Physica D 127 (1999) 198–232

A fractal model for large eddy simulation of turbulent flow

A. Scotti1 , C. Meneveau ∗
Department of Mechanical Engineering, Johns Hopkins University, Baltimore, MD 21218, USA
Received 1 September 1997; received in revised form 10 June 1998; accepted 23 September 1998
Communicated by R. Temam

Abstract
A new class of subgrid closures for large eddy simulation (LES) of turbulence is developed, based on the construction of
synthetic, fractal subgrid-scale fields. The relevant mathematical tool, fractal interpolation, allows to interpolate the resolved
velocity with fields that have fluctuations down to much smaller scales and to compute the required stresses explicitly. In one
dimension, the approach is used in the context of the coarse-grained Burgers equation. Then, fractal interpolation is extended
to three dimensions and is used to formulate a subgrid model for the filtered Navier–Stokes equations. The model is applied to
LES of both steady and freely decaying isotropic turbulence. We find that the assumption of fractality per sè is not enough to
yield physically meaningful results, and we explore several variants of the model in which the rules to generate the synthetic
fields explicitly incorporate the condition that energy dissipation take place. In one dimension, this is accomplished by means
of an additional transport equation that allows to dynamically determine the fractal dimension. In three dimensions, good
results are obtained only once the fractal dimension is allowed to vary in different eigendirections of the resolved strain-rate
tensor so as to (nearly) maximize energy dissipation. c 1999 Elsevier Science B.V. All rights reserved.
Keywords: Large eddy simulation; Fractal; Turbulence modeling

1. Introduction and overview

1.1. Statement of the problem

Many flows of interest in science and engineering display complex spatial and temporal structures (eddies)
spanning a wide range of scales. The ratio between the largest (L) and smallest (η) scale can easily exceed 104 in
typical engineering applications, and can be as high as 106 or higher in geophysical applications. Since the nonlinear
interaction between eddies of different sizes eludes even the most sophisticated analytical approaches, one must
resort to either extensive experimentation or direct numerical simulation (DNS) of the governing equations. The
latter approach has gained strength by the rapid increase in the power of digital computers during the past 20 years.
Despite this fact, DNS of flows for which the ratio L/η is much larger than 102 are still prohibitive, as they require at
∗ Corresponding author; e-mail: [email protected]
1 Present address: Department of Physical Oceanography, Woods Hole Oceanographic Institution, Woods Hole, MA 02543, USA.

0167-2789/99/$ – see front matter c 1999 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 7 - 2 7 8 9 ( 9 8 ) 0 0 2 6 6 - 8
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 199

least O[(L/η)3 ] operations per time step, not to mention the memory requirements. Hence DNS has been confined
to relatively simple flows at small to moderate Reynolds numbers.
It is to be noted that for many applications only the large eddies are of importance. Hence, beginning with the
pioneering work of Smagorinsky [64] and Deardorff [16], considerable attention has been devoted to the problem
of finding effective equations for the motion of large eddies without involving the small eddies (see [27,38,44,54]
for a comprehensive review). This approach is called large eddy simulation (LES). The problem is formulated
mathematically in the following way: let u(x, t) be the velocity field of fluid whose density is ρ and viscosity ν
satisfying the Navier–Stokes (NS) equations [34]
∂u
+ (u · ∇)u = −(1/ρ)∇p + ν∇ 2 u, ∇ · u = 0. (1)
∂t
p(x, t) denotes the pressure field. The field describing the motion of eddies of size larger than ∆ is obtained by
convolving the velocity field u(x, t) with an appropriate kernel (filter) G∆ (x, x0 ) which eliminates the fluctuations
below scale ∆, according to
Z
ũ(x, t) = G∆ (x, x0 )u(x0 , t) dx0 . (2)

It can be shown by inspection that, with mild assumptions of regularity and homogeneity for the filter, away from
boundaries the filtered field ũ(x) obeys the following set of equations derived from the NS equations
∂ ũ
+ (ũ · ∇)ũ = −(1/ρ)∇ p̃ − ∇ · ␶ + ν∇ 2 ũ, ∇ · ũ = 0, (3)
∂t
where
τij = ug
i uj − ũi ũj . (4)
The tilde, now and in what follows, denotes a quantity filtered according to Eq. (2). As they are, these equations are
not useful, as the term τij , the sub grid scale (SGS) stress tensor responsible for the interactions between the eddies
larger than ∆ with those smaller than ∆, requires the knowledge of the unfiltered velocity field to be computed.
However, the universality of small eddies in high Reynolds number turbulence [31,32,50,5,37] is often invoked to
argue that it is possible to write a model τij = τij [ũ] in a way that applies independently from how the flow is set
and kept in motion. In the following, we briefly review the SGS models currently available.

1.2. Review of current models

Current SGS models can be loosely divided in two classes: eddy viscosity models and synthetic field models. Eddy
viscosity models write τij = − 13 τkk δij = −νt (∂i ũj + ∂j ũi ) ≡ −2νt S̃ij , where S̃ij is the resolved rate-of-strain
tensor, and then assign an expressionqfor νt = νt [ũ]. The most widely used are variants of the original Smagorinsky
model [64], which writes νt = Cs2 ∆2 2S̃ij S̃ij . Cs is an empirical, dimensionless coefficient. Eddy viscosity models
were the first to be developed, have been studied and tested extensively and currently are the ones used in most
of the applications where simplicity and robustness are sought. Their drawback is that they correlate poorly with
the real stress tensor. Correlation coefficients of only 0.1–0.2 are typically obtained when comparing real stresses
with the model [2,40,66]. In other words, eddy viscosity closures represent the real physics of the subgrid scale
motion very poorly. Moreover, they are unable to predict the phenomena of backscatter of energy, where energy is
transferred from small to large eddies, which may be important in some applications [21,35,53]. Such effects have
been modeled by stochastic forcing of the equations (e.g. [35]).
Fundamental shortcomings of the eddy viscosity assumption are not surprising, since it implies a wide separation
and spectral gap between the energy containing scales and the dissipative scales (see, for example, the derivation
of the Navier–Stokes equations from the Boltzmann equation [12]). No such separation exists in turbulence.
200 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Synthetic field (or velocity estimation) models introduce a synthetic field u0 (x, t) which is supposed to capture
the essential features of the small scales of the unfiltered field u(x, t), and then compute τij from its definition,
Eq. (4). Synthetic field models have been introduced more recently and usually are based on more or less clearly
defined physical assumptions regarding the small eddies. In the following we give a brief review of models based
on synthetic fields.

1.2.1. Similarity models

Bardina et al. [2] and Liu et al. [40] have proposed a model based on the idea of scale similarity. According to
this model, the real field is approximated with the filtered field, i.e. uj ' ũj , which then leads to write

τij = CL (ũg ˜ ˜
i ũj − ũi ũj ). (5)
(In [40] the second filter is performed at scale 2∆). In a priori tests these models show a correlation of about 0.5–0.6
with the real stress (using a Gaussian filter and appropriate coarse-grained sampling [40]), which is significantly
larger than the correlation of the Smagorinsky model (about 0.1–0.2). However, it turns out that similarity models
underestimate the dissipation of energy once implemented into a LES (see below). To correct for this an ad hoc
approach was proposed, in which to the similarity model a conventional Smagorinsky term is added (mixed model,
see [2,41,67]).

1.2.2. Stretched vortex model

In this recent model by Misra and Pullin [49], the subgrid structure of the turbulence is assumed to consist of
stretched vortices (which represent local solutions to Navier–Stokes equations), with a preferred orientation. The
SGS stress is then analytically computed from such a distribution and from the subgrid kinetic energy, together with
an assumed Kolmogorov energy spectrum for the subgrid scales. A relationship between the vortices’ orientation
and the resolved field must be postulated.

1.2.3. Phase estimation model

Domaradzki and Saiki [19] have recently proposed a model where the unfiltered velocity is approximated by
expanding the resolved large-scale velocity field to subgrid-scales two times smaller than the grid scale. This is
motivated by observations that the dominant contributions to the SGS stress come from the scales in a band between
∆ and ∆/2 [40]. The estimation procedure then adjusts the magnitudes of the newly created modes by utilizing
properties of the filtering operation. The phases associated with the small scales are adjusted in order to correspond
to the small scale phases generated by nonlinear interactions of the large-scale field. The SGS stress is then computed
according to its definition.

1.3. Other attempts to generate synthetic turbulence

At large Reynolds numbers, SGS velocity fields can extend over a wide range of scales between ∆ and the
Kolmogorov scale η. Indeed, some efforts to generate broad-band synthetic velocity fields for small-scale turbulence
have been made in the past. These include the random-phase Gaussian fields of Fung et al. [25], the linear eddy
model of Kerstein [30], and several fractal models [6,29] etc.)
These schemes are able to provide fluctuations over a wide range of scales, but are not particularly well suited to
parametrize the SGS stresses since it seems difficult to evaluate them analytically from the synthetic fractal fields.
Moreover, SGS fields so generated do not always include correlations with the large-scale field. This will be shown
to be a crucial property of a model, necessary for sufficient energy dissipation to take place.

1.4. Fractal model

In this paper, we focus on developing fractal models with fluctuations over a wide range of SGS scales which
are amenable to be implemented in LES of turbulent flows. As motivation, we recall that abundant empirical ob-
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 201

servations have been made about fractality [23,42] in hydrodynamic turbulence as the Reynolds number increases.
In particular, many works have investigated the fractal properties of various sets of physical interest in turbulence,
such as isoconcentration surfaces of advected scalars [57,65] and vectors [58], or the structure of the spatial dis-
tribution of dissipation [47,56]. More directly related to the present subject matter, [63] considers the problem of
how to characterize geometrically the graph of a turbulent velocity signal, say u1 (x1 ). By analyzing a variety of
experimental data, it was shown that the signals display fractal scaling, with a fractal dimension D = 1.7 ± 0.05.
However, despite the attention devoted to this kind of investigations, the crucial task of translating the observa-
tions of fractality into a workable framework which is based on the dynamical equations, i.e. employing fractals
in a predictive fashion, has turned out to be an elusive goal. In this paper, we try to fill this void by describing
an approach that uses the assumption of fractality for the small unresolved velocity field to close the filtered NS
equations.
In Section 2 we begin by describing the basic idea in a general setting. There, details of the mathematical
framework required for the model are given: first for 1D problems and then for 3D problems. With the aid of these
tools, the fractal model is applied to the forced Burgers equation in 1D (Section 3) and to the NS equations at high
Reynolds number in 3D (Section 4). While describing different physical settings, both cases are known to display
a scale-invariant behavior for the small scales. Each problem separately is also the prototype of a large class of
problems.

2. The fractal interpolation technique

In an earlier short paper [62] the method and an application to the study of a 1D case have been described. In
this section, the main features of the approach are recalled in order to make the paper self-contained. Basically, the
proposed method reconstructs the unknown small-scale eddies of the velocity field u(x, t) by “extrapolating” eddies
of the coarse-grained field ũ(x, t) to smaller and smaller scales. A mathematical tool which allows to synthesize
such an artificial small-scale field is the so-called fractal interpolation technique (FIT) [3,4]. It is based on a
mapping W[·] which transforms ũ at scales even coarser than ∆ onto the complete signal ũ. More concretely, if
ũ represents the field ũ at some resolution ∆0 > ∆, coarser than the basic resolution ∆, the mapping produces
ũ = W[ũ]. In order to generate a “synthetic” small-scale field, this mapping is simply iterated many more times,
i.e. we create a fractal field uf (x, t) = limn→∞ W(n) [ũ] ≡ limn→∞ W[W[W[...W[ũ]...]]] (n times). Once we have
generated the full field, the SGS stress is evaluated according to Eq. (4). One important feature in our proposed
approach is that we shall be able to compute this term analytically from the parameters of the mapping W (which
depend on ũ and the fractal dimension), without having to explicitly construct the small-scale field. Once the
unknown term τij is expressed in terms of ũ, it is substituted into the filtered NS equations (3), which may then be
solved.
In the following we give an intuitive exposition of the fractal interpolation technique. We first briefly review the
case of one-dimensional interpolation. The reader interested in formal proofs is referred to [3,4]. The subsequent
extension to higher dimensions is original and the formal proofs are given in some detail.

