Chaotic Development of Turbulent Fluid Flow

Matthew Lee
North Carolina School of Science and Math

1. Introduction
With all the research that has been done to understand the motion of fluids, turbulence is
a concept that has evaded the comprehension of physicists. Not only is it difficult to accurately
simulate computationally, but it is also challenging to explain the theory behind the onset of
turbulence and identify the general behavior of turbulent flows. Much of the problem stems from
the nonlinearity of the systems involved, which can lead to unpredictable time evolution under
specific parameters. There are also complex tendencies in fluids, such as the dispersion of kinetic
energy to smaller and smaller length scales until released as heat due to internal stresses. These
internal stresses can be quantified as viscosity, which often complicates our understanding
depending on the assumptions we make for our system.
However, turbulence is chaotic not just in the colloquial sense but mathematically, the
key idea being a disproportionate sensitivity to the initial state. There are a plethora of other
systems that exhibit similar behavior across the sciences, many of which share a few key
characteristics. With the tools of chaos theory at our disposal, there is valuable insight regarding
periodicity and self-similarity and their parallels across certain classes of motion. Though it still
does not fully explain the conundrum at hand, chaos and the field of nonlinear dynamics is an
important framework through which we can attempt to unravel the mysteries of turbulence.

2. Fluid Dynamics
2.1. Assumptions
When working with fluid dynamics, the variable of interest is the velocity field and how
it changes with time. There are many parameters within the fluid which determine this evolution,
such as the pressure and density. To achieve a general theory, we let these vary over space and
use ideas from vector calculus to quantify physical relationships.
One of the fundamental laws of physics we utilize is the conservation of mass. In three
dimensions, the conservation of mass can be written as
 ∂ρ
− ∂𝑡
= ∇ · (ρ𝑣),

where ρ and 𝑣 are the density and velocity of the fluid. It is often a good approximation to
assume that the fluid is incompressible, where the density is constant, as long as we are working
with velocities much less than the speed of sound (Feynman). With this, the conservation of
mass reduces to
∇ · 𝑣 = 0, (2.1.1)
which is an equivalent definition for an incompressible fluid and will come up throughout the
rest of this paper. Intuitively, if the divergence of the velocity field is zero, then there is no
expansion or contraction of the field, which aligns with the idea of an incompressible fluid.
The other fundamental law of physics we consider is Newton’s second law of motion,
which we apply by considering the effect of forces on small volumes of the fluid.. There are a
number of forces to account for, which we can classify as internal if it is the result of part of the
fluid acting on another part of the fluid and external otherwise. The internal forces consist of
pressure and force due to compressibility, which are the internal normal forces, and viscosity,
which are the internal shear forces. The external forces consist of conservative forces, like that of
a gravitational or electrostatic field. There are also external non-conservative forces, such as
force due to sustained contact with a solid, which can be thought of as constraints and are not
factored into the equations of motion derived here.

2.2. Equations of Motion

The defining property of a fluid is
that any surface in the fluid will move when
a shear stress, or force per unit area parallel
to the surface, is applied (Feynman).
Consider a fluid at rest. There cannot be
any forces parallel to any surface in the
fluid or else it would not be at rest. If we
consider a small unit cube of volume in the
fluid, we only need to consider the pressure
𝑝, or normal force per unit area, on opposite faces. Without loss of generality, we can define a
rectangular coordinate system with directions along the edges of the cube. The net force on the
cube can be calculated by summing the forces on two opposite faces that are perpendicular to
each coordinate direction (Figure 1):
∂𝑝 ∂𝑝
Σ𝐹𝑥 = (𝑝0)(∆𝑦∆𝑧) − (𝑝0 ∂𝑥
∆𝑥)(∆𝑦∆𝑧) = − ∂𝑥
∆𝑥∆𝑦∆𝑧
∂𝑝 ∂𝑝 ∂𝑝
Σ𝐹 = (− ∂𝑥
∆𝑥∆𝑦∆𝑧, − ∂𝑦
∆𝑥∆𝑦∆𝑧, − ∂𝑧
∆𝑥∆𝑦∆𝑧)

Therefore, the force per unit volume ∆𝑥∆𝑦∆𝑧 due to pressure is simply − ∇𝑝.
Suppose there are conservative forces acting on a fluid that is not necessarily at rest. For
the sake of simplification, these are usually written in terms of ϕ, the potential energy per unit
mass that causes such forces. By the work-energy theorem and the fundamental theorem of
calculus, the force per unit mass is equal to − ∇ϕ, which implies that the force per unit volume
due to conservative forces is − ρ∇ϕ.
For everything in our equations of motion to be in terms of the velocity field, we must
rewrite the acceleration vector. This is not as simple as taking the time derivative of the velocity
vector at every point since this ignores how the velocity changes with space. Remember that we
are applying Newton’s second law to a small unit volume of the fluid, which will move to
locations with different velocity vectors. By the multivariable chain rule,
𝑑𝑣(𝑡,𝑥,𝑦,𝑧) ∂𝑣 𝑑𝑥 ∂𝑣 𝑑𝑦 ∂𝑣 𝑑𝑧 ∂𝑣 𝑑𝑡 ∂𝑣 ∂𝑣 ∂𝑣 ∂𝑣
𝑎= 𝑑𝑡
= ( ∂𝑥 𝑑𝑡
+ ∂𝑦 𝑑𝑡
+ ∂𝑧 𝑑𝑡
+ ∂𝑡 𝑑𝑡
) = 𝑣𝑥 ∂𝑥
+ 𝑣𝑦 ∂𝑦
+ 𝑣𝑧 ∂𝑧
+ ∂𝑡

