Turbulent Flows: General Properties

“I am an old man now, and when I die and go to heaven there are
two matters on which I hope for enlightenment. One is
relativity/quantum mechanics/quantum electrodynamics [in various versions],
and the other is turbulent motion of fluids. About the former I am rather optimistic.”
Werner Heisenberg and/or Horace Lamb

“If you thought that science was certain —

well that is just an error on your part.”
Richard Feynman

Turbulence is ubiquitous in natural fluids: atmosphere, ocean, lakes, rivers, Earth’s interior,
planetary atmospheres and convective interiors, stars, space gases (neutral and ionized), and per-
haps even galaxies. From a mathematical perspective, its essential behaviors arise because the
governing equations of fluid dynamics are a nonlinear partial differential equation (PDE) system,
the Navier-Stokes equations. Apart from exotic materials with a non-Newtonian molecular rheol-
ogy and diffusion and/or peculiar equation of state, the validity of Navier-Stokes is not in serious
dispute, albeit the nature of its solutions is still open to imaginative conjectures. From an accom-
panying physical perspective, the advection (i.e., movement following fluid parcels moving with
the velocity field) causes the generic behavior of the entanglement of neighboring material parcels;
this causes chaotic evolution, transport, and mixing.
The subject of this course is turbulence in Earth’s tropospheric atmosphere and ocean on spatial
scales less than the general circulation and climate. Because this is an old subject — sometimes
called the last unsolved problem of classical (i.e., human-sized) physics — there is a long history
of contributions by many able and insightful scientists. The core content of this course is about
this culture of turbulence, in communion with our ancestors and being conveyed to the progeny.
It is even a respectable conjecture that, if the fundamental problem were solved, we would not be
able to understand the answer.
Turbulent flows manifest a complexity that has thus far exceeded scientists’ abilities to mea-
sure, theorize, or simulate comprehensively. Nevertheless, much has been and is being learned
about the behaviors of turbulence, and these lectures provide a survey of the fruits of this research.
The major themes are
• statistical symmetries and observed statistical regularities
• power of dimensional reasoning (a.k.a. similarity or scaling theory)
• patterns, processes, and effects of coherent structures in different physical regimes
• interaction of turbulence with mean flow and density stratification in statistical equilibrium
regimes
• utility of simple conceptual and parameterization models of turbulence for the use in simu-
lating larger-scale flows.
• power of modern computational simulations as approximate solutions.
1 Incompressible Fluid Dynamics
The basic equations of fluid dynamics are the Navier-Stokes Equations for parcel acceleration in
the presence of forces1 . In this course we will further make the approximation of incompressibil-
ity, not because it is a very accurate approximation in the ocean and atmosphere but because it
allows a simpler dynamical characterization of the essential turbulent phenomena2 The result is
the Boussinesq Equations:
Du h i
= −∇φ + ν∇2 u + ẑb − f ẑ × u
Dt
∇·u = 0
h Db i
= κ∇2 b , (1)
Dt
where u is the velocity; p is the pressure (n.b., Appendix A); φ = p/ρo ; f is the Coriolis frequency
associated with (planetary) rotation; ν is the molecular viscosity; and κ is the molecular diffusivity.
b is the buoyancy field,  
ρ
b = g 1− , (2)
ρ0
where g is the gravitational acceleration constant; ρ is the density; and ρ0 is a constant reference
value. The vertical direction ẑ is assumed to be parallel to both gravity and the axis of rotation
for simplicity. A simple thermodynamic equation of state is implicit in (1) whereby density is
adiabatically conserved following a parcel when κ = 0. (This is too simple a model for moist
convection in the later chapter on Planetary Boundary Layer Turbulence.) The square-bracketed
terms in (1) include the effects of rotation and gravitation, which are not part of the “classical” tur-
bulence problem (a.k.a. homogeneous or shear turbulence; Batchelor (1967) and Tritton (1988)).
These equations must be completed by forcing and boundary and initial conditions for a well-posed
problem.
We can also analyze the evolution of a passive tracer (i.e., a material concentration field). It
satisfies the same equation as b above, but it has no influence on the evolution of u. Reactive tracers
will have additional source or sink terms apart from their boundary fluxes.
The essential nonlinearity is contained in the advective time derivative,
D ∂
= +u·∇. (3)
Dt ∂t
Thus D/Dt is the time derivative following the flow, i.e., in the Lagrangian reference frame. The
second term in (3) is called advection. The essential effect of ν and κ is to spatially smooth the
fields of u and b, respectively, through diffusion, often referred to as molecular diffusion because
it arises from inter-molecular interactions that are averaged over to obtain a continuum fluid repre-
sentation.
1
If viscosity is neglected, they are called the Euler Equations.
2
There is a useful generalization of the incompressible approximation for flows that extend over a significant
vertical range in the presence of a background density stratification, ρ(z). It is called the analastic approximation,
with a mass balance equation, ∇ · (ρ(z)u) = 0. It includes some effects of compressibility while still excluding
extraneous acoustic motions.

2
 These equations have conservative (i.e., with ν = κ = 0) spatial integral invariants for the en-
ergy and all powers or functionals of the buoyancy and passive tracer fields. For non-conservative
dynamics, the energy and scalar variance principles are
dE dV
= −E & = − Eb , (4)
dt dt
where Z Z Z
dx e, b2 , ε, χ ,
   
E, V, E, Eb = (5)

and
1 2
u − zb, ε = ν ∇u : ∇u, χ = 2κ(∇b · ∇b) .
e = (6)
2
e is the energy density, comprised of kinetic and potential energy components (internal energy is
neglected in the simple thermodynamics assumed here). The right-hand-side terms in (4), ε and
χ, are the dissipation terms that act to decrease the kinetic energy and scalar variance. They are
written in tensor notation in (6), or they can alternatively be written in index notation as
∂ui ∂ui ∂b ∂b
ε = ν , χ = 2κ ,
∂xj ∂xj ∂xj ∂xj
where a repeated index implies index summation. In deriving (4), it is assumed that there are no
boundary fluxes of energy or scalar variance (which certainly is not always true). In the absence
of forcing, sources, or boundary fluxes, these integral measures of the flow can only decrease with
time through the action of molecular viscosity and diffusivity. The right-hand-side terms in (4), ε
and χ, are the dissipation terms that act to decrease the kinetic energy and scalar variance.
Another parcel invariant for conservative flows governed by (1) is the potential vorticity, de-
fined by
Q = (f ẑ + ζ) · ∇b , (7)
where ζ = ∇ × u is the 3D vorticity. Note that Q becomes trivial (i.e., dynamically meaningless)
in the absence of gravitational effects3 . Its integral variance is called potential enstrophy. A related
quantity is the integral variance of ζ, called enstrophy, whose dissipation rate, analogous to the
relations in (4), we denote by Eζ .

2 Transition to Turbulence
If we nondimensionalize (1) by characteristic scales,
L
x ∼ L, u ∼ V, t ∼
V
(the latter assumes that advection dominates the evolution), then the momentum and scalar diffu-
sion terms will have nondimensional prefactors that are the inverse of
VL VL
Re = & Pe = , (8)
ν κ
3
In a 2D or barotropic fluid with ∂z = 0, f + ζ z plays the role of potential vorticity without gravity influence.

3
respectively. Re is the Reynolds number, and P e is the Peclet number. Thus, the physical statement
that advection dominates diffusion is equivalent to Re, P e  1. For historical reasons turbulent
flows dominated by gravitational instability, called convection, have a different parameter that
expresses the degree of advective dominance, the Rayleigh number,
BL3
Ra = , (9)
νκ
where B (e.g., B = αg∆T in thermal convection) is a characteristic scale for the buoyancy dif-
ference forcing the flow.√Again, Ra  1 is the advectively dominated, turbulent regime. Note
that for ν ∼ κ and V ∼ BL (often called the free-fall velocity, associated with the velocity of a
parcel gravitationally accelerated over a distance L),

Ra ∼ Re2 .