2.1. One dimensional interpolation

Assume that a simulation in one spatial dimension is carried out in the unit interval. Let us discretize the interval
in ∆−1 segments, each centered at position xi . For simplicity, let us define a local coordinate ξ = (x − xi−1 )/(2∆)
which goes from ξ = 0 to ξ = 1 between points xi−1 to xi+1 (see Fig. 1) . On the interval [xi−1 , xi+1 ] parametrized
by ξ we consider the space Ĉ[0, 1] = {u ∈ C[0, 1] | u(0) = ũi−1 , u(1) = ũi+1 } and the map Wi defined as

Wi [u](ξ ) = di,1 u(2ξ ) + qi,1 (2ξ ) if ξ ∈ [0, 1/2],

Wi [u](ξ ) = di,2 u(2ξ − 1) + qi,2 (2ξ − 1) if ξ ∈ [1/2, 1], (6)

202 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 1. Different stages during the construction of a fractal function which interpolates between ũi−1 , ũi and ũi+1 . If u0 (ξ ) = (ũi+1 −
ũi−1 )ξ + ũi−1 , then the dashed line is Wi [u0 (ξ )], with Wi defined as in Eq. (6). The dotted line is Wi [Wi [u0 (ξ )]] and the convoluted sig-
nal is Wi10 [u0 (ξ )]. The circles mark the points to be interpolated. In this example D = 5/3(di,1 = −2−1/3 , di,2 = 2−1/3 ).

where qi,j (ξ ) are polynomials and di,j are stretching parameters. It can be verified by inspection that if

u0 (ξ ) = (ũi+1 − ũi−1 )ξ + ũi−1 , (7)

then Wi [u0 ](ξ ) is piecewise linear and

Wi [u0 ](0) = ũi−1 , (8)

Wi [u0 ](1/2) = ũi , (9)

Wi [u0 ](1) = ũi+1 , (10)

provided that the polynomials qi,1 (ξ ) and qi,2 (ξ ) are defined as follows:

qi,1 (ξ ) = [ũi − ũi−1 − di,1 (ũi+1 − ũi−1 )]ξ + ũi−1 (1 − di,1 ) = ai,1 ξ + bi,1 , (11)

qi,2 (ξ ) = [ũi+1 − ũi − di,2 (ũi+1 − ũi−1 )]ξ + ũi − di,2 ũi−1 = ai,2 ξ + bi,2 , (12)

where

ai,1 = [ũi − ũi−1 − di,1 (ũi+1 − ũi−1 )], bi,1 = ũi−1 (1 − di,1 ), (13)

ai,2 = [ũi+1 − ũi − di,2 (ũi+1 − ũi−1 )], bi,2 = ũi − di,2 ũi−1 . (14)

That is to say, the map brings a piecewise interpolation on scale 2∆ onto the piecewise interpolation on scale ∆.
The parameters di,1 and di,2 determine the vertical stretching of the left and right segments at each iteration and are
as yet unspecified parameters.
The main facts concerning Wi are summarized in the following [3]:
Theorem 2.1. For any two continuous functions u, g ∈ Ĉ[0, 1] let d(u, g) = supx∈[0,1] |u − g|. Then
(i) Ĉ[0, 1] equipped with the distance d is complete
(ii) Wi : Ĉ[0, 1] → Ĉ[0, 1].
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 203

(iii) if |di,j | < 1, j = 1, 2 then ∃ s, 0 < s < 1 such that

d(Wi [u], Wi [g]) ≤ s d(u, g).
Wi is thus a contraction on Ĉ[0, 1]. Let uf i be the unique solution of u = Wi [u]. Then uf i (1/2) = ũi .
(iv) If uf i and ufj are two fixed points relative to two different sets of coefficients in Eqs. (11) and (12) then
d(Wi [uf i ], Wj [uf i ])
d(uf i , ufj ) ≤ .
1−s
(v) If 2 > |di,1 | + |di,2 | > 1 the graph of the fixed point is a fractal with Hausdorff dimension D given by
log(|di,1 | + |di,2 |)
D =1+ .
log 2
If |di,1 | + |di,2 | ≤ 1, the fractal dimension of the graph is one.

A pictorial way to construct the fixed point of Wi is shown in Fig. 1. We start from the linear interpolation of the
field at scale 2∆ given by Eq. (7) and we successively apply the map Wi . Shown are the first two steps as well as
the 10th iteration of the map.

2.1.1. Fractal calculus: integration

Interestingly, it is now possible to evaluate, by recurrence, a variety of integrals of the fractal function. In particular,
R1
all the integrals 0 ξ m unf (ξ ) dξ can be computed in closed form. Details can be found in [3]. We are interested in
integrals of powers of uf (ξ ), since the evaluation of SGS stresses requires us to integrate (filter) nonlinear terms.
The following formulae summarize the results in 1D. For notational simplicity, in what follows in this section we
use u(ξ ) for uf i (ξ ). If we write
Q(ξ ) = q1 (2ξ ) if ξ ∈ [0, 1/2],
Q(ξ ) = q2 (2ξ − 1) if ξ ∈ [1/2, 1] (15)
and
Z 1
Qm = ξ m Q(ξ )dξ, (16)
0
Z 1
Um = ξ m u(ξ )dξ, (17)
0
Z 1
Ul,m = ξ m ul (ξ )dξ, (18)
0

then we have
 
Pm−1 m
Qm + 2−(m+1) d2 k=0 Uk
k
Um = . (19)
1 − 2−(m+1) (d1 + d2 )
Furthermore, introducing the coefficients K(l, m, p, j ) defined by
  l+m−p
X
−(m+1) l p p
2 d1 [q1 (ξ )]l−p ξ m + d2 [q2 (ξ )]l−p (ξ + 1)m ≡ K(l, m, p, j )ξ j , (20)
p
j =0
204 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

we can write
 
Pl−1 Pl+m−p Pm−1 m
p=0 j =0 K(l, m, p, j )Up,j + 2−(m+1) d2 j =0 Ul,j j
Ul,m = . (21)
1 − 2−(m+1) (d1l + d2l )
R1
For instance, solving for U0,1 = 0 u(ξ )dξ yields
Z 1 ũi−1 + 2ũi + ũi+1 − (di,1 + di,2 )(ũi+1 + ũi−1 )
U0 = udξ = . (22)
0 2(2 − di,1 − di,2 )
In the following we set di,1 = −di,2 = d, because then Eq. (22) shows that the integral (i.e. the synthetic field
box-filtered at scale 2∆) reduces to the integral of the piecewise linear interpolation, i.e. of the velocity ũ filtered at
scale 2∆. This bears some similarity with the kinematic condition used in [19]
These formulae are readily generalized to the case of integrals over subsets I ⊂ [0, 1] of the form I = [p2−m ,
(p + 1)2−m ], p, m ∈ N involving two or more distinct fractal functions. Interpolations on such sub-intervals will
be later shown to be of interest when computing the SGS stress. Interpolations involving two different functions will
become useful when extending the fractal calculus to 3D (and stresses of the form τij with i 6= j need to be evaluated).
Due to the linear dependence of the parameters specifying the qi ’s, we can introduce coefficients
(m,p)
Aα1 ,... ,αn (du, . . . , dz), αj ∈ {−1, 0, 1}, j = 1, . . . , n so that
Z (p+1)2−m X
1
u(ξ ) · · · z(ξ ) dξ ≡ A(m,p)
α1 ,... ,αn (du, . . . , dz)ũi+α1 · · · z̃i+αn . (23)
p2−m | {z }
n α1 ,... ,αn =−1

For example, from Eq. (22) we can write immediately

 
1 1 1
A0,0
α1 (d) = , , . (24)
4 2 4
A list of the coefficients that will be used in the rest of the paper is given in Appendix A.
In applications of fractals to real physical systems, there always exists a cut-off scale, η, below which the field
becomes smooth (e.g. the Kolmogorov scale in turbulence). The formulae derived in this section apply strictu sensu
only to fields that are fractal at all scales. One may then wonder if the introduction of a cut-off will alter the results
and how. The answer is that provided η  ∆ these formulae are a good approximation, in the following sense:
Let u(n) (ξ ) be an approximation of u obtained applying the map W n times; then all the moments of u(n) approach
the moments of u exponentially fast in n. Hence, provided η  ∆, replacing the real field with the fully fractal
interpolated field does not introduce significative errors when the moments are concerned. In other words, the effect
of the smallest scales becomes negligible as n becomes large for the computation of the moments. We conclude
this section with a discussion of what happens when the presence of the cut-off cannot be ignored, namely when
derivatives of the fields are involved.

2.1.2. Fractal calculus: integration of gradients

Some terms that will need to be modeled contain derivatives of the fields, for instance terms like
g2
∂u
.
∂x
Strictly speaking the fractal functions are not differentiable, not even piecewise, therefore such an expression is
ill-defined. However, if for physical reasons it is expected that the scale-invariance behavior ceases below a certain
cut-off scale η ∼ 2−N ∆ for some N, then it is still possible to work in the framework of the fractal interpolation
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 205

technique, provided that at most N iterations of the map are considered. Also note that unlike the previous case
the result depends on the choice of u(0) . Again, an example can illustrate the idea. To simplify the discussion, we
discuss the special case d = di,1 = −di,2 . Let u(n) be the n-th iterate of the map W starting from u(0) as defined in
Eq. (7). Then the derivative map is given by
u0(n+1) (ξ ) = 2du0(n) (2ξ ) + 2ai,1 , ξ ∈ [0, 1/2],
0(n+1) 0(n)
u (ξ ) = −2du (2ξ − 1) + 2ai,2 , ξ ∈ [1/2, 1]. (25)
R1
Note that the stretching ratio is now 2d. It is easy to show that Dn1 ≡ 0 u0(n) (ξ )dξ = ai,1 + ai,2 . Then
Z 1 Z 1/2 Z 1
Dn2 ≡ (u0(n) (ξ ))2 dξ = (2du0(n−1) (2ξ ) + 2ai,1 )2 dξ + (−2du0(n−1) (2ξ − 1) + 2ai,2 )2 dξ
0 0 1/2
= 4d 2 Dn−1
2
+ 4d(ai,1 − ai,2 )Dn1 + 2(ai,1
2
+ ai,2
2
).
It follows that
1 − (2d)2n
Dn2 = (2d)2n D02 + (4d(ai,1 − ai,2 )Dn1 + 2(ai,1
2
+ ai,2
2
)) . (26)
1 − (2d)2
This result shows that if d > 1/2, Dn2 diverges, as expected for a fractal function. Notice that the divergence is not
due to the fact that by using piecewise linear functions we introduce jumps, but rather because of the factor 2 that
appears in front of the stretching factor. We will make use of Eq. (26) in Section 3.3.

2.2. Interpolation in higher dimensions

In this section we examine the two-dimensional case. It is assumed that a field is defined on a 2D Cartesian grid
of mesh size ∆x , ∆y . As before we introduce local coordinates ξ = (x − xi−1 )/2∆x and ζ = (y − yi−1 )/2∆y
so that the grid element is mapped into Ω = [0, 1] × [0, 1]. As a first step we generate three 1D interpolat-
ing functions along the x direction that interpolate the filtered field ũ on the three levels (j − 1), j and j + 1,
i.e. the triplets (ũi−1,j −1 , ũi,j −1 , ũi+1,j −1 ), (ũi−1,j , ũi,j , ũi+1,j ) and (ũi−1,j +1 , ũi,j +1 , ũi+1,j +1 ) with stretching
dux (i, j ). Let uj −1 (ξ ), uj (ξ ) and uj +1 (ξ ) denote these interpolating functions. Then for a fixed ξ we interpolate
(uj −1 (ξ ), uj (ξ ), uj +1 (ξ )) in the y direction with stretching duy (i, j ) to obtain u(ξ, ζ ). Fig. 2 shows a schematic
of the procedure. Of course one could start by interpolating along the y-axis first and then along the x-axis. We will
show that this leads to the same function. But first let us investigate the properties of u(ξ, ζ ). We consider the space
Ĉ[Ω] of the continuous functions on Ω equipped with the usual metric d(u, g) = supx∈Ω |u − g| which interpolate
the points {ũi+σx ,j +σy , σx , σy ∈ {−1, 0, 1}}. We have the following theorems:

Theorem 2.2. Let u(ξ, ζ ) be defined as above. Then u(ξ, ζ ) ∈ Ĉ[Ω].

