(S) iff
p ∈ cX ({x|(x, q) ∈ S}) or q ∈ cY ({y|(p, y) ∈ S}) [45, 17.D.7], i.e.,

cl(S) = (cX (S −1 [q]) × {q}) ∪ ({p} × cY (S[p])) . (8)

In the case of connected graphs, i.e., finite pretopological spaces, the canonical
and the inductive product reduce to the strong and Cartesian graph products,
respectively; see [71] for their properties.
In the case of graphs the factorizations w.r.t. both the strong product
and the Cartesian product are unique (up to automorphisms) [72, 73]. Poly-
nomial time algorithms for computing the prime factor decomposition of an
undirected graph are known [74, 75]. This is probably true also for finite
neighborhood spaces although we do not yet have a formal proof for this
claim. A “unique prime factor theorem” for closely related structures is given
in [76, 77].

9 Character Identity and Homology

In [23, 69] we argue that quasi-independence is a local rather than a global
notion. Local factorization means that the variational neighborhood of a phe-
notype can be described only in a neighborhood of a given phenotype by
the combination of character states, i.e. the coordinates of “dimensions” or
factors. The biological meaning of “locally factorizable” is that there are
no variational limitations on realizing all possible combinations of adjacent
character states. The range of phenotypes that can be described as a combi-
nation of states of a given set of character, on the other hand, may of course
be limited. For instance, it may be possible to describe all fish species by a
combination of a character states of the set of “fish characters”, but there is
no such set of characters which would describe the phenotypic disparity of
all metazoans.
In mathematical terms the problem becomes to identify a factorizable
subspace Q of (X, cl). In terms of graphs this means that we have to find
induced subgraphs that are factorizable. Ideally, given a graph G one would
like to find maximal induced subgraphs that have non-trivial prime factor de-
compositions. Unfortunately, an efficient algorithms for this Induced Subgraph
Factorization Problem is not known at present.
The original definition of homology by Owen identified two characters as
homologous if they are “the same” in some unspecified way. The meaning of
“sameness” was implicitly defined through the morphological criteria used to
distinguish between superficial and essential similarity, i.e. between analogy
and homology. This notion was re-interpreted by Darwin with reference to a
common ancestor. It can be called the “historical homology concept” since it
is defined solely on the basis of historical, genealogical relationships. It does
 The Topology of Evolutionary Biology 15

X
x
x
N(x) z
N(x) B
N(y) A N(z)
y
H y
X N(y)

(a) (b)

Fig. 7. A factorizable region H established the identity of characters between two

points x, y ∈ int(H). In a second step overlapping factorizable regions A and B can
mediate character identity via points in their common interior z ∈ int(A) ∩ int(B).

not, however, clarify what “character identity” means [78]. Alternatively, one
can define homologues as clusters of observable attributes that remain stable
during adaptive evolution by natural selection. This “biological homology
concept” [8, 9] is, in its definition, independent of relatedness by common
of descent and thus has an unclear relationship to the historical homology
concept.
Both homology concepts and their relationship can be accommodated in
the topological theory of character identity [69]: In the previous section we
have identified characters with local factors in phenotype space. The question
of character identity hence can be rephrased as follows: Suppose we are given
two points x and y and factorizations of their neighborhoods: can we identify
factors of the neighborhoods that correspond to each other? The first step
towards answering this question is to clarify the relationship between differ-
ent subspaces. More formally, consider a neighborhood space (X, cl) and let
(U, cU ) be a factorizable subspace, i.e., there are spaces (U1 , c1 ) and (U2 , c2 )
such that (U, cU ) = (U1 , c1 )×(U2 , c2 ). Then every set of the form A = A1 ×A2
with A1 ⊆ U1 and A2 ⊆ U2 is also a factorizable subspace [69]. In partic-
ular, every point x ∈ int(U ) is locally factorizable, i.e., every neighborhood
of x contains a factorizable neighborhood. This property can then be used
to establish the identity of characters (i.e., factors of local factorizations) at
different points x and y, provided there is a connected factorizable set H
such that x, y ∈ int(H), see Fig. 7a. The assumption of connectivity is crucial
for the uniqueness of the factorization [71]. Overlapping factorizable regions
may then be used to extend character identity to pairs of points that are not
contained in a common factorizable region. The necessary condition is that
there is a point z ∈ int(A) ∩ int(B) such that the local restrictions of the
factors of A and B to sufficiently small neighborhoods of z coincide. For the
technical details we refer to [69].
16 B. M. R. Stadler & P. F. Stadler

The identity of variational characters is therefore well defined (via over-

lapping regional factorizations) and determines a class of (in most cases) vari-
ationally connected phenotypes sharing a certain factor. Phenotypes which
share a certain factor/character can therefore evolve into each other without
going through states where the character is not defined. The notion of char-
acter identity based on quasi-independence is thus fully consistent with the
historical homology concept: Continuity of descent is sufficient to establish
character identity. It is not a necessary conditions, however, because nothing
in the general theory prevents two lineages from evolving phenotypes which
have the same variational character — however unlikely this scenario might
be.

