A Field Guide to Bifurcation Theory

James Todd
May 9, 2019

1 Introduction
One way to understand mathematics is to envision each topic as a set of
tools, where one ought to know what tool is best suited for what job; for exam-
ple, in the same way that it is easiest to pound in a nail with a hammer, it is
easiest to approach an area under a curve problem with integral calculus. With
this in mind, what follows is a brief overview of a particularly useful tool for the
analysis of functions, known as ”bifurcation theory.”

Bifurcation theory is a small but integral part of an overarching set of such

tools known as nonlinear dynamics. Nonlinear dynamics involve systems that
evolve over time, where a change in input does not necessarily correspond to
a constant multiple of the same change in the output[8]. A common way to
handle these systems that can be difficult to work with is to “linearize” them,
which allows more classical techniques to work. However, this fails to analyze
complex behaviors such as singularities (in which the behavior of the function
may be poorly defined), chaos (in which a linearization loses essentially all infor-
mation), or in the case of solitons (in which dissipative and nonlinear behavior
are balanced in some way). Clearly, a different set of tools is needed in these
types of situations, which is where bifurcation theory comes in.

2 Bifurcations and Their Properties

Small changes to function parameters in nonlinear systems can result in
so-called “bifurcations,” sudden or erratic changes in behavior that have certain
properties and follow certain rules [3]. Bifurcation theory gives us a way to de-
tect and consider these properties. In a nonlinear dynamical system, there are
many ways bifurcations can occur, and many consequences for different types of
bifurcation, which is why bifurcation theory is an interesting field of study. This
tool is ideal in that it works regardless of what causes the onset of the bifurca-
tion. Bifurcations can be classified into species based on certain properties (the
number of variables, codimension, continuity, etc.). Certain details can both
change what types of bifurcation can occur, as well as in what manner. For

1
example, some bifurcations exist in general forms in higher dimensions, where
a smaller number of variables can ‘flatten’ the effect of a bifurcation, such as a
saddle-node bifurcation shown in two dimensions (this will be explored in more
detail later). The so-called “codimension” of a bifurcation is how many param-
eters must be ‘tuned’ to a certain range for the behavior to occur. Most of the
ones discussed in this paper will be in the context of codimension-1 systems that
only require a single variable to be changed to view how the bifurcation occurs
inside a function, with the exception of the monkey-saddle bifurcation, which
is an example of a bifurcation with minimum codimension-2. Continuity is an-
other property of particular interest. As Wiens discusses, some models, such as
the Trygve Haavelmo growth model, can have different dynamical behaviours
depending on whether the model is expressed as either differential or difference
equations; solutions to the continuous version converge to a fixed point, whereas
the discrete version exhibits small-scale chaotic behavior[10]. Part of the diffi-
culty about viewing bifurcations in dynamical systems is that often the relevant
information is stored along the time axis, rather than just the spatial axes.

Figure 1: These are the types of attractors, visualized in phase space: a strange
attractor, that pulls towawrds a fractal set (top left), a point attractor that
pulls towards a single point (top right), a torus attractor (that pulls towards a
torus in 3D space), and a limit cycle (that pulls towards a cycle.) [1]

Before moving further, a couple definitions may be useful. Over time, if a

function trends toward a value, or set of values over a domain, then it is said

2
to have an attractor. The easiest type to consider is that of a “fixed-point” at-
tractor, where the trajectory of a function will eventually converge to a specific
value, and these are what are typically what are shown with bifurcation dia-
grams. Another type of attractor is that of a “limit cycle”, where a specific path
in phase space attracts, rather than a point; these are typically shown with a
Van der Pol phase portrait. Some also consider a torus attractor to be a unique
type, that pulls into complex shapes that can be described with the geometry of
a torus. Finally, perhaps the most complicated type of attractor is the mythi-
cal “strange attractor”, which is a set of attractive points described by a fractal.