Proof. u(ξ, ζ ) interpolates the given points by construction. Let (ξ0 , ζ0 ) ∈ Ω and let δ > 0. Then
|u(ξ, ζ ) − u(ξ0 , ζ0 )| ≤ |u(ξ0 , ζ ) − u(ξ0 , ζ0 )| + |u(ξ, ζ ) − u(ξ0 , ζ )|.
From point (ii) of Theorem 2.1 we know that ∃  such that
δ
|u(ξ0 , ζ ) − u(ξ0 , ζ0 )| ≤
2
if |ζ −ζ0 | ≤ . Let Wξ and Wξ0 be the maps associated to (uj −1 (ξ ), uj (ξ ), uj +1 (ξ )) and (uj −1 (ξ0 ), uj (ξ0 ), uj +1 (ξ0 )),
respectively. Then by virtue of point (iv) of Theorem 2.1 we know that
d(Wξ [u(ξ, ζ )], Wξ0 [u(ξ, ζ )])
|u(ξ, ζ ) − u(ξ0 , ζ )| ≤ ,
(1 − s)
with 0 < s < 1.
206 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 2. Fractal interpolation in two dimensions. First three 1D interpolations are generated along the x-axis. Then for each triplet
(ui−1 (ξ ), ui (ξ ), ui+1 (ξ )) the interpolation along the y-axis is generated.

Again by the continuity of the fixed point in 1D (uj −1 (ξ ), uj (ξ ), uj +1 (ξ ))→ξ →ξ0 (uj −1 (ξ0 ), uj (ξ0 ), uj +1 (ξ0 ))
and since Wξ is continuous in the parameters [3] there is  0 such that if |ξ − ξ0 | ≤  0
d(Wξ [u(ξ, ζ )], Wξ0 [u(ξ0 , ζ )]) δ
≤ ,
(1 − s) 2
which proves the continuity in ξ0 , ζ0 and hence in Ω. This concludes the proof. 
The extension of the idea to 3D and higher dimensions is trivial. Let the third coordinate be parametrized
by η = (z − zl−1 )/(2∆z ). Once the three 2D interpolations along the ξ ζ planes at η = 0, 1/2, 1 have been
generated, u(ξ, ζ, η) is defined as the interpolation of (ul−1 (ξ, ζ ), ul (ξ, ζ ), ul+1 (ξ, ζ )) and so on for every dimension
above 3. This leads to:
Theorem 2.3. The function u(ξ1 , . . . , ξn ) defined by iterating in Rn the procedure defined above is continuous.

2.2.1. Integration
The Fubini theorem reduces the calculation of multidimensional integrals to the calculation of 1D integrals that
we have discussed at length in the previous section. In particular the following examples show that the coefficients
(p,n)
Aσ1 ,... ,σn (du1 , . . . , dun ) are all one needs in order to compute integrals of products of fractal interpolating functions.
For clarity we work in 2D, the extension to three and higher dimensions being trivial. Let Ωl,m,n,p = [p2−n ,
(p + 1)2−n ] × [l2−m , (l + 1)2−m ]. Then
Z Z (p+1)2−n Z (l+1)2−m ! Z (p+1)2−n
X 1
u(ξ, ζ ) dξ dζ = u(ξ, ζ )dζ dξ = A(l,m)
σy (du y ) uj +σy (ξ )dξ
Ωl,m,n,p p2−n l2−m σy =−1 p2−n

X
1
= A(p,n) (l,m)
σx (dux )Aσy (duy )ũi+σx ,j +σy . (27)
σx ,σy =−1
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 207

Similarly,
Z X
1
u(ξ, ζ )v(ξ, ζ )dξ dζ = A(p,n) (l,m)
σx ,αx (dux , dvx )Aσy ,αy (duy , dvy )ũi+σx ,j +σy ṽi+αx ,j +αy . (28)
Ωl,m,n,p σx ,σy ,αx ,αy =−1

Now we are in a position to show that it does not matter which direction one chooses to start the interpolation. Let
us assume that g(ξ, ζ ) has been computed interpolating first in the y direction and then in the x direction for the
nine points ũi+σx ,j +σy . Let Ωl,m,n,p be defined as before,
Z Z Z ! Z
(l+1)2−m (p+1)2−n X
1 (l+1)2−m
g(ξ, ζ ) dξ dζ = g(ξ, ζ )dξ dζ = A(p,n)
σx (dux ) gi+σx (ζ )dζ
Ωl,m,n,p l2−m p2−n σx =−1 l2−m

X
1
= A(p,n) (l,m)
σx (dux )Aσy (duy )ũi+σx ,j +σy , (29)
σx ,σy =−1

that is to say ∀(l, m, p, n) ∈ N4 , l < 2m , p < 2n

Z Z
u(ξ, ζ ) dξ dζ = g(ξ, ζ ) dξ dζ. (30)
Ωl,m,n,p Ωl,m,n,p

But this is a sufficient condition to ensure that the continuous functions u and g are equal everywhere, as the
following lemma proves.
Lemma 2.1. Let u, g ∈ C[[0, 1] × [0, 1]] such that ∀(l, m, p, n) ∈ N4 , l < 2m , p < 2n
Z Z
u(ξ, ζ ) dξ dζ = g(ξ, ζ ) dξ dζ,
Ωl,m,n,p Ωl,m,n,p

where Ωl,m,n,p = [p2−n , (p + 1)2−n ] × [l2−m , (l + 1)2−m ]. Then u = g.

Proof. By reductio ad absurdum, assume (without loss of generality), that ∃(ξ0 , ζ0 ) such that u(ξ0 , ζ0 )−g(ξ0 , ζ0 ) ≥
δ > 0. Then by the hypothesis of continuity ∃ > 0 such that if (ξ − ξ0 )2 + (ζ − ζ0 )2 <  2 , u(ξ, ζ ) − g(ξ, ζ ) >
δ/2 > 0. Let m0 = n0 = [−log/log2] + 1 where [·] denotes the integer part and let p0 = [2n0 ξ0 ], l0 = [2m0 ζ0 ].
Then
Z Z
δ
0= (u(ξ, ζ ) − g(ξ, ζ )) dξ dζ = |u(ξ, ζ ) − g(ξ, ζ )| dξ dζ ≥ 2−(m0 +n0 ) > 0,
Ωl0 ,m0 ,n0 ,p0 Ωl0 ,m0 ,n0 ,p0 2

which is absurd. Notice that the requirement of continuity cannot be relaxed. 

2.3. Closing remarks

The method outlined in this section relies on the assumption of scale-invariance and leads to an assumed fractal
SGS field. Fractal interpolation opens the possibility of evaluating exactly all the relevant moments of the broad-
band fields (and even of derivatives of the fields under some assumptions) without explicitly generating the subgrid
field.
However, the approach has only partial a priori justification. For instance, the observations of fractality of velocity
signals [63] do not imply that the real subgrid field coincides with the one obtained from fractal interpolation. Indeed,
later on it will be shown using DNS data that real and modeled stresses do not coincide exactly (although they do
display a general resemblance). Moreover, the model is a deterministic fractal, while the real small-scale field should
208 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

be regarded as significantly more random. Also, the stretchings d1 and d2 are as yet undetermined, free parameters.
Hence there is a need to study the behavior of the proposed modeling approach in various applications and to
compare the results with data coming from both experiments, from DNS and from LES using traditional models.

3. Fractal model in 1D: LES of Burgers equation

In this section we focus on a simple 1D hydrodynamic problem, the randomly forced, viscous Burgers equation
with periodic boundary conditions in the interval [0, 2π ]
 
∂u ∂ 1 ∂ 2u
+ uu = ν0 2 + f (x, t). (31)
∂t ∂x 2 ∂x
This equation is the prototype for a family of equations that govern a range of phenomena, from stochastic growth
of interfaces to directed polymers [1]. Recent work [14] has shown that if the forcing has a spectrum of the form
hfˆ(k, ω)fˆ(k 0 , ω0 )i = A2 k −1 δ(k + k 0 )δ(ω + ω0 ) (i.e. white in time with spatial correlations), Kolmogorov-like
scaling of the resulting velocity is obtained (h|û(k)|2 i ∼ k −5/3 ) for k  η−1 . Here η is the ultraviolet viscous
cut-off scale. It follows that the second-order structure function obeys a scaling

h[u(x + r) − u(x)]2 i ∼ r 2/3 ,

which suggests that the fractal dimension of the signal u(x) is D = 5/3 (not unlike velocity signals in 3D hy-
drodynamic turbulence, see [63]). Therefore, this is an application where we strongly suspect that asymptotically
(ν0 → 0) the solution at small scales is scale-invariant. The fractal model explicitly uses this information to mimic
the small-scale features of the system.
The effective equation for the coarse-grained 1D velocity ũ(x, t) is found by convolving Eq. (31) with a top-hat
filter of size ∆ (G∆ (x) = ∆−1 if |x| < ∆/2, zero otherwise):
 
∂ ũ ∂ 1 1 ∂τ ∂ 2 ũ
+ ũũ = f˜(x, t) − + ν0 2 , (32)
∂t ∂x 2 2 ∂x ∂ x

where τ = ue2 − ũ2 is the unknown subgrid stress.

3.1. Modeling

While one can certainly obtain numerical solutions to Eq. (32) by using an effective viscosity model, e.g.
τ = −νeff ∂ ũ/∂x, such a closure can only be justified for cases with significant scale separation and has little
basis in the case of Burgers equation. Instead, let us employ the basic idea underlying fractal interpolation, which
is to replicate on smaller and smaller scales features of ũ present at a coarse scale. According to its definition the
unknown subgrid stress τ produced by the synthetic field around discrete position xi (see Fig. 1) can be written as
Z Z !2
3/4 3/4
τi = [u(ξ )] dξ −
2
u(ξ )dξ
1/4 1/4
 2
X
1 X
1
= (A(1,2) (2,2)
σ1 ,σ2 (d, d) + Aσ1 ,σ2 (d, d))ũi+σ1 ũi+σ2 −
 (A(1,2) (2,2)
σ1 (d) + Aσ1 (d))ũi+σ1

σ1 ,σ2 =−1 σ1 =−1

1 di (8 − 3di2 ) 2 1 + 15di2− 24di4 + 12di6 2 2

= (δi ũ)2 + δi ũδi ũ + (δi ũ) . (33)
12 48 192(1 − di2 )
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 209

Above,

δi ũ = (ũi+1 − ũi−1 )/2, δi2 ũ = ũi+1 − 2ũi + ũi−1 ,

and we have made use of the coefficients A(p,n) defined in Section 2.1.1 together with the constraint that di,1 =
−di,2 = di . The vertical stretching di must still be prescribed. In the next sections several options will be described.
Another property of Eq. (33) concerns the limit when the grid size approaches the viscous scale: in this limit the
viscous term in the equation will smooth the velocity signal and the SGS stresses will vanish, as it should (a Taylor
series expansion of Eq. (33) shows that the stress vanishes as ∆−2 for a smooth velocity field).

3.2. Zero equation models

As a first attempt, we can prescribe the magnitude of the vertical stretching factors by fixing the fractal dimension
of the interpolating field to the value implied by the power spectrum exponent, which is also close to the value
obtained directly from experimental velocity signals (see [63]) in 3D turbulence, namely D = 5/3. This yields
stretching factors of d = ±2D−2 ' ±0.8. A choice must still be made for the sign. Two approaches are explored.
In the first, the sign is randomly chosen with equal probability for positive or negative values. We will denote this
model ZE1 (shorthand for zero equation model, type 1).
Another approach to select the sign is to consider the dissipative properties of the model. The local SGS dissipation
is given by
∂ ũ
εSGS = − τ.
∂x
One of the most important properties of the subgrid scales is to exchange, and (typically) on average to extract
kinetic energy from the resolved scales. Physically this process represents the forward cascade of energy. Numer-
ically, it helps in stabilizing the simulation. To reproduce this effect, εSGS should typically be positive. In other
words, a model should consistently reproduce a negative correlation between τ and ∂ ũ/∂x. By choosing the signs
of the stretching factors randomly, as in ZE1, it is unlikely that such correlation will exist.
Replacing with the fractal model for τ we have
" #
1 d i (8 − 3di
2) 1 + 15di
2 − 24d 4 + 12d 6
i i
εSGS ∼ − (δi ũ)3 + δi2 ũ(δi ũ)2 + (δi ũ)(δi2 ũ)2 . (34)
12 48 192(1 − di2 )

The average of the first term,h(δi ũ)3 i, should be negative, since in “Burgers turbulence” the third moment of the
derivative is negative [9]; the third term is generally smaller than the other two and its sign does not depend on
the sign of d. The second term can always be made negative if the sign of d is chosen such that di δi2 ũ < 0. Thus,
choosing sign[di ] = −sign[δi2 ũ] maximizes the dissipation rate within the constraints of the model. We will call
this second version of the zero equation model ZE2.
Finally, for comparison’s sake, we also consider a constant eddy viscosity model obtained by replacing ν0 with
νeff  ν0 and setting τ ≡ 0 in Eq. (32). The numerical value of the eddy viscosity will be chosen by matching the
energy decay as given by a DNS of the same problem. This eddy viscosity model is denoted ED1.