∂𝑣
𝑎 = 𝑣 · ∇𝑣 + ∂𝑡
.

Plugging this into Newton’s second law and dividing by some unit volume 𝑉, we see that
(𝑚 · 𝑎)/𝑉 = ρ · 𝑎 = 𝐹/𝑉
∂𝑣
ρ(𝑣 · ∇𝑣 + ∂𝑡
) = − ∇𝑝 − ρ∇ϕ (2.2.1)
∂𝑣
∂𝑡
+ 𝑣 · ∇𝑣 = − ∇𝑝/ρ − ∇ϕ. (2.2.2)

Note that we have left out the forces due to compressibility and viscosity, which will be
discussed in section 2.4.
2.3. Vorticity
Another way to think about fluid flow is through the curl of the velocity field, known as
the vorticity. We define the vorticity as the vector field Ω where
Ω = ∇ × 𝑣.
Through manipulation with vector calculus identities, (2.2.2) can be rewritten in terms of Ω as
∂𝑣 1 2
∂𝑡
+ Ω ×𝑣 + 2
∇𝑣 = − ∇𝑝/ρ − ∇ϕ.

Assume an incompressible fluid, where ρ is constant. If we take the curl of both sides, every
gradient term vanishes and we are left with
∂Ω
∂𝑡
+ ∇ × (Ω × 𝑣) = 0.

There is no longer any information about the density or pressure of the fluid, but we have
reduced our system to the following system of equations:
∂Ω
∂𝑡
+ ∇ × (Ω × 𝑣) = 0 (2.3.1)

Ω= ∇ ×𝑣
0=∇ ·𝑣
Since the divergence and curl uniquely determine a vector field, the second and third equations
can be used to find the velocity field given the vorticity. The first equation gives us an expression
for how the vorticity changes over time due to the velocity field. From the initial conditions, this
can be used to find the vorticity after some small increment of time ∆𝑡, which can be used to find
the velocity field, which can be used to find the vorticity after another small increment of time.
This process can continue to determine the flow at all times, which is our main motivation for
doing all of this (Feynman).

2.4. Viscosity and Compressibility

Before static equilibrium is reached, a fluid in motion can experience shear forces on
surfaces in the fluid. These are described by viscosity, a measure of internal friction. The basis
for our theory will be one of the foundational principles of fluid dynamics, which was developed
experimentally: the velocity of a fluid is exactly zero at the surface of a solid relative to it
(Feynman).
 As a model, consider water between two horizontal plates. Suppose we move the upper
plate with a force 𝐹 so that it moves with velocity 𝑣 (Figure 2). The water between the two plates
would have a velocity of zero at the bottom plate and 𝑣 at the top plate. In this case, there
certainly seems like some
shearing frictional force is
acting parallel to successive
layers of water as you go
down the fluid region. It
turns out that the force
required to balance this
frictional force and keep the
upper plate moving is
proportional to the area of the
plates 𝐴 and 𝑣/𝑑, where 𝑑 is
the separation between the
plates. This gives the
following relation:
𝐹/𝐴 = η𝑣/𝑑
This can be generalized by considering a very small layer of fluid between two other layers of
fluid (Figure 2). The velocity would then be given by the relative difference in velocity between
the top and bottom layers of fluid. By approaching an infinitesimal scale, this relation becomes
∂𝑣
𝐹/𝐴 = η ∂𝑑

The constant of proportionality η is known as the coefficient of viscosity. Because this considers
forces 𝐹 perpendicular to the surface 𝐴 that maintain velocities 𝑣 perpendicular to displacement
𝑑, this can be interpreted as a proportional relationship between shear stress and the rate of
change of shear strain (Feynman).
In compressible fluids, there is also a force to account for expansion and contraction of
volumes in the fluid. This is described by internal normal stress (separate from pressure), which
is proportional to the divergence of the velocity field. To account for this we define a new
proportionality constant η', sometimes called the “second coefficient of viscosity,” and δ𝑖𝑗, which

returns 1 if 𝑖 = 𝑗 and 0 otherwise. Together with the general stress tensor for the viscous forces,
we can derive a complete description of the forces per unit volume in the fluid (Feynman):
∂𝑣 ∂𝑣𝑗
𝑆𝑖𝑗 = η( ∂𝑥𝑖 + ∂𝑥𝑖
) + η'δ𝑖𝑗(∇ · 𝑣)
𝑗