3 Bifurcation Diagrams

Figure 2: This is the bifurcation diagram for the chaotic logistic map: xn+1 =
r(1 − xn ). Notice the bifurcation that occurs at r = 3, and entirely chaotic
behavior that arises later by r = 4. [9]

One way to view bifurcations is with a bifurcation diagram. Interpreting

a bifurcation diagram is one of the most major hurdles in getting a fundamen-
tal understanding of bifurcations, but it need not be. A simple, codimension-1
bifurcation diagram may have the vertical axis be the variable value (e.g. x),
whereas the horizontal axis is the parameter that is being adjusted. Each point
on the graph represents an equilibrium (not necessarily stable). A codimension-
2 diagram may be a 3D plot, with the two parameters as a flat base, and the
x axis as the vertical. The reason bifurcations are useful to express this way
is because the difference from the change in parameters becomes clearly visi-
ble. In other words, any split that appears in such a diagram is a bifurcation!
For example, in the diagram above, a bifurcation occurs at (r,x) approximately
(3, 0.7). The graph above shows a standard period doubling bifurcation. This

3
method is beautiful in its simplicity, because bifurcations are difficult to visual-
ize in a vacuum.

4 Classification of Bifurcations
As organizing things into groups is a useful way to better understand sys-
tems as a whole, bifurcations are classically divided into separate categories.
One typically divides systems by property, such as by local and global, homo-
clinic and heteroclinic, or by codimension; or by directly named species as they
were discovered in different fields, such as the saddle-node bifurcation, trans-
critical bifurcation, and pitchfork bifurcation. Some instead choose to divide
them over what type of fixed points the system has, such as equilibria, periodic
orbits, and global. However, that classification only works for codimension-1
bifurcations. Local codimension-2 bifurcations may be classified into Bautin
bifurcations, Bogdanov-Takens bifurcations, cusp bifurcations, Fold-Hopf bifur-
cations, and Hopf-Hopf bifurcations.

Figure 3: This is a critical point in a global homoclinic Shilnikov bifurcation,

where values above this the entire loop depicted here disappear. At smaller
parameter values, this loop is ”looser”, and increasing the parameter pushes the
loops tighter and tighter. [2]

The most prototypical division is to separate bifurcation types into local

or global. A local bifurcation occurs when a slight modification of a parameter
causes the stability of equilibria to change. A global bifurcation occurs when
larger parts of the function collide with equilibria by changing parameters, such
as when constricting loops tighter and tighter causes them to eventually over-
lap. Global bifurcations also give rise to crises, a phenomenon that occurs when
strange attractors appear or disappear from a dynamical system. Local bifur-
cations are much easier to study, because the change is generally less dramatic,

4
whereas global ones can causes widespread change over the entire function.

Another common division is over homoclinic versus heteroclinic bifurca-

tions. To momentarily take a step back: Heteroclinic cycles are equilibria, limit
cycles, or chaotic invariant sets joined by saddle-sink connections. Homoclinic
orbits, on the other hand, are ones which join a saddle equilibrium point to it-
self. Naturally, then, a heteroclinic bifurcation is one that contains heteroclinic
orbits linked by heteroclinic cycles, and likewise a homoclinic bifurcation is one
that contains homoclinic orbits. However, perhaps a more natural (or at least,
more informative) division is by the name the bifurcation is generally called in
literature, which is reflective of its discovery.

5 Examples of Common Bifurcations

Figure 4: The above graph is shows the normal form of the saddle-node bifur-
cation as the parameter r crosses the critical threshold. This graphic, as well as
several of the following, come from Strogatz’ work, which is definitive on this
topic. [8]

An example of a specific bifurcation is the so-called, “saddle-node bifurca-

tion.” It is described this way because when taken in 3 dimensions the resultant
shape appears to look like a saddle. It’s easiest to describe as a single stable and
unstable fixed point meeting and mutually annihilating if depicted in one dimen-
sion. The homoclinic, global version of this bifurcation is very similar. There
exist both stable and unstable fixed points on a limit cycle that disappear af-
terward, leaving a much simpler loop. There are some similar bifurcations that
some authors refer to by the same name, such as the fold, turning point or blue
sky bifurcation. This last one likely seems like the most fanciful and bizarre way
to reference this behavior; it is called this way because a new fixed point appears
“out of the clear blue sky,”–in other words, as if by magic. Worth mentioning
here is that all bifurcations have a prototypical example mathematically–or,

5
more formally, a “normal form.” This is important, because regardless of the
complexity of a system, a bifurcation of a given type can always be written in
terms of its normal form. The normal form for a saddle-node bifurcation is:
dx/dt = r ± x2 . This is the most simple type of bifurcation, and it is very com-
mon to see this behavior in addition to other bifurcations which are discussed
more later.