3.3. A one equation model

Instead of prescribing D (or di ) a priori, a dynamical equation for τ can be combined with Eq. (33) to self-
consistently solve for τ and d. From Burgers equation, we find
∂τ ∂τ ∂ ũ ∂Q
+ ũ = −ε − 2 τ − 2 + 2(ffu − ũf˜), (35)
∂t ∂x ∂x ∂x
210 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

where
" 2   #
g
∂u ∂ ũ 2 (ue3 − ũ3 )
ε = 2ν0 − , and Q= − ũτ. (36)
∂x ∂x 3

The terms on the rhs of Eq. (35) represent molecular dissipation, SGS production of τ , flux of τ , and the effect of
forcing, respectively. The flux Q is computed in a straightforward fashion with the fractal interpolation technique
outlined in section 2.1.1, yielding
h
Qi [ũ, di ] = δi2 ũ 4(2 − di2 )(4 − 2di2 + 3di4 )(δi ũ)2 + 8di (8 − 14di2 + 13di4 − 6di6 )δi ũδi2 ũ
i
+di2 (28 − 72di2 + 69di4 − 36di6 )(δi2 ũ)2 /(3072(2 − di2 )). (37)

To compute the small-scale forcing (ffu − f˜ũ), f˜ is interpolated fractally with vertical stretching determined by
the forcing spectrum, which is proportional to k −1 . For the graph of a random variable with Gaussian distribution
and power spectrum E(k) ∼ k −α , Orey’s theorem [52,43] implies a fractal dimension D = (5 − α)/2, i.e. D = 2.
Therefore, we use |d| = 1 to model unresolved scales of the random forcing. Using the results of Section 2.1.1, we
obtain
n
ffu − f˜ũ = 16(1 − di,f di )δi f˜δi ũ + di,f (16 − 6di,f
2
)(1 − di,f di )δi2 f˜δi ũ + di (16 − 6di2 )(1 − di,f di )δi2 ũδi f˜
o
+[(1 − di,f di ) + di,f di (16 + 6(di,f + di )2 − 12di,f2
di2 )]δi2 f˜δi2 ũ /[192(1 − di,f di )], (38)

with di,f = ±1 with the sign chosen at random. The dissipation due to molecular viscosity, ε, requires more care
since the fractal signal is not differentiable when d > 1/2 (or when 1 < D < 2, which is of course the interesting
case). However, if the map’s iterations are terminated after n steps, i.e. if a cut-off scale

η = 2−n ∆, (39)

is introduced, we can evaluate integrals of (∂u/∂x)2 as explained in Section 2.1.2. The result is
" #
(4d 2 )n+1 − 1
1 i
εi = 2ν0 + 4di4 (δi2 ũ)2 . (40)
4 4di2 − 1

For forced Burgers equation with Kolmogorov scaling, the usual dimensional argument yields η = C(ν03 /ε)1/4 ,
with C = O(1). Besides the geometric rules of our fractal interpolation scheme, this step is the only physical
modeling required. Combining this value for η (with C = 1) together with Eqs. (40) and (39) we obtain an equation
for the dissipation solely in terms of d and ν0 .
Note that as explained at the end of Section 2.1.1, the formulae for the integration of the moments of the function
(as opposed to its derivative) depend only weakly on the number of steps (for large n). Thus, for simplicity τi and
Qi are evaluated using the asymptotic values for n → ∞, with negligible errors.
In summary, we have a system of four equations (32),(33),(35) and (40) in the four unknowns (ũ, τ, d, ε). This
one equation model will be denoted as ME1. In this case the sign of both di and di,f will be chosen at random.
As an aside, it is interesting to examine the asymptotic properties of the dissipation provided by the fractal signal,
in the limit of small viscosity. It can be easily shown that

lim ε(d, ν0 ) = O(ν0 ) if d < 1/2,

ν0 →0

−(3/2)log2 (21/3 d)
lim ε(d, ν0 ) = O(ν0 ) if d > 1/2, (41)
ν0 →0
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 211

which implies that d = 2−1/3 (or D = 5/3) is the cross-over exponent at which the fractal signal dissipates energy
even in the inviscid limit [22]. Different dissipation mechanisms (such as hyperviscosity) would lead to different
expressions for ε in Eq. (40). This would affect the estimate for η and in turn may affect the value of d.

3.4. Numerical aspects

Now we can explore numerical solutions of the models ZE1, ZE2, ED1 and ME1. We use a pseudo-spectral code
on the interval [0, 2π ] to solve for ũ(x, t) and τ (x, t), using 27 modes (i.e. a filter scale ∆ = 2π/256). This scale
is expected to be small enough for a scaling range to develop in the resolved field. The√time advancement is based
on the second-order Adam–Bashforth scheme. The magnitude of the forcing is A = 2 × 10−3 , the time step is
1t = 5.0 × 10−5 (resulting in a CFL number much less than unity), and the viscosity is ν0 = 1.0 × 10−5 . This
value implies an expected cut-off scale for the real Burgers equation of about η ∼ 2π/212 , much smaller than our
coarse resolution.
√ The random forcing is generated in Fourier space by superposing Fourier modes with amplitude
equal to A/ ∆t, and phase θ chosen randomly at each time step with a uniform weight in [0, 2π ] [14]. The model
expressions are evaluated by means of Fourier collocation derivatives. The nonlinear terms are dealiased by the
2-rule (zero padding) [10], and the minimum attainable value of τ (realizability) is enforced numerically as follows.
The value of τ is constrained to remain positive by adding a term which is nonzero only when 0 < τ < τ 0 . τ 0 is the
“trivial” value of the stress when d = 0, i.e. τ 0 = (δi ũ)2 /12 + (δi2 ũ)2 /192. The term added is τ 0 /τ when τ < τ0 ,
zero otherwise. During simulations, the condition τ < τ 0 was observed to occur very infrequently, only 5% of the
time. When it occurred during this 5% of the time, it occurred on at most 3% of the grid-points.
Eq. (33) for di is, once τi is specified, a sixth-order equation. It was found that when ni > 0 a solution always
existed. Moreover this solution was in the interval [0.5, 1] that corresponds to the fractal behavior, and only a single
solution existed in that interval. When ni = 0, it was found that the solution, when it exists, always lied in the
interval [0, 0.5], i.e. in the smooth, non-fractal region. In cases when a solution was not found (for ni = 0), di was
assigned the value 0.

3.5. Results

The results of LES using the different models are compared to DNS. The latter is a direct simulation (of Eq.
(31)) with no coarse-graining or modeling, using 213 modes. For the forcing, the same realization for the large-scale
modes, common to the LES, are used. The DNS uses the same viscosity ν0 = 1.0 × 10−5 and a time step of
∆t = 5.0 × 10−5 . Fig. 3 shows a realization of the DNS field u(x). Notice the presence of large shocks as well as
many fine structures.
Next, several large eddy simulations with different initial conditions, modifying the random phases but keeping
the same Kolmogorov spectrum, have been performed. Beginning with qualitative observations, in Fig. 4 we show
a realization from ZE1 and ZE2. The signal arising from ZE1 is extremely noisy and little is left of the large-scale
features that are visible in the DNS signal. The signal based on model ZE2 retains some of the structures that are
visible in the DNS. Still, the signal is somewhat noisy. The signal arising from ED1 (Fig. 5(a)) shows excessive
damping of many features at scales comparable to the cut-off ∆. Finally the signal from ME1 (Figure 5(b)) shows
(qualitatively) large-scale features that are present in the DNS, but with a reduced level of noise compared to ZE2.
R1
Next, the evolution of global resolved kinetic energy K(t) = 1/2 0 ũ2 dx is considered. The DNS solution shows
that after a transient period during which K(t) decays, a plateau is reached where forcing and dissipation are in
approximate equilibrium. In Fig. 6 we show the kinetic energy vs. time plot for the different models considered,
including a run without any model. As already apparent in Fig. 4, ZE1 does not dissipate enough energy, hence it
does not level off and grows as much as the kinetic energy of a simulation without a model (setting τ ≡ 0). The
latter keeps increasing without reaching a steady state, showing that the presence of a SGS term is indeed necessary
to simulate the flow at this level of filtering. All the other models follow the decay from the DNS. Notice that in
ED1 the effective viscosity was chosen equal to 5.0 × 10−3 to match approximately the decay of energy (triangles
212 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 3. DNS velocity fields. (a) Sample of DNS field, displaying several large quasi-shocks. (b) Magnified view of portion of the DNS field,
displaying some smaller shocks.

in Fig. 6(a)). As a next step, energy spectra are computed (Figure 6(b)). Ideally, the spectrum arising from LES
would match the DNS spectrum up to about the cut-off wave number kc = π/∆. The almost insignificant amount
of dissipation provided by ZE1 is reflected in the spectrum showing equipartition after some time. ZE2 displays a
better agreement, although there is still a marked pile-up at wave numbers above π/(5∆). This is consistent with
the significant amount of noise observed on the signal in Fig. 4.
The model ME1 displays good agreement with the DNS up to a wave number k ∼ π/2∆, about half of the cut-off
wave number. At the smallest octave of resolved scales, the spectrum of the coarse-grained simulation tends to fall
somewhat above that of the direct simulation. As will be shown later, this overshoot is spurious and of no concern.
The results of ME1 can be contrasted with the strong departure of the eddy viscosity simulation. The latter
excessively damps the solution. Of course the spectrum of ED1 could be made to approach the DNS by adjusting
the eddy viscosity to smaller values. However, then the energy decay rate becomes significantly smaller and no
longer matches that of the DNS.
In ME1, the points where the fractal dimension exceeds 2−1/3 , that is to say the region where most of the
dissipation takes place, covers on average only 25% of the space. There, the computed η reaches down to 1.25 ×
10−4 which is consistent with the fact that for the present level of forcing, ∆dns ∼ 1.25 × 10−4 is the largest
mesh size that can be used to solve Eq. (31) without modeling. Finally, in Fig. 7 we show a segment of the
coarse-grained solution ũ(x, t), represented by the symbols (values at grid-points), together with the fractal field
(solid line).
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 213

Fig. 4. Representative signals from LES of forced Burgers equation. (a) Sample LES field simulated with model ZE1. Notice the absence of
large-scale structures. (b) Sample of LES field simulated with model ZE2. The large-scale structures are present, although significant ripples are
still noticeable.

The latter is an explicit construction of the synthetic small-scale signal using the fractal interpolation technique
(iterating the map Wi for n times, with n and d set to the locally computed values). Values of n and d (and thus )
fluctuate greatly from one location to another, indicative of a high level of intermittency.
We stress again that during the simulation, there was no need to explicitly construct such a small-scale signal,
because the fractal interpolation technique allows us to analytically evaluate their effect on the coarse scales directly.
As ũ evolves in time, the synthetic small-scales are “slaved” to the resolved ones, and evolve accordingly. The
simulation proceeds at the coarse resolution, but the simulated dynamics of ũ are as if the signal contained all the
small-scale fluctuations shown in the figure.
The spectrum for ME1 was recomputed for the full interpolated signal (such as Fig. 7(a)). The resulting spectrum,
shown in Fig. 7(b), displays an excellent agreement with DNS up to the cut-off scale. Note that now the pile-up near
the cut-off has disappeared, since the addition of the small-scale portion of the signal effectively eliminates aliasing
of the spectrum. Below kc , the ME1 LES spectrum is lower than the DNS spectrum, although its upper envelope
qualitatively follows the decay of the DNS spectrum. The periodic modulation is related to the dyadic nature of the
fractal interpolation technique. To study the dependence on filter size, another simulation with ∆0 = 2π/64 and
ME1 is performed. The spectrum of the fully interpolated solution, shown in Fig. 8, leads to the same conclusions
as Fig. 7(b).
Next, we compute the global fractal dimension of the fully interpolated signal, using the variational method of
Dubuc et al. [20]. In Fig. 9 we show the logarithm of the area of the envelope of thickness r, V (r), required to cover
214 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 5. More LES signals. (a) LES field simulated using ED1. Notice the almost complete absence of all but the largest structures. (b) Same as
(a) but this time using ME1. The large-scale structures and jumps appear realistic, and compared to the fields simulated with ZE2 there are less
ripples.

the graph as function of log(r) for both resolutions. As expected, the resolved range is not inconsistent with a fractal
dimension of D = 5/3. We observe that in both cases, below the cut-off scale ∆ or ∆0 , there is a slight curvature
which precludes us from measuring a well-defined fractal dimension. We believe this is due to the combination of
fractal signals with locally different fractal dimensions. The leveling off of the last two points for the simulation on
the finer grid is probably due to the effect of the viscous cut-off, as discussed in [63].
Finally, we describe the computational cost required for the different LES models compared with the cost of the
full DNS. For the present value of ν0 , the full simulation requires roughly 10 times the memory as compared with
the models (more if compared to ED1 since in this case one does not need to store τ ). The full simulation takes
about 20 times longer per time step than ED1, 15 times longer than ZE1 and ZE2 and four times longer per time
step than ME1. Of course these ratios can be made arbitrarily large (more favorable to the SGS models) when the
Reynolds number is increased.