3 ∂𝑆𝑖𝑗 3 ∂𝑣 ∂𝑣𝑗
(𝑓𝑣𝑖𝑠𝑐)𝑗 = ∑ ∂𝑥𝑖
= ∑ η( ∂𝑥𝑖 + ∂𝑥𝑖
) + η'δ𝑖𝑗(∇ · 𝑣)
𝑖=1 𝑖=1 𝑗

2
𝑓𝑣𝑖𝑠𝑐 = η∇ 𝑣 + (η + η')∇(∇ · 𝑣)

Adding these forces to equation 2.2.1 gives

∂𝑣 2
ρ( ∂𝑡
+ 𝑣 · ∇𝑣) = − ∇𝑝 − ρ∇ϕ + η∇ 𝑣 + (η + η')∇(∇ · 𝑣) (2.4.1)

which is the general Navier-Stokes equation for a compressible viscous fluid. For an
incompressible viscous fluid, we can use equation 2.1.1 and simplify to get:
∂𝑣 ∇𝑝 η 2
∂𝑡
+ 𝑣 · ∇𝑣 = − ρ
− ∇ϕ + ρ
∇𝑣 (2.4.2)

3. Observations
3.1. Vortex Lines
Another way to think about fluid flow is through vortex lines. Similar to gravitational,
electric, and magnetic field lines, vortex lines are always tangent to the vorticity at any given
point and are denser near
regions of higher
vorticity. Additionally,
since vorticity is the curl
of a vector field, its
divergence everywhere is
zero. This implies that
there are no areas that
“produce” or “remove”
vorticity and, more
importantly, that vortex lines never start or stop. They always form closed loops, or do so in the
infinite limit. Recall the system of equations that governs the flow of an incompressible
non-viscous fluid from section 2.3. With our model of vortex lines, equation 2.3.1 can be
interpreted geometrically as saying that vortex lines move with the fluid. Essentially, they will
trace through the same bits of fluid as those bits move according to the velocity field (Feynman).
Consider a region of vorticity, modeled by a small cylinder of rotating fluid that is not
necessarily incompressible. By definition, the vortex lines here will run parallel to the axis of the
cylinder. If this cylinder were to shrink in diameter, perhaps due to motion along its axis that
pulls its surface inwards, then the vortex lines will move closer together accordingly. With a
higher density of vortex lines, the vorticity will have increased. This phenomenon is known as
vortex stretching (Figure 3) and is a behavior that has been studied in experiments on turbulent
flows because of how it demonstrates a change from larger to smaller scales of motion. (Johnson,
2021)

3.2. Energy Cascade

Though we have yet to define turbulence rigorously, one of the most common behaviors
that has been observed in turbulent systems is an increase in energy dissipation, known as the
energy cascade. Generated kinetic energy is dispersed to smaller scales of motion until viscous
forces are able to dissipate the energy into heat, since they act as the frictional force for fluids.
Vorticity does not directly contribute to the energy cascade since viscosity is a shear force and
does not resist rotation, but it can do so indirectly through the reduction of scale from vortex
stretching. Another process that has been found to pass down energy is strain self-amplification.
High-velocity fluid enters regions of strong compression quicker than low-velocity fluid, which
increases the compression further. This also reduces scale as the region of compression is
squeezed tighter and the kinetic energy of the fluid encompasses less space (Johnson 2021).
Recent research has found a more precise description of the transfer of energy between
large and small scales. There are five contributions to this energy cascade rate. The first two are
vortex stretching and strain self-amplification at the specified scale. The next two are their
multiscale analogs, that being the amplification of small-scale strain and stretching of small-scale
vorticity due to large-scale strain. Finally, there is the distortion of small-scale vorticity
covariance, which simulations have found can be practically omitted. Numerically calculated
turbulent solutions to the Navier-Stokes equation have shown that the four interactions stated
above contribute to the energy cascade in a ratio of around 3:1:2:2 in the order they were stated.
It has also been found that there is a tendency for an increase in the rate of vortex stretching and
strain self-amplification in turbulent flows, which increases the complexity of the velocity field
(Johnson 2021).