As a traditional shorthand, Re is used as the representative parameter among Re, P e, and Ra; all
three parameters are essentially related for fluids like air and water, in which ν and κ have similar
values.
A basic perspective for fluid dynamics is to view Re as a control parameter: as Re varies the
nature of the flow varies. The sequence of flow regimes that occur as Re increases is called the
transition to turbulence.
If Re  1, the advective terms can be neglected, and the governing dynamics of the fluid
become linear. In the case of steady flow, solutions are simpler yet, with a simple balance between
pressure gradient and viscous momentum diffusion (here using index notation):
∂φ ∂ 2 ui
=ν 2 , (10)
∂xi ∂xj

which is called creeping flow. Equation (10) has been solved for a variety of situations and geome-
tries. It is generally of no geophysical interest, with two possible exceptions: laminar Poiseuille
Flow and Stokes Flow. The former describes the flow profile in pipes or channels and can be impor-
tant in observational instruments. The latter describes the flow around an object and is important in
understanding the hydrodynamics of small hydrometeors in suspension (cloud droplets) and other
aerosols and aquasols. For larger hydrometeors such as rain-drops and ice particles, inertial effects
(i.e., involving the finite mass of the particle) become increasingly important in modifying (10).
For almost every other flow of interest in the atmosphere and ocean, Re is very much greater
than one. For instance, in the atmospheric boundary layer (ABL), L ≈ 1000 m, U ≈ 1 ms−1
(or more), and ν ≈ 1.5 × 10−5 m2 s−1 , leads us to estimate Re ≈ 108 (or more). This means
that, if one were to measure the contribution of various terms in (1) to the momentum balance
of ui ,, one would expect to find that the viscous term is entirely negligible — fully eight orders
of magnitude smaller than the other terms. And while this is true for components of the flow of
scale L this is not to say that viscosity plays no essential role. Paradoxically, even when Re is
large, the dissipation terms in (1) are usually not small. The paradox disappears when one realizes
that simple scale estimation or dimensional analysis is not reliable in this matter, since turbulence
comprises an exceedingly broad range of scales: for larger scales of motion, which may contain
most of the energy or scalar variance, Re is large and the evolution is essentially conservative over
a characteristic advective (or eddy turnover) timescale of L/V ; whereas for small enough scales

4
of motion, which may contain most of the variance of vorticity and scalar gradients, the relevant
Re value is small and the dissipation rate is dynamically significant. Thus, there can be profound
differences in solution behavior between the asymptotic tendency as Re → ∞ and the Euler limit,
Re = ∞ or ν = 0, and one must be very careful about this situation. A hallmark of high-Re
flows is their ability to generate small scale fluctuations on which molecular effects can operate
efficiently.
Viewing Re as a control parameter, we can ask what will happen to a flow as Re is steadily
increased, from a small number less than one, to a very large value, such as is characteristic of the
ABL. In doing this we are asking how does the flow change as advective effects become increas-
ingly dominant and viscosity becomes increasingly relegated to smaller and smaller scales. We
find that as Re becomes large relative to unity the flow behaves increasingly erratically. This is il-
lustrated with the help of Fig. 1 for the wake flow behind a cylinder. Similar figures, differing only
in the details, could be constructed for most other flow regimes. In each case, once Re becomes
sufficiently large, we will find a completely erratic flow field, with no evidence of global order; we
call fully developed turbulence.

Re = 85

Re = 185

Re = 280

Re = 640

Figure 1: Illustration of the transition to turbulence as a function of Re for a cylindrical wake.

Note that the timescale of the last panel has been increased by a factor of about 3. Adapted from
Tritton (1988, Fig. 3.10).

The sequence of steps flows go through on their way to becoming turbulent is called the tran-
sition to turbulence. Exactly how flows transition to turbulence is a wide area of study, partly
motivated by practical devices and partly by an interest in chaos for systems with few degrees of
freedom (e.g., the marginally unstable modes near the first bifurcations in Re). In Fig. 1 the route
is through something called frequency doubling. This is evident because the flow goes from being

5
completely steady for Re < 1 (not shown) to oscillatory at Re = 85. When Re increases to 185
the period of the oscillations doubles, and there is some evidence of low-frequency modulation.
Increasing Re further leads to yet another doubling and to even less uniformity in the amplitude
of the signal. By the time Re increases to 640, there is no longer evidence of a single dominant
period, but rather the time-series of the flow velocity is rather erratic and seemingly unpredictable.

Figure 2: Illustration of the transition to turbulence as a function of Ra for Rayleigh-Benard con-

vection. Shown is the power spectra of velocity for: (a) Ra/Rac = 21, (b) Ra/Rac = 26, (c)
Ra/Rac = 27, (d) Ra/Rac = 37. (Rac ∼ 103 is the “critical” value at which the conducting state
is unstable to the onset of steady convective cellular circulations.) In each panel the abscissa is
linear in frequency and logarithmic in power. Adapted from Gollub et al. (1980); see also Tritton
(1988, Fig. 24.8).

This process can also be illustrated by considering the frequency spectrum of the signal. In
Fig. 2 we do just that, except it is the case of Rayleigh-Benard convection, which is the flow
between two horizontal plates with an imposed destabilizing buoyancy gradient between them
(e.g., a hot plate below a cold one). Instead of looking at the transition in terms of Re, we look
at things in terms of Ra, as discussed above. For Ra < Rac = 1700 we have a motionless
state with diffusive heat flux between the plates, and the first instability of this state develops at
Rac with further instabilities (i.e., bifurcations) for Ra > Rac . At Ra = 21Rac we find the
dominant unstable mode and its harmonics. As Ra increases further to 26Rac sub-harmonics with
frequencies of f /2 appear, indicating period doubling. A slightly greater increase leads to further

6
harmonics, 2f , 3f /2, 3f /4, f /4, etc. before the onset of turbulent flow (i.e., the absence of any
dominant periodicity, but rather fluctuations at all frequencies) by Ra = 37Rac .
A recurring idea in the above is that flows become turbulent through a well ordered sequence of
steps, but once they are turbulent (or if you like, once Re is large enough) further increases in Re
lead to no qualitative changes in the flow, in effect implying that the flow becomes independent of
the exact value of viscosity in such a regime. This state of affairs is associated with fully developed
turbulence and is often referred to as Reynolds number similarity.
The following represent typical Re values for successive regimes:

Re → 0 is laminar flow, which is smoothly varying in space and time, hence predictable in prin-
ciple.

Re ≥ Rec ∼ 10 − 103 is the transitional flow regime, in which there is instability in any hypoth-
esized laminar (or mean) flow and often some degree of approximate periodicity of fluctu-
ations in space and/or time. Therefore, the wavenumber and/or frequency spectra usually
show sharp peaks. This dynamics can often be analyzed linearly, either in terms of normal
mode instabilities or non-normal (algebraic) growth, or in terms of low-order, nonlinear or-
dinary differential equations in time (sometimes called amplitude equations). The spatial
patterns in this regime are usually rather simple and often have a global (i.e., domain filling)
extent; however, the temporal behavior can be more complex. Often there are several differ-
ent Rec values associated with different bifurcations (e.g., in non-rotating Rayleigh-Benard
convection), and temporal complexity occurs only after multiple bifurcations.