3.6. Final comments

The results show quite clearly that the simple geometrical assumption of scale-invariant behavior, upon which
ZE1 is based, is not enough to reproduce the correct dynamics. Some dynamical constraint has to be imposed
on the SGS scales. Already trying to maximize the dissipation provided by the model significantly improves the
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 215

Fig. 6. Evolution of kinetic energy. (a) Kinetic energy K versus time: fully resolved equation (solid line), ZE1 (solid circles), ZE2 (pluses), ME1
(squares), ED1 (triangles) and coarse-grained solution without modeling (dash–dotted line). (b) Energy spectra (same symbols as (a)), averaged
in time between 60 < t < 200, and over different realizations of the coarse-grained simulation.

simulation. However, it was only by introducing a second dynamic equation for τ (and d) that the simulation began
to reproduce realistic dynamics. By doing so, we have reduced the burden that the SGS model has to carry, since
now τ is computed from its evolution equation. The fractal model only affects it through (i) the predicted flux of τ ,
(ii) the interaction with the forcing which is quite small, and (iii) the modeled dissipation of energy experienced by
the unresolved scales, which is a dominant term in the equation for τ .

4. Fractal LES model in 3D: applications to forced and decaying isotropic turbulence

In this section we apply the ideas developed in the previous sections to the problem of large eddy simulations
of 3D isotropic turbulence in real fluids. As in the 1D case, considered in the previous section, the “kinematic”
condition based only on the fractal dimension will be shown to be insufficient to reproduce the proper physics.
Several different approaches to endow the fractal small scales with the properties necessary to yield good results
will be considered. To discriminate among these various approaches, they are implemented in LES of forced and
decaying, isotropic turbulence. Results are compared with experimental data.
216 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 7. The fully interpolated field. (a) Example of large-scale field (circles) as computed from coarse-grained Burgers equation using the fractal
model, and explicitly constructing the small-scale field (solid curve) using d and n as obtained from the simulation (ME1). (b) Spectrum of the
fractally interpolated signal (shown in (a)), dotted line, compared to the spectrum of the DNS, solid line.

Fig. 8. Spectrum of the fractally interpolated signal from LES on a coarser grid. The signal comes from a simulation with ME1 with cut-off at
∆ = 2π/64.
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 217

Fig. 9. Logarithm of the area of the envelope of thickness r required to cover the graph vs. logr. The circles refer to the simulation with
∆ = 2π/256, while the pluses to the simulation on a coarser grid with ∆0 = 2π/64. The slope of the solid line is 1.66. The dotted line indicates
the filter scale ∆, while the dashed line indicates the scale ∆0 .

4.1. Formulation of the model

Let the resolved velocity field ũ at node (l, m, n) be denoted by ũ(l, m, n). Assuming that at a given location
j j j
(l, m, n) on the grid the nine vertical stretching parameters (du1 , du2 , du3 ), j = 1, 2, 3 are given, the SGS stress
τij (l, m, n) can be obtained by integration over (ξ, η, ζ ) ∈ [1/4, 3/4] × [1/4, 3/4] × [1/4, 3/4]. Utilizing the
integration formulae developed in Section 2.2, the result can be conveniently written as

X
1 X
1 X
1 X
1 X
1 X
1
j j j
τij (l, m, n) = [Aσx ,αx (duix , dux )Aσy ,αy (duiy , duy )Aσz ,αz (duiz , duz )
σx =−1 σy =−1 σz =−1 αx =−1 αy =−1 αz =−1
j j j
−Bσx (duix )Bσy (duiy )Bσz (duiz )Bαx (dux )Bαy (duy )Bαz (duz )]
×ũi (l + σx , m + σy , n + σz )ũj (l + αx , m + αy , n + αz ), (42)

where the coefficients Aσ,α and Bσ are the ones introduced in Section 2.1.1 and listed in detail in Appendix A. Thus,
the stress at point (l, m, n) is expressed in terms of all possible bilinear combinations of the resolved velocity at all
27 neighboring grid points. A number of variants of the model can be formulated depending on the choice of the
stretching factors d’s.
We begin by noticing that if all the stretching factors are set to zero, the interpolation becomes a simple bilinear
interpolation of ũ(l, m, n) between the grid points. Evaluating the stress based on this “smooth” resolved field we
thus obtain, essentially, the Bardina model described in the introduction (Section 1.2.1). There is a minor difference
with the traditional Bardina model in the numerical treatment of the term ũg i ũj . The usual implementation of the
Bardina model [67], denoted here as BARD L, linearly interpolates the function ũi ũj in order to compute the
integral represented by the second filtering. On the other hand, in the present formalism, first the ũj are interpolated
linearly and then the products (which are now polynomial functions in between the given points) are integrated. This
version will be denoted as BARD NL. The Bardina model (BARD L) has already been investigated extensively in
the past, therefore we will only use it as a reference case.
Clearly, setting all the stretching factors to zero is typically an unphysical approximation for the SGS velocity
field, since at high Reynolds number it is broad-band even at scales below ∆. Thus, the next version of the fractal
218 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

interpolation model consists in choosing the stretching factors all equal, in magnitude equal to the “Kolmogorov”
value found in [63], D = 5/3. From point (v) of Theorem 2.1, we see that to generate signals with a fractal dimension
j
D = 5/3 we need to set |duα | = 2−1/3 . Finally, we choose the signs randomly. This is model FRACT 1.
Recall that it was found that for Burgers equation (Section 2.2), specifying only the fractal dimension and selecting
the signs of the stretching factors randomly was not enough to capture the proper dissipation of resolved kinetic
energy. We anticipate a similar problem for the 3D case, and will consider several options of introducing elements
in the construction of the fractal SGS fields to capture energy dissipation. For this purpose, we begin by observing
that the principal axes of the strain-rate tensor S̃ij represent locally a natural frame of reference, in which the SGS
dissipation is simply written as

0 0 0
sgs = −(τ11 λ1 + τ22 λ2 + τ33 λ3 ). (43)

In this expression, the λj ’s are the eigenvalues of S̃ij and τij0 is the SGS stress tensor expressed in the frame
P
of references of the principal axis of S̃ij . Due to incompressibility 3j =1 λj = 0, hence at least one eigenvalue
is negative and one is positive. Without lack of generality we assume that λ1 ≥ λ2 ≥ λ3 with corresponding
eigenvectors e1 , e2 and e3 . Due to the mechanism of vortex stretching, it is expected that small-scale vorticity in
the direction of e1 is stretched most, by the expanding strain λ1 . This will induce small-scale velocity fluctuations
in the e2 and e3 directions, while velocities in the e1 direction typically decrease. If λ2 < 0 this dynamical effect
should lead to low values of τ11 0 and to high values of τ 0 and τ 0 . If instead the intermediate eigendirection is also
22 33
expanding (λ2 > 0), it makes sense to assume that τ22 0 is decreasing. Based on these considerations, the signs of

the stretching factors can be determined as follows: for the velocity component in the e1 direction, among the eight
possible configurations (±2−1/3 , ±2−1/3 , ±2−1/3 ) for the stretching factors the one that minimizes the magnitude
0 is chosen. Let it be du01 . For the velocity component in the e direction, the configuration that maximizes
of τ11 3
0
τ33 is selected. Let it be du3 . Finally, for the case λ2 > 0 (λ2 < 0) the configuration of the stretchings for the
velocity in the e2 direction that minimizes (maximizes) τ22 0 . Let it be du02 . Once the nine values (du01 , du02 , du03 )
0
are determined, all components of τij , including the off-diagonal ones, can be computed according to Eq. (42).
This model will be denoted FRACT 2. Notice that the selection of signs described above maximizes the dissipation
[17,51] (Eq. (43)) within the constraint of using the same magnitude for the stretchings in the different directions.
Finally, a third version of the 3D fractal model, which is a refinement of the previous one, can be obtained by
relating the magnitude of the stretching to the magnitude of the eigenvalues, while still choosing the signs as in
FRACT 2 to achieve maximum dissipation. In particular, as a simple first choice, we set the stretching factors for
the velocity components in the direction of positive strain-rate eigenvalues to zero. Along directions of negative
eigenvalues the stretchings are set to an appropriate value to be decided based on simulation results. We will call
this model FRACT 3.
The analog of the successful model for 1D Burgers equation, ME1, would entail solving additional transport
equations for each of the SGS stress elements, and then adjusting the nine stretching factors accordingly. At present,
such an approach is deemed too complicated to be developed and tested.

4.2. Numerical implementation of fractal models

The numerical implementation of BARD NL and FRACT 1 is straightforward. At each node the SGS stress is
computed from Eq. (42). According to Eq. (42), the evaluation of τij requires O(272 ) operations. As argued in
Section 4.1, we wish to select the stretching factors to represent some dynamical features of how the SGS vorticity
responds to the strain-rate which can be achieved by selecting them so that the SGS dissipation is maximized.
However, evaluating

sgs = −τij S̃ij (44)

 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 219

Fig. 10. Schematic of grids used for constructing τij0 at point P. The velocity field is known on the computational grid x1 , x2 (open circles). e1
and e2 are the eigenvectors of S̃ij at P. The value of the velocity field at the grid nodes of the rotated grid (filled circles) is computed by bilinear
interpolation. τij0 is computed based on the velocity field on the rotated grid nodes and finally is rotated back to the original frame.

in the frame of reference of the simulation requires the evaluation of τij over 29 different configurations at every
node, which is very inefficient. In order to implement numerically the models FRACT 2 and FRACT 3, it is
computationally efficient, in addition to being physically sound, to work locally in the frame of reference of the
principal axes of the strain-rate tensor.
This much more efficient way, which requires us to compute τij only 3 × 23 = 24 times, is implemented as
follows. At each node, we find the similarity transformation (a rotation) Rij that diagonalizes the rate of strain
tensor S̃ij ; since S̃ij is a symmetric matrix, Rij always exists and is a unitary operator. Next, we express the resolved
velocity field in the new coordinate system
ũ0j (x0 ) = Rj i ũi (R · xt ). (45)

In order to apply the fractal interpolation formulae, we require ũ0j (x0 ) to be defined on a lattice aligned with the
strain-rate eigenvectors. We compute ũi (R · xt ) on the nodes of this locally rotated lattice by means of bi-linear
interpolation. Fig. 10 shows a sketch illustrating the procedure in 2D. In the rotated coordinate system S̃ij is diagonal,
therefore sgs reduces to the sum of three terms which do not involve products of different component of the SGS
velocity field. In other words we can write

X
3
0 0
sgs = − S̃jj τjj (du0j ) (46)
j =1

(the primes indicate quantities in the rotated frame, no summation over repeated indexes) and each term can
j
be maximized separately requiring only 23 evaluations for each. Let dumax , j = 1, 2, 3 be the three optimal
220 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

configurations for the stretching factors. The stress τij0 can be computed according to

X
1 X
1 X
1 X
1 X
1 X
1
0j 0j
τij0 (l, m, n) = [Aσx ,αx (du0imax,x , dumax,x )Aσy ,αy (du0imax,y , dumax,y )
σx =−1 σy =−1 σz =−1 αx =−1 αy =−1 αz =−1
0j
×Aσz ,αz (du0imax,z , dumax,z )
0j 0j 0j
−Bσx (du0imax,x )Bσy (du0imax,y )Bσz (du0imax,z )Bαx (dumax,x )Bαy (dumax,y )Bαz (dumax,z )]
×ũi (l + σx , m + σy , n + σz )ũj (l + αx , m + αy , n + αz ), (47)

and finally τij0 can be rotated back into the original frame of reference τij = Ril τlm
0 R . Given the efficiency of the
mj
routines available to find the eigensystem of a 3 × 3 symmetric matrix the whole process requires much less time
(roughly 20 times faster) than if we were to optimize in the frame of reference of the computation. Notice that a
priori the dissipation can still be negative (which corresponds to backscatter).

