From SIAM News, Volume 38, Number 7, September 2005

Computational Models of Turbulence:

The LANS Model
and the Role of Global Analysis
By Darryl D. Holm and Edriss S. Titi
Over the last 50 years, researchers have proposed numerous computational models of turbulence for obtaining closure. The objective of closure is to capture certain statistical features of the physical phenomenon of turbulence at computably low resolution by mimicking the average
effects of the small scales on the larger ones, without calculating the former explicitly.
The Lagrangian-averaged NavierStokesalpha (LANS) model (also known in the literature as the viscous CamassaHolm equations, or
the NavierStokes model) is the first turbulence closure model produced by Lagrangian averaging. The LANS model combines
Lagrangian-averaged nonlinearity with NavierStokes viscosity. In modifying the nonlinearity of the NavierStokes equation, instead of its dissipation, LANS provides a way to reach closure without enhancing viscosity. Derivation from these first principles implied mathematical theorems for its solutions, thereby guaranteeing that the most basic features and statistical properties of the flow (energy transport, circulation, variability, instability, dissipation anomaly, and intermittency) at length scales above the effective cutoff scale of alpha are all modeled accurately.
It has been proved analytically that the LANS solutions converge to certain solutions of the three-dimensional Navier
Stokes solutions. Moreover, the models solutions for nonzero alpha possess a global attractor whose fractal and Hausdorff dimensions are finite.
The finite number of degrees of freedom in this model guarantees that the solutions are computable in finite resolution. Details can be found in
[4], [1011].
The theorem-based approach of the LANS model development raises new mathematical possibilities for the derivation and analysis of
other computational models of turbulence. Application of the alpha model is still in its infancy, but results so far suggest that this new approach
will complement, and in some cases subsume, earlier approaches for modeling turbulence in real-world applications.

What Do the NavierStokes Equations Say about Turbulence?

Turbulence is an outstanding unsolved multiscale problem of classical physics. It occurs spontaneously in a fluid, when forcing by stirring
at the large scales is transferred by convection into slender, swirling circulations in the flow. These coherent swirling blobs of fluid, pierced
by vortex lines and bounded by material circulation loops, are called eddies. The eddies are Lagrangian structuresthat is, they travel with the
flow, stretching into extended shapes (sheets or tubes) as they follow the flow induced by the vortex lines that pierce them. The coherent eddies,
sheets, and tubes of vorticity, stretching into finer and finer shapes, can be thought of as the sinews of turbulence.
The characteristic features of a turbulent flowthe distribution of eddy sizes, shapes, speeds, vorticity, circulation, nonlinear convection, and
viscous dissipationcan all be captured with the exact NavierStokes equations. The NavierStokes equations correctly predict how the cascade of turbulent kinetic energy and vorticity accelerates and how the sinews of turbulence stretch out to finer and finer scales, until their motions
reach scales of only a few molecular mean free paths, where they can finally be dissipated by viscosity into heat. Further details can be found
in [9], [1213], [28]. The fidelity of the NavierStokes equations in capturing the cascade of turbulence, however, is also their downfall for direct
numerical simulations of turbulence.
The Need for a Computational Turbulence Model
Turbulence is a paradigm for nonlinear phenomena. Indeed, without the nonlinear term representing convection in the NavierStokes equations, the dynamics of the time-dependent Stokes problem would be trivial. For this reason, the dimensionless Reynolds number Re, a measure
of the intensity of the nonlinear inertial term in the NavierStokes equations in comparison with the viscous linear effects, is usually used as an
indicator of the complexity of a turbulent flow. Heuristic physical arguments attributed to Landau suggest that the number of active degrees of
freedom required to simulate the turbulent cascade in high-Reynolds-number flows increases as Re 9/4. Because turbulent flows often have
Re > 10 6, simulation requirements quickly outstrip the numerical resolution capabilities of even the largest, most powerful computers.
To make turbulence computable, then, it is necessary to forsake computing some of the fine details of turbulent flows. To do so without
compromising the main statistical features of the physical phenomena, scientists have developed various approximate models that halt the cascade into smaller, faster eddies. Most models accomplish this by causing eddies below a certain size to dissipate computationally into heat. This
dissipative imperative causes errors, however, because it damps out the variability (known as intermittency) in the larger-scale flow that is
caused by the myriad motions at small scales interacting nonlinearly in the fields of the larger motion. Thus, computational turbulence closure
models based on reducing Reynolds number by enhancing viscous dissipation over its NavierStokes value run the risk of producing unrealistically low variability.