Figure 5: The above graph is shows the normal form of the transcritical bifur-
cation as the parameter r crosses the critical threshold. [8]

A transcritical bifurcation is another common example. Here, the stability

of a pair of fixed points flip (a previously stable point becomes unstable and a
previously unstable point becomes stable). The normal form of a transcritical
bifurcation is: dx/dt = rx − x2 . One place where this is commonly found is
the producer-consumer problem in computing (also called the bounded buffer
problem). This is a problem surrounding the idea of two processes, sharing a
common memory space, where one is creating data and adding it to the space
while the other removes it. Problems arise when data is attempted to be added
when the memory is full, or removed when the memory is empty. If the amount
consumed is a function of how much is currently in memory, then one might
model the amount of stored memory as a system involving a transcritical bifur-
cation.

Pitchfork bifurcations are a bit of a special case relative to the previous

examples in that they come in two forms–a supercritical case and a subcritical
case. In either case, pitchfork bifurcations are common in systems that have
large amounts of symmetry (meaning left-right symmetry in the go-to one di-
mensional case). An example of such a system might be a heavy weight sitting
atop a load-bearing pole. If the pole is directly below the weight’s center of
mass, it will be balanced, but a slight displacement in either direction will cause
the pole to buckle, forcing the system to into new equilibria. The supercrit-
ical case, which is simpler, is given by the normal form in one dimension as
dx/dt = rx − x3 . This system exhibits strong symmetry, because it is invariant

6
Figure 6: The above graph is shows the bifurcation diagram of the pitchfork
bifurcation in the supercritical case (top) and subcritical case (bottom). [8]

under the change of variables x=-x (this simply means that substitution of -x
in place of x does not change the system, i.e., it is exactly mirrored about y).
This is characterized by a single stable equilibria splitting apart into two stable
equilibria with an unstable equilibria between them. This type of model occurs
frequently in systems that are similar to the Lorentz model of dielectrics (which
deals with a positive and negative charge attached via a spring), or the sigmoid
layer in neural networks, which also has a “bounciness” that must be satisfied.
This system is stabilizing, meaning it works as a restorative force, whereas the
more complicated destabilizing version of this system is the subcritical pitchfork
bifurcation. The normal form for this type of bifurcation is dx/dt = rx + x3 .
Note that this equation quite clearly blows up very quickly, because the pre-
viously restorative force is now working outward. In physical models, this is
normally opposed by a very large term that (by necessity of keeping a symmet-

7
ric system) is either of the forms −x5 or −x7 , such as in the case of electron
degeneracy for small r. This results in two stable fixed points that drift away
from each other for large x. This causes a large problem in ecological systems
that use models that have this behavior, as shifting between two increasingly
disparate stable points can wildly impact large scale systems.

Figure 7: The above shows both the subcritical and supercritical case of the
Neimark-Sacker bifurcation. These are purely example cases, and not from a
specific function. The limit cycle is shown in green in the supercritical case, and
red in the subcritical case. [5]

Another bifurcation that has a supercritical and subcritical case is the

Neimark-Sacker bifurcation. Both cases appear as a function that spirals inward
inside a limit cycle. The supercritical case appears to unfurl as the parameter is
adjusted, pulling outward into looser loops. This results in the stable fixed point
in the center becoming unstable after the parameter passes through some criti-
cal value. The subcritical case tightens near the center of the limit cycle, with
loops pulling tighter and tighter until they overlap and are lost. This results in
the unstable central fixed point becoming stable as the parameter is adjusted.
This extrapolates very easily into 3D as a wrapping torus. The normal form
for this function is complex and stated differently by different authors; for that
reason, it is beyond the scope of this paper.