4.3. Forced isotropic turbulence at Re = ∞

To study steady-state turbulence at very high Reynolds number we set ν = 0.0 in Eq. (1). Also, since the
interaction with the unresolved scales removes energy from the large resolved ones,
R we add a forcing f(x, t) on the
rhs of Eq. (1) acting on the largest modes such that the rate of energy input ε = f(x, t) · u(x, t) dx is kept constant,
equal to 1.0 [7]. The equations are solved pseudo-spectrally on a 323 grid with periodic boundary conditions [24].
The computational domain D includes all the modes (k1 , k2 , k3 ) such that |kj ∆/π| < 1, j = 1, 2, 3. The Fourier
transform of a field g(x) is defined as
X
g(x) = ĝ(k)eik·x .
k∈D

In the following, we will omit the ˆ· to denote quantities in Fourier domain. The equations solved are:
∂[ũj (k)]
= [j il ũi (k) ∗ ω̃l (k) − ikj P̃ (k) − iki τij (k) + fj (k)], (48)
∂t
ki ũi (k) = 0, (49)

where 2 P̃ = p̃ + 1/2ũi ũi . Above, ∗ denotes the convolution product, and we have used the equivalent rotational
form of the N–S equations, which has some advantages from the numerical point of view [10]. The convolution
product is evaluated pseudo-spectrally by transforming both velocity and vorticity field to physical space, taking the
product and transforming back to Fourier space. Dealiasing is achieved by zero padding with the 2-rule. The physical
grid associated to the domain D is a cube of side L. The grid element has sides (∆, ∆, ∆). Time discretization is
achieved by a second-order Adam–Bashforth scheme.
To generate an initial condition a LES is run using the standard Smagorinsky model (Cs = 0.16) [64] for several
eddy turnover times until a steady state is achieved. Then, a realization of the field is used as initial condition for
the LES with the fractal models. Statistics are accumulated over a period of 10 eddy turnover times.
BARD NL and FRACT 1 fail to achieve a steady state. Fig. 11 shows the evolution of resolved kinetic energy,
K(t), for these two simulations. In both cases, after some time, sgs becomes negative, making the simulation
unstable (as well as unphysical). The analysis of the energy spectra (not shown) shows that as the simulations
proceed, the scales which are not forced tend towards a spectrum proportional to k 2 , i.e. they achieve equipartition.
This result is not surprising, since a similar behavior was already observed in the 1D case [61,62]. Also shown in
√
2 Notice the (obvious) difference between i = −1 and the index i.
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 221

Fig. 11. Total resolved kinetic energy vs. time of forced isotropic turbulence. Circles, BARD NL; dotted line, FRACT 1; dashed, FRACT 2;
and solid line, FRACT 3. All simulations have the same rate of energy injection .

Fig. 12. Spectra of FRACT 2 (circles) and FRACT 3 (broken lines). d = 0.90 dot–dashed line; d = 0.93, dotted line; and d = 0.95, dashed
line. The solid line shows the Kolmogorov spectrum E(k) = 1.6  2/3 k −5/3 .

Fig. 11 are the results for FRACT 2 and FRACT 3. Energy spectra are computed from the calculated field in the
following fashion:
X
E(k, t) = ũj (k, t)ũj (−k, t) g(k) ≡ hũj (k)ũj (−k)ishell
k−1/2<|k|≤k+1/2

4πk 2
g(k) = , (50)
No. of modes in shell k
where g(k) is a correction function for the discrete distribution of Fourier modes which considerably smoothes the
graph, especially at low k. The results in Fig. 12 show that FRACT 2 achieves a steady state characterized by a
222 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 13. 3D view of the simulated and synthetic field. (a) Velocity field at the surface of the computational domain. Notice the presence of large
vortical structures. Within each cell, the stress is computed from the synthetic SGS field. (b) Shows the SGS field associated to the shaded cell
in (a) in the frame of reference of the resolved rate-of-strain tensor. To better appreciate the fine-scale structures, the average speed within the
cell has been subtracted.

large pile up of energy at large wave numbers, which in turn depletes the energy of the intermediate scales. The
small scales do not achieve equipartition, as the last portion of the spectrum grows slower than k 2 . On the other
hand, FRACT 3 with d = 0.93 shows a spectrum very close to the −5/3 spectrum. A larger d produces a marked
dipping of the spectrum at large scales. A lower value results in energy piling up at large wave numbers while in the
inertial range the spectrum falls below the Kolmogorov spectrum. Fig. 13(a) shows a 3D view of the velocity field
as simulated by FRACT 3. Fig. 13(b) shows the synthetic SGS field associated to the cell shaded in (a) in the frame
of reference of the principal axis of S̃ij . To give a better appreciation of the small-scale structures the average speed
within the cell has been subtracted in (b). As can be seen, rich structure exists to (arbitrarily) small scales. Some
limitations of the present fractal model are also apparent: The tensor-product type construction of the 3D SGS field
yields certain cubic symmetries. Also, the small-scale vector field is not divergence-free. Notice, however, that its
filtered version is divergence-free.
Returning to the value of d, from point (v) of Theorem 2.1 we know that the fractal dimension of a signal with
d = 0.93 is D = 1.895. This does not imply that the fractal dimension of the graph of signals measured in the
original frame of reference is 1.895, since that will result from the combination of regions where the subgrid structure
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 223

Fig. 14. PDF of ν(x). (a) DNS (solid line), FRACT 2 (boxes) and FRACT 3 (dotted line). (b) SMAG (solid line) and FRACT 3 (dotted line).

of the field is smooth with regions where it is fractal. Hence we expect that the effective fractal dimension, if it
exists, is lower than 1.895.
To analyze the results of FRACT 2 and FRACT 3 in more detail we have considered the PDF of the “local” eddy
viscosity defined as

τij (x)S̃ij (x)

ν(x) = − . (51)
2|S̃|2 (x)

In the past, many authors have used DNS data (confined at moderate Rλ ) to perform a priori and a posteriori
tests of turbulence models [19,21,53]. Moin and Mahesh [48] have recently discussed the extent to which it is
possible to extrapolate results obtained from DNS at moderate Reλ to higher values. Acknowledging that there
can be quantitative differences between high and low Reynolds number flows, it is still of interest to qualitatively
compare our Re = ∞ LES results with data coming from a DNS of forced steady turbulence at Rλ = 93. The
DNS is performed on a 1283 grid [11]. First the DNS field is filtered with a top-hat filter with ∆ = 2π/32 and
subsequently is sampled on a 323 grid (see Section 4.5 for more details). We consider the nondimensionalized eddy
viscosity ν ∗ = ν/(hS̃11
2 i1/2 ∆/π). As shown in Fig. 14(a) , the PDF of the DNS case has a rather broad distribution

and exponential decay both for ν ∗ > 0 and ν ∗ < 0. Backscatter (i.e. ν ∗ < 0) occurs over roughly 20% of the volume
occupied by the flow. On the other hand, the PDF for the case FRACT 3 is strongly peaked and shows backscatter
over only about 1% of the volume (Fig. 14(a)), but it does display the asymmetry. The qualitative comparison
between DNS at Rλ = 93 and LES at Rλ = ∞ can be justified as follows: At higher Reynolds numbers one would
expect wider and more intermittent tails in the pdf. Thus, the trends exhibited in Fig. 14(a) (LES tail narrower than
the DNS tail) are expected to hold even if the DNS had been done at higher Reynolds numbers.
224 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

For further comparison, we have computed the same PDF for a LES using the standard Smagorinsky model
(SMAG) [64]. The distribution, shown in Fig. 14(b), is similar to the one obtained by using FRACT 3, although it
lacks the backscatter tail and is even less broad. One important feature that should be noted is that FRACT 3, unlike
SMAG, qualitatively captures the exponential decay of the PDF of ν ∗ > 0. Finally, the PDF of FRACT 2, shown
in Fig. 14(a), displays a narrow peak near the origin and is almost symmetrically distributed.

4.4. LES of decaying isotropic turbulence

In this section we consider LES of decaying isotropic turbulence. Since the decay rate must be properly reproduced,
this flow represents a stringent test for LES. The classic experiment of Comte–Bellot and Corrsin [15] of grid
generated decaying turbulence provides a well documented test case, which has been used to test LES models before
[46]. In the original experiment, turbulence generated downstream of a grid of mesh size M placed at the beginning
of the test section of the Johns Hopkins University Corrsin wind tunnel is measured at three different downstream
stations. Taylor’s hypothesis is then used to (i) convert time series collected at each station into information in the
space domain and (ii) to relate information taken at the three different locations in space as data gathered at the
same spot but at different times according to

Ut = x

where U is the mean streamwise velocity. The flow is characterized by a microscale Reynolds number ranging
between 60 and 70 and an integral scale l < M/2 at all times. 1D energy spectra E11 (k1 ) of the velocity fluctuations
in the streamwise direction are measured. As they are, these spectra cannot be compared to the output of a LES,
since they contain contributions from fluctuations in the plane normal to k1 from all scales, even those smaller than
∆. However, assuming that the turbulence is isotropic, we can relate the 1D spectrum E11 (k1 , t) to the 3D spectrum
E(k, t) according to [5,28]
 
1 3 ∂ 1 ∂
E(k, t) = k E11 (k, t) . (52)
2 ∂k k ∂k

Once the 3D spectrum is known, it can be filtered at scale ∆. Thus the results from the LES can be compared to the
ones measured during the experiment. Following the notation of the original paper [15], we denote the three stations
by means of the nondimensionalized time t ∗ = tU /M. The measurement locations (times) are t ∗ = 42, 98 and
171. Next, one has to consider how to approximate the experimental condition numerically. Suppose that a cubic
domain of length L is chosen, the velocity field rms speed is Urms and that the kinematic viscosity of the fluid is ν.
As in [46], we choose a computational box of side L = 11M, where M = 5.08 cm is the size of the grid generating
the turbulence. Periodic boundary conditions are imposed around the box. The initial condition is generated as a
signal with random phases and a spectrum that matches the spectrum measured at the first station, at t ∗ = 42. This
choice implies that the rms velocity of the filtered signal at t ∗ = 42 is Urms = 17.5 cm/s. No forcing is imposed. The
viscosity of air at standard condition ν = 0.145 cm2 /s is used. The viscous term is now included in Eq. (49), and
integrated implicitly by means of an integration factor. The equations of motion are solved with the same technique
explained earlier (Section 4.3).
Fig. 15 shows energy decay and spectra for BARD NL, FRACT 1 and FRACT 2. As in the steady case, BARD NL
and FRACT 1 yield poor results. Most of the dissipation comes from the molecular viscosity term and very little
energy is dissipated by the model during the simulation. Despite maximizing dissipation, FRACT 2 also does not
provide enough dissipation, as was the case for simulations of steady turbulence. The kinetic energy remains above
the observed one at all times (Fig. 15(a)). The spectra of all three models (BARD NL, FRACT 1 and FRACT 2)
show pile up of energy at large wave numbers. On the other hand at large scales the spectra fall below the measured
one.
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 225

Fig. 15. Fractal LES of decaying turbulence. (a) Resolved kinetic energy vs. time using BARD NL (squares), FRACT 1 (dashed line), FRACT 2
(solid line); the circles are the experimentally measured values. (b) Energy spectra at the three stations: t ∗ = 42 circles (exp.) and solid line (initial
condition); t ∗ = 98 squares (exp.), stars plus solid line (BARD NL), crosses plus solid line (FRACT 1), diamonds and solid line (FRACT 2);
t ∗ = 171 triangles (exp.), stars plus dotted line (BARD NL), crosses plus dotted line (FRACT 1) and diamonds plus dotted line (FRACT 2).
Note that the spectra of BARD NL and FRACT 1 are almost identical.

Fig. 16 shows the results for FRACT 3. As in the forced case, FRACT 3 with d = 0.93 shows good agreement
with the experimental data. Fig. 16(b) shows the spectrum at the three different stations. Good overall agreement is
seen, although at the small scales the spectrum contains slightly more energy.

4.5. A priori tests of fractal SGS models

To further characterize the different models considered, we have performed so-called a priori tests. In a priori
testing (e.g. [45,53]) the predicted SGS stress tensor is compared against the real SGS stress tensor computed from
data coming from a DNS or from experiments. In our case, we use the data coming from a 1283 DNS of steady
turbulent flow at Reλ = 93 [11] filtered at scale ∆ = 2π/32. The top-hat filter is defined in Fourier space as

F (k) = f (kx )f (ky )f (kz ), (53)

sin(1k/2)
f (k) = . (54)
1k/2
To compute τij from the data at scale δ = 2π/128 we take the product of the fields in real space, transform to
Fourier space, multiply by the filter kernel and finally transform back into physical space. Due to the fact that the
signal was well resolved, no dealiasing was deemed to be necessary. Once τijdns and ũi are known, we can compute
226 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

Fig. 16. Fractal LES of decaying turbulence. (a) (top) Resolved kinetic energy vs. time using FRACT 3 (solid line); the circles are the exper-
imentally measured values. (b) (bottom) Energy spectra at the three stations: t ∗ = 42 circles (exp.) and solid line (initial condition); t ∗ = 98
squares (exp.) and dotted line (FRACT 3); t ∗ = 171 triangles (exp.) and dashed line (FRACT 3).