Clues from Mathematical Analysis

Remarkably, one of the clues for understanding how turbulence closure models can be developed without enhancing viscous dissipation came
from the great analyst Leray. In the first regularization of the NavierStokes equations, Leray [22] modified their nonlinearity to the well-known
form

v + u v + p
t
= v + F, div v = 0,

with v = 0 on the boundary. Here, is the (constant) kinematic viscosity coefficient, F is the prescribed external force, and u = G * v
is a filtered version of the regularized velocity v . The filtering operation is defined by G * v = G(x,y) v (y) d 3y for a radially symmetric
smooth kernel G(x,y) of characteristic width . The NavierStokes equations for v are recovered in the limit as 0, so that u v.
In an insightful review of the Leray regularization of the NavierStokes equations [14], G. Galovotti made the point that the Leray regularization no longer satisfies the Kelvin circulation theorem. Galovotti [14] challenged the turbulence community to produce a regularization of
the NavierStokes equations that does satisfy a Kelvin circulation theorem. Remarkably, the combination of Lagrangian averaging (time-averaging at fixed Lagrangian coordinates) and Taylors hypothesis (that the fluctuations are of low enough power to be regarded as carried along
by the mean flow) leads to the LANS model, which produces a regularized equation set that answers Galovottis challenge. These regularized equations constitute the LANS model:

1
v + u v + uT v
t
2

( u 2 + 2

u 2 )+ p

= v + F, div u = 0.

(1)

Here, is a constant of length; the filtering relation u = G * v for the LANS model is specified to be v u 2 u .

Applying Filtering in
Kelvins Circulation Theorem
The filtering kernel G for the LANS model turns out to be the Greens function for the Helmholtz operator, (1 2).
As stated earlier, the LANS motion equation satisfies the Kelvin circulation theorem:

d
dt

c u
(

v dx

c u ( t v + u v uT v ) dx

c u ( v + F ) dx.

The circulation theorem tells us that the rate of change of momentum per unit mass v around a closed material loop c( u ) moving with
velocity u = G * v is given by the integral around that loop of the tangential component of the sum over forces (viscous and external) acting on the fluid.
This statement of the circulation theorem can also serve as a mnemonic, allowing the derivation of other regularized turbulence models of
the LANS type simply by specifying a different filtering kernel G.

A Brief History of the LANS Model

Although we can now write the LANS model directly from its circulation theorem, the approach used historically [4] for deriving the
closed Eulerian form (1) of the LANS motion equation was based on a combination of two earlier results. First, the Lagrangian-averaged
variational principle of [16] was applied to derive the inviscid averaged nonlinear fluid equations, obtained by averaging Hamiltons principle
for fluids over the rapid phase of their small turbulent circulations at fixed Lagrangian coordinates. (This step had its own precedent in the earlier work on Lagrangian-averaged fluid equations of [1].)
Second, the EulerPoincar theory for continuum mechanics of [20] was used to produce the Eulerian form of the equations resulting from
the Lagrangian-averaged fluid variational principle. This step determined the relation between the momentum per unit mass v , the velocity of
the Lagrangian-averaged fluid u , and the Lagrangian fluctuation statistics. Next, Taylors hypothesis of frozen-in turbulence circulations was
invoked for closing the Eulerian system of Lagrangian-averaged fluid equations, by obtaining the explicit relation v u 2 u . Finally,
NavierStokes Eulerian viscous dissipation was added, so that viscosity would cause diffusion of the newly defined Lagrangian-averaged
momentum and monotonic decrease of its total Lagrangian-averaged energy. The theory can be either isotropic or anisotropic. Details of the
derivation are provided in [1718]. An alternative derivation is given in [24].
Turbulence Modeling: A New Role for Leray Analysis
At this point, we can consider the LANS model as simply a regularization of the NavierStokes equations and re-examine its properties

from the viewpoint of Lerays analysis. It turns out, moreover, that the same ideas that restore Kelvins circulation theorem to Lerays regularization of the NavierStokes equations also provide a basis for proposed computational models of turbulence, including large eddy simulations.
The converse is also true: Any proposed model of turbulence will lead to likely candidates for application of Lerays analysis. Consequently,
the classical Leray analysis of the NavierStokes equations has a new role in the study of the analytical properties of turbulence models. Indeed,
the Leray model itself was recently found to be a viable candidate for computational modeling of turbulence [8], [15].

