What Isnt Complexity?

Christopher R. Stephens
C3 Centro de Ciencias de la Complejidad and
Instituto de Ciencias Nucleares,
Universidad Nacional Autonoma de Mexico
Circuito Exterior, A. Postal 70-543

Abbreviated Spanish translation published in: Encuentros con la Complejidad, eds.

J. Flores Valdes and G. Martinez-Mekler, Siglo XXI (2012).

1. Introduction
The question What is Complexity? has occupied a great deal of time and paper over the
last 20 or so years. There are a myriad different perspectives and definitions [1, 2, 3]
but still no consensus. But what does the question mean? What do we expect from
addressing it? Many think of the goal as finding an intentional definition, whereby
necessary and sufficient conditions are specified, according to which anything can be
uniquely classified as complex or not. On the other hand, an extensional definition takes a
more phenomenological approach, characterizing the set of complex systems by trying to
name its members. The intentional route faces the difficulty of either being too restrictive
or too general. For example, the notion of computational complexity [4] is
mathematically quite rigorous but is too restrictive and, given that maximally complex
things are random bit strings, certainly does not capture the intuitive notion of what
complexity is. On the other hand, defining complex systems as having many degrees of
freedom and non-linear interactions is completely vacuous given that, basically,
everything is like that, from a salt crystal to a zebra or from a simple atom to the human
brain. One cannot argue that these conditions are not necessary, but they are certainly not
sufficient. However, they do indicate two features that we should be aware of What are
things made of? and, What are their interactions?
The extensional definition runs into the problem of having to decide which systems are in
the set of complex systems and which not. Unfortunately, there is a definite, subjective
disciplinary bias in answering this question, a physicist and a biologist often having quite
different perspectives. However, instead of trying to determine a precise boundary to the
set of complex systems we can take a more pragmatic approach of starting off by listing
some systems that everyone would agree are complex, and therefore in the set of
complex systems, and others that everyone would agree are not complex, and therefore

not in the set. We can then try to determine what properties discriminate between what
we have definitely put in the set and what we have definitely left out. In taking this
approach it will be useful to also pose the complementary, question: What isnt
Complexity? For instance, complexity isnt many degrees of freedom and non-linear
interactions. If it were, then basically everything would be a complex system and it
wouldnt make any sense in trying to distinguish complexity as something different.

2. An example of a Complex System Human language

As stated, instead of first making an a priori definition of complexity and complex
system, we will take the path of first considering some systems that definitely would be
accepted (at least by the vast majority) as being complex and compare them to some
examples of systems that definitely would be accepted (at least by the vast majority) as
being not complex. So, starting off with what is complex two classes of system that
immediately spring to mind, that certainly intuitively qualify as complex, are biological
systems and human languages. We can think of the former as representing physical
complexity and the latter symbolic complexity. Let us address first the question of
human language.
To be, or not to be--that is the question:
Whether 'tis nobler in the mind to suffer
The slings and arrows of outrageous fortune
Or to take arms against a sea of troubles
And by opposing end them. To die, to sleep-No more--and by a sleep to say we end
The heartache, and the thousand natural shocks
That flesh is heir to. 'Tis a consummation
Devoutly to be wished. To die, to sleep-To sleep--perchance to dream: ay, there's the rub,
For in that sleep of death what dreams may come
When we have shuffled off this mortal coil,
Must give us pause.
Figure 1: Hamlets famous soliloquy from Shakespeares play of that name
In Figure 1 we see an excerpt from Hamlets famous soliloquy in the Shakespeare play of
the same name, while in Figures 2 and 3 we see two sequences extracted from a protein
and the human genome respectively. There are several questions we can pose: are all
three symbolic sequences complex? If so, is one more complex than the other? Naturally,
our understanding of the text is greater - we speak the lingo meaning that we
understand both its grammatical and semantic content.

Figure 2: An amino acid sequence from a

protein

Figure 3: A nucleotide sequence from

the human genome

Imagine for a moment though, that Figure 1 is a mystery, as to a large extent are the
others. We will analyze it from the point of view of a statistical physicist. What would
probably strike one first, at least in the case of non-pictographic languages like English, is
that there is a fundamental microscopic degree of freedom the letter (by which we
also include some other elementary symbols such as a space, number, punctuation marks
etc.) and that language consists of an ordered sequence of such letters. The first question
one might ask concerns the relative frequency of the different letters, <xi>, in a given
sample of text, noting that these are highly non-random. For instance, in English, the
vowels a and e, with frequencies of about 8% and 13% respectively, are much more
common than consonants, such as x and z, which have frequencies of 0.15% and
0.07%.
What would we do next look for correlations? Taking text as a linear ordered sequence,
we can look for correlations between pairs of letters located at positions l and l in the
sequence and evaluate the two-point correlation function <xi(l)xj(l)>, which will be a
function of (l-l) only. What would we observe? In the case of l = l+1 we are considering
digraphs; then, we would find, for instance, once again in English, that <t(l)h(l+1)> >>
<q(l)z(l+1)> due to the fact that h often follows t but z does not follow q, th
being the most common digraph in English. Wed also find that the correlations are noncommutative, in that <xi(l)xj(l)> <xj(l)xi(l)>, e.g., th is much more common than
ht. Proceeding, we could consider three-point functions, <xi(l)xj(l)xk(l)>, finding for
l = l +1 and l = l + 2 that the trigraph the was the most frequent. What is more, we
could observe that frequently the trigraph the has a space symbol on either side in the
linear sequence, as had other combinations of letters.
A perspicacious statistical physicist, with an insight into how effective degrees of
freedom that describe collective behavior can emerge from combinations of underlying
microscopic degrees of freedom, might be led to posit the existence of an effective degree
3

of freedom of the type, SPACE xi(l)xj(l+1)xk (l+n) SPACE, and might be led to call
this new emergent degree of freedom a word. The most common word in English is
the which comprises about 2% of all words in a typical text. Armed with this
knowledge of the existence of a bound state of letters the word our inquisitive
statistical physicist following the lead of an analysis of the relative frequency of different
letters might be inclined to now think about the relative frequency of these words. In
Figure 4 we see a plot in log-log coordinates of the frequency of a word versus its rank
for Wikipedia web pages. Lo and behold, one finds a straight line signifying that the
relation between word frequency as a function of frequency rank is a power law of the
form f(r) =A/ra, where a=1 for r < rc, and rc ~ 10,000. This is the famous Zipfs law [5].
Interestingly, there appears to be a crossover at rc such that for r > rc f(r) = B/r2. This is all
very exciting and surprising for our statistical physicist. The question is: does it tell us
anything about language and its associated complexity?

Figure 4: A plot of word frequency in Wikipedia (November 27, 2006). The plot is in
log-log coordinates. x is rank of a word in the frequency table; y is the total number of
the words occurences. Most popular words are the, of and and, as expected. Zipf's
law corresponds to the upper linear portion of the curve, roughly following the green
(1/x) line (Victor Grishchenko)
Now, all this analysis for text in natural language can also be carried out for the systems
of Figures 2 and 3. In fact, many of the results are somewhat analogous. Genes, indeed,
are somewhat analogous to words. Beyond that though, we really dont have much
understanding of the effective degrees of freedom of genetic systems. In fact, there is
currently ample debate as to what degree a gene is an isolatable effective degree of
freedom [6].