8
Figure 8: The above graph is a typical monkey saddle bifurcation shown in
(x,y,z) space. The particular functions that define this one are x(u, v) = u,
y(u, v) = v, and z(u, v) = u3 − 3uv 2 . [4]

The monkey-saddle bifurcation is a codimension-2 bifurcation defined by

the normal form z = Re[(x + iy)k ], where k is generally listed as 3. This is
because it takes a normal saddle bifurcation in three dimensions and adds a
third depression. By setting k, however, one can create a k-depression saddle.
It is generally referred to as a monkey-saddle bifurcation because a monkey
would require a place in a saddle that also had room for a tail (hence the third
depression).

Period doubling bifurcations are a special type of bifurcation that arise

typically from difference equations or iterative equations such as the logistic
equation (see Figure 2). The particularly unique property that these functions
have is that they have bifurcations at a doubling rate as a parameter is shifted.
At some point, these equilibria get too close together, which is the onset of more
chaotic behavior.

6 Related Topics
6.1 Chaos
Bifurcation theory is also particularly helpful in the analysis of “chaotic”
functions, another subfield of nonlinear dynamics. Chaotic systems are systems
that are extraordinarily sensitive to initial conditions. To discover how chaotic

9
Figure 9: The above graph is the Lyapunov exponent calculated for the Hénon
map (xn+1 = (1 − Cx2n + Bxn−1 ), B = 0.3, 0 < C < 2). It is important in
dynamical systems as a simplified Poincaré section of the Lorenz model, a more
complicated and useful system. The horizontal line here demarks the x axis.
Any values plotted above this are values at which the function exhibits chaotic
behavior. [7]

a dynamical system is, one might calculate the Lyapunov exponent (which is a
single numerical value that measures divergence of a function) of the function
at a specific parameter value, averaged over the attractor. By calculating many
such values while slightly varying the parameter, the result is a function where
the x is given by the parameter value and the y is given by the calculated ex-
ponent value. Where the Lyapunov exponent is positive, the function is said
to be chaotic. The function of Lyapunov values is considered to be a potential
function, calculated by a somewhat complicated Taylor-series expansion. Con-
structing a bifurcation diagram is one method of visualizing what values this
chaotic attractor causes the function to approach. An important side note here
is that the unpredictability of chaotic functions in no way makes the function
“random,” and the signal during chaotic behavior is by no means “noise.” At
all points, a chaotic function is deterministic, even if what causes it to behave
in a certain way is esoteric. Chaotic functions and noise can exist in the same
system however.

Chaotic functions are a powerful and ubiquitous tool, and have found uses
in many different fields and applications. In cryptography they are used for
hashing, streaming, watermarking, steganography, and all sorts of applications

10
Figure 10: The above graph is a typical solution to the Lorenz equations in 3D,
which forms a striking butterfly shape. [6]

that need a way to generate information that is both predictable with some
information and nearly unpredictable without it. In biology, various different
populations have been shown to exhibit chaotic behavior in their population
models. Chaotic blinking of stars caused by planets passing in front of them
have allowed astronomers to identify many planets that were previously un-
known. It’s been used to model human psychology and physiology of the brain.
One of the most visually striking chaotic systems is the Lorenz attractor that
resembles a butterfly when graphed on the (x,y,z) coordinate space. This one
is of great use in its own right in fluid dynamics, lasers, dynamos, circuits, and
chemical reactions. Out of all of bifurcation theory, chaos theory has shown the
most promise, and appears in a striking number of different fields.

6.2 Ergodic Theory

A tangentially related topic to bifurcation theory is that of ergodic theory.
At its most basic, ergodic theory considers what happens to a dynamical sys-
tem after a sufficiently long time. For example, one might consider a domain,
a function that moves from one point to the next point in the domain, and a
measure function, that determines how close those two points are in the phase
space. Over many iterations, the function may visit the same part of the do-
main repeatedly, or may get stuck in a certain region. This behavior mimics the
understanding of bifurcation theory, in that ergodic theory considers a specific
type of nonlinear dynamical system, where the behavior near the fixed points
must be necessarily convergent, but to different values depending on the initial
conditions position relative to the function’s fixed points.