3.2. Reynolds Number

Consider incompressible flow of a viscous fluid past a long cylinder of diameter 𝐷, with
the boundary conditions that the velocity at large distances is 𝑉 and zero at the surface of the
cylinder. Similar to the process in section 2.3, we can disregard some of the information about
the fluid while still describing the flow by taking the curl of our equation of motion, which in
this case is equation 2.4.2. If we do so, we are left with:
∂Ω η 2
∂𝑡
+ ∇ × (Ω × 𝑣) = ρ
∇Ω (3.2.1)

If we use a coordinate system with the z-axis parallel to the axis of the cylinder and the x-axis in
the direction of 𝑉, then the following equations define the system:
Ω= ∇ ×𝑣 (3.2.2)
0=∇ ·𝑣 (3.2.3)
2 2 2
𝑣𝑥 = 𝑣𝑦 = 𝑣𝑧 = 0 for 𝑥 + 𝑦 = 𝐷 /4 (3.2.4)
2 2 2
𝑣𝑥 = 𝑉𝑥, 𝑣𝑦 = 𝑣𝑧 = 0 for 𝑥 + 𝑦 + 𝑧 ≫ 1 (3.2.5)

This system is described in terms η, ρ, 𝐷, and 𝑉. However, by rescaling our spatial coordinates
and our unit of time in terms of 𝐷 and 𝑉, we can eliminate them from the boundary conditions.
With the substitutions
𝑥𝑖 = 𝑥𝑖'𝐷 (which implies ∇ = ∇'/𝐷) and

𝑡 = 𝑡'𝐷/𝑉,
equation 3.2.2 becomes
Ω = ∇ × 𝑣 = (∇'/𝐷) × (𝑉𝑣') = (𝑉/𝐷)(∇' × 𝑣') = (𝑉/𝐷)Ω'.
Along with the rest of the equations, we end up with the following system of equations:
 ∂Ω' η 2
∂𝑡
+ ∇' × (Ω' × 𝑣') = ρ𝑉𝐷
∇' Ω'

Ω = ∇' × 𝑣'
0 = ∇' · 𝑣'
2 2
𝑣𝑥' = 𝑣𝑦' = 𝑣𝑧' = 0 for 𝑥' + 𝑦' = 1/4
2 2 2
𝑣𝑥 = 1, 𝑣𝑦 = 𝑣𝑧 = 0 for 𝑥' + 𝑦' + 𝑧' ≫ 1

This substitution collapses all of the independent parameters in the equation of motion into a
single factor, η/ρ𝑉𝐷, meaning we can describe the flow of a fluid in terms of one parameter. The
convention is to define 𝑅𝑒 = ρ𝑉𝐷/η, leaving the following equation of motion:
∂Ω 1 2
∂𝑡
+ ∇ × (Ω × 𝑣) = 𝑅𝑒
∇Ω (3.2.6)

𝑅𝑒 is known as the Reynolds number and turns out to be dimensionless (Feynman).

The Reynolds number is a critical development in fluid dynamics. Because it
single-handedly determines the flow of any incompressible fluid, any two systems will exhibit
the same behaviors as long as they have the same Reynolds number. This can be achieved, for
instance, by keeping the density and coefficient of viscosity constant and varying the fluid
velocity to compensate for a change in relative size. This is the idea behind wind tunnels, where
a small-scale model of some flow can accurately mirror a much larger flow by adjusting the wind
speed, as long as the spatial and temporal coordinates are scaled accordingly. There is a
limitation to this theory that must be taken into account for fluid velocities near or greater than
the speed of sound since this will make incompressibility a poor approximation. In these cases,
both the Mach number and the Reynolds must be the same for two fluids to behave in the same
way, where the Mach number is given by the ratio of 𝑉 to the speed of sound (Feynman).

3.3. Variation of Flow

For the system described in the last paragraph, different ranges of Reynolds number
produce quantifiably different flows. First, note that we can rewrite equation 3.2.6 as
∂Ω 2
(𝑅𝑒)( ∂𝑡
+ ∇ × (Ω × 𝑣)) = ∇ Ω.
 2
When 𝑅𝑒 is small, the equation of motion simplifies to ∇ Ω = 0. Without the time derivative
term, the flow is steady, which means that it does not vary with time (Figure 4a). Equivalently,
∂𝑣/∂𝑡 = 0. For a sphere in such a flow, Stokes found that the drag force is equal to 6πη𝑟𝑉,
where 𝑟 is the radius of the sphere (Feynman).
When 𝑅𝑒 is somewhat greater than 1, there seems to be the spontaneous generation of a
pair of circulating vortices behind the cylinder (Figure 4b). When 𝑅𝑒 is greater than 40, the
vortices get long enough to break off from each other and there is the creation of a new vortex
further down in the flow when the fluid curls back on itself. This process repeats itself creating
what is known as a Kármán vortex street, made up of vortices with alternating vorticity (Figure
5c). The flow is now unsteady but laminar, or generally smooth, varying with time in a regular,
periodic manner. As 𝑅𝑒 reaches several hundred, the velocity gradient between the stationary
flow near the surface of the cylinder and the fast-moving fluid around it increases, leaving less
and less time for the vorticity caused by this gradient to diffuse. This tight layer of vorticity
around the cylinder eventually becomes irregular and there is a transition from laminar behavior
to a sort of pseudo-periodicity to complete turbulence (Figure 4d, 4e) (Feynman).
4. Chaos Theory
4.1. Logistic Mapping
It turns out that this sort of descent from predictable to unstable behavior is not unique to
turbulence or even to the field of physics as a whole. In the pursuit of a general theory, consider a
simple mathematical model for a population of some sort. Suppose it starts initially with 𝑁0