10 Concluding Remarks
In this contribution we have discussed a mathematical framework that is
capable of describing important aspects of macroevolution within the Dar-
winian framework. The language is general enough to be applicable in the
same way to both the discrete setting of sequences and the continuum domain
of population genetics; nevertheless it is powerful enough e.g. to construct
a theory for continuity in evolution and to yield a meaningful notions of
character and homology. Another interesting consequence to developmental
evolution is discussed in [79].
We have focused on the formal aspects of this language rather than on
real or potential applications. The reason is that the kind of variational data
that would be necessary for a non-trivial example are not readily accessible
and that a crucial technical step, namely the factorizable induced subgraph
problem, remains yet to be solved.

Acknowledgments

The theory outlined in this contribution is joint work with Walter Fontana
and Günter P. Wagner.

References
1. Futuyma, D.J.: Evolutionary Biology. Sinauer Associates, Sunderalnd, Mas-
sachusetts (1998)
2. Graur, D., Li, W.H.: Fundamentals of Molecular Evolution. Sinauer Associates,
Sunderland, Massachusetts (2000)
3. Schlichting, C.D., Pigliucci, M.: Phenotypic Evolution: A Reaction Norm Per-
spective. Sinauer Associates, Inc., Sunderland, Massachusetts (1998)
4. Eldredge, N., Gould, S.J.: no title. In Schopf, T.J.M., ed.: Models in Paleobi-
ology. Freeman, San Francisco (1972) 82–115
 The Topology of Evolutionary Biology 17

5. Maynard-Smith, J., Burian, R., Kauffman, S.A., Alberch, P., Campbell, J.,
Goodwin, B., Lande, R., Raup, D., Wolpert, L.: Developmental constraints
and evolution. Quart. Rev. Biol. 60 (1985) 265–287
6. Schwenk, K.: A utilitarian approach to evolutionary constraint. Zoology 98
(1995) 251–262
7. Müller, G.B., Wagner, G.P.: Novelty in evolution: Restructuring the concept.
Annu. Rev. Ecol. Syst. 22 (1991) 229–256
8. Wagner, G.P.: The biological homology concept. Ann. Rev. Ecol. Syst. 20
(1989) 51–69
9. Wagner, G.P.: The origin of morphological characters and the biological basis
of homology. Evolution 43 (1989) 1157–1171
10. Fontana, W., Buss, L.W.: ”the arrival of the fittest”:towards a theory of bio-
logical organisation. Bull. Math. Biol. 56 (1994) 1–64
11. Lewontin, R.C.: The Genetic Basis of Evolutionary Change. Columbia Univer-
sity Press, New York, New York (1974)
12. Wagner, G.P., Altenberg, L.: Complex adaptations and the evolution of evolv-
ability. Evolution 50 (1996) 967–976
13. Fontana, W., Schuster, P.: Continuity in Evolution: On the Nature of Transi-
tions. Science 280 (1998) 1451–1455
14. Fontana, W., Schnabl, W., Schuster, P.: Physical aspects of evolutionary opti-
mization and adaption. Phys. Rev. A 40 (1989) 3301–3321
15. Schuster, P., Fontana, W., Stadler, P.F., Hofacker, I.L.: From sequences to
shapes and back: A case study in RNA secondary structures. Proc. Roy. Soc.
Lond. B 255 (1994) 279–284
16. Huynen, M.A., Stadler, P.F., Fontana, W.: Smoothness within ruggedness: The
role of neutrality in adaptation. Proc. Natl. Acad. Sci. USA 93 (1996) 397–401
17. Fontana, W., Schuster, P.: Shaping Space: The Possible and the Attainable in
RNA Genotype-Phenotype Mapping. J. Theor. Biol. 194 (1998) 491–515
18. Ancel, L., Fontana, W.: Plasticity, evolvability and modularity in RNA. J.
Exp. Zoology (Molecular and Developmental Evolution) 288 (2000) 242–283
19. Spiegelman, S.: An approach to experimental analysis of precellular evolution.
Quart. Rev. Biophys. 4 (1971) 213–253
20. Lenski, R.E., Travisano, M.: Dynamics of adaptation and diversification: A
10,000-generation experiment with bacterial populations. Proc. Natl. Acad.
Sci. USA 91 (1994) 6808–6814
21. Szostak, J.W., Ellington, A.D.: In Vitro selection of functional RNA sequences.
In Gesteland, R.F., Atkins, J.F., eds.: The RNA World. Cold Spring Harbor
Laboratory Press, Plainview, NY (1993) 511–533
22. Cupal, J., Kopp, S., Stadler, P.F.: RNA shape space topology. Artificial Life
6 (2000) 3–23
23. Stadler, B.M.R., Stadler, P.F., Wagner, G.P., Fontana, W.: The topology of the
possible: Formal spaces underlying patterns of evolutionary change. J. Theor.
Biol. 213 (2001) 241–274
24. Stadler, B.M.R., Stadler, P.F., Shpak, M., Wagner, G.P.: Recombination
spaces, metrics, and pretopologies. Z. Phys. Chem. 216 (2002) 217–234
25. Stadler, B.M.R., Stadler, P.F.: Generalized topological spaces in evolutionary
theory and combinatorial chemistry. J. Chem. Inf. Comput. Sci. 42 (2002)
577–585
26. Maynard-Smith, J.: Natural selection and the concept of a protein space. Na-
ture 225 (1970) 563–564
18 B. M. R. Stadler & P. F. Stadler