3. Bored on the Edge of Chaos

Having seen the appearance of a power law and an associated scale invariance in the
context of human languages our statistical physicist is reminded of what has been an
important concept in the study of complexity - the Edge of Chaos [7, 8]. The idea is
that complex systems, in analogy to the porridge of the three bears, are neither too
ordered nor too disordered but finely balanced between the two. The original
application of the Edge of Chaos idea was to the very interesting subject of information
processing in cellular automata, where it was shown that, in order to promote complex
computation, an adequate balance between information storage and information
transmission was required; information transmission requiring a low entropy (low noise)
regime while information storage, which required many available states, is associated
with a high entropy regime.
Another very interesting example of a balance between ordering and disordering
tendencies is associated with the error threshold [9] in molecular evolution, where there
can be a fine balance between mutation and selection. If the mutation rate is too high then
selection cannot act, while if it is too low the system cannot easily evolve. There is, in
fact, evidence that some viruses do operate with mutation rates near the error threshold
[10] in order to strike a balance between exploration and exploitation. In these different
examples, as in the case of word frequency and Zipfs law, one encounters power laws
characterizing the relationship between certain variables. These power laws are a
consequence of the absence of a characteristic scale and the associated phenomenon of
scale invariance.
However, this idea of balancing order and disorder is an old one in statistical mechanics,
being associated with the idea of a phase transition, especially a second order one, where
physical quantities exhibit power law behavior, just like Zipfs law, and the system
displays criticality. For instance, the magnetization of a ferromagnet, , is related to the
temperature deviation from the critical temperature, t = TcT, as t , where is a
characteristic critical exponent. Near the critical temperature there is a fine balance
between the disordering tendencies of thermal fluctuations and the ordering one due to
the fact that the magnetic dipoles in the ferromagnet energetically prefer to be in an
aligned state. In this case, the temperature has to be tuned quite delicately to be in the
vicinity of the critical temperature in order to see scale-invariant power law-type
behavior. This requirement of tuning is very common.
There are systems, though, where the analog of the temperature tunes itself without
external influence. These systems exhibit self-organized criticality, and the most famous
example is that of a sand-pile [11]. One imagines slowly dropping sand grains onto a pile
on a table. The pile eventually reaches the edge and a dynamic equilibrium is reached
whereby the amount of sand arriving and the amount leaving are equal. In such
circumstances the angle between the sides of the pile and the table is fixed. On the sides
of the pile there take place avalanches of sand grains. The frequency, f, of such

avalanches depends on their size, n, such that f (n) n . Once again, we see the
ubiquitous power law behavior characteristic of scale invariance. Similar results exist for
earthquakes, species extinctions, fundamental laws of physics, such as electromagnetism
and gravity, metabolic rates, city populations, income distribution, flying velocities of
insects and stock prices, to name just a few!
What are we to make of a phenomena that is so omni-present? Is it something very deep
or very trivial? Of course, no one could deny that on a case by case basis all these
phenomena are extremely interesting, but remember, we are here trying to find
discriminating characteristics of complex systems. The ubiquity of power law behavior,
applying equally to simple systems, such as ferromagnets and Newtons universal law
of gravitation, as to complex ones, such as cities populations and stock prices, puts the
Edge of Chaos in its most general meaning in the same class as other characteristics,
such as many degrees of freedom and non-linear interactions, vacuous as a means of
saying what truly discriminates complex systems from simple ones. Thus, although the
Edge of Chaos may be a characteristic of complex systems, it is not a defining
characteristic, in that it does not discriminate between what is definitely complex and
what is definitely not complex.
I mentioned above that there are certainly disciplinary biases associated with
considerations of what is complex and what not. This is also true for whether or not one
thinks that scale invariant behavior is normal, or surprising, or not. For particle
physicists, what is surprising is the existence of masses, i.e. preferred scales, rather than
scale invariance. For a solid state physicist however, seeing scale invariant behavior near
a second order phase transition seems more surprising in a system that possesses
important fixed scales, such as the inter-atomic spacing in a crystal lattice.

4. Structural Building Block Hierarchies

So, not everything exhibits power law behavior. On the contrary, there are just as many
instances of effective degrees of freedom that do have a characteristic scale massive
elementary particles, nucleons, nuclei, atoms, molecules and macromolecules, cell nuclei
and other organelles, cells, tissues and organisms. For instance, there is no scale invariant
distribution of elephant sizes, or number of nucleotides in the DNA of a given species. In
particular, systems that manifestly exhibit physical complexity, such as biological
organisms, are associated with a myriad of quite distinct characteristic scales, at each of
which the characteristic effective degrees of freedom are radically different. What is
more, such effective degrees of freedom exhibit a hierarchical relationship loosely
speaking, nucleons are composed of quarks, nuclei of nucleons, atoms of nuclei and
electrons, molecules of atoms, cell nuclei and other organelles of biological
macromolecules, cells of organelles, tissues of cells and organisms of tissues.

There is a very powerful, simple argument as to why nature should favor such
hierarchical structures. If we try to construct a human by throwing atoms together we
would be waiting a very long time. However, if we try to form molecules by throwing
atoms together then that is much more feasible. Then, we can try and form polymeric
macromolecules from encounters between their monomeric constituents, and so on. In
other words, it is much easier to construct a more complex system via interactions
between a not too large a number of slightly less complex potential building blocks,
than from a very large number of very simple building blocks. One then develops a
nested hierarchy of building blocks at different scales such that blocks at one scale are
constituents for blocks at a larger scale, which in their turn etc. Nucleosynthesis in the
sun works in such a way. A salient feature of building blocks is that they have
characteristic sizes. I know of no manifestly complex system that is constructed from
scale invariant blocks.
Basically all progress in science has come from being able to study systems by restricting
attention to phenomena associated with a fixed scale, thereby isolating a given
characteristic type of building block whose interactions can be studied using an
appropriate effective theory at that scale. In fact, the compartmentalization of science into
a large number of separate sub-disciplines, such as particle, nuclear, atomic and
molecular physics, are a direct consequence of this property of nature. The question is: to
what extent can complexity be understood by taking this approach?
To think on this further, lets return to human languages. Certainly, languages are
associated with a hierarchy of building blocks. The most basic, in written language, is the
letter, followed by syllable, then word, then phrase, then sentence, then paragraph etc.
These different building blocks are certainly associated with characteristic sizes. The
sizes of words and sentences certainly do not follow power law distributions. A
characteristic of language then is the existence of non-trivial structure on many different
scales, each scale being associated with a different effective degree of freedom or
building block. Can we now argue that the presence of a hierarchy of building blocks
with characteristic scales is a sufficient condition for complexity? Certainly, this seems to
be a much more discriminating criterion than many degrees of freedom, or non-linear
interactions, or Edge of Chaos. But is it enough? Certainly biological systems and
human languages exhibit such hierarchies. What about other systems?
Most people would not consider a hydrogen atom to be a complex system. A hydrogen
atom is a bound state of a proton and an electron. The proton is, in its turn, a bound state
of quarks. Perhaps, these in their turn are low energy excitations of a more fundamental
degree of freedom, such as a string or a membrane. So, even at the level of the lowly
hydrogen atom, there exists a hierarchy of building blocks. Are building block hierarchies
a red herring then? In the introduction, I mentioned that, although many degrees of
freedom and non-linear interactions in no way indicated complexity, they were associated
with two important properties what elements are things made of and how do those
elements interact. We are trying to posit that complex systems are composed of
hierarchical building blocks, but have hit the obstacle that some non-complex systems

also exhibit such hierarchies. So, we pose the question: Do these hierarchies exhibit
similar interactions in complex and non-complex systems?