members, increases if the population is small, and decreases if the population is too large for the
environment to support. A simple quadratic recurrence relation can quantify this, such as
𝑁𝑛+1 = 𝐴𝑁𝑛(𝑁𝑚𝑎𝑥 − 𝑁𝑛)

where the population updates after every iteration of 𝑛 based on the current population. 𝐴 is just
a scaling factor, perhaps indicating a sort of reproductive rate. The equation can be normalized so
that the variable is instead a proportion of the maximum population,
𝑥𝑛+1 = 𝐴𝑥𝑛(1 − 𝑥𝑛), (4.1.1)

which is called the logistic map, and is a useful form of the relation since the domain is
0 ≤ 𝑥𝑛 ≤ 1 (Hilborn, 1994).

We can make a few observations through inspection. For 𝑥𝑛 close to 0, the population

increases in proportion to itself (𝑥𝑛+1 ~ 𝐴𝑥𝑛) while for 𝑥𝑛 close to 1, the population decreases in

proportion to itself (𝑥𝑛+1 ~ 𝐴(1 − 𝑥𝑛)). The evolution of the population, which we can call the

trajectory of the system, depends on the value of the parameter 𝐴. For instance, with 0 < 𝐴 < 1
, all starting values 0 ≤ 𝑥𝑛 ≤ 1 will result in 𝑥𝑛 → 0 as 𝑛 → ∞. Here, 𝑥𝑛 = 0 is considered an

attractor with a basin of attraction of 0 ≤ 𝑥𝑛 ≤ 1. For 1 < A < 3, 𝑥𝑛 = 1 − 1/𝐴 becomes the

attractor for 0 < 𝑥𝑛 < 1. Both 𝑥𝑛 = 0 and 𝑥𝑛 = 1 − 1/𝐴 are fixed points, that is, values that

don’t change when applying the recurrence relation because 𝐴𝑥𝑛(1 − 𝑥𝑛) = 𝑥𝑛. As the

parameter 𝐴 increases from 3, there is no longer one population that is approached as we

continually iterate. Instead, no matter what initial population we start with between 0 and 1, 𝑥𝑛

will eventually oscillate between the same two population values as long as 𝐴 is fixed. Even
more interesting is that starting at 𝐴 = ~3. 449, the population approaches an oscillation
between four population values. The period of oscillation continues to double at an exponential
rate until 𝐴 = ~3. 5699 when there is no longer any discernible period for the trajectory of the
population (Hilborn, 1994).

4.2. Onset of Chaos

In certain ranges of parameter values, the
population approaches not one but a repeating
sequence of many values. We can call this sequence a
limit cycle, an oscillatory trajectory that is reached in
the limit. Initially, for 𝐴 < 3, there was just one
value that the population approached. As soon as we
had a value of 𝐴 > 3, we saw what we could
describe as period-2 behavior where the limit cycle
contained two values. For 𝐴 > ~3. 449, we had
period-4 behavior, and so on and so forth. At each of
these 𝐴 values where the period of oscillation
suddenly doubles, we say that a period-doubling
bifurcation occurs (Hilborn, 1994).
This is one way that we can transition into
chaotic behavior, though others will be discussed in
section 5. As we vary one of the parameters of a
system, successive period-doubling bifurcations
occur at an exponential rate until we have trajectories
with effectively infinite periods that will take on a
continuous range of values. Such chaotic behavior
appears completely random and aperiodic as if
disturbed by external noise, even though the
underlying mathematical model describing the
system does not have many degrees of freedom. This is because the equations are nonlinear,
with the dynamic variables in nonlinear terms, which essentially implies that changes in the input
are not proportional to changes in the output. Because of this, chaos theory is often used
synonymously or in direct relation with the field of nonlinear dynamics (Hilborn, 1994).