Ergodic theory has applications in statistical mechanics, number theory,

11
differential geometry, and functional analysis. It has been described as a way of
avoiding getting one’s feet wet without turning to constantly gauge approach-
ing waves. In other words, some processes are difficult to judge at any exact
moment (such as the ocean), but observing long term behavior allows one to
avoid any negative consequences in the short term. In statistical mechanics, er-
godic theory finds its place in metal folding, gas mixing, and impurity reduction.
Number theorists have used it to look at lensing for prime numbers, and other
similar applications. It also has applications in computing, as with complicated
loops in cellular automata that must eventually find an end case.

6.3 Catastrophe Theory

Catastrophe theory is an application of of bifurcation theory. It considers
only the case in which the Lyapunov function is smooth and well-defined. In
particular, catastrophe theory is used to analyze degenerate critical points of
the Lyapunov function where the derivative and one or more higher derivatives
are zero. Taking the Taylor Series expansion of the Lyapunov function makes
it clear how many parameters make that point occur, and whether those are
active or passive parameters (x versus some constant c). René Thom, a famous
mathematician, categorized all of the lower dimensional catastrophes that can
occur into seven types. Similar to bifurcations, catastrophes also have normal
forms, which will be listed alongside their counterpart. The fold catastrophe
has a potential function given by: x3 + ax. The a parameter here is a tipping
point, whereafter no stable solution exists once it becomes positive. Shifting
a beyond this point can cause the function to have spontaneous and massive
value changes. The cusp catastrophe is what happens when a new parameter
is given to a fold catastrophe, and has the form x4 + ax2 + bx. For a > 0,
the system is still unstable, but for a < 0, the function can jump back and
forth in increasingly tight loops as b is increased. This leads to a wildly un-
stable function with a rigorous underlying framework. In other words, once
it reaches a critical point, even if the function falls back to lower x values, it
stays in the higher portion of the loop. Moving further, one can layer these
catastrophes together into increasingly complex frameworks referred to as the
swallowtail and butterfly catastrophes respectively, given by: x5 +ax3 +bx2 +cx
and x6 + ax4 + bx3 + cx2 + dx. These types of functions have such bizarre and
interlocked behavior that they become incredibly difficult to analyze. Finally,
there are 3 other types of catastrophe, referred to as hyperbolic umbilic, elliptic
umbilic, and parabolic umbilic catastrophes whose behavior is difficult to de-
scribe, given that they described as motions over two active variables and three
to four static parameters.

Catastrophe theory is useful for a number of applications such as structural

fracture mechanics, electron transfer between outer shells, and modelling bub-
bles in real estate, in addition to other systems in which catastrophic failure
is common. The effects of this type of catastrophic collision with black holes

12
allows physicists to model resultant gravitational waves, and other useful prop-
erties without being able to directly observe the collision.

7 Conclusion
While bifurcations mainly rear their heads in complicated or chaotic map-
pings, wildly different fields, with varying degrees of complexity, the underlying
premise is the same, and isn’t something to be afraid of. In fact, with the under-
standing given here, one can extrapolate and learn about just about any other
part of this field without difficulty, including wild and untested waters and new
territories that have yet to be explored.

References
[1] Adermei Adewumi, Jimmy Kagamba, and Alex Somto Alochukwu. Appli-
cation of chaos theory in the prediction of motorised traffic flows on urban
networks. 2016.
[2] P. Aguirre, B. Krauskopf, and H. M. Osinga. Global invariant manifolds
near a shilnikov homoclinic bifurcation, 2011.

[3] Paul Blanchard, Robert L. Devaney, and Glen R. Hall. Differential Equa-
tions. Brooks/Cole Thomson Learning, 2006.
[4] Weiqing Gu. The monkey saddle. 2003.
[5] Yuri A. Kuznetsov and Robert J. Sacker. Neimark-sacker bifurcation. 2008.

[6] Geoffrey Milburn. Lorenz and the butterfly effect. 2011.

[7] J. C. Sprott. Numerical calculation of largest lyapunov exponent, 2015.
[8] Steven H. Strogatz. Nonlinear dynamics and chaos: with applications to
physics, biology, chemistry, and engineering. Perseus Books Publishing,
1994.
[9] Muhammad Usama and Nordin Zakaria. Chaos-based simultaneous com-
pression and encryption for hadoop. 2017.
[10] Elmer G. Wiens. Egwald mathematics–nonlinear dynamics, 2003.

13