5. Hierarchical interactions
If we consider the air that you are currently breathing, it exhibits a hierarchy of effective
degrees of freedom. It is composed of a set of mainly simple molecules, principally
nitrogen and oxygen. These in their turn are composed of pairs of covalently bonded
oxygen and nitrogen atoms, which in their turn are composed of a fixed number of
electrons and nucleons of two types protons and neutrons. The nucleons themselves are
composed of a fixed number of quarks. Thus, in this system there is a hierarchy of
effective degrees of freedom, or building blocks, passing from one scale to another. The
air as a non-ideal gas in approximate thermal equilibrium represents the ultimate level of
aggregation for this system.
In terms of interactions, the energy scales that measure the degree of interaction between
different building blocks are much higher at the level of quarks than they are at the level
of nucleons, which in turn are higher than the electromagnetic interactions between
nucleus and electrons, which in their turn are higher than the covalent interaction between
the atoms that make a molecule, while the Van der Waals interactions between the
molecules themselves are even weaker still. Thus, the interaction strength systematically
weakens as we go to larger scales. Put another way, it is easier to separate two nitrogen
molecules, than two oxygen atoms, while separating the electrons is even harder and the
nucleons even more so. Interestingly, the interaction strength at the level of quarks in a
nucleon is so high that they cannot be separated at all without creating new ones! We can
also think of the degree of interaction in terms of the ease with which the system can be
perturbed. Separating the constituents is obviously one way to do this.
As another example, consider a spin chain, where, for simplicity, we take the spins to be
binary valued, as in the Ising model, though any alphabet would do - 4 for the nucleotides
of DNA, or 32, thinking of the English alphabet along with a space symbol and five
punctuation marks. We consider nearest-neighbor ferromagnetic interactions so that
aligned spins is the energetically preferred state. In this system there are two ways of
characterizing the degrees of freedom in terms of the individual spins, or in terms of
domains as contiguous sets of aligned spins. Obviously, a domain can be simply
represented in terms of the underlying spins. The same is true of a sand pile, where the
analogs of spin and domain are sand grain and avalanche. In this case there is no
hierarchy or, rather, the hierarchy is of depth two in scale the domain/avalanche
effective degree of freedom being composed of the microscopic spin/sand grain degrees
of freedom. The domains/avalanches are building blocks for the whole system sand
pile/spin chain - but in between the micro and the macro there are no more intermediate
layers. Of course, if we went beyond the above description, and considered how the

above abstract spins actually represent properties of atoms and that, again, atoms are
composed of electrons and nucleons etc. then we would see a richer hierarchy.
So, how do these hierarchical interactions differ from those of a complex system?
Language clearly is also composed of a rich hierarchy of different building blocks. How
do these blocks interact? What will be our notion of interaction strength? We can
consider different complementary measures. We could for instance, just demand that the
words in the system exist in a given lexicon irrespective of grammar or semantics, for
example giving a fitness f to a word that exists and a fitness f << f to one that doesnt.
We could demand that the words obey the rules of grammar without necessarily being in
the lexicon, such as occurs in nonsense verse. We could also demand that the words are
fully consistent both grammatically and semantically, this corresponding to the lowest
energy or highest fitness state.
Lets consider a famous sentence in English, from the point of view of a statistical
physicist, not a linguist, I hasten to add. The quick brown fox jumped over the lazy dog.
contains all the letters of the English, alphabet so we can see all the microscopic degrees
of freedom displayed. This sentence is completely self-consistent, both from the
grammatical and semantic point of view.
We can note that there are certainly strong interactions between letters within words.
The quick brown ofx jumped over the lazy dog doesnt makes sense as ofx does not
exist as a word in the standard English lexicon. However, it does make good grammatical
sense if we accept ofx as a noun. Words are, in general, very sensitive to the
introduction of noise by mutating letters. A mutation of fox to fzx leads to
something not in the lexicon. However, a mutation to fix gives something that is in the
lexicon. Indeed, different words have different degrees of robustness in the presence of
mutations. For instance, the word dead is much more robust to mutations bead, deed,
lead etc. - than the analogous word defunct. It is really the constraint that a word is in
the lexicon that makes the interactions between letters in a word so strong. The
corresponding fitness landscape for letter combination is rugged but not random as there
are preferred non-random combinations of letters that accord with the sounds that the
human voice is capable of making.
Passing now to words: any permutation of the sentence is fine if we are only concerned
with having words in the lexicon. In this case there are no interactions between the
words. However, the constraints of grammar induce interactions between them. Thus, if
we think of grammatical correctness as a measure of fitness then: The quick, brown
jumped fox over the lazy dog. is of lower fitness than the original sentence as the
placement of the verb jumped is grammatically incorrect, thereby indicating that the
fitness of the sentence is sensitive to the order of the words. On the other hand, the
permutation: The lazy dog jumped over the quick, brown fox. is perfectly fine
grammatically, though it does sound a little counterintuitive, while The quick, brown
lazy jumped over the dog fox. makes no sense as we have only a string of adjectives in
the place of the subject of the sentence.

There are two important points that emerge here: first, that there are different classes of
words nouns, verbs, adjectives, prepositions, articles etc. and second, that word position
matters. The degree of interaction between two words in a sentence depends sensitively
on both their type and position. Once again though, we must ask: Is this different to what
we see in non-complex physical systems? After all, type and position often count for a lot
in physics too. Back to our considerations of air! Interchanging the positions of two
nitrogen molecules in the air makes no difference to its macro-state. Nor does
interchanging the positions of an oxygen and a nitrogen molecule. However, if we
interchange a neutron in one of the nitrogen nuclei with a proton in another, we now get a
carbon 14 nucleus. This would make an important change in the energy of the nucleus
which would affect its interactions at higher levels, though clearly it wouldnt affect the
macro-energy state. So, what are we missing? Or, maybe, language isnt complex after
all.