τijmod according to the different models. Note that to compute the modeled stresses τijmod , ũi is sampled on a 323
grid and only information on this grid is used. Given two random processes we compute the correlation coefficient
habi − haihbi
ρ(a, b) = p , (55)
h(a − hai)2 ih(b − hbi)2 i
where h·i stands for the average (in our case space average). The comparison between “real” and “modeled”
SGS variables can be done at three levels. At the tensorial level we consider ρ(τijdns , τijmod ), at the vector level
ρ(Fjdns , Fjmod ) where Fj = ∂i (τij − δij (∂l ∂m /∇ 2 )τlm ) is the solenoidal component of the force acting on the large
scales due to the interaction with the small scales. Finally at the scalar level we have ρ(Fjdns ũj , Fjmod ũj ) and
ρ(τijdns S̃ij , τijmod S̃ij ). Fjmod ũj represents the local rate of change of energy due to both advection by small scales and
exchange with unresolved scales, while τij S̃ij considers only exchange. Note that when the quantities appearing in
Eq. (55) are tensors (vectors) the products are intended to be tensor (vector) products.
Results for the a priori tests are reported in Table 1.
For comparison, we have added results using a classical Smagorinsky model. Our results confirm the rather large
correlation found at every level for both versions of the Bardina model (BARD L and BARD NL). The correlation
of FRACT 1 is rather low at the vector and scalar level. FRACT 2 shows an improved correlation at vector and
scalar levels with respect to FRACT 1. Finally we consider FRACT 3 for two different values of the stretching
factors. Notice how FRACT 3 with d = 0.90 achieves a correlation level at the vectorial and scalar level only
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 227

Table 1
Correlation coefficients of different models
Model τij Fj Ff ũj τij S̃ij
Smag 0.21 0.44 0.49 0.76
BARD L 0.86 0.77 0.78 0.84
BARD NL 0.85 0.80 0.81 0.82
FRACT 1 0.64 0.22 0.21 0.57
FRACT 2 0.62 0.34 0.33 0.72
FRACT 3 (d=0.95) 0.60 0.32 0.32 0.62
FRACT 3 (d=0.90) 0.69 0.40 0.39 0.68

Fig. 17. Spectral eddy viscosities. DNS data (solid line), Smagorinsky model with an adjusted coefficient (dotted line), Bardina model (dashed
line), FRACT 2 (squares) and FRACT 3 (circles). Commonly quoted values (the fit of Eq. (57)) are indicated by the long-dashed line.

slightly smaller than the one found for the Smagorinsky model. At the tensorial level however the correlation is
significantly higher.
Finally, we consider the spectral eddy viscosity defined as
∗
hτ̂ij (k)S̃ˆ ij (k)ishell
ν(k) = − , (56)
2k 2 E(k)
which gives an indication of how much energy is drained from the eddies at scale r ∼ k −1 . (For a definition of
h·ishell see Eq. (50)).
Results are shown in Fig. 17. The “real” eddy viscosity shows a plateau at intermediate wave numbers and a cusp
near the cut-off (the first two shells are influenced by the presence of the forcing and should not be considered). The
cusp is a well documented feature of the eddy spectral viscosity (see e.g. [33,68]). As a comparison with previously
reported values, we show (long-dashed line) a working fit to the closures and some DNS results that has been given
by Briscolini and Santangelo [8]. It is
ν(k, k∆ ) = (0.15 + 5 · e−3.03k∆ /k ) (E(k∆ )/k∆ )1/2 (57)
−5/3
with E(k∆ ) = 1.6 2/3 k∆ , k∆ = 16,  = 1. It is of the same order, but slightly higher than our DNS results. The
reason may be that present results use a box-filter as opposed to a cut-off filter. Next, we show results for the models
228 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

considered. The Smagorinsky model shows a flat eddy viscosity. On the other hand, the eddy viscosity implied by
the Bardina model decreases with increasing wave number. This feature is probably what makes the Bardina model
alone unsuitable as a candidate for LES, even though a priori it shows a remarkable correlation, even concerning
the value of the total dissipation. A similar, albeit less pronounced dipping at large scales is shown by the spectral
viscosity of FRACT 2. Finally, provided that the stretching is chosen large enough, the eddy viscosity of FRACT 3
becomes close to flat, except very near the cut-off scale. It is still larger than BARD L and BARD NL.
The result from the a priori tests confirm what we have found during the simulations. Failure to effectively damp
the smallest scales of the simulation leads to an accumulation of energy at these scales that, in the worse cases,
makes the simulation unstable. It also confirms earlier observations [59] that a model displaying a high correlation
coefficient not necessarily performs well when implemented in an actual simulation.

4.6. Summary

From the simulations of forced 3D turbulence we conclude that FRACT 3 is the only version of the fractal models
considered which yields acceptable results. Comparison with turbulence statistics more detailed than energy spectra
shows that FRACT 2 does not reproduce the asymmetry in PDF of local eddy viscosity required for the energy
cascade to unresolved scales to be taking place at the correct rate. While FRACT 3 yields results that are much
better than FRACT 2, significant differences with the DNS remain, which must await further improvements in the
model. At any rate, the local eddy viscosity of FRACT 3 is closer to that of the DNS than the Smagorinsky model.
Simulations of decaying isotropic turbulence essentially yield the same conclusions: of all versions of the fractal
models considered, only FRACT 3 is able to adequately reproduce the energetics of this flow. The results for
d = 0.93 are slightly less accurate than those of the traditional Smagorinsky model, but they can be improved, if
desired, by a slight increase in d.

5. Conclusions and open questions

In this paper, we have developed an approach for modeling the effects of turbulence at small scales by combining
a kinematical assumption of fractality (using the fractal interpolation technique) with a dynamical hypothesis based
on energy dissipation. The approach was applied to 1D forced Burgers equation and to 3D forced and decaying
Navier–Stokes turbulence. Our overriding conclusion is that simply specifying a fractal dimension for the small-scale
velocity field is not enough to provide a realistic subgrid model.
For the 1D Burgers equation, the simulation was simple enough that the problem could be remedied by incorpo-
rating the required physics through an additional transport equation for the stress (and hence the fractal dimension).
For 3D formulations, this approach would entail solving six additional stress transport equations. Besides inherent
difficulties, such an effort would require modeling the pressure–strain correlation. This would require prescription
of the fractal dimension of the pressure signal, a subject that has not been addressed so far. Therefore, it was decided
to concentrate on simpler, zero equation, models for 3D turbulence.
A physical argument based on vortex stretching motivated a simple choice for the stretching factors. By minimizing
SGS turbulence in the direction of the largest positive strain-rate eigenvalue, and maximizing it in the direction
of the most negative eigenvalue, the resulting SGS stress tensor had a strong enough alignment with the resolved
strain-rate tensor so as to dissipate enough energy. This approach (FRACT 3) yielded good results in LES of forced
and decaying isotropic turbulence. In comparing FRACT 3 with the models used in the 1D problem, FRACT 3 is
qualitatively intermediate between models ZE2 and ME1, because as ZE2 it maximizes dissipation rate, but also
allows for the magnitudes of the stretching factors to vary locally, as in ME1. However, in contrast to ME1 where
the stretching factors were determined from a dynamical equation, in FRACT 3 the possible stretching factor values
are tuned.
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 229

The fractal model inherits the strength of the eddy viscosity model (effective dissipation during simulation),
but what about its weakness, namely the small correlation of the modeled SGS tress tensor with the real one.
This question was examined using a priori tests based on DNS data. The main result is that FRACT 3 exhibits
a significant correlation with the real SGS stress, between 0.6 and 0.7, while the correlation coefficient for the
traditional Smagorinsky model is typically only about 0.2. We conclude that the fractal model displays correlations
that approach those of the similarity models (e.g. [40] found ρ = 0.6), but allows for sufficient energy dissipation to
take place. Thus, the approach appears to combine naturally the advantages of similarity and eddy viscosity models.
Moreover, in passing we mention that fractal interpolation extends the idea of the phase estimation model [19] in
the sense that eddies are generated at all scales below the cut-off (and not only between ∆ and ∆/2).
It is also worthwhile mentioning recent results by Dubois et al. [21], who show explicitly that it is important to
include the interactions between resolved and subgrid motions, and not only the purely subgrid motions. In our case,
by defining and modeling the SGS stress in its entirety, without decomposing it into resolved, cross and Reynolds
small-scale terms [36], we are (arguably) including the interactions between resolved and subgrid motions.
With respect to the Smagorinsky model, the fractal model FRACT 3 has some additional advantages: it does not
require as input a length and it can be thus extended without modification to nonisotropic grids; it allows backscatter,
although in the present formulation only a small fraction of the volume experiences backscatter. However, compared
to the Smagorinsky model, at this stage the major drawback for practical application is the significant increase in
required CPU time. On a single R8000 processor of a SGI Power Challenge, the version FRACT 3 requires 23 s per
time step, compared to 2 s taken by the Smagorinsky model. A detailed profiling of the Fortran code reveals that the
bottleneck is the evaluation of the quadratic form required to compute τij which is invoked either 13 or 20 times per
node per iteration. In the code used to produce the results showed, we have not made use of any particular optimized
routine (such as the BLAS package [18] available as precompiled library), so that improved performance is likely
to occur in future developments. As it is, the model is still unsatisfactory also because it requires the specification
beforehand of the stretching factors. Instead of having a binary state model (d = 0 or 0.93) one could, possibly,
employ the Germano identity [26,39] to dynamically solve for the stretching factors.
Logically the next step would be to apply the model to LES of nonisotropic turbulence, such as bounded channel
flow, and then transitional flows, where backscatter is an important feature [55]. We remark that the fractal interpo-
lation technique may prove particularly well suited in the context of numerical methods that are themselves based
on a hierarchy of scales, such as the method of incremental unknowns [13] or wavelet-based methods [60]. Finally,
it may be possible to use the technique to model multiscale features of scalars, such as temperature or concentration
fields, being transported by turbulence, of convoluted interfaces and flamelets, and possibly to simulate the effects
of rough surfaces that often bound turbulent flows.

Acknowledgements

We are indebted to Profs. O. Knio and M. Robbins for useful discussions. We would like to thank Mr. S. Cerutti
for providing us the DNS field. The present work has been supported by NSF Grant CTS-9408344. Computer
resources were made available through NSF Equipment Grant CTS-9506077.

Appendix A. Coefficients for integration of fractal functions

To compute τ both in 1D and in 3D we need the following coefficients.