Relation of LANS Inertial Subrange to Taylors Turbulence Microscale

Of the various remarkable properties of the LANS system of equations (1), we discuss here only one: the two different scalings of its inertial regime, depending on whether the eddies are larger or smaller than alpha. In fact, the KrmnHowarth theorem for the LANS system,
discussed in [19], implies that the translational kinetic energy spectrum E(k) (i.e., the energy spectrum of the translational velocity as a function of the wave number k) changes from E(k) ~ k 5/3 for large scales, corresponding to wave numbers k  1, to E(k) ~ k 3 for small scales,
corresponding to wave numbers k  1. A dimensional argument justifying this change of scaling in the inertial regime for the LANS model
was first given in [10].
With this wave-number scaling, the inertial range is shortened for circulations with high wave numbers (i.e., k > 1) in the LANS model.
With fixed, the wave number k at the end of the second, steeper, k 3 regime of the LANS inertial range is determined in [10] to be

(1 )

1/ 3

k ~

2/3
.
kKo

This is the wave number k at which dissipation balances nonlinearity in turbulence described by the LANS equations. In this formula,
kKo is the Kolmogorov dissipation wave number, at which dissipation balances nonlinearity in turbulent solutions of the NavierStokes equations. Because kKo scales with integral-scale Reynolds numbers as LkKo Re 3/4, with L denoting the integral scale (or domain size), we find that
dissipation balances nonlinearity for the LANS model at Lk Re 1/2. Remarkably, the wave number for the well-known Taylor microscale
also scales as Re 1/2 [27]. Thus, for the three progressively larger wave numbers:
L/ < Lk Re 1/2 < L kKo Re3/4.
Shortening the inertial range for the LANS model to Lk < Lk Re 1/2, rather than the Lk < LkKo Re3/4 of the NavierStokes equations,
implies fewer active degrees of freedom in the solution for LANS, which, as discussed below, makes LANS much more computable than
NavierStokes at high Reynolds numbers.

Counting Degrees of Freedom

For turbulence that is extensive in the thermodynamic sense, we might expect the number of active degrees of freedom Ndof for alphamodel turbulence to scale as

N dof
( Lk )3 ~ ( L / )

( LkKo )2 ~

L 3/2
Re ,

where k is the end of the LANS inertial range and Re = L4/3 1/3/ is the integral-scale Reynolds number (with total energy dissipation rate
and viscosity ). Because the corresponding number of degrees of freedom for NavierStokes with the same parameters is
NS
N dof
( LkKo )3 ~ Re 9 / 4 ,

a possible trade-off emerges in the relative Reynolds number scaling of the two models, provided resolution down to the Taylor microscale. (In
practice, users of the LANS model often obtain acceptable results by setting the resolution scale at just half the size of alpha.)
Should these estimates not prove overly optimistic, the implication would be a two-thirds power scaling advantage for use of the LANS
model. In other words, in needing to resolve only the Taylor microscale, the LANS model could compute accurate results (at scales larger
than alpha) by using two decades of resolution in situations that would require three decades of resolution for the NavierStokes equations, at
sufficiently high Re. This is because the number of degrees of freedom for the two models scales as

N dof
( Lk )3 ~ ( L / ) ( LkKo )2

L 3/2
NS
~ (N dof
Re

2/3

Re 3/2 Scaling Estimate for the Hausdorff Dimension of the LANS Global Attractor
These dimensional arguments were substantiated by a slightly better estimate with the same Re3/2 scaling when the fractal and Hausdorff
dimensions of the global attractor for the LANS model were estimated in [11]. In addition, the well-posedness of the LANS model in a
bounded domain was confirmed in [23].

Numerical Speed-up of LANS over NavierStokes

A numerical argument for the speed-up advantage of the LANS model in comparison with direct simulations of the NavierStokes equaNS / N )1/3 in relative increased computational speed for each spatial dimentions goes as follows. The LANS model gains one factor of (Ndof
dof
sion and another factor (at least) for the accompanying reduced CourantFriedrichsLevy (CFL) time-step restriction. Altogether, the gain in
speed would be
NS
N dof

N dof

4 / 3
~ Re.