Re → ∞ (or at least Re  Rec ) is fully developed turbulence, which is rapidly varying in space
and time, hence unpredictable in practice over many fluctuation cycles. Its spatial and tem-
poral patterns are complex and lack global order, although they typically have clearly ev-
ident locally ordered patterns, called coherent structures. Therefore, the wavenumber and
frequency spectra are broadband and without sharp peaks. This dynamics is fundamentally
nonlinear, and this regime will be the focus of this course.

It is very much an open question whether there are usually any important bifurcations beyond
Re ∼ 104 or Ra ∼ 108 , and as yet there is not convincing evidence that this is so. This question
has been extensively studied in the Rayleigh-Benard (convection) problem, where experimental
apparati are capable of producing very large Ra values using liquid helium. The way the problem
has been posed for these flows is to ask how the scaling properties, rather than the flow patterns,
change as a function of Ra. Specifically the question has been asked as to how the nondimensional
heat flux (called the Nusselt number, N u, and defined to be the ratio of the total heat flux to
the diffusive heat flux), scales with Ra. The hope has been that once Ra has been increased far
enough the N u − Ra scaling ceases to change, and then we can obtain a universal, essentially
Ra-independent regime. Experiments have shown that for Ra values that are large but less than
108 , N u ∝ Ra1/3 ; however, for larger values of Ra the scaling regime changes and N u ∝ Ra2/7 .
It was originally thought that the 1/3 regime was the asymptotic regime, but the discovery of the
new regime led to revisions in that thinking. Because both regimes are turbulent yet distinct,
they are often referred to as soft and hard turbulence, respectively. Recently this thinking as had
to be revised yet again: at increasingly large values of Ra there is evidence that the 2/7 regime
also ceases to persist and the scaling exponent reverts back toward 1/3. Thus the question as

7
to whether truly hard turbulence, and by implication Reynolds number similarity, actually exists
remains unanswered.
The lesson above illustrates that attempts at a strict definition of turbulence can be problem-
atic. We have taken the opposite approach, loosely associating turbulence with erratic behavior in
the flow. One can be a bit more precise without being restrictive. Perhaps the most general and
useful definition of a flow as turbulent is if it exhibits spatial and temporal complexity through
advective dynamics. This is a concise way of saying that turbulent flows are fundamentally non-
linear, and that their nonlinearities lead to broad-band wave-number and frequency spectra. This
is particularly evident in Fig. 2 where we note with increasing Ra an increasing range of scales
becomes evident in the velocity power spectrum. By this broader definition turbulence is not lim-
ited to smaller scales in the atmosphere and ocean, and we can usefully interpret winter storms and
oceanic mesoscale eddies as manifestations of turbulence on the planetary scale.

3 Turbulent Cascades
The naive scaling of the momentum equation above suggests that the viscous terms would be
negligible. But how can that be? For instance we know that even high Re fluids obey no-slip
boundary conditions at a surface. For the case of bounded flows, Prandtl developed the concept of
a boundary layer; a thin layer attached to the surface wherein viscosity is important. For viscosity
to be important implies that this layer has a scale,
p
λ ∝ νL/U . (11)
This relation is an example of what might be called dimensional analysis or dimensional reasoning,
i.e., representing dynamical balances in the governing PDE by scale estimates. It follows from the
statement that a viscous diffusion time on the scale λ, viz., λ2 /ν, is balanced against an eddy
turnover time on the dominant flow scale L/U . Such layers are indeed observed in real flows. For
the ABL numbers above, λ ≈ 10−1 m, which is quite small. Thus, we must understand the role
of viscosity in flows with large Re values in a singular perturbation sense: viscosity is mostly
unimportant to the dominant flow patterns, but there are (∼ singular) locations where the spatial
scale is small enough and the higher-order derivatives in the diffusion operator are large enough
that viscous effects are significant in comparison with advection.
The development of thin layers, over which viscosity acts, is not confined to boundaries of the
flows. They are also readily apparent in unbounded regions of the flow. If this were not the case,
fuel would not mix efficiently in combustion chambers, and chemical reactions and dissipation
would only occur at the boundaries. Thus, despite the startlingly small magnitude of the molecular
terms in our ’scaled’ governing equations, they can never be neglected: the flow always contrives
to produce scales on which molecular processes operate efficiently. For this reason, rather than
viewing the Reynolds number as a measure of the relative importance of viscous vs. advective
effects, it is perhaps better to view Re = (L/λ)2 as a measure of the ratio of the scales at which
each is commensurate. The quantity λ appears often enough to be given a separate name; it is
called the Taylor microscale. Often it is defined based on the magnitude of the fluctuating velocity
u0 and its shear: s
h u02 i
λ= , (12)
h (∂x u0 )2 i

8
where the angle brackets indicate an average over many fluctuation cycles. The dimensional rea-
soning behind this definition of λ is that u0 characterizes the most energetic scale while ∂x u0 is as
large as it can be, given u0 , limited only by the action of viscosity to set a smallest scale with sub-
stantial associated variance. The subsidiary quantity Reλ = L/λ is called the microscale Reynolds
number. (The relation between Re and Reλ is more fully explained in 3D Homogeneous Turbu-
lence.)
Above we have claimed that some hallmarks of turbulent flows are the wide range of scales
they allow and the associated presence of fine-scale structure. In a flow where only one scale is
forced, which for instance is the case in the Rayleigh-Benard problem where energy is put in at the
scale H, the additional scales are primarily associated with non-linear interactions and secondary
instabilities. The filling up of the spectrum, which we see in Fig.2, reflects the development of
something we call the turbulent cascade. This is the process whereby all scales of the fluid are
excited, and through a sequence of non-linear interactions, inviscid invariants of the flow (e.g.,
energy), are carried to scales where molecular processes operate efficiently. This cascade is illus-
trated schematically in Fig. 3 for the case where energy is inviscidly converted from larger scales
to smaller scales.

Decreasing Scale
Energy Density

Forcing Dissipation

Figure 3: Cartoon of a forward cascade whereby energy or some other scalar variance (the ordi-
nate) is systematically transferred by conservative advection from larger scales where it is forced
to smaller scales where it is dissipated by molecular viscosity or diffusivity (the abscissa).

Vortex Stretching: One classical conception of why a cascade occurs is closely related to the
process of 3D vortex stretching, although as illustrated below in a purely horizontal flow this is
not the only cascade mechanism. The vortex-stretching mechanism for the energy cascade can
be illustrated mathematically by considering small amplitude vorticity superimposed on a large-
scale straining field. Recall that the strain rate Sij is simply the second-order tensor describing the
symmetric part of the velocity gradients:
 
1 ∂ui ∂uj
Sij = + , (13)
2 ∂xj ∂xi

9
and the vorticity, ζi ≡ −ijk ∂uj /∂xk , is the curl of the velocity field (with ijk = 1 for an even
permutation of the indices (123), = −1 for an odd permutation, and = 0 for any other combination,
i.e., with any index repeated). An equation for ζi can readily be derived by taking the curl of the
momentum equations. In the absence of viscosity, rotation, and stratification, this equation takes
the form:
dζi ∂ui
= ζj . (14)
dt ∂xj
A simple straining flow can be represented in two dimensions as follows:

∂ui s
 if i = j = 1
(u, v, w) = (sx, −sy, 0) =⇒ = −s if i = j = 2 (15)
∂xj 
0 otherwise


for constant s, as illustrated in Fig. 4. Thus we take ui as given above as our (large-scale) back-
ground flow. Note that the vorticity of the background flow is zero, so that the vorticity equation
simply describes the effect of the background straining flow on the vorticity of some other (small-
scale, 3D) embedded flow component. Non-trivial relations exist only for the first two components
of the vorticity field:
dζ1 dζ2
= ζ1 s and = −ζ2 s. (16)
dt dt
These equations imply exponential growth for ζ1 and damping for ζ2 ; i.e., growth of the compo-
nent of vorticity ζ1 that is aligned with the component of background flow with positive divergence,
∂x u = ∂x1 u1 = s > 0. Because the magnitude of a damped solution is lower-bounded by zero,
overall the vorticity in the system will grow. Physically we interpret this result as the manifes-
tation of vortex tubes aligned with the flow being stretched while those opposed to the flow are
compacted. This vortex stretching leads to energy being associated with smaller scales (Fig. 4),
and it is one way to think about the energy cascade in three-dimensional (3D) flows. Although we
have illustrated stretching with a particular flow and vorticity orientation, we can imagine that a
full range of orientations occur in a statistically isotropic 3D flow. However, note that this mecha-
nism does not exist for strictly two-dimensional (2D) flows because of the geometrical restriction
that the only non-trivial vorticity component is ζ3 , and ζ3 has a trivial evolution equation; vortex
stretching does not occur; and vorticity is an inviscid parcel invariant for the flow.