6. Hierarchical Emergence
Above, I argued that lexicographic and grammatical constraints could be viewed as
imposing interactions between letters and words respectively. For words, grammatical
considerations essentially restrict the interactions to phrases and sentences. Are there
longer range interactions? Yes. And these interactions are associated with an important
aspect of language that we have not much touched on up to now semantics. We have
assumed that our erstwhile statistical physicist is ignorant of any meaning associated with
any particular symbol strings, even though he/she has been able to determine an
extremely non-trivial structure with the emergence of many different collective degrees
of freedom or building blocks. Semantics is to do with the map between a particular
ordered sequence of symbols and an associated set of concepts. The quick brown fox
jumped over the lazy dog. has a clear semantic meaning for any English speaker. This is
not so for a non-English speaker. Thinking of both as statistical physicists however, the
non-semantic statistical content is identical. So, can the semantic content be thought of in
statistical terms? Yes. The semantic content is associated with a different set of
correlations than those intrinsic to the letter sequences themselves correlations between
letter sequences and concepts. The sequence fox refers to an animal with a certain set
of phenotypic characteristics. I would bet that the patterns of neural activity are quite
distinct between an English and non-English speaker when presented with this sentence.
Semantics induces interactions beyond the scale of a sentence. Consider:
1. The quick, brown fox jumped over the lazy dog. The dog woke up,
startled.
2. The quick brown fox jumped over the lazy dog. Please pick up milk on the
way home from the office.

10

In the first combination, both sentences are grammatically and semantically selfconsistent in themselves. However, the second sentence is also a continuation in the
narrative associated with the first. For that reason there is an interaction between the two.
For the second combination, once again, both sentences are grammatically and
semantically fine. In this case though, the second sentence is completely logically
separate from the first in a semantic sense and therefore there is no interaction between
them.
But now, having arrived at the scale of two sentences we begin to come to the crux of the
matter, where we can see some property that human language, as a representative of the
set of complex systems, has that other simpler systems dont. The semantic content of the
two sentences is a multi-scale phenomenon. What do I mean by that? Each word has a
semantic content, as does each phrase and each sentence (and each paragraph, section
etc.). If we think that every one of these building blocks has its own characteristic scale,
then the full semantic meaning involves an integration across all these different scales. If
we restrict our attention to the building blocks of any one scale without knowing how
those blocks integrate into others at a higher scale then we lose the full semantic
meaning.
For instance, just looking at combination 1) above; at the level of individual words there
are two nouns, 3 adjectives, one verb, one preposition and two articles. We can see that
we are talking about, a fox (Any of various carnivorous mammals of the genus Vulpes
and related genera, related to the dogs and wolves and characteristically having upright
ears, a pointed snout, and a long bushy tail.), a dog (A domesticated carnivorous
mammal (Canis familiaris) related to the foxes and wolves and raised in a wide variety of
breeds.), the action of jumping (To spring off the ground or other base by a muscular
effort of the legs and feet.), something that is lazy (Resistant to work or exertion;
disposed to idleness.) etc. Looking at just the words we have no means to extend the
semantic content to higher order building blocks. In terms of phrases, 1) has four: The
quick, brown fox is a noun phrase, jumped a verb phrase, over a prepositional
phrase and the lazy dog another noun phrase. Irrespective of how these phrases are
joined together as building blocks to form a sentence we can see that there is now new
semantic content at this scale. The first noun phrase tells us that its the fox that is quick
and brown while the last one tells us that it is the dog that is lazy. Apart from that we still
dont know who jumped over whom. If we go up to the sentence level of building block
however now the ordering of the phrases tell us that it was the quick brown fox that
jumped over the lazy dog. A similar analysis of the second sentence tells us that it was
the dog that woke up and that its condition was one of startled. However, we can now
do one further integration step and combine the two sentences together to understand that
the dogs waking up and startled state was a direct causal result of the fox jumping over
it.
To recap: what distinguishes language as a complex system isnt just that there is a
hierarchy of building blocks but, rather, the existence of emergent properties that depend
on all levels of this hierarchy. Meaning is such a property that transcends any given fixed
level of building block. It is present at every level to a given degree but at each higher

11

level a new integrated form emerges that depends on the lower levels. Thus, the meaning
of a word depends on the letters that form it; the meaning of a phrase depends on the
meaning of the words that it is composed of; the meaning of a sentence depends on the
meaning of the phrases that constitute it. There is no meaning to a building block at a
given level without having the meaning of its constituent lower level building blocks. It is
this that I claim is a true hallmark of complexity.
What about physical complexity? Biological systems clearly display hierarchical building
blocks. Are they more like language or more like air? The case of language illustrated
that it was not the building block hierarchy per se that was associated with complexity
but rather how a particular observable meaning induced interactions between
different levels of that hierarchy. I would argue that there exists at least one observable
in biological systems that has the same property fitness, thought of as reproductive
success. With meaning I argued that it was manifest at every building block level, but
became transformed passing from one level to a higher one. Clearly, with meaning the
whole is not the sum of the parts, even though the parts clearly contribute. Fitness in the
same way has a meaning at each level but, again, the whole is not just the sum of the
parts.
To give an example: start with a microscopic scale, that of the important biological
macromolecules, such as DNA, RNA, proteins etc. These obviously contribute to fitness.
There are many micro-events, such as mutations, that have important macroconsequences, such as sickle-cell disease. There, a mutation alters the nature of
haemoglobin proteins so that red blood cells form with anomalous sickle-like shapes
which change their oxygen carrying capacity. However, the consequences are restricted
in their scope. The sickle-cell mutation does not directly affect other body functions but
does affect overall fitness as it reduces life expectancy in non-malarial environments.
Essentially, cells with a gene with the sickle-cell mutation arent doing what they should
and this can be reflected in a lower contribution to the overall fitness of the organism
from the cell reproduction mechanism associated with those cells. This in turn can be
reflected at a higher level by a lower fitness contribution from the red blood cells
themselves, which, at an even higher level, can be thought of as a lowered contribution
from the red bone marrow that produces them and even up to a lower contribution from
the entire cardio-vascular system.
Contributions to fitness originate in all building block scales, just as contributions to
meaning do in the case of language. In fact, it is probably not too fanciful to think that
meaning in language is closely associated with an analogous concept of fitness or utility,
as would accrue, for example, from being warned by a companion of a risk, such as a
dangerous predator. In more general terms utility can stem from many things: following a
cooking recipe, building an engine or solving differential equations, even being
entertained by a book.