4.3. Properties of Chaos

With the terminology established, we can finally begin classifying chaos. The defining
characteristic of a system that exhibits chaotic behavior is the divergence of nearby trajectories.
Two nearly identical initial states, no matter how close in phase space they may be to each other,
will quickly end up evolving in different ways as time goes on. In fact, the difference in their
trajectories will increase exponentially until they are entirely unrecognizable to each other
(Figure 7). This is one way to distinguish a chaotic system from one that is just inherently
complex or noisy since the latter will result in a cluster of nearby trajectories indicative of local
variation rather than a broad, smeared range of trajectories indicative of randomness and
aperiodicity. Keep in mind that this concept of sensitivity to initial conditions is distinct from
spontaneous bifurcations due to a change in parameter, though they are both observed in
nonlinear systems (Hilborn, 1994).
A key consequence of this nature of chaos is that it is fundamentally impossible to predict
the trajectories of such systems. Since even the smallest imprecision will be quickly amplified
and completely alter the course of the trajectory, it would require infinitely precise initial
conditions to know the evolution of a system in the long-term. This is infeasible not only because
measurements always have some level of imprecision but also because most nonlinear systems
do not have a closed form solution and therefore must be approximated with numerical methods.
However, these systems are deterministic, meaning that knowing the initial state is enough to
know exactly the trajectory that follows from it. The result is another characteristic of chaos,
irregular behavior derived from simple and deterministic equations (Hilborn, 1994).
With the period-doubling cascade into chaos, there is motivation to visualize how the
limit cycle changes with respect to the varying parameter. To that end, we can construct a
bifurcation diagram, a plot of some characteristic of a system against the parameter. For the
logistic mapping, this could be the population values in the limit cycle plotted against 𝐴, but this
can vary depending on the system of study. This bifurcation diagram would show a continuous
line from 𝐴 = 0 to 𝐴 = 3, at which point the line would split into two branches as there are now
two points instead of one in the limit cycle. At each period-doubling bifurcation, each branch
splits into two more branches until they fully take up the entire output space past around
𝐴 = ~3. 5699. This is when the trajectory never seems to repeat itself and there is not a finite
limit cycle that is approached (Hilborn, 1994).
Within this bifurcation diagram is a third characteristic of chaos, self-similar structure.
The divergence of nearby trajectories is exponential, so there should be some common ratio that
quantifies the magnitude of the descent to chaos. In the logistic map, Feigenbaum found two
such ratios, one describing the distance between successive bifurcation points and one describing
the “amplitude” of the limit cycle (defined more precisely as the distance from 1/2 to the nearest
value in the limit cycle). These constants are known as the Feigenbaum δ and α respectively and
quantify the scaling of the branches of the bifurcation diagram, since each bifurcation “looks”
the same as the previous one but at a smaller scale. It turns out that this fractal-like behavior is
found in other chaotic systems, and some even share the same proportionality constants (Argyris,
1993).
This leads us to the most important aspect of chaos theory, which motivates the study of
the field in the first place. Chaos is universal, and many features that show up in one particular
system will show up in other systems. For instance, the Feigenbaum constants turn out to be the
same in certain other chaotic systems despite their differences in physical configuration. With
this, we can apply the techniques for analysis that we develop in chaos theory to all sorts of
complex systems across math, science, and even the humanities (Hilborn, 1994).

5. Turbulent Flow
5.1. Routes to Chaos
In the pursuit of explaining the chaotic nature of turbulence, a quantitative model of the
transition, or route, from laminar to turbulent flow was highly sought after. One of these early
routes to chaos was developed by Lev Landau. According to his theory, an stationary initial state
becomes unstable as the Reynolds number increases. Upon becoming unstable, there is the
creation of a limit cycle in the phase space of the system, which corresponds to periodic motion
of a certain frequency. When dealing with fluid mechanics, we often talk about trajectories
through phase space rather than output space since it is a useful way to view the properties of the
system as a whole. Landau goes on to state that continuing to increase the Reynolds number will
superimpose another periodic motion of another frequency onto the existing trajectory. Further
increase of the Reynolds number will impose more and more frequencies to the evolution of the
system until its trajectory is too complex and becomes reminiscent of turbulent behavior.
Landau’s route to chaos would eventually be proven inaccurate as such classes of motion were
not found to be sensitive to initial conditions and do not entail mixing in the phase space.
However, this idea of a trajectory made up of infinite frequencies provokes a similar idea as the
trajectory with an irreconcilable period due to the period-doubling route to chaos (Argyris,
1993).
Another route that is worth discussing is chaos from intermittency, which is when regular
motion is occasionally and sporadically interrupted by irregular motion. For fluid dynamics, this
corresponds to turbulent outbreaks in stretches of laminar flow. What can happen is that a change
in the parameter causes the period of time that laminar flow dominates to become shorter and
shorter until the entire flow is dominated by turbulence. Of course, there are other mathematical
models for chaos that utilize different types of bifurcations besides period-doubling bifurcations,
but that will be left for future research (Argyris, 1993).
Before moving on to physical models, notice that the chaos that has been described was
with respect to time. Nearby trajectories diverge as time changes, so there is temporal irregularity
in the system. However, there can also be spatial chaos not only as a fractal-like structure in a
bifurcation diagram but as the physically disorderly flow patterns that are present in turbulence.
There is the mixing of flow lines, repeated stretching and folding that can result in vortex
formation, for example. The analysis of chaos with respect to space is usually disregarded when
analyzing that with respect to time, but both are aspects that exist in physical systems (Argyris,
1993).