Line (or Contour) Stretching: More generally, turbulent flows exhibit cascades for many flow
properties, not just kinetic energy, even in the absence of vortex stretching. Nevertheless, the
distinctive roles of vorticity and strain in cascades is fundamental. To explore further some of the
roles of these velocity-gradient fields, we will consider a special type of flow field that is purely
horizontal in direction but fully 3D in spatial variation,

u = ẑ × ∇ψ(x, y, z, t) , (17)

where ψ is a scalar potential called the streamfunction. This is often a good approximation for
large-scale geophysical flows because of the strong influences of the Earth’s rotation and stable
stratification (i.e., geostrophic flows). For this flow the vertical vorticity is

ζ = ζ3 = ∂x v − ∂y u = ∇2h ψ , (18)

10
 y

flow streamline

x
vortex stretching

Figure 4: A vortex tube aligned with the axis of the strain-rate tensor stretches its length, shrinks
its diameter (i.e., reduces its spatial scale), preserves its circulation, and increases its vorticity
magnitude.

where the subscript h denotes purely horizontal components. (Unlike the preceding case, no vor-
tex stretching can occur here because there is no velocity in the direction of the vorticity.) The
magnitude s of the strain rate associated with horizontal shear is given by

s2 = (∂x v + ∂y u)2 + (∂x u − ∂y v)2 = (∂xx ψ)2 + (∂yy ψ)2 − 2(∂xx ψ)(∂yy ψ) + 4(∂xy ψ)2 . (19)

The advective dynamics in this special case is essentially 2D since

D D ∂ ∂
= = + uh · ∇ = + Jh (ψ, ) , (20)
Dt Dth ∂t ∂t
where J is the horizontal Jacobian operator:

Jh (A, B) = (∂x A)(∂y B) − (∂y A)(∂x B) .

For ν = 0, the curl of the horizontal momentum equations gives

Dq
= 0, (21)
Dth
where q = f + ζ3 is the potential vorticity that is conserved following the flow. Equation (21)
also holds for a more generally defined potential vorticity q in other circumstances, e.g., quasi-
geostrophic flow. Equation (21) also has the same form as a material tracer equation, assuming
(17) and κ = 0.

11
 Taking the horizontal gradient of (21),
Dq
∇h =0
Dth
D
∇h q = −Jh (∇h ψ, q)
Dth
d
∇h q = M · ∇ h q , (22)
dtL
where d · /dtL indicates the Lagrangian derivative following a fluid parcel and M is a 2 × 2 matrix
of second spatial derivatives of ψ,
 
∂xy ψ − ∂xx ψ
M = . (23)
∂yy ψ − ∂xy ψ
If we assume that M is slowly varying in the Lagrangian frame compared to ∇h q (e.g., because
the spectrum has more power on small scales in the latter than in the former, as certainly is true for
∇h q ∼ ∇h ∇h ∇h ψ and M ∼ ∇h ∇h ψ), then we can solve (22) as if it were a linear ODE system
with constant coefficients. In particular, there are eigensolutions of the form,
∇h q = (∇h q)o eσt , (24)
where the local growth rate σ in the Lagrangian frame is an eigenvalue of the matrix M and thus
satisfies
(∂xy ψ − σ)(−∂xy ψ − σ) + (∂xx ψ)(∂yy )ψ = 0
σ 2 = −Jh (∂x ψ, ∂y ψ)
1 2
σ2 = [s − ζ 2 ] , (25)
4
using (18)-(19) and the steps shown in (50). Thus, the temporal behavior of the gradient of a con-
servative scalar is also governed by the competition between vorticity and strain rate magnitudes.
Furthermore, one can show by horizontally integrating by parts and discarding boundary integral
terms (assuming periodicity or zero velocity) that
Z Z Z Z Z Z
2 2
dx dy σ = 0 and dx dy ζ = dx dy s2 . (26)

Thus this competition is an equal one between dominance by vorticity or strain-rate, averaged over
the domain.
Now consider the evolutionary consequences of this competition:
s2 > ζ 2 (strain-rate dominance) This ⇒ σ 2 > 0 ⇒ σ is ± a positive real number. Thus there
is exponential growth and decay of initial scalar gradients. General forcing or initial con-
ditions will project equally onto the two eigenmodes, and after some time the solution will
be dominated by the growing modes. This implies an impermanence of initial patterns in
q. Growth of gradients is also a representation of the turbulent cascade to smaller scales,
hence eventually to dissipation with small but finite (ν, κ). It occurs when a local Taylor
series expansion of ψ has a hyperbolic pattern at second order (e.g., as in ψ ∼ xy), and since
parcel trajectories follow isolines of ψ, the flow in this regime has locally diverging parcels.
The local topology of ψ is a saddle node.

12
s2 < ζ 2 (vorticity dominance) This ⇒ σ 2 < 0 ⇒ σ is ± a positive imaginary number. Thus
the time behavior of scalar gradients is oscillatory. Spatial patterns of q will be termporally
recurrent. Lack of growth of gradients ⇒ a lack of turbulent cascade, hence a lack of dissi-
pation whenever Re  1. It occurs when a local Taylor series expansion of ψ has an elliptic
pattern at second order (e.g., as in ψ ∼ x2 + y 2 ), and the flow in this regime has locally
confined parcels. The local topology of ψ is a central extremum.

If we assume that a trajectory is random over a long time interval, it will wander through
different ψ environments, experiencing intervals of growth, decay, and oscillation in its scalar
gradient. However, the net effect is growth (e.g., the expected value of eat , for a a Gaussian
2
random variable with zero mean and unit variance, shows growth with t like et /4 ). Thus on
average, scalar variance is cascaded toward smaller scales and dissipation. In contrast with the 3D
vortex stretching cascade mechanism discussed above, the mechanism embodied in this example
can be considered as line stretching, i.e., elongation of a material contour (labeled in this case by
particular values of the tracer field q). Both mechanisms contribute to cascades in turbulent flows.
However, if there is a long-time correlation between the trajectory and its ψ environment—in
particular, if a trajectory remains trapped within a region of vorticity dominance—then there need
not be a cascade. As we will see later in the discussion of 2D homogeneous turbulence, a 2D
coherent vortex has an extremum in ζ and a minimum in s at its center, hence a region of vorticity
dominance in its core, and it sustains a pattern of recirculating trajectories and trapped passive
scalar concentrations for a very long period of time (much longer than L/V ). On the other hand,
it is more common for strain rate to be dominant outside of the cores of 2D coherent vortices; this
is where scalar gradients grow most strongly, and the cascade of scalar variance is more clearly
evident.
For general flows (i.e., without the restricted form of u in (17)), one cannot make such a simple
analysis of the roles of vorticity and strain rate4 . Nevertheless, experience shows that there are
often associations between regions of strong vorticity and persistent flow patterns and between
regions of strong strain rate and pattern deformation and cascade to smaller scales. Think of this a
rule of thumb based on experience, but not a theorem.