12

7. The tyranny of physical law and the difference between

being and doing
Besides being characterized by a rich hierarchy of building blocks at different scales that
interact in an integrated fashion, what else characterizes complex systems? Once again
we return to the logic of what properties do manifestly complex systems exhibit that one
does not tend to find elsewhere? In physics and chemistry we have worked within the
same paradigm for nearly 500 years that the world can be described by a state and that
there is a unique dynamical law that evolves that state in time. What has changed over
time is our understanding of that unique dynamical law, passing from the laws of
classical mechanics to those of quantum mechanics and incorporating relativity. In fact,
not withstanding the quasi-theological pursuit of a theory of everything one could
sensibly take quantum electrodynamics as the fundamental microscopic theory from
which just about everything up to terrestrial scales should be derivable. Biological
systems in principle fall into this category. Of course, no one seriously believes that
quantum electrodynamics can be sensibly used to understand biology. It is not facile,
however, to ask why not? Am I hidden as a collective excitation somewhere deep in the
generating function for quantum electrodynamics? Given that every atom that composes
an amoeba has to obey quantum electrodynamics then it follows that the amoeba itself
obeys quantum electrodynamics. The latter is sufficient to describe and predict the
behaviour of individual atoms, but not the enormously complex dynamics of the huge
number of interacting macromolecules that characterize the amoeba.
But is it just a question of complication? For example, the energy levels of hydrogen can
be solved for analytically at least in the confines of an approximate model the one
particle Schroedinger equation. The energy levels of Uranium on the other hand cannot.
However, numerically, the levels of Uranium can be solved for very accurately. So, what
does an amoeba do that a uranium atom doesnt? The question itself, in fact, contains the
answer the amoeba does things. An atom passively obeys Schroedingers equation,
while a tennis ball passively obeys Newtons laws. A cat necessarily obeys both - but not
passively. A cat falling upside down will right itself in mid-air so as to land on its feet.
Both cats that fall on their head and cats that land on their feet obey all the restrictions
imposed by the laws of physics and chemistry. Clearly then, these cannot help us decide
why we only see cats that land on their feet. They cannot be used to distinguish between
head-first and feet-first cats. The point is, the cat or the amoeba have a huge number of
potential internal states that are all equally compatible with the fundamental laws of
physics. What is more, that internal state can change dynamically in reaction to what is
happening in the environment. A word that can be applied to such dynamically changing
internal states is strategy.
Two competing strategies that falling cats can adopt are: head-first strategy and feetfirst strategy. As the former leads to more injuries and a lower survival rate, the latter
strategy has been selected as a useful adaptation. What is a strategy in this context? It is a
rule for updating the state of the system. The fact that different strategies exist means that
there are different dynamical rules for updating a state. This also seems to be an
13

important characteristic of complexity in the biological regime: That a more adequate

description of the systems dynamics is in terms of both states and update rules or
strategies.
Is this phenomenon only present in biological systems? A wonderful illustration of it in
an artificial life, as opposed to real life, context, can be found in Karl Sims [12] stunning
work on evolving virtual creatures to do different tasks. What Sims did was to create a
creature by combining a simple morphology, obtained by combining together a hierarchy
of three-dimensional rigid parts quite literally building blocks. These building blocks
were represented genetically as a directed graph, where each node in the graph contains
information about the dimensions of the corresponding block and how it is joined to its
parent block. A neural network was used as a brain that indicated how one block should
move relative to another. The creatures were set a task swim in a virtual sea, or move
on a virtual surface etc. A population of such creatures was then evolved over time to see
what strategies emerged for tackling the given problem. In Figure 5 we see a one such
evolved solution for swimming. Go to
http://www.archive.org/details/sims_evolved_virtual_creatures_1994 to see the
impressive movie.
This represents much more than just a striking
visualization. Great care and attention went into
this work to assure that all the elements of the
simulation were physically sound, i.e. that both
internal movement and movement through the
media were all completely consistent with the
dynamics of rigid bodies in viscous media or in a
gravitational field. Why do I take pains to
emphasize this? Because the results of this
simulation - the evolution of fitter and fitter
Figure 5: A virtual creature evolved organisms adapted to a given environment
cannot be understood in terms of the laws of
for swimming in a virtual sea.
physics
What do I mean by this? The space of possible strategies for a given complex system is,
of course, restricted by the laws of physics. For instance, humans cannot fly unaided
because, among other things, we cannot generate enough lift without wings. But the laws
of physics only bound what the possibilities are, not which ones are chosen or why. This
is completely different to what happens for physical systems themselves. In that case the
tyranny of the laws of physics is such that there is only one relevant update rule, either at
the micro- or the macro-level. Thus, a molecule, a pollen grain or a baseball are all
constrained to obey update rules that tells us what their state will be later given that we
know it now. It is for that reason there is no meaning to the concept of adaptation in the
physical world. Adaptation is to do with being able to adopt a new update rule (strategy)
within a given environment.
Physical systems only have to do one thing be. There is no requirement that an atom
evolve in order to survive in a hostile environment. When air is heated to sufficient
temperature, the atoms that form the molecules in it will start to dissociate. The
14

molecules do not develop an adaptation enabling them to survive in this hostile thermal
environment. If we keep raising the temperature we will start to ionize the atoms,
separating off some of the electrons. There is no possibility of an adaptive change in the
atoms that enables them to stop losing electrons. However, if it gets cold I put on a coat.
The origin of this is that in human prehistory, those who could prevent heat loss by
covering themselves would have higher survival rates than those who didnt. Just as with
the falling cats, there are two strategies cover up and reduce heat loss, or stay
uncovered and risk hypothermia. Unlike the atoms, a change in environment in this case
can lead to more than one response.
The emergence of strategy as a description of a complex adaptive system is a truly
emergent phenomenon. But, where does it emerge? If we think of chemical evolution, say
in the context of an RNA world, where, as an example, different RNA molecules can
compete in the context of the enzymatic catalysis of a particular interaction, it does not
seem very natural to think of the RNA molecules as having a strategy. At this level, the
tyranny of the laws of physics is still manifest.
As emphasized physical systems only have to do one thing be. They do not have
choices. Biological systems however do have choices, different strategies, all
consistent with the restrictions imposed by the laws of physics. What is more, different
strategies are often associated with different elements and levels of the building block
hierarchy, all of them exquisitely choreographed by the requirements of evolution to
work in harmony.
This discussion naturally leads us to consider - what is adaptation? Does a population of
RNA molecules adapt searching for an optimal configuration for catalyzing some
reaction? We can paint this at another level, in the context of artificial evolution, such as
in Genetic Algorithms or Artificial Life. In Genetic Algorithms, the relevant dynamical
equations are almost identical to those long familiar from population biology, where a
population of chromosomes evolve under the action of selection and genetic mixing
operators, such as mutation and recombination. The equations that direct the dynamics in
these systems is very familiar from physics and chemistry: Stochastic dynamics of
Markov chains in the case of finite populations, while in the infinite population limit, the
relevant equations are a set of deterministic, non-linear first order difference equations
fiendishly difficult to work with, but conceptually within the same paradigm as any other
traditional differential equation based approach. But it is precisely such systems that I
argue are subject to the tyranny of law. In the case of a Genetic Algorithm, the
restrictions do not come from any physical requirement but rather the restrictions of the
model itself.
Is a Genetic Algorithm really adapting or, more provocatively, is population genetics a
model for adaptation? In both cases the dynamics takes place in a space of states, but
there is no explicit dynamics for any update rule. Rather, an explicit fitness function
(viability in the context of population genetics) is used as a proxy for how well the
system does. In Genetic Algorithms this paradigm has been successfully used in the
context of combinatorial optimization, where there does exist a concept of best solution

15

and it is reasonable what a solution is rather than what it does. In the more interesting
case of Genetic Programming, this is not the case, an evolved solution representing a
computer program and therefore a solution that does something as opposed to just is.
For these reasons I would argue that biological adaptation really should be thought of as
taking place in the space of strategies and states, not in the space of states alone. In fact, I
believe that adaptation cannot emerge from any paradigm where a fitness function has
been specified a priori.