5.2. Taylor-Couette Flow

Consider two thin concentric cylinders, oriented with their axes in the vertical direction,
with a fluid between them. Similar to how a fluid between two horizontal plates will flow when
one of them moves relative to the other, turning the cylinders causes the fluid to begin rotating
around the axis of the cylinders. The resultant flow is known as Taylor-Couette flow and
demonstrates complex characteristics despite being a relatively simple model.
When the inner or outer cylinder is turned slowly, then we see a velocity gradient similar
to the case of two horizontal plates. The flow is laminar, steady, and well-defined (Figure 8a).
Increasing the speed of the inner cylinder past a threshold causes the fluid to suddenly break into
horizontal bands (Figure 8b). However, turning the outer cylinder past this threshold speed does
not have the same result and instead leads to a similar symmetric laminar flow as when the inner
cylinder is spun slowly. This is despite the fact that the torque required to keep the cylinders
rotating at constant angular velocity relative to each other is the same if the inner or outer
cylinder is the one in motion. The reason for this is that the layers of fluid with higher velocity
are those in contact with the cylinder in motion, and the centrifugal force is more effective on
these layers. When the outer cylinder is rotating, the fluid layers far from the axis are inclined to
move outwards, which builds up pressure against the outer cylinder but does not disturb the
equilibrium of the velocity gradient. On the contrary, when the inner cylinder is rotating, the
fluid layers close to the axis are inclined to move outwards which forces the outer layers to shift
inwards. This
process creates
circulating
convection cells,
analogous to
those that form
when warmer air
sits below cooler
air and wants to
rise, pushing the cooler air out and downwards. Geometrically, these are toroidal vortices and
can be thought of as the oriented solid of revolution formed if a circulating vortex of fluid is
revolved about the axis of the cylinder. Adjacent toroidal vortices spin in opposite directions and
generally do not mix fluid with each other (Feynman) (Nemri, 2016).
As we continue to increase the speed of the inner cylinder, the amount of horizontal
bands increases, and then they start to become wavy past another threshold (Figure 8c). This is
often referred to as wavy vortex flow since the vortices are essentially oscillating as you trace
around the circumference of the cylinder. These waves travel around the cylinder with a speed
that is around 1/3 that of the inner cylinder and do entail mixing between layers. Further
increase past a third threshold leads to modulated wavy vortex flow, where another frequency is
superimposed on top of the vorticial oscillation. More frequencies are overlaid, though not
necessarily one at a time, and there is also the generation of turbulent spots in the flow.
Eventually, the superimposed turbulent flow overtakes the regular oscillation frequencies of the
vortices and the entire flow is left in turbulence (Feynman) (Nemri, 2016).
When the cylinders rotate in opposite directions, there is another interesting behavior that
develops. From flat circumferential motion, the flow pattern starts to break up into wavy regions
and quieter regions that trace out a spiral along the surface of the cylinder (Figure 8d). These
quiet regions are turbulent, but as we increase the angular velocities of the cylinders the wavy
regions also become chaotic until the entire flow is turbulent (Feynman).
5.3. Rayleigh–Bénard Convection
Consider a fluid between two large horizontal plates, with the top one kept at a colder
temperature than the bottom one and a constant temperature difference between them. For small
differences in temperature, the fluid will be viscous enough to disperse the heat and resist the rise
of the lower warmer fluid,
maintaining a state of rest. In
this case, the temperature
gradient between the layers is
linear. For large enough
differences in temperature,
however, gravity will
overcome viscosity and pull
the cool dense fluid
downwards causing the
formation of convection cells
called Rayleigh–Bénard cells (Figure 9). The circulation is steady until the temperature
difference exceeds a critical value, at which point the flow will become unsteady and vary with
time, settling into turbulence (Hilborn, 1994).
Lorenz analyzed this system by making an approximation to simplify the partial
differential equations of motion into three ordinary differential equations in terms of three
variables 𝑋, 𝑌, and 𝑋:
𝑑𝑋
𝑑𝑡
= σ(𝑌 − 𝑋)
𝑑𝑌
𝑑𝑡
= 𝑟𝑋 − 𝑋𝑍 − 𝑌
𝑑𝑍
𝑑𝑡
= 𝑋𝑌 − 𝑏𝑍

These variables do not represent spatial coordinates, rather they encapsulate the intensity of
convection, difference in temperature between rising and falling fluid, and deviation from the
linear temperature gradient respectively. The three parameters σ, 𝑟, and 𝑏 are positive
dimensionless constants. σ is the Prandtl number, the ratio between energy dissipated by
viscosity to energy dissipated by heat conduction; 𝑟 describes the vertical temperature difference
through the Rayleigh number; 𝑏 is related to the geometry of the convection cells. The parameter
that is typically varied to observe the route to turbulence is 𝑟, while σ and 𝑏 are kept constant
(Vulpiani, 2010).
We will assume, as is prototypical for introduction, σ = 10 and 𝑏 = 8/3. For 𝑟 < 1, the
point (𝑋, 𝑌, 𝑍) = (0, 0, 0) is an attractor with all of 𝑋𝑌𝑍 space as its basin of attraction. This
corresponds to the steady unmoving flow that occurs for low temperature differences. At 𝑟 = 1,
there is a bifurcation in behavior that results in two stable stationary attractors,