Cascade Inertial Range: With these ideas we can now return to Fig. 3 and imagine some station-
ary state, in which case energy must be put into the flow at the same rate at which it is dissipated.
If we denote this rate of viscous dissipation (forcing) by ε, we note that by dimensional argument
that a new length scale, η, where
 3 1/4
ν
η= , (27)
ε
is implied. This length
pscale is called the Kolmogorov scale. It is related to the Taylor microscale by
the square root, η = λν/U . It is more directly characteristic of the dissipative scales, as opposed
to the scales where dissipation starts becoming important. In terms of η, the Reynolds number
scales as Re ∝ (L/η)4/3 . This dissipation scaling also affords us the possibility of introducing
dissipation time and velocity scales:
p
τη = ν/ε , uη = (νε)1/4 . (28)
4
As another characterization of their distinctive roles, Appendix A shows their relation to the pressure field.

13
The Kolmogorov scale is also of fundamental importance because it bounds the range of scales
necessary to accurately represent all the scales in a high Reynolds number flow, i.e., L/η ∝ Re3/4 .
Thus to simulate the atmospheric boundary layer we, in principle, should represent six decades of
scales, from millimeters to kilometers.
If Re  1 then L  η in which case we can posit an inertial range of scales l such that
L  l  η (n.b., inertial is sometimes used as a synonym of advective). Because these scales are
not directly influenced either by the forcing or by the dissipation, we can posit that their dynamics
are predominantly advective. Furthermore, in the well-known phenomenological cascade theory of
Kolmogorov for 3D homogeneous turbulence, it is often assumed that the behavior of the flow in
this regime is local in wavenumber space (Sec. 4), depending on only the non-linear interactions
among commensurate scales. In this view, the dissipation rate ε occurring on small scales on
average must be equal to the large-scale energy generation rate by forcing, as well as to the cascade
rate at all intermediate scales in the inertial range.
We can now complete the schematic of Fig. 3. Doing so for the case of 3D turbulence we
find an inertial range whose scaling is that proposed by Kolmogorov, which is used to connect
the dissipation range and energy-containing ranges thereby giving a more complete and somewhat
generic description of the spectral energy density of 3D turbulent flows.
A way to summarize these characteristics is that turbulence generally has the properties of
cascade and stirring on the outer and intermediate scales that connect to dissipation and mixing
on the finest scales where molecular viscosity and diffusivity are important. The outer scale is
defined as the largest scale in the turbulent regime where the advective tendency dominates the
pointwise evolution; this is opposed to the even larger scales of the turbulent environment where
other evolutionary influences (e.g., solar heating) are at least as important. Cascade is defined as the
transfer of the variance of a fluid property from one spatial scale to another one (usually smaller).
Dissipation is defined as the removal of variance at small scales. Stirring is defined as increasing
the variance of gradients through stretching and folding of isolines of a fluid property, through its
cascade. Mixing is defined as the removal of inhomogeneities when the scale of property gradients
reach the dissipation scale.

4 Statistical Descriptions and Dynamics

Because of the complexity of turbulence, it is usual to seek some kind of statistical description as a
condensation of the excessive amount of information involved. A statistical measure is an average
of some kind. It can be over the symmetry coordinates if any are available (e.g., a time average
in a stationary regime; a spatial average in a homogeneous regime; or a directional average in
an isotropic regime). Or it can be over multiple realizations (i.e., an ensemble), taking advantage
of the property of sensitive dependence to assure that the outcomes will differ among ensemble
members. Or it can be over the phase space of the solutions if the dynamics is ergodic so that all
possible outcomes will be realized. In practice, exact averages are not obtainable in measurements
or computations because of finite samples sizes and because exact symmetries do not occur in
nature, so compromises and approximations are required.
The governing equations are deterministic, not random. So fundamentally turbulent fields
are not random because their generating dynamics is not random. Nevertheless, it is useful to
use statistical methods for random variables in order to describe turbulence, specifically averages,

14
Fourier transforms, spectra, and covariance functions. The view of turbulence as akin to random
motion is not incommensurate with its characterization as erratic behavior, and it is perhaps most
clearly stated in the Introduction of Batchelor’s (1953) book on homogeneous turbulence:

It is a well-known fact that under suitable conditions, which normally amount to a

requirement that the kinematic viscosity ν be sufficiently small, some of these motions
are such that the velocity at any given time and position in the fluid is not found to be
the same when it is measured several times under seemingly identical conditions. In
these motions the velocity takes random values which are not determined by the os-
tensible, or controllable, or ’macroscopic’ data of the flow, although we believe that
the average properties of the motion are determined uniquely by the data. Fluctuating
motions of this kind are said to be turbulent . . . The problem is to understand the me-
chanics, and to determine analytically the average properties, of this kind of motion.

The language of random fields is probability theory, whose basic unit is the probability distri-
bution function (PDF). Here is a brief summary (but also see Lumley and Panofsky, 1964, for a
more extensive discussion). If u denotes a fluid property that fluctuates with zero mean, then we
denote the PDF of u by p(u). To be a PDF, p(u) must satisfy two properties: first it is everywhere
positive, p ≥ 0, and second Z Z
p(u)d u = d pu = 1 . (29)

Given p(u) the nth moment of u is

Z
n
hu i ≡ un p(u)d u . (30)

The statistical mean is denoted by the angle brackets. Hence p(u) determines all of the moments of
u. By working with a centered distribution, i.e., one with zero mean, these moments are called the
central moments. They tell us something about the shape of the distribution. Roughly speaking,
even ordered moments describe symmetric properties of the distribution (i.e., its width) and odd
order moments describe asymmetric aspects of the distribution. In addition to the mean, the second,
third and fourth moments are encountered frequently. h u2 i is called the variance and sometimes
denoted as σu2 . The normalized third and fourth order moments,

h u3 i h u4 i
Sk(u) ≡ and Ku(u) ≡ , (31)
σu3 σu4

are called the skewness and kurtosis, respectively, The skewness is indicative of asymmetry be-
tween the positive and negative fluctuations, and the kurtosis is indicative of the likelihood of
occurrence of extreme amplitudes. Just as the PDF determines the moments of a distribution, the
infinity of moments determine the PDF. Particularly useful are those families of distributions that
can be determined in terms of only a small number of moments, the best known such example
being the Gaussian or normal distribution,
1 2 2
p(u) = √ e− (u−µ) /(2σ ) , (32)
σ 2π