Because /L  1 and Re  1, the two factors on the right side do compete; the Reynolds number should eventually win out, however,
because Re can continue to increase, while /L is expected to tend to a constant value, say /L = 1/100, at high (but experimentally attainable)
Reynolds numbers, at least for simple flow geometries. Empirical indications of this tendency were found in [46] in comparisons of steady
LANS solutions with experimental mean-velocity-profile data for turbulent flows in pipes and channels.
Thus, according to this scaling argument, a factor of 10 4 in increased speed for accurate computation of scales greater than could be
achieved, by using the LANS model at the Reynolds number for which kKo /k = 10. An early indication of the feasibility of obtaining such
increases in computational speed was realized in the direct numerical simulations of homogeneous turbulence reported in [7], in which
kKo /k  4 and the full factor of 4 4 = 256 in computational speed was obtained with spectral methods in a periodic domain at little or no cost of
accuracy in the statistics of the re-solved degrees of freedom, i.e., those with k < 1.

Outlook
Further steps are being taken at Los Alamos National Laboratory, the National Center for Atmospheric Research, and elsewhere to test
whether the LANS model will continue to live up to its promise for fast accurate numerical simulations of turbulence when additional physical processes are included in these computations. At Los Alamos, the LANS model has been extended to include rotation, topography, and
buoyancy stratification for applications in ocean and atmosphere circulation studies for global climate modeling. At NCAR, the LANS model
has been extended to include magnetic fields, so that researchers there now stand at the threshold of being able to model the effects of turbulence on the dynamics of the geodynamo and the solar dynamo.
Several other variants of the LANS model have also been investigated analytically and numerically for incompressible turbulence, e.g., in
pipes and channels. The steady solutions of all of these variants compare well with the measurements of mean velocity in turbulent flows in
pipes and channels over a wide range of Reynolds numbers and for a constant value of alpha that is small (about one percent of the pipe diameter or channel width). Numerical results for two of the primary variants, the Leray model [8], [15] and the Clark model [2], are promising; analytical estimates prove that everything said here about the computability, energy spectrum, and finite-dimensional global attractor for
the LANS model also holds true for these two alternative models. The characteristic preservation of the Kelvin circulation theorem for
NavierStokes of the LANS model, however, is not a feature of the Leray and Clark models.
The challenge first enunciated in [14] of developing a regularization of the NavierStokes equations that preserves Kelvins circulation theorem was answered (accidentally) in the development of the LANS model. After a promising beginning, the eventual roles of both Kelvins circulation theorem and the global analysis of PDEs in developing and analyzing computational turbulence models remain to be fully determined.
Acknowledgments
The authors are grateful to friends and colleagues for their comments, encouragement, and constructive suggestions, especially the Turbulence
Working Group at Los Alamos and its advisory committee (S.Y. Chen, A.J. Domaradzky, R. Donnelly, G. Eyink, U. Frisch, R.M. Kerr, A.
Pouquet, and S.R. Sreenivasan), as well as the participants in the groups annual turbulence workshops during the past four years.
The work described in this article was supported by the U.S. Department of Energy (by contract W7405ENG36 and by the AMR/MICS
program of the Office of Science) and the National Science Foun-dation. The work of E.S.T. was supported in part by NSF (DMS0204794)
and by the U.S. Civilian Research and Development Foundation (RM12343MO-02).

References
[1] D.G. Andrews and M.E. McIntyre, An exact theory of nonlinear waves on a Lagrangian-mean flow, J. Fluid Mech., 89 (1978), 609646.
[2] C. Cao, D.D. Holm, and E.S. Titi, On the Clark model of turbulence: global regularity and long-time dynamics, submitted;
http://arxiv.org/abs/nlin.CD/0412007.
[3] C. Cao and E.S. Titi, Analytical study of the long-term dynamics of the Smagorinksy model of turbulence, submitted.
[4] S. Chen, C. Foias, D.D. Holm, E.J. Olson, E.S. Titi, and S. Wynne, The CamassaHolm equations as a closure model for turbulent channel and pipe flows, Phys. Rev. Lett., 81 (1998), 53385341.
[5] S. Chen, C. Foias, D.D. Holm, E.J. Olson, E.S. Titi, and S. Wynne, A connection between the CamassaHolm equations and turbulence
in pipes and channels, Phys. Fluids, 11 (1999), 23432353.
[6] S. Chen, C. Foias, D.D. Holm, E.J. Olson, E.S. Titi, and S. Wynne, The CamassaHolm equations and turbulence in pipes and channels, Physica
D, 133 (1999), 4965.
[7] S.Y. Chen, D.D. Holm, L.G. Margolin, and R. Zhang, Direct numerical simulations of the NavierStokes alpha model, Physica D, 133
(1999), 6683.
[8] A. Cheskidov, D.D. Holm, E.J. Olson, and E.S. Titi, On a Leray model of turbulence, Proc. Roy. Soc. London A, to appear, 2005.
[9] C. Foias, What do the NavierStokes equations tell us about turbulence? in Harmonic Analysis and Nonlinear Differential Equations
(November 35, 1995, University of California, Riverside), Contemporary Mathematics: 208, American Mathematical Society, Providence, 1997,
151180.