8. Specialist vs. Generalist: Evolution as the development

of multi-tasking
A key element of biological systems then is that their dynamics is naturally described in a
space of states and update rules. This, of course, is reminiscent of game theory [13]. In
the latter, payoffs are assigned to individuals or groups implementing a certain strategy in
the context of other players playing their strategy. A familiar example would be the
childrens game rock-paper-scissors. For two players, a payoff of 1 would be given to a
player playing rock, while 0 would be given to the player playing scissors. It is not clear
how useful the game theory paradigm is for physical complex systems. First of all, there
is the question of the tremendous number of possible strategies. Worse still, we have no
idea in what space were working in. How can we talk about the utility of one strategy
versus another when we have no idea about the possibilities? Imagine 3.5 billion years
ago trying to wonder about how to assign a payoff to lion, cockroach or oak tree
as the analogs of rock/paper/scissors! How would one begin to even imagine the
possibility that such strategies could exist? The only option that we have is to observe the
discovery of certain regions of the space of possible strategies as evolution moves
forward and infer their relative benefits from observation.
We also have to address the question of strategy to do what? Game theory is usually
couched in the language of one strategic objective, where there is a clear notion of
winning or losing. One could argue that survival in an evolutionary context is a game.
However, there is a very important difference between evolutionary survival and rockpaper-scissors that can be couched in the following way: Really, an organism is
involved in a vast number of games and has to develop strategies for all of them. It also
has the pressure of having to do well in all these different games in order to have a high
survival probability. An organism that has a successful strategy in almost all areas save
one can still be fatally flawed, as the overall fitness of the organism gets contributions
from the payoffs of all these different simultaneous games.
For instance, an organism must play a game against potential predators, and another
against potential territorial rivals, and yet another against potential mates. However, this
is just the tip of the iceberg. Inside the organism its immune system is also playing many
simultaneous games, with white blood cell and antibody players competing against
invading pathogens for example. Some of these games are carried out in series and some
16

in parallel. Our immune system implements a strategy to ward off dangerous microorganisms. It can do this while at the same time, the brain and cardio-vascular system are
functioning trying to ward off a potential predator for example. On the other hand, an
individual cannot simultaneously fight off a predator and look for a mate. So, in some
problems a specialist approach is taken the heart pumps blood, the lungs transfer
oxygen, the stomach digests food etc. while in others a generalist approach is taken.
The latter is the rule when there is no sub-system that can specialize, such as in the case
of seeking a mate and warding off predators. A human can choose to do one or the other,
but not both at the same time.
The specialist approach allows for multi-tasking, a property that clearly lends a
tremendous evolutionary advantage. Imagine if your immune system worked only by
explicit conscious effort while you werent distracted doing something else! Evolution is
very much a phenomenon driven by the development of abilities to multi-task. In the
context of multi-tasking there can be many combinations of strategies that lead to more or
less the same fitness with compensations between one element and another. The multitasking inherent in biological systems is hierarchical in nature, with the goal of
evolutionary survival being associated with an overall strategy that itself is composed of
a number of building block strategies, which in their turn are composed of other less
functionally complex strategies. In a dynamic where novel building block strategies can
be combined at a higher level then it is possible to have an emergence of truly novel
strategies that were not originally present.

9. Functional and Structural Modularity and Building Blocks

Above I emphasized that the overall evolutionary strategy of biological organisms can be
thought of as being composed of a very large number of building block strategies that
work together to give an overall one. These building block strategies however, are, in
general, carried out in parallel not in series. This requires modularity, a manifestation of
the philosophy of divide-and-conquer. Modularity allows for parallel strategy
implementation and for specialized strategies. Thus, lungs take care of the fact that
oxygen is necessary for cellular metabolism, converting fuel molecules, such as glucose,
into biochemical energy, while the stomach and intestines take care of digestion of food,
breaking it down into more easily useable molecular forms. To locate food sources,
sensory organs are used, among other things, to process environmental stimuli, the brain
and nervous system interpreting it and the muscles then being used to move to a food
source. Each of these modules is highly specialized to a particular task. The cell itself
also has its corresponding specialized modules all working in parallel nucleus,
mitochondria, centriole, vesicle, lysosole etc.
Thus, in this context, a module is really just a building block where we can readily its
function. It shouldnt be surprising that there should be such a close relationship between
building blocks as discrete structures and building blocks of strategies, a structural

17

building block in complex systems being associated with a particular function. In fact the
modules are really just the effective degrees of freedom of the system. A characteristic of
a complex system is that there exists a hierarchy of modules at different scales where
modules at larger scales are composed of more microscopic building block modules.
Usually, we think of effective degrees of freedom as being characterized as collective
excitations, being composed of combinations of the microscopic degrees of freedom.
Here, I am saying that this hierarchy can and should also be thought of in terms of what
the effective degrees of freedom do, i.e., the strategy that they represent. A strategy at a
higher level is then an emergent phenomenon that results from a combination of
strategies at a lower level. Imagine a system and two strategies A and B that both
enhance the fitness of the system. The question then is: Is it better to spend some time
doing A and some time doing B? Or to develop the capacity to carry out both
simultaneously by modularizing and using part of the system to do one and part the
other? The evidence from biology is that there is no uniform answer, as examples of both
occur, but that it is clearly the case that with base level functions, such as respiration,
metabolism, excretion, reproduction, temperature control, immune response etc. the
modular route is the only feasible one.

10. Can we measure complexity?

I have argued that, although manifestly complex systems are associated with a rich,
emergent hierarchy of building blocks at different scales, this fact, in and of itself, is not
sufficient to identify them as being complex. Instead, I argued that complexity was
manifest in the special way these building blocks interacted and that this in its turn was a
consequence of the system being required to fulfill a certain goal give meaning in the
context of language (symbolic complexity), and fitness in the context of biological
systems (physical complexity).
If complexity is to be a meaningful scientific concept, however, then it should be
measurable, or at least related to measurable quantities. What measuring device could
detect hierarchical interactions between building blocks across multiple length scales?
One possible measurement apparatus that springs to mind is illustrated in Figure 5 and is
in itself highly complex the human brain.

18

To be or not to
be that is the
question?

Figure 5: Is this the only apparatus capable

of measuring complexity?

Returning to the subject of language: Imagine that

we take the sentence: To be or not to be that is the
question and we use different brains to measure
it. Of course, there will be a significant difference
in the response between an English speaker and a
non-English speaker that surely would be manifest
in the internal configurations of the measuring
apparatus i.e., the neural states of the two
different brains when presented with the sentences.
However, it is not difficult to imagine that there
would also be a difference in the responses
between a native English speaker and a non-native
English speaker, or differences due to cultural or
socio-demographic factors.
Why would there be differences? Due to two
sources: non-semantic structural differences and,
more importantly, semantic differences. The
former is associated with the fact that an English
speaker would recognize a coherent structure in
the sentence irrespective of its semantic content.
Beyond that, the differences in how different
measuring apparatus react are more subtle, being
associated with the present state of the apparatus,
which, in its turn, depends on its past history.