15
that is completely determined by its mean µ and variance σu2 . Because a Gaussian distribution is
symmetric its skewness is identically zero, Sk(u) = 0, while its kurtosis is Ku(u) = 3. This is the
common comparison standard for the measured or computed PDFs of turbulent quantities.
Another useful distribution for inherently positive quantities, u ≥ o, is the log-normal distri-
bution,
1 2 2
p(u) = √ e− (ln u−ln ug ) /(2σg ) , (33)
uσg 2π
that is also described by two parameters ug and σg related to the mean and standard deviation,
respectively. It has the property that extreme values of u are very much more likely than they are
for Gaussian PDF. This PDF is often used, e.g., to describe the statistics of the dissipation ε.
One interesting aspect of turbulent flows is that strong, or extreme, events are more common
than one would expect based on the variance of the distribution, i.e., relative to a Gaussian PDF.
This is referred to as intermittency, and is evident by large kurtosis, or a flattening of the PDF
away from the origin u = 0. An example in terms of the velocity tendency (i.e., , the difference in
velocity at two times close together, and thus related to the acceleration) exemplifies this behavior
(Fig 5). (A survey of PDFs from different geophysical turbulent flows is in McWilliams, 2007.)
Broadly speaking, the higher the value of Re, the broader and more non-Gaussian are the PDFs,
and the more intermittent are the largest fluctuations.
For coupled nonlinear systems this type of statistical description of the flow requires not only a
determination of the PDFs of the individual variables, but also the joint PDFs among variables. For
instance, in attempting to write an equation for the expected value of some flow field, the advective
terms in the governing equations naturally give rise to expected values of products h uv i, which by
decomposing the flow into its expected values and fluctuations according to
u = h u i + u0 , (34)
takes the form
h uv i = h u ih v i + h u0 v 0 i. (35)
Hence, determining h uv i requires not only knowledge of p(u) and p(v) but also p(u, v), the prob-
ability of some u given some v. When we say that u and v are independent, we mean p(u, v) =
p(u)p(v), which is another way of saying that h uv i = h u ih v i. However, in almost all prob-
lems of interest these higher order covariances (equivalently, joint PDFs describing the correlation
among flow properties) are important to the overall evolution of the flow.
Because most geophysical flows are neither homogeneous, nor stationary, the PDFs of flow
properties depend on space and time, i.e., we must adopt the language of conditional probabil-
ity. That is we speak of p(u; x, t), the probability of some u occurring at some point x and
time t. In practical situations, where the PDFs of flow properties are not known we often as-
sume homogeneity or stationarity on some scale x0 , respectively τ . By which we mean that
p(u; x, t) = p(u; x + r, t) for r  x0 , or p(u; x, t) = p(u; x, t + s) for s  τ. In such situa-
tions the fluid properties on large-scale cease to depend on time t or position x, and the expected
values of fluid properties may, as a practical matter, be replaced by spatial or temporal averages.
The common means of representing the scale distribution of a field is through its Fourier trans-
form, spectrum, and covariance function. For example, the Fourier transform for a spatially varying
field ψ(x) in an infinite, homogeneous domain is
Z
ψ(x) = dk ψ̂(k)eik·x , (36)

16
 tions ¯atness, shown in the inset to Fig. 3, characterizes the intermittency
time, of the acceleration, and would be 3 for a gaussian distribution. These
ulent ¯atness values are consistent with direct numerical simulation
his is (DNS) at low Reynolds number14 and exceed 60 at the highest
onal, Reynolds numbers.
olent The prediction by Heisenberg and Yaglom of the variance of an
3 ms. acceleration component based on Kolmogorov '41 theory8,9 is
ht, is
y to a
ation 100 100
trates Rλ = 200
<a4>/<a2>2

ation 10–1 50
Rλ = 690
e and Rλ = 970
ces of 10–2 0
0 500 1,000
Probability

nolds 10–3 Rλ
ution
et by 10–4
s in a
ment 10–5
ertex
ornell 10–6
ng in
ional –20 0 20
data a/<a2>1/2
everal
es per Figure
Figure 35:Acceleration
Measurementsdistribution. Probability
from a laboratory flow of density functions
the normalized of thevelocity
transverse transverse
derivative
in time, or acceleration, PDF at different values of the microscale Reynolds number, Reλ = Rλ .
acceleration normalized by its standard deviation at different Reynolds numbers. The
From La Porta et al. (2001, Fig. 3).
given acceleration is measured from parabolic ®ts over 0.75th segments of each trajectory. The
solid line is a parameterization of the highest Reynolds number data using the function
P a† ˆ C exp 2 a 2 = 1 ‡ jab=jjg †j2 ††, with b ˆ 0:539, g ˆ 1:588 and
1† j ˆ 0:508, and the dashed line is a gaussian distribution with the same variance. The
inset shows the ¯atness of the acceleration distribution, (ha4 i/ha2 i2, evaluated using 0.5th
ocity parabolic ®ts) as a function of Rl.
17

Illumination
b
together with the inverse transform relation,
Z
1
ψ̂(k) = dx ψ(x)e−ik·x . (37)
(2π)3

k is the vector wavenumber, and ψ̂(k) is the complex Fourier transform coefficient. With this
definition the spectrum of ψ is
S(k) = h |ψ̂(k)|2 i . (38)
S(k) can be interpreted as the ensemble-mean variance of ψ associated R with a spatial scale,
L = 1/k, with k = |k|, such that the total variance, h ψ 2 i, is equal to dk S (sometimes called
Parceval’s Theorem). The positional covariance function is defined by

C(x) = h ψ(x0 )ψ(x0 + x) i , (39)

where the assumption of homogeneity implies that C is independent of x0 . In this formula x is the
difference in position between the two quantities being multiplied and is called the spatial lag (so
C can also be called the spatial lag covariance function). Note that C(0) is the variance h ψ 2 i. It
can also be shown that S(k) and C(x) are related by a Fourier transform and its inverse, e.g.,
Z
C(x) = dk S(k)eik·x .

Finally, sometimes turbulent fields are represented by their spatial-coordinate structure function,

St(x) = h (ψ(x0 ) − ψ(x0 + x))2 i = 2( C(0) − C(x) ) . (40)

Such transforms may be performed for variables which depend on time, in which case we speak
of the frequency spectra S(f ) or S(ω) and C(t) the temporal lag covariance function. Figure 2
shows examples of frequency spectra associated with a Fourier transform for a function of time.
For the S(f ) with sharp peaks at smaller Ra values, the associated C(t) are oscillatory functions
of the time lag, while for broad-band S(f ) at larger Ra, C(t) decays to zero as t increases.
If we now bring a statistical description to the dynamics of turbulence, it is natural to construct
various averages from the governing equations. For example, governing equations for the mean
fields (i.e., h u i and h b i) can be obtained by integrating (1) over the PDFs of u and b respectively:

∂h u i
+ h u i · ∇h u i = −∇h φ i + ν∇2 h u i + ẑh b i − f ẑ × h u i − ∇ · h u0 u0 i
∂t
∇ · hui = 0
∂h b i
+ h u i · ∇h b i = κ∇2 h b i − ∇ · h u0 b0 i . (41)
∂t
These equations are isomorphic to the unaveraged equations except for the additional, final terms
in the momentum and tracer equations. These terms are divergences of the eddy momentum flux
(or Reynolds stress) and eddy tracer flux, respectively, where the word eddy refers to the fact that
they are associated with the fluctuations. Equations like (41) are a central framework for the theory
of eddy – mean interaction, but they obviously need to be combined with some prescription for the
eddy dynamics in order to be a complete system.

18
Figure 6: Trajectories for acoustically tracked neutrally buoyant floats at a nominal depth of 700
m in the Northwest Subtropical Gyre in the North Atlantic Ocean. Arrows along trajectories occur
at 30 day intervals. Note the general mixing and dispersion behaviors due to stirring by mesoscale
eddies. Adapted from Owens (1991, Fig. 2).

19
 This procedure can suffice if the goal is only to determine the mean fields. Implicit here is an
assumption that the solutions of the averaged equations do not exhibit fluctuation behavior, and at
the least a posteriori consistency must be demonstrated. Note, however, that to determine the mean
fields, the eddy flux fields must be known, so the mean-field equations are incomplete. If one seeks
to complete them by calculating the eddy fluxes, then an average must be taken of the products of
a fluctuating field and a fluctuating equation, and the result will depend on third moments; to
calculate the third moments requires knowledge of the fourth moments; etc.Escaping this infinite
regression is called the closure problem, and, because it is simply a reformulation of the original
problem, it is not really any easier to solve. The usual computational approach is the more direct
one of calculating the unaveraged PDE system and taking averages of the solution. This involves
a truncation in spatial scale (and a redefinition of the effective Re, often through modeling, or
parameterization, of the effects of the deleted scales of motion) rather than a truncation or modeling
in the statistical moment hierarchy. The former is referred to as Large-Eddy Simulation (LES) and
the latter as direct closure modeling, but even LES must invoke some form of closure modeling
for the effects of its subgrid-scale turbulence. Almost all computations of oceanic and atmospheric
turbulence are LES, because the target phenomena are on scales too large to connect to the ∼
1 mm scale where molecular diffusion begins to dominate advection. Even climate and general
circulation models are LES, albeit ones quite distantly separated from the scales on which micro-
and mesoscale turbulence motions occur.