27. Eigen, M., Schuster, P.: The Hypercycle. Springer-Verlag, New York, Berlin
(1979)
28. Gitchoff, P., Wagner, G.P.: Recombination induced hypergraphs: a new ap-
proach to mutation-recombination isomorphism. Complexity 2 (1996) 37–43
29. Shpak, M., Wagner, G.P.: Asymmetry of configuration space induced by un-
equal crossover: implications for a mathematical theory of evolutionary inno-
vation. Artificial Life 6 (2000) 25–43
30. Changat, M., Klavžar, S., Mulder, H.M.: The all-path transit function of a
graph. Czech. Math. J. 51 (2001) 439–448
31. Stadler, P.F., Wagner, G.P.: The algebraic theory of recombination spaces.
Evol. Comp. 5 (1998) 241–275
32. Stadler, P.F., Seitz, R., Wagner, G.P.: Evolvability of complex characters:
Population dependent Fourier decomposition of fitness landscapes over recom-
bination spaces. Bull. Math. Biol. 62 (2000) 399–428
33. Page, W.: Topological Uniform Structures. Dover Publications, Inc., Mineola,
NY (1994)
34. Gaal, S.A.: Point Set Topology. Academic Press, New York (1964)
35. Steen, L.A., Seebach, Jr., J.A.: Counterexamples in Topology. Holt, Rinehart
& Winston, New York (1970)
36. Davey, B.A., Priestley, H.A.: Introduction to Lattice and Order. Cambridge
Univ. Press, Cambridge UK (1990)
37. Kuratowski, C.: Sur la notion de limite topologique d’ensembles. Ann. Soc.
Polon. Math. 21 (1949) 219–225
38. Eckhardt, U., Latecki, L.: Digital topology. Technical Report 89, Hamburger
Beitr. z. Angew. Math. A (1994)
39. Smyth, M.B.: Semi-metric, closure spaces and digital topology. Theor. Com-
puter Sci. 151 (1995) 257–276
40. Pfaltz, J.: Closure lattices. Discrete Mathematics 154 (1996) 217–236
41. Galton, A.: Continuous motion in discrete space. In Cohn, A.G., Giunchiglia,
F., Selman, B., eds.: Principles of Knowledge Representation and Reasoning:
Proceedings of the Seventh International Conference (KR2000), San Francisco,
CA, Morgan Kaufmann Publishers (2000) 26–37
42. Marchand-Maillet, S., Sharaiha, Y.M.: Discrete convexity, straightness, and
the 16-neighborhood. Computer Vision & Image Underst. 66 (1997) 316–329
43. Largeron, C., Bonnevay, S.: A pretopological approach for structural analysis.
Information Sciences 144 (2002) 169–185
44. F. LeBourgeois, M. Bouayad, H.E.: Structure relation between classes for su-
pervised learning using pretopology. In ???, ed.: Fifth International Conference
on Document Analysis and Recognition. (1999) 33
45. Čech, E.: Topological Spaces. Wiley, London (1966)
46. Jablonka, E., Lamb, R.M.: Epigenetic Inheritance and Evolution. Oxford
University Press, Oxford, UK (1995)
47. Schultes, E., Bartel, D.: One sequence, two ribozymes: Implications for the
emergence of new ribozyme folds. Science 289 (2000) 448–452
48. Mills, D., Peterson, R., Spiegelman, S.: An extracellular darwinian experiment
with a self-duplicating nucleic acid molecule. Proc. Nat. Acad. Sci. USA 58
(1967) 217
49. Klug, S.J., Famulok, M.: All you wanted to know about SELEX. Mol. Biol.
Rep. 20 (1994) 97–107
 The Topology of Evolutionary Biology 19