(± 𝑏(𝑟 − 1), ± 𝑏(𝑟 − 1), 𝑟 − 1), which correspond to the two directions that the
convection cells can circulate. The (0, 0, 0) stationary point becomes unstable, so trajectories
tend to move away from it and towards the two stable attractors. There is another bifurcation at
𝑟 = 13. 927... which results in the spontaneous creation of unstable limit cycles around the
nontrivial stationary points.
Since they are unstable and
repel trajectories, a trajectory
that begins within these limit
cycles will remain within them
and approach the stable
attractors. For trajectories
outside of the limit cycles, they
will be unable to fully attract to
the stable attractors but will
instead oscillate around. In fact,
trajectories that are initially
repelled from the limit cycle
around one attractor will switch over and circle around the limit cycle of the other attractor.
Beyond 𝑟 = 24. 74, these unstable limit cycles collapse into the stable attractors, which become
unstable. Though the trajectories no longer approach the two nontrivial stationary points, they do
approach a new attractor infamously known as the Lorenz attractor. Trajectories now circle
around an unstable point, cross over to circle around the other unstable point, and then return,
repeating the process over and over again. They will never get close to the stationary points, but
they will also never diverge and instead remain bound within a region in the 𝑋𝑌𝑍 phase space
(Figure 10). The Lorenz attractor is an example of a strange attractor, where the attracting set of
points has a fractal structure and displays self-similarity in some way. They typically exhibit
chaotic behavior, as indicated by the divergence of nearby trajectories, and are difficult to
understand beyond a numerical analysis (Vulpiani, 2010).

6. Conclusion
Though they are often considered distinct fields of study, turbulence and systems that
exhibit turbulent flow under certain conditions exist in an overlap between fluid dynamics and
chaos theory. It is not straightforward to define turbulence without the framework of chaos, and
there is already enough nuance in the mechanics of vorticity and viscosity without adding on
even more mathematical rigor. However, the Navier-Stokes equations are not sufficient to grasp
what exactly turbulence is and how it comes about without concepts such as the exponential
divergence of nearby trajectories or bifurcations that alter the asymptotic behavior of a system.
There are several factors to consider when analyzing the equations of motion that
describe fluid dynamics, such as the density, compressibility, and viscosity. It turned out that
assuming incompressibility was not only generally valid for most systems but also collapsed all
of the factors into a single parameter known as the Reynolds number. Even varying this one
parameter for a simple system like flow around a cylinder resulted in dramatic, discontinuous
changes in the generated flow pattern. Such spontaneous changes can be classified into different
types of bifurcation depending on the change in behavior, such as the creation of a limit cycle
around an attractor. With the theory of bifurcations, we can think of different mechanisms that
disrupt laminar flow to produce turbulence and examine how they manifest in physical systems.
Understanding turbulence is important in accurately modeling all sorts of naturally
occurring aerodynamic and hydrodynamic systems. From atmospheric convection currents to
ocean tides, fluids are everywhere and often do not move with predictable laminar flow. Weather
forecasting, for instance, involves the analysis of an inherently chaotic system and requires
precise consideration of the behavior of turbulent flows in order to come up with accurate
predictions. It is also important to be able to quantify the transition to chaos in order to model
man-made systems, such as the generation of lift on the wing of an airplane. Without proper
simulations of air flow that account for turbulence, there would be a catastrophic lack of control
over all sorts of aerial vehicles (Jolie Duquene, personal communication, March 2023). Due to
the dependence of behavior on the dimensionless Reynolds number and the universality of chaos,
further observation of reference systems could prove vital in advancing our current models for
fluid systems.
 Works Cited

Argyris, J., Faust, G., & Haase, M. 1993. Routes to Chaos and Turbulence. A Computational
Introduction. Philosophical Transactions: Physical Sciences and Engineering, 344(1671),
207–234. http://www.jstor.org/stable/54163
Feynman, Richard P., et al. California Institute of Technology, 1963, The Feynman Lectures on
Physics, vol. 2. https://www.feynmanlectures.caltech.edu/info/
Hilborn, Robert. 1994. Three Chaotic Systems. Pages 3-41 in Chaos and Nonlinear Dynamics:
An Introduction for Scientists and Engineers. Oxford University Press. NY.
Johnson, Perry. 2021. The squeezes, stretches, and whirls of turbulence. Physics Today 74, 46-51
(2021). https://doi.org/10.1063/PT.3.4725
Nemri, Marouan, et. al. 2016. Mixing and axial dispersion in Taylor–Couette flows: The effect of
the flow regime. Chemical Engineering Science, Volume 139, pg. 109-124,
https://doi.org/10.1016/j.ces.2015.09.022.
Vulpiani, Angelo, et al. Chaos: From Simple Models To Complex Systems. World Scientific,
2010. EBSCOhost,
https://login.proxy122.nclive.org/login?url=https://search.ebscohost.com/login.aspx?direct=tr
ue&db=nlebk&AN=340555&site=ehost-live