5 Characteristics of Turbulence
Let us now list, in a highly qualitative way, some of the essential characteristics of turbulence that
apply in almost all physical situations. The first three entries have been discussed above. The next
two entries, dispersion and loss of correlation, are certainly central aspects of the historical view
of turbulence. The remaining six entries are perhaps the more modern elements in the evolving
understanding of turbulence.

• Cascades

• Scale breadth in space and time

• Dissipation of fluctuation variance for various quantities

• Dispersion of material concentrations and entangling of parcel trajectories (Fig. 6))

• Aperiodicity and loss or correlation with increasing spatial and temporal lags

• Deterministic chaos (vs. randomness and ergodicity)

• Sensitive dependence and limited predictability from imperfect initial data

• Intermittency and occurrence of extreme events

• Irreversibility of outcome — due both to mixing and dissipation and due to evolutionary
complexification – even though the Euler equations have a time-reversal symmetry (Pri-
gogine, 1980)

20
 • Differentiable but non-smooth fields on all but the microscale (vs. fractal: continuous but
non-differentiable)

• Locally ordered (coherent structures; Fig. 7) but globally disordered (chaotic) (Prigogine &
Stengers, 1984)

Figure 7: Oceanic (left, in the marginal ice zone) and atmospheric (right, in a stratus cloud deck)
coherent vortices in Davis Strait (north of the Labrador Sea, west of Greenland) during June 2002.
Both vortex types are mesoscale vortices with horizontal diameters of 10s-100s km. Courtesy of
Jacques Descloirest (NASA Goddard Space Flight Center).

6 Theoretical Foundations and Models

The mathematical characterization of the Navier-Stokes Equation, and a fortiori the Boussinesq
Equations, is incomplete, essentially because of the advection operator in (1). There are no math-
ematical proofs of the long-term existence and uniqueness of PDE solutions for large Re starting
from smooth initial conditions and forcing. And certainly there is no general analytic solution
method for these equations. It is still very much an open question whether singularities can arise in
the PDE solutions. Nevertheless, most fluid dynamicists are persuaded by the cumulative number
of particular successes comparing measurements with analytical and computational solutions that
these equations are physically reliable.

21
 Because of the lack of a general theory, turbulence is primarily an experimental problem in
the broad sense that we know much more about it from experience than from fundamental theory.
There is a great deal of data from turbulence experiments (in controlled laboratory conditions),
measurements (in uncontrolled natural conditions), and computations. The role of theory is to
provide concepts which organize the facts. Many, perhaps most, of the concepts are mathematically
embodied in much simpler models than the full fluid equations. For example, a random walk is a
paradigm for the spreading of material tracers (i.e., dispersion), and a random walk is a solution of
a linear PDE with stochastic excitation (rather than deterministic fluid dynamics). Even though we
lack a full theory, we are able to make rather accurate calculations in many turbulent situations; an
example is in the design of aircraft, which obviously works rather well.
As remarked above there are no deep or complete theories of turbulence. Nevertheless there
are many models of turbulence which have simpler mathematical structure than the Navier-Stokes
or Euler Equations, some have solutions that skillfully mimic turbulent behaviors under certain
circumstances, hence are useful either as full-turbulence representations (a.k.a. Reynolds-Averaged
Navier-Stokes models, RANS) or as partial subgrid-scale (SGS) parameterizations of the effects
of unresolved, smaller-scales in Large-Eddy Simulations (LES).
Merely as an introduction of terminology, the following are some common model types:

Eddy Diffusion: Replacing turbulent advection by a diffusion operator under the hypothesis that
turbulence mixes the quantities that comprise its environment, as if the fluid evolution were laminar
with an increased diffusivity.

Instantaneous Adjustment: Assuming that turbulence mixes with such great efficiency that the
environmental distributions instantaneously adjust to a condition of marginal stability (e.g., con-
vective adjustment, a well-mixed boundary layer). (Also called a stability-bounds model.)

Moment Closure: Making closure assumptions for the averaged dynamical equations for low-
order statistical moments, e.g., (41), to break the infinite regress of moment coupling. (Mean-field
closures, second-moment single-point closures, turbulent kinetic energy equations, etc.)

Stochastic Dynamics: Devising dynamical equations in which difficult nonlinear advection is

replaced by a stochastic forcing term (e.g., inducing mixing).

Vortex Population Dynamics: Devising dynamical equations for the evolution of a population
of coherent structures, such as point- or line-vortices.

Rapid-Distortion Theory: Making a quasi-linearization of the Navier-Stokes Equation based

on the presumption that small-scale fluctuations u in the presence of large-scale background flow
U evolve in a mean-Lagrangian frame primarily through the straining effect of the latter on the
former, without significant nonlinear self-interaction, viz.,
 
∂
+ U · ∇ u + ∇φ − ν∇2 u = −(∇U) · u, ∇ · u = 0 . (42)
∂t

22
The essential eddy – mean interaction implied here is a combination of bulk advection by the mean
velocity and differential advection by the mean shear; both have the effect of putting the fluctuation
dynamics into a mean Lagrangian reference frame. Implicit in (42) is the possibility of unstable
normal modes, as well as transient (non-modal) fluctuation amplification (Farrell, 1988).
Finally there are mathematical relations, sometimes confusingly called models, that are ady-
namical representations of the empirically determined statistical structure of turbulent fields. In
this category are (multi-)fractals (Frisch, 1995), log-normal distributions (e.g., of the single-point
PDF of dissipation rate), and inertial-range scaling laws (She & Leveque, 1994; Lundgren, 2008).
We will revisit all these types of turbulence models in particular situations. The reader is
referred to, e.g., Lesieur (1997), Frisch (1995), Monin & Yaglom (1971, 1975), and Pope (2000)
for extensive elaboration on models.

7 Atmospheric and Oceanic Regimes of Turbulence

There are many different combinations of physical conditions for the turbulence in the ocean and
atmospheric troposphere (Fig. 8). Here we survey the distinctive principal regimes, though all
combinations of conditions can and do occur. In all regimes the assumption of Re  1 is im-
plicit. (A particularly hot topic is the mechanisms for transition out of geostrophic turbulence into
smaller-scale, less rotationally constrained flows; McWilliams, 2008.)

a⊕ L

10 4 km 10−103 km 10−103 m 10 −2−1 m

submesoscale
balanced instability: ageostrophic
barotropic, baroclinic instability stratified,
shear turbulence
general geostrophic isotropic 3D
circulation turbulence turbulence
planetary
boundary layer
(analogous to turbulence
2D turbulence)
astronomical dissipation
forcing and mixing

cascade

Figure 8: Schematic diagram of the regimes of turbulence in the atmosphere and ocean in a broad
sweep of energy from the astronomically-forced planetary scale down to the microscale where
mixing and dissipation occur.

23
Homogeneous Turbulence: Spatially isotropic and homogeneous (hence without boundaries)
with f = g = 0 (hence without buoyancy forces). (Note the dual meanings of homogeneous:
uniform in space and uniform in density.) The Reynolds number is
VL
Re =  1,
ν
where V and L are characteristic scales for the turbulence. Usually boundary effects are ignored.