50. Zuker, M., Sankoff, D.: RNA secondary structures and their prediction. Bull.
Math. Biol. 46 (1984) 591–621
51. Hofacker, I.L., Fontana, W., Stadler, P.F., Bonhoeffer, L.S., Tacker, M., Schus-
ter, P.: Fast folding and comparison of RNA secondary structures. Monatsh.
Chem. 125 (1994) 167–188
52. Shelton, V.M., Sosnick, T.R., Pan, T.: Applicability of urea in the thermody-
namic analysis of secondary and tertiary RNA folding. Biochemistry 38 (1999)
16831–16839
53. Mathews, D., Sabina, J., Zucker, M., Turner, H.: Expanded sequence depen-
dence of thermodynamic parameters provides robust prediction of RNA sec-
ondary structure. J. Mol. Biol. 288 (1999) 911–940
54. Babajide, A., Hofacker, I.L., Sippl, M.J., Stadler, P.F.: Neutral networks in
protein space: A computational study based on knowledge-based potentials of
mean force. Folding & Design 2 (1997) 261–269
55. Stadler, P.F.: The genotype phenotype map. (2002) Konrad Lorenz Institute
Workshop on Biological Information.
56. Hammer, P.C.: Extended topology: Continuity I. Portug. Math. 25 (1964)
77–93
57. Gnilka, S.: On continuity in extended topologies. Ann. Soc. Math. Pol., Ser. I,
Commentat. Math. 37 (1997) 99–108
58. Wright, S.: The roles of mutation, inbreeding, crossbreeeding and selection
in evolution. In Jones, D.F., ed.: Int. Proceedings of the Sixth International
Congress on Genetics. Volume 1. (1932) 356–366
59. Wright, S.: “surfaces” of selective value. Proc. Nat. Acad. Sci. USA 58 (1967)
165–172
60. Reidys, C.M., Stadler, P.F.: Combinatorial landscapes. SIAM Review 44 (2002)
3–54
61. Flamm, C., Hofacker, I.L., Stadler, P.F., Wolfinger, M.T.: Barrier trees of
degenerate landscapes. Z. Phys. Chem. 216 (2002) 155–173
62. Stadler, P.F.: Fitness landscapes. In Lässig, M., Valleriani, A., eds.: Biological
Evolution and Statistical Physics, Berlin, Springer-Verlag (2002) 187–207
63. Rechenberg, I.: Evolutionstrategie. Frommann-Holzboog, Stuttgart (1973)
64. Heidrich, D., Kliesch, W., Quapp, W.: Properties of Chemically Interesting
Potential Energy Surfaces. Volume 56 of Lecture Notes in Chemistry. Springer-
Verlag, Berlin (1991)
65. Mezey, P.G.: Potential Energy Hypersurfaces. Elsevier, Amsterdam (1987)
66. Schuster, P.: Evolution in silico and in vitro: The rna model. Biol. Chem. 382
(2001) 1301–1314
67. Fontana, W.: Modelling “evo-devo” with RNA. BioEssays 24 (2002) 1164–1177
68. Kauffman, S.A.: The Origin of Order. Oxford University Press, New York,
Oxford (1993)
69. Wagner, G., Stadler, P.F.: Quasi-independence, homology and the unity of
type: A topological theory of characters. J. Theor. Biol. 220 (2003) 505–527
70. Lewontin, R.C.: Adaptation. Sci. Am. 239 (1978) 156–169
71. Imrich, W., Klavžar, S.: Product Graphs: Structure and Recognition. Wiley,
New York (2000)
72. Dörfler, W., Imrich, W.: Über das starke Produkt von endlichen Graphen.
Österreich. Akad. Wiss. Math.-Natur. Kl. S.-B. II 178 (1970) 247–262
73. McKenzie, R.: Cardinal multiplication of structures with a reflexive multipli-
cation. Fund. Math. 70 (1971) 59–101
20 B. M. R. Stadler & P. F. Stadler

74. Feigenbaum, J., Schäffer, A.A.: Finding the prime factors of strong direct
products of graphs in polynomial time. Discr. Math. 109 (1992) 77–102
75. Imrich, W.: Factoring cardinal product graphs in polynomial time. Discr.
Math. 192 (1998) 119–144
76. Lovász, L.: Operations with structures. Acta Math. Acad. Sci. Hung. 18 (1967)
321–328
77. Lovász, L.: Unique factorization in certain classes of structures. In: Mini-Conf.
Univers. Algebra, Szeged 1971, Bolyai Janos Math. Soc. (1971) 24–25
78. Wagner, G.P.: Homology and the mechanisms of development. In Hall, B.K.,
ed.: Homology: The Hierarchical Basis of Comparative Biology. Academic
Press, San Diego, California (1994) 273–299
79. Wagner, G.P.: What is the promise of developmental evolution? Part II: a
causal explanation of evolutionary innovations may be impossible. J. Exp.
Zool. (Mol. Dev. Evol.) 291 (2001) 305–309