Someone with an intimate knowledge of Hamlet and the works of Shakespeare, and the
historical and cultural context of the play, would have a different neural response to
someone who was unknowledgable on all those counts.
So, does such subjectivity make a nonsense of using the brain as a measuring apparatus
for language? Of course not! Measurement is to do with examining correlations between
a property of a system and a property of a measuring device. One chooses a measuring
device carefully for a given phenomenon so that this correlation is clear and readily
interpretable and reproducible. In this context the brain of a non-English speaker is not
that useful for measuring the semantic content of a text in English. Similarly, the brain of
a non-Japanese speaker is not so useful for measuring the semantic content of a Japanese
text. However, in both contexts the measuring apparatus, the brain, will respond. What
differs is the nature of that response. The difference is that the individual measuring
apparatus have been calibrated to respond to different signals one to English and the
other to Japanese. Is that so different to what happens in physics though?
In Figure 6 we illustrate this by considering the case of two physical phenomena: boiling
water and the electrical impulses of the brain as manifest in an electro-encephalogram.
The two corresponding measuring apparatus are a thermometer and a voltmeter. The
thermometer is an apparatus that has been calibrated to respond to temperature and the

19

voltmeter to potential differences. However, if we throw the voltmeter in the boiling

water it will certainly respond! Just not in a way that has been calibrated and gives us the
information we want or expect about the underlying system.

good
bad
bad
good
Electricity, eeg

Figure 6: Correlations between different physical phenomena and some potential

measuring apparatus.

The key property of the brain in terms of measuring structure and meaning in text is that
it does so at multiple scales while simultaneously integrating information from those
different scales. It does this by utilizing memory. Thus, the brain detects patterns
irrespective of semantic meaning as, for instance, the sentient scientist sensitive to several
significant sources of signal will notice in the alliterative tone. This ability to measure
and integrate data across multiple scales using memory is what allows the brain to cope
with the complexity inherent in the semantic content of human language and therefore
measure meaning.
It is important to distinguish between the hierarchical building block structure intrinsic to
language that is independent of semantics, as measured, for example, in terms of
correlation functions etc., and the extrinsic complexity associated with its semantic
meaning, which is an implicit relation between the system, text say, and a measuring
apparatus, a brain. However, as emphasized, different brains have different calibrations.

20

Of course, brains are plastic and can be recalibrated, in order to measure meaning in other
languages. Is there anything else that can measure meaning besides the human brain?
That meaning requires a complex measuring apparatus such as the brain is manifest in the
difficulties of automatic translation. Below we see different translations of the phrase To
be or not to be that is the question. as generated by an on-line translation engine. In
order to test whether the system is generating any sense of meaning one would hope
that its output from English to language X when fed in as input for translation from
language X to English should return the same phrase, or a very close approximant.
1) To be or not to be that is the question.
2) Para ser o no ser que es la pregunta.
3) Om te zijn of te zijn niet dat de vraag is.
4)

5) Because of a certain or because it is not, it is question?

6) Because or it is not for the sake of, that having asked and being convinced?
7) Being not to be for the sake of, or that that, you ask, are convinced?
8) It is that without having for the sake of, or, you ask, are convinced?
2) and 3), along with 1), form a closed cycle for Spanish and Dutch respectively. This
happens because in this case a literal translation is possible, as the structures of the
languages are sufficiently similar that a purely grammatical map in a single fairly simple
sentence is sufficient to generate something that approximates the same semantic content.
Rest assured this would not be the case if the full soliloquy was fed into the system. 4) is
the translation of 1) into Japanese and 5) the output from feeding 4) into the Japanese to
English direction. 6), 7) and 8) are further iterations of this process. Obviously, an
unstable state has been reached!

21

11. Can we model complexity and is it the same as modeling

complex systems?
Lets consider a simple mathematical model [14]: a group of point particles with position
vectors ci and velocities vi and a force due to interactions with other particles which has a
direction di = d i / | d i | for the ith particle, where

d i =
j i

c j (t ) ci (t )
| c j (t ) ci (t ) |

+
j =1

v j (t )
| v j (t ) |

(1)

The first term on the right hand side represents a repulsion between the particles, while
the second term an attraction that tries to align their motion. The force is made stochastic
by adding a small random number to it. The effect of the force is to align the direction of
a particle with di.
We can ask whether a simple mathematical model for point particles based on the
interactions of (1) exhibits any of the characteristics that we have argued discriminate
between complex and non-complex systems. There is certainly no manifest appearance of
any of the symptoms, such as a hierarchy of effective degrees of freedom, or the
emergence of strategy as a more meaningful description of an element of the system. On
the contrary, the model seems to describe a very simple system with competing shortrange repulsion and longer-range attraction familiar in physics. However, equation (1)
has been successfully used to model the dynamics of fish shoals! Now, no one would
think that fish shoals do not represent a complex system, so how can this simple system
model a complex system when we have argued that the model shows none of the most
discriminating features of complexity? Maybe the criteria are too restrictive? What if we
use the model to describe interactions between point particles rather than fish? Does the
model still represent a complex system? If so, then there are many simple physical
systems that we should now classify as complex.
The resolution to this conundrum is the following: describing complexity is not the same
as describing a particular facet of a complex system. The facet the above model describes
is the mechanistic dynamics of how pelagic fish shoal, which just so happens to be the
same type of model as describes how particles with repulsive and attractive forces
interact. But fish are not mechanistic in the same sense as particles in a force field. Fish
do many interesting things besides shoal. The above model cannot describe any of these
other characteristics. Of course, it wasnt meant to. The model provides a mechanistic
description of a behavior, a strategy fish shoaling. We can hypothesize, given the nature
of the system, that the fish do this for a reason. That it has some evolutionary advantage,
such as helping reduce the predation rate. Do molecules that can be described by a similar
model employ a strategy too? Of course not. One could imagine other fish behaviors that
might also yield to such a mechanistic description based on a simple model. Imagine that
we could, in fact, do that for any fish behaviour. Would that mean that the fish were

22

describable in terms of a set of simple models, one per behavior? Once again, of course
not. Such models do not and cannot give us any insight into complexity or the notion of
what is complex in a complex system. They can however, give us quantitative models
of certain aspects of complex systems that, importantly, can, in principle, be compared
with experiment.
So complex systems can be modelled, in the above sense, but that does not mean we can
model complexity. Is this possible? Well, we certainly do not have any existing theory
that can do the job. What ingredients would such a model need? Well, for physical
complexity, i.e. biological systems, first, it would have to be a model that worked at the
level of strategies and states not just at the level of a unique dynamical law evolving a
state. Second, it would have to function at a level where the strategies were not a priori
known, nor was their payoff. This also has to be an emergent property. The systems that
most closely approach this paradigm are agent-based systems, such as are used to model
financial markets [le baron]. In such a case, one models strategies in a context where
there is no explicit pre-specified payoff function. Rather, the success of an agent strategy
is implicitly dependent on the strategies of all other agents, and therefore cannot be
calculated until at a given moment of time all the other strategies have been specified.
To understand these systems, except in the case of the very simplest strategies, such as
random trading, inevitably, a simulation has to be run. Although such systems exhibit
several features that I would claim are important to the development of complexity they
also leave several important ones out. For instance, in such systems there is no building
block hierarchy of strategies. A reason for this is that there is no real need for modularity,
i.e. no associated multi-tasking.