Shear Turbulence: In the presence of a mean shear flow, ∇u, with

V
Sh = . 1, (43)
SL
where S = | ∇u |, V is a characteristic velocity of the turbulent flow, and L is a characteristic
scale across the mean shear.
With stable density stratification and a buoyancy frequency N defined by

db ∂log[ρ]
N 2 (z) = = −g > 0, (44)
dz ∂z
shear turbulence can also occur if the Richardson number,

N2
Ri = . 1.
S2

Convection: In the presence of an unstable density stratification, with ∂z b < 0 (e.g., Rayleigh-
Benard convection). If the convecting region is bounded by a stably stratified region, then it is
called penetrative convection because it will tend to spread into the stable region by entrainment.

Cloud Turbulence: radiation and phase changes of water are important.

Stably Stratified Turbulence: In the presence of a stable buoyancy stratification, with

V
Fr = ≤ 1, (45)
NL
where L is a vertical length scale. This implies that Ri ∼ F r−2  1.

Boundary Layer Turbulence: A layer near a material boundary through which fluid parcels
mostly do not pass but fluxes of momentum and/or buoyancy do pass. pThe primary types of
boundary layer turbulence are (a) shear, with friction velocity u∗ = |τ |/ρo and boundary
1/3
tangential stress τ ; (b) convection, with convective velocity w∗ = [BH] , surface buoyancy flux
B, and boundary layer depth H;p and (c) Langmuir in the ocean, with Stokes drift velocity from the
surface gravity waves, uSt = gk 3 a2 ∼ u∗ , where a and k are the dominant wave amplitude
and wavenumber plus additional mixing from breaking waves.

24
Geostrophic Turbulence: In the presence of rotation and stable stratification, with F r ∼ Ro
and
V
Ro = ≤ 1. (46)
fL
This regime also includes influences of the spatial variation of the planetary vorticity, β =
df (y)/dy, and topographic variations of the fluid depth δH, with
βL δH
, ∼ Ro . (47)
f H
It also includes interactions with the general circulation u, with V 0 ∼ V .

Rotating Turbulence: Ro  1 with constant ρ.

Weak Wave Turbulence: In the presence of weakly nonlinear waves (e.g., surface or internal
gravity, inertial, coastal, or Rossby), with ak  1 and V /σL  1, where a is wave amplitude, k
is wavenumber, and σ is wave frequency. (This topic is not otherwise discussed in this course.)

Appendix A: Role of Pressure

We can simplify the turbulence problem by formally eliminating pressure as a dependent variable.
This is possible because of the incompressibility condition in (1). We take the divergence of the
momentum equations in (1) and note that ∇ · ∂t u = 0. This yields a relation called the Pressure
Poisson Equation:
∇2 φ = ∇ · [ −(u · ∇)u + ν∇2 u + ẑb − f ẑ × u ]. (48)
Since there are no time derivatives in (48) , p ( = ρo φ) is a purely diagnostic field which is wholly
slaved to u. After solving (48) for φ, it can be substituted for the pressure-gradient force in the
momentum equations. Its role is to maintain incompressibility under the action of all other forces.
Therefore, it is merely a question of mathematical or computational convenience whether φ is
explicitly retained as a dependent variable.
We can illustrate this for a the special flow in (17). Since w = ẑ · u = 0 here, the vertical
momentum equation is simply hydrostatic balance,
φz = b,
and the incompressibility condition is only horizontal,
∇h · uh = 0 ,
where the subscript “h” denotes a horizontal vector. (Note that horizontal incompressibility is
automatically satisfied by the form (17).) Thus, the important remaining constraint on φ in this
case comes from taking the horizontal divergence of the horizontal momentum equations; the
result is
∇2h φ = ẑ · [∇h × f uh − 2∇h u × ∇h v]
= ∇h · [f ∇h ψ] + 2Jh (ψx , ψy ) . (49)

25
The relation (49) is a simplified form of the general pressure relation (48), and it is called gradient-
wind balance. It says that the pressure Poisson equation is forced by the divergence of the Coriolis
force and an advective force that can be interpreted as a generalized centrifugal force in a curvilin-
ear coordinate frame defined by the streamlines (i.e., isolines of ψ).
We can alternatively express the generalized force divergence in (49) as
2
2J(ψx , ψy ) = 2(ψxx ψyy − ψxy )
1
= [ (ψxx + ψyy )2 − (ψxx − ψyy )2 − 4ψxy
2
]
2
1
= [ (vx − uy )2 − {(vx + uy )2 + (ux − vy )2 } ]
2
1 2
= [ ζ − {s2 } ] , (50)
2
where ζ is the vertical component of vorticity and s is the magnitude of the horizontal strain rate,
here defined by reference to the terms in the preceding line. Thus, the sense of curvature (i.e., the
topology of the level surfaces) of φ is established by a competition between the magnitudes of the
vorticity and strain rate in this special case, apart from the influence of the Coriolis force. This
is another indication of the important and competing roles of vorticity and strain rate in advective
fluid dynamics.
If we further simplify (49) to axisymmetric flow and f = fo , i.e.,

ψ(x, y, z, t) = ψ(r, z, t)

for r2 = x2 + y 2 , then (49) becomes

 
1∂ ∂φ 1∂ 
fo rV + V 2 ,

r = (51)
r ∂r ∂r r ∂r
or
∂φ V2
= fo V + , (52)
∂r r
for V = ∂r ψ the azimuthal velocity, which makes explicit the preceding interpretation of the force
terms in (49).

References
Batchelor, G., 1953: The Theory of Homogeneous Turbulence, Cambridge.
Batchelor, G., 1967: An Introduction to Fluid Dynamics, Cambridge.
Farrell, B.F., 1988: Optimal excitation of perturbations in viscous shear flow. Phys. Fluids 31,
2093-2102.
Gollub, J.P., S.V. Benson, and J.F. Steinman, 1980: Ann. N.Y. Acad. Sci. 357, 22- .
Frisch, U., 1995: Turbulence, Cambridge.

26
La Porta, A., G.A. Voth, A.M. Crawford, J. Alexander, and E. Bodenschatz, 2001: Fluid particle
accelerations in fully developed turbulence. Nature 409, 1017-1019.
Lesieur, M., 1997: Turbulence in Fluids, Klewar.
Lumley, J.L., and H. Panofsky, 1964: The Structure of Atmospheric Turbulence, Interscience.
Lundgren, T.S., 2008: Turbulent scaling. Phys. Fluids 20, 031301 – 1-10.
McWilliams, J.C. [with unpublished contributions from A. Bracco, X. Capet, J.-P. Laval, M.J.
Molemaker, & P.P Sullivan], 2007: Extreme events in geophsycial turbulence and waves: What
populates the tails of the distribution functions? In: ’Aha Huliko’a Proceedings: 2007, P. Muller,
ed., U. Hawaii, Honolulu, 73-80.
McWilliams, J.C., 2008: Fluid dynamics on the margin of rotational control. Environmental Fluid
Mechanics 8, 441-449.
Monin, A.S., and A.M. Yaglom, 1971 & 1975: Statistical Fluid Mechanics: Mechanics of Turbu-
lence 1 & 2, MIT Press.
Owens, W.B., 1991: A statistical description of the mean circulation and eddy variability in the
northwestern Atlantic using SOFAR floats. Prog. in Oceanog. 28, 257-303.
Pope, S.B., 2000: Turbulent Flows, Cambridge.
Prigogine, I., 1980: From Being to Becoming, Freeman.
Prigogine, I., and I. Stengers, 1984: Order Out of Chaos, Bantam.
She, Z.S., and E. Leveque, 1994: Phys. Rev. Lett. 72, 336-339.
Tritton, D., 1988: Physical Fluid Dynamics, Oxford.

27