Aαx ,αy (d1 , d2 ) = A(1,2) (2,2)
αx ,αy (d1 , d2 ) + Aαx ,αy (d1 , d2 )

and
Bαx (d) = A(1,2) (2,2)
αx (d) + Aαx (d).
230 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

The computation is rather tedious, and it is best accomplished by using a symbolic manipulator, such as Mathemat-
ica(TM). A copy of the Mathematica(TM) notebook used to prepare these coefficients is available at
http://www.me.jhu.edu/∼scotti/notebook
In particular we have

A−1,−1 (d1 , d2 ) = (−8 + 8d2 − 6d22 − 3d23 + 8d1 − 8d2 d1 − 8d22 d1 + 12d23 d1 + 3d24 d1 − 6d12 − 8d2 d12
−3d13 + 12d2 d13 + 3d2 d14 )/(192(d1 d2 − 1)),
A−1,0 (d1 , d2 ) = (−16 − 16d2 + 12d22 + 6d23 + 48d2 d1 + 16d22 d1 − 24d23 d1 − 6d24 d1 − 36d12 + 24d2 d13 )/
×(192(d1 d2 − 1)), A−1,1 (d1 , d2 ) = (8d2 − 6d22 − 3d23 − 8d1 − 16d2 d1 − 8d22 d1
+12d23 d1 + 3d24 d1 − 6d12 + 8d2 d12 + 3d13 + 12d2 d13 − 3d2 d14 )/(192(d1 d2 − 1)),
A0,−1 (d1 , d2 ) = (−16 − 36d22 − 16d1 + 48d2 d1 + 24d23 d1 + 12d12 + 16d2 d12 + 6d13 − 24d2 d13 − 6d2 d14 )/
×(192(d1 d2 − 1)),
A0,0 (d1 , d2 ) = (−112 + 72d22 + 48d2 d1 − 48d23 d1 + 72d12 − 48d2 d13 )/(192(d1 d2 − 1)),
A0,1 (d1 , d2 ) = (−16 − 36d22 + 16d1 + 48d2 d1 + 24d23 d1 + 12d12 − 16d2 d12 − 6d13 − 24d2 d13 + 6d2 d14 )/
×(192(d1 d2 − 1)),
A1,−1 (d1 , d2 ) = (−8d2 − 6d22 + 3d23 + 8d1 − 16d2 d1 + 8d22 d1
+12d23 d1 − 3d24 d1 − 6d12 − 8d2 d12 − 3d13 + 12d2 d13 + 3d2 d14 )/(192(d1 d2 − 1)),
A1,0 (d1 , d2 ) = (−16 + 16d2 + 12d22 − 6d23 + 48d2 d1 − 16d22 d1 − 24d23 d1 + 6d24 d1 − 36d12 + 24d2 d13 )/
×(192(d1 d2 − 1)),
A1,1 (d1 , d2 ) = (−8 − 8d2 − 6d22 + 3d23 − 8d1 − 8d2 d1 + 8d22 d1
+12d23 d1 − 3d24 d1 − 6d12 + 8d2 d12 + 3d13 + 12d2 d13 − 3d2 d14 )/(192(d1 d2 − 1)),
and
 
1 + 2d 2 6 − 4d 2 1 + 2d 2
B(d) = , , . (58)
8 8 8

References

[1] A.-L. Barabasi, G.T. Dewey, H.E. Stanley, J. Am. Chem. Soc. 117 (1995).
[2] J. Bardina, J.H. Ferziger, W.C. Reynolds, Improved subgrid scale models for large eddy simulation, AIAA paper (1980) 80–1357.
[3] M.F. Barnsley, Fractal functions and interpolation, Constr. Approx. 2 (1986) 303–329.
[4] M.F. Barnsley, Fractals Everywhere, Academic Press, Boston, 1988.
[5] G.K. Batchelor, The Theory of Homogeneous Turbulence, Cambridge University Press, Cambridge, 1953.
[6] R. Benzi, L. Biferale, A. Crisanti, G. Paladin, M. Vergassola, A. Vulpiani , Physica D 65 (1993) 352–358.
[7] M. Briscolini, P. Santangelo, Numerical simulation of three-dimensional homogeneous isotropic flows, in: P. Messind, A. Murli (Eds.),
Proc. Conf. on Parallel Computing, Elsevier, Amsterdam, 1992.
[8] M. Briscolini, P. Santangelo, The non-Gaussian statistics of the velocity field in low-resolution large-eddy simulations of homogeneous
turbulence, J. Fluid Mech. 270 (1994) 199–217.
[9] J.M. Burger, The Nonlinear Diffusion Equation. Asymptotic Solutions and Statistical Problems, Reidel, Dordrecht, 1974.
[10] C. Canuto, M.Y. Hussaini, A. Quarteroni, T.A. Zang, Spectral Methods in Fluid Dynamics, Springer, Berlin, 1987.
[11] S. Cerutti, C. Meneveau, Intermittency and relative scaling of subgrid – scale energy dissipation in isotropic turbulence, Phys. Fluids 10
(1998) 928–937.
[12] C. Cercignani, The Boltzmann Equation and its Applications. Springer, Berlin, 1988.
[13] M. Chen, A. Miranville, R. Temam, Incremental unknowns in finite differences in 3 space dimensions, Comp. Appl. Math. 14 (1995)
219–252.
[14] A. Chekhlov, V. Yakhot, , Phys. Rev. E 51 (1995) R2739.
 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232 231

[15] G. Comte-Bellot, S. Corrsin, Simple Eulerian time correlation of full and narrow-band velocity signals in grid-generated “isotropic”
turbulence, J. Fluid Mech. 48 (1971) 273–337.
[16] J.W. Deardorff, A numerical study of three-dimensional turbulent channel flow at large Reynolds numbers, J. Fluid Mech. 41 (1970) 453.
[17] C. Doering, P. Constantin, Physica D 82 (1995) 221–228.
[18] J. Dongarra, J. DuCroz, S. Hammarling, R. Hanson, An extended set of Fortran Basic Linear Algebra Subroutines, ACM Trans. Math. Soft.
14 (1998) 1–32.
[19] J.A. Domaradzki, E.M. Saiki, A subgrid-scale model based on the estimation of unresolved scales of turbulence, Phys. Fluids 9 (1997)
1–17.
[20] B. Dubuc, J.F. Quiniou, C. Roques-Carmes, C. Tricot, S.W. Zucker, Evaluating the fractal dimension of profiles, Phys. Rev. A 39 (1989)
1500.
[21] T. Dubois, F. Jauberteau, Y. Zhou, Influences of subgrid scale dynamics on resolvable scale statistics in large-eddy simulations, Physica D
100 (1997) 390–406.
[22] G. Eyink, Energy dissipation without viscosity in ideal hydrodynamics I. Fourier analysis and local energy transfer, Physica D 78 (1994)
222–240.
[23] J. Feder, Fractals, Plenum press, New York, 1988.
[24] C.A.J. Fletcher, Computational Techniques for Fluid Dynamics, Springer, New York, 1988.
[25] J.C.H. Fung, J.C.R. Hunt, N.A. Malik, R.J. Perkins, Kinematic simulation of homogeneous turbulence by unsteady random Fourier modes,
J. Fluid Mech. 236 (1992) 281–318.
[26] M. Germano, Turbulence: the filtering approach, J. Fluid Mech. 238 (1992) 325.
[27] J.R. Herring, Subgrid scale modeling – an introduction and overview, in: F. Durst, B.E. Launder, F.W. Schmidt, J.H. Whitelaw (Eds.),
Turbulent Shear Flows I, Springer, Berlin, 1979.
[28] J.O. Hinze, Turbulence, McGraw–Hill, New York, 1975.
[29] A. Juneja, D.P. Lathrop, K.R. Sreenivasan, G. Stolovitzky, Synthetic turbulence, Phys. Rev. E 49 (1994) 5179–5194.
[30] A.R. Kerstein, A linear-eddy model of turbulent scalar transport and mixing, Combust. Sci. Tech. 60 (1988) 391–421.
[31] A.N. Kolmogorov, The local structure of turbulence in incompressible viscous fluid for very large Reynolds number, C.R. Acad. Sci.
U.S.S.R. 30 (1941) 301.
[32] A.N. Kolmogorov, A refinement of previous hypothesis concerning the local structure of turbulence in a viscous incompressible fluid at
high Reynolds number, J. Fluid Mech. 62 (1962) 82.
[33] R.H. Kraichnan, Eddy viscosity in two and three dimensions, J. Atmos. Sci. 33 (1976) 1521.
[34] L.D. Landau, E.M. Lifshitz, Fluid Mechanics, Pergamon Press, Oxford, 1987.
[35] C.E. Leith, Stochastic backscatter is a subgrid-scale model: plane shear mixing layer, Phys. Fluids A 2 (1990) 297–299.
[36] A. Leonard, Energy cascade in large-eddy simulations of turbulent fluid flows, Adv. Geophys. 18 (1974) 237 .
[37] M. Lesieur, Turbulence in Fluids, Martinus Nijhoff, Dordrecht, 1987.
[38] M. Lesieur, O. Métais, New trends in large eddy simulations of turbulence, Annu. Rev. Fluid. Mech. 28 (1996) 45–82.
[39] D.K. Lilly, A proposed modification of the Germano subgrid scale closure method, Phys. Fluids A 4 (1992) 633.
[40] S. Liu, C. Meneveau, J. Katz, On the properties of similarity subgrid-scale models as deduced from measurements in a turbulent jet, J.
Fluid Mech. 275 (1994) 83–91.
[41] S. Liu, C. Meneveau, J. Katz, Experimental study of similarity subgrid-scale models of turbulence in the far-field of a jet, Appl. Sci. Res.
54 (1995) 177–190.
[42] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, New York, 1982.
[43] B.B. Mandelbrot, J.W. Van Ness, Fractional brownian motions, fractional noises and applications, SIAM Rev. 10 (1968) 422–437.
[44] P.J. Mason, Large-eddy simulation: A critical review of the technique, Q.J.R. Meteorol. Soc. 120 (1994) 1–26.
[45] C. Meneveau, Statistics of turbulence subgrid-scale stresses: Necessary conditions and experimental tests, Phys. Fluids A 6 (1994) 815–833.
[46] C. Meneveau, T. Lund, W. Cabot, A Lagrangian dynamic subgrid-scale model of turbulence, J. Fluid Mech. 319 (1996) 353–385.
[47] C. Meneveau, K.R. Sreenivasan, The multifractal nature of turbulent energy dissipation, J. Fluid Mech. 224 (1991) 429.
[48] P. Moin, K. Manesh, Direct numerical simulation: A tool in turbulent research, Annu. Rev. Fluid Mech. 30 (1998) 539–578.
[49] A. Misra, D.I. Pullin, A vortex-based subgrid stress model for large-eddy simulation, Phys. Fluids 8 (1997) 2443–2454.
[50] A.M. Oboukhov, Some specific features of atmospheric turbulence, J. Fluid Mech. 62 (1962) 77.
[51] L. Onsager, Reciprocal relations in irreversible processes. I, Phys. Rev. 37 (1931) 405–426.
[52] S. Orey, Gaussian sample functions and the Hausdorff dimension of level crossings, Z. Wahrscheinlichkeitstheorie verw. Geb. 15 (1970)
249–256.
[53] U. Piomelli, W. Cabot, P. Moin, S. Lee, Subgrid-scale backscatter in turbulent and transitional flows, Phys. Fluid A 3 (1991) 1766–1771.
[54] U. Piomelli, S. Ragab (Eds.), Engineering applications of Large Eddy Simulations, ASME–FED 162 (1993).
[55] U. Piomelli, Y. Yunfang, R.J. Adrian, Subgrid-scale energy transfer and near-wall turbulence structure, Phys. Fluids 8 (1996) 215–224.
[56] R.R. Prasad, C. Meneveau, K.R. Sreenivasan. The multifractal nature of the dissipation field of passive scalars in fully turbulent flows,
Phys. Rev. Lett. 61 (1988) 74–77.
[57] A.A. Praskovsky, W.F. Dabberdt, E.A. Praskovskaya, W.G. Hoydysh, O. Holynskyj, Fractal geometry of isoconcentration surfaces in a
smoke plume, J. Atmos. Sci. 53 (1996) 5–21.
232 A. Scotti, C. Meneveau / Physica D 127 (1999) 198–232

[58] I. Procaccia, A. Brandenburg, M.H. Jensen, A. Vincent, The fractal dimension of iso-vorticity structures in 3-dimensional turbulence,
Europhys. Lett. 19 (1992) 183–187.
[59] W.C. Reynolds, The potential and limitations of direct and large eddy simulations, in: J.L. Lumley (Ed.), Whither Turbulence? or Turbulence
at Crossroads, 1990, p. 313.
[60] K. Schneider, N.K.-R. Kevlahan, M. Farge, Comparison of an adaptive wavelet method and nonlinearly filtered pseudospectral methods
for two-dimensional turbulence, Theoret. Comput. Fluid Dynamics 9 (1997) 191–206.
[61] A. Scotti, On scale-invariance based models of hydrodynamic turbulence, Ph.D. Thesis, Baltimore, 1998.
[62] A. Scotti, C. Meneveau, Fractal model for coarse-grained nonlinear partial differential equations, Phys. Rev. Lett. 78 (1997) 867–870.
[63] A. Scotti, C. Meneveau, S. Saddoughi, Fractal dimension of velocity signals in high-Reynolds number hydrodynamic turbulence, Phys.
Rev. E 51 (1995) 5594–5608.
[64] J. Smagorinsky General circulation experiments with the primitive equations. I. The basic experiment, Mon. Weather Rev. 91 (1963) 99.
[65] K.R. Sreenivasan, C. Meneveau, The fractal facets of turbulence, J. Fluid Mech. 173 (1986) 357–386.
[66] G.S. Winckelmans, T.S. Lund, D. Carati, A.A. Wray. A priori testing of subgrid-scale models for the velocity–pressure and vorticity–velocity
formulations, in: Studying turbulence using numerical simulation databases – VI, CTR Summer Program Proceedings, 1996.
[67] Y. Zang, R.L. Street, J.R. Koseff, A dynamic mixed subgrid scale model and its application to turbulent recirculating flows, Phys. Fluids 5
(1993) 3186–3196.
[68] Y. Zhou, G. Vahala, Reformulation of recursive–renormalization-group-based subgrid modeling of turbulence, Phys. Rev. E 47 (1993)
2503–2519.