12. From symbolic to physical complexity

I have used throughout as paradigms of complexity human languages and biological
organisms. In section 2. I compared a text of human language to some texts of genetic
language, posing the question of whether the latter was also complex. It seems fairly
clear that it is genetic language that leads to physical complexity through development.
How then can such physical complexity be reached if the language that underlies it is not
itself symbolically complex? As argued in the case of human language, this implies that it
is complex from the demands of genetic semantics not genetic grammar. In other
words strands of DNA, on their own, in solution, say, are not complex, just as Japanese is
not complex for me, as I cannot measure the complex interactions inherent in it due to the
demands that it have meaning. My brain has not been calibrated for Japanese. As stated,
meaning comes from an interaction between a text and a brain understood as a measuring
apparatus. What is the brain for genetic texts? DNA in solution has no meaning relative
to its surroundings the solvent. However, DNA in the context of a cell does have
meaning, as is manifest by the complex activity of the cell. In this case the cell
environment acts as the equivalent of the brain, a measuring apparatus giving meaning to
what is written in the genetic texts. Once this meaning is established, then an action can

23

be taken, such as to produce a particular protein with an eventual macro-characteristic at

the level of a phenotypic trait. This is precisely the genotype-phenotype map.
Unfortunately, our understanding of it is minimal.
So, in the cell there is an instruction booklet which needs to be read and, more
importantly, needs to be understood. This instruction booklet exhibits a high degree of
complexity in the context of a cell machinery that is capable of understanding it. These
instructions are subsequently acted upon and this eventually leads to some phenotypic
traits, which in their turn can be understood as facilitating certain strategies. Thus, in the
case of birds, there are instructions for making wings and feathers and these allow the
bird to implement the evolutionarily useful strategy of flying.

13. Conclusions
In this contribution I have given a very personal account of what I believe complexity,
and the related notion of complex system, to be. Starting off with the premise that
biological systems and human languages are definitely complex I tried to determine what
non-tautological properties distinguished those systems from others. The idea was not to
provide a rigorous definition of these concepts but rather see what phenomenological
properties discriminated most. Properties such as many degrees of freedom, non-linear
interactions and Edge-of-Chaos definitely do not discriminate. Neither does the more
sophisticated concept of a hierarchy of effective degrees of freedom building blocks
as a function of scale, as simple physical systems also exhibit such hierarchies.
What does seem to distinguish the complex systems I consider is the property that there
exist emergent characteristics meaning and fitness that induce interactions across
different levels of building block and are such as to require an integration of the
contributions from the different building blocks across different levels. This is quite
distinct to physical systems where building block structure on one scale is effectively
frozen out at others. It is for that reason that the physical sciences have been so successful
when compared to the biological ones. In the physical world pretty much things are either
homogeneous, crystals etc, or random, e.g. glass or a gas. It is the stability of the physical
world that leads to a relative lack of diversity in structure and function. It takes a lot of
energy to break up nuclei, atoms, molecules etc. For complexity we need to be able to
construct a hierarchy of building blocks that is neither too stable nor too unstable. We
need low energy to do that.
We also need to avoid the tyranny of physical law. Complex systems were argued to be
characterized more by what they do rather than what they are. Physical systems, on the
other hand, are completely constrained, in a usually quite transparent way, as to what
they can do. Complex systems however, are best described in terms of different strategies
that they can implement. Put metaphorically - complex systems are verbs while physical
systems are nouns. The tyranny of physical law is strongly related to the simple

24

requirements that such systems have to obey find the state of least energy, or least
action, for example. In contrast, complex systems are characterizable by a building block
hierarchy of function wherein physical law acts only as a constraint to the possible
dynamics not as an explanation of the dynamics per se as it does in physical systems.
The escape from the tyranny of physical law has come about due to the challenge of
evolving in a complex environment. This has driven systems to develop structural
building block hierarchies that, in their turn, represent functional building block
hierarchies. Each functional building block can then be tasked with solving a part of the
puzzle of survival.
The conclusion then is that one can isolate properties of biological systems and human
languages that do seem to truly distinguish them when compared to any physical system.
In that sense we can take these properties as characteristic of complexity and complex
systems. Whether there are systems, other than the ones mentioned here remains to be
seen. Of course, the reader may disagree with my usage of the word complexity as it is
far more restrictive than previously used meanings of the word. It is, after all, just a word.
What is more important than the word, is the set of properties that I have put under the
rubric of this word, and our ability to measure and model them.

References
[1] Edmonds B. (1996): "What is Complexity?", in: F. Heylighen & D. Aerts (eds.), The
Evolution of Complexity (Kluwer, Dordrecht).
[2] Johnson, Neil F. (2007). Twos Company, Three is Complexity: A simple guide to the
science of all sciences. Oxford: Oneworld.
[3] Sol, R. V.; B. C. Goodwin (2002). Signs of Life: How Complexity Pervades
Biology. Basic Books.
[4] Goldreich, Oded. (2008) Computational Complexity: A Conceptual Perspective,
Cambridge University Press.
[5] Wentian Li (2003) Zipf's Law Everywhere, Glottometrics 5, 14-21.
[6] Shapiro, JA. (2005) A 21st Century View Of Evolution: Genome System
Architecture, Repetitive DNA, And Natural Genetic Engineering. Gene 345: 91-100.
[7] Langton, Christopher G. (1990) "Computation at the edge of chaos". Physica D, 42, .
[8] Lewin, Roger (1992). Complexity: Life at the Edge of Chaos. New York:
Macmillan Publishing Co.

25

[9] Eigen, M. (1971). "Selforganization of matter and evolution of biological

Macromolecules". Naturwissenschaften 58 (10): 465.
[10] Holmes, E.C. (2003). Error thresholds and the constraints to RNA virus evolution
Trends in Microbiology, Volume 11, Issue 12, 543-546.
[11] Bak, Per, Chao Tang, and Wiesenfeld, K., (1988) Self-organized criticality, Phys.
Rev. A 38, 364 374.
[12] Sims, K. (1994). "Evolving Virtual Creatures", Computer Graphics (Siggraph '94
Proceedings), pp.15-22.
[13] Osborne, Martin J. (2004), An introduction to game theory, Oxford University
Press.
[14] Couzin, I.D., Krause, J., Franks, N.R. & Levin, S.A.(2005) Effective leadership and
decision making in animal groups on the move, Nature 433, 513-516.

26