[10] C. Foias, D.D. Holm, and E.S. Titi, The NavierStokesalpha model of fluid turbulence, Physica D, 152 (2001), 505519.
[11] C. Foias, D.D. Holm, and E.S. Titi, The three dimensional viscous CamassaHolm equations, and their relation to the NavierStokes equations and turbulence theory, J. Dyn. and Diff. Eqns., 14 (2002), 135; http://xxx.lanl.gov/abs/nlin.CD/0103039.
[12] C. Foias and R. Temam, The connection between the NavierStokes equations, dynamical systems and turbulence theory, in Directions
in Partial Differential Equations, M.C. Randall, R.H. Rabinowitz, and M.E. Turner, eds., Academic Press, Boston, 1987, 5573.
[13] U. Frisch, Turbulence: The Legacy of A.N. Kolmogorov, Cambridge University Press, Cambridge, UK, 1995.
[14] G. Galovotti, Some rigorous results about 3D NavierStokes, in Les Houches 1992 NATO ASI Meeting on Turbulence in Extended
Systems, R. Benzi, C. Basdevant, and S. Ciliberto, eds., Nova Science, New York, 1993, 4581.
[15] B.J. Geurts and D.D. Holm, Regularization modeling for large-eddy simulation, Phys. Fluids, 15 (2003), L13L16.
[16] I. Gjaja and D.D. Holm, Self-consistent wave-mean flow interaction dynamics and its Hamiltonian formulation for a rotating stratified
incompressible fluid, Physica D, 98 (1996), 343378.
[17] D.D. Holm, Fluctuation effects on 3D Lagrangian mean and Eulerian mean fluid motion, Physica D, 133 (1999), 215269.
[18] D.D. Holm, Lagrangian averages, averaged Lagrangians, and the mean effects of fluctuations in fluid dynamics, Chaos, 12 (2002),
518530.
[19] D.D. Holm, KrmnHowarth Theorem for the Lagrangian averaged NavierStokes alpha (LANS ) model, J. Fluid Mech., 467
(2002), 205214.
[20] D.D. Holm, J.E. Marsden, and T.S. Ratiu, The EulerPoincar equations and semidirect products with applications to continuum theories, Adv. in Math., 137 (1998), 181.
[21] A.N. Kolmogorov, Dissipation of energy in a locally isotropic turbulence, Dokl. Akad. Nauk SSSR, 32 (1941), 141143. (English translation in Proc. Roy. Soc. London A, 434 (1991), 1517.)
[22] J. Leray, Sur le mouvement dun liquide visqueux emplissant lespace, Acta Math., 63 (1934), 193248.
[23] J.E. Marsden and S. Shkoller, Global well-posedness for the Lagrangian averaged NavierStokes (LANSalpha) equations on bounded domains, Philos. Trans. Roy. Soc. London A, 359 (2001), 1449.
[24] J.E. Marsden and S. Shkoller, The anisotropic Lagrangian averaged Euler and NavierStokes equations, Arch. Rational Mech. Anal.,
166 (2003), 2746.
[25] L.F. Richardson, Atmospheric diffusion shown on a distance-neighbour graph, Proc. Roy. Soc. London A, 110 (1926).
[26] G.I. Taylor, Diffusion by continuous movements, Proc. London Math. Soc., 20 (1921), 196212.
[27] G.I. Taylor, The spectrum of turbulence, Proc. Roy. Soc. London A, 164 (1938), 476490.
[28] M.I. Vishik and A.V. Fursikov, Mathematical Problems of Statistical Hydrodynamics, Kluwer Academic Press, Dordrecht, 1980,
709737.
Darryl D. Holm ([email protected]) is a researcher in the Computational and Computer Science Division at Los Alamos National Laboratory
and a professor of mathematics at Imperial College, London. Edriss S. Titi ([email protected]) is a member of the Departments of Mathematics
and of Mechanical and Aerospace Engineering at the University of California, Irvine, and the Department of Computer Science and Applied
Mathematics at the Weizmann Institute of Science.