See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/258479794

On the local nature of the pressure Hessian in fluid turbulence

Article · November 2011

CITATIONS READS

0 52

6 authors, including:

Francesco Taddia Huidan Yu

Stockholm University Indiana University-Purdue University Indianapolis
112 PUBLICATIONS   1,228 CITATIONS    52 PUBLICATIONS   687 CITATIONS   

SEE PROFILE SEE PROFILE

C. Rosales
Universidad Técnica Federico Santa María
15 PUBLICATIONS   244 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Fluid Mechanics View project

All content following this page was uploaded by Huidan Yu on 13 January 2016.

The user has requested enhancement of the downloaded file.

 PHYSICS OF FLUIDS 23, 095108 (2011)

Local and nonlocal pressure Hessian effects in real and synthetic

fluid turbulence
Laurent Chevillard,1 Emmanuel Lévêque,1 Francesco Taddia,1 Charles Meneveau,2
Huidan Yu,2 and Carlos Rosales3
1
Laboratoire de Physique de l’E´cole Normale Supérieure de Lyon, CNRS, Université de Lyon,
46 allée d’Italie F-69007 Lyon, France
2
Department of Mechanical Engineering and Center for Environmental and Applied Fluid Mechanics,
The Johns Hopkins University, 3400 N, Charles Street, Baltimore, Maryland 21218, USA
3
Department of Mechanical Engineering, Technical University Federico Santa Marı́a, Avenida España 1680,
Valparaiso 2340000, Chile

(Received 6 June 2011; accepted 19 August 2011; published online 21 September 2011)
The Lagrangian dynamics of the velocity gradient tensor A in isotropic and homogeneous
turbulence depends on the joint action of the self-stretching term and the pressure Hessian. Existing
closures for pressure effects in terms of A are unable to reproduce one important statistical role
played by the anisotropic part of the pressure Hessian, namely the redistribution of the probabilities
towards enstrophy production dominated regions. As a step towards elucidating the required prop-
erties of closures, we study several synthetic velocity fields and how well they reproduce aniso-
tropic pressure effects. It is found that synthetic (1) Gaussian, (2) multifractal, and (3) minimal
turnover Lagrangian map incompressible velocity fields reproduce many features of real pressure
fields that are obtained from numerical simulations of the Navier Stokes equations, including the
redistribution towards enstrophy-production regions. The synthetic fields include both spatially
local, and nonlocal, anisotropic pressure effects. However, we show that the local effects appear to
be the most important ones by assuming that the pressure Hessian is local in space, an expression
in terms of the Hessian of the second invariant Q of the velocity gradient tensor can be obtained.
This term is found to be well correlated with the true pressure Hessian both in terms of eigenvalue
magnitudes and eigenvector alignments. V C 2011 American Institute of Physics.

[doi:10.1063/1.3638618]

I. INTRODUCTION time singularities, both the anisotropic part of the pressure

Hessian and the viscous diffusion term have to be modeled.
The study of the velocity gradient tensor in fully devel-
This was the subject of former works.5–7 In particular, clo-
oped turbulence has led to interesting findings and has con-
sures were proposed in Ref. 7 for P and DA in terms of the
tributed to improved understanding of many statistical and
local value of A. The local closures of Ref. 7, when inserted
geometrical properties of turbulent flows. In particular,
into the dynamics generated by Eq. (1) under the action of a
recent progress has been made in the study of the Lagrangian
stochastic forcing term, lead to stationary statistics of A along
dynamics and modeling of the velocity gradient tensor (see
Lagrangian trajectories which compare well with those
Ref. 1 for an overview of the subject). This tensor is given
obtained from direct numerical simulations (DNS) of the
by Aij ¼ @ui=@xj, where u is the velocity vector. Taking a
Navier-Stokes equations at moderate Reynolds numbers.8 At
spatial gradient of the Navier-Stokes equations, the follow-
higher Reynolds numbers, predictions of the stochastic model
ing transport equation for A is obtained:
proposed in Refs. 7 and 8 turn out to become unrealistic,
dA mainly because of the weakness of the closure for the aniso-
¼ A2  P þ DA; (1) tropic part of the pressure Hessian.
dt
Indeed, the pressure Hessian is related to the spatial dis-
where d=dt stands for the Lagrangian time derivative,  is the tribution of the velocity gradient using singular integral
kinematic viscosity, and Pij ¼ @ 2p=@xi@xj is the pressure Hes- operators9–13 according to
sian. The first term A2 is the self-stretching term. The re- ð
stricted Euler (RE) approximation, which assumes an @2p dij
¼ trðA2 Þ  P:V: kij ðx  yÞtrðA2 ÞðyÞdy: (2)
isotropic pressure Hessian Pij ¼ tr(A2)dij=3 and neglects vis- @xi @xj 3
cous effects, leads to an autonomous set of coupled ordinary
differential equations.2 The intrinsic dynamics of the RE sys- In the above equation, the integral is understood as a Cauchy
tem leads to a finite time divergence of the components of A principal value (P.V.), and kij is the Hessian of the Green’s
during which the vorticity x ¼ r ^ u gets aligned with the function for the Laplacian operator, namely

eigenvector of the rate of strain, S ¼ A þ A> =2, associated
with the intermediate eigenvalue, as often observed in real tur- @2 1 jxj2 dij  3xi xj
kij ðxÞ ¼ ¼ : (3)
bulence.1,3,4 To prevent the development of unphysical finite @xi @xj 4pjxj 4pjxj5

1070-6631/2011/23(9)/095108/9/$30.00 23, 095108-1 C 2011 American Institute of Physics

V

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-2 Chevillard et al. Phys. Fluids 23, 095108 (2011)

One can see from Eq. (2) that only the isotropic part of the features of the pressure Hessian. In particular, they will be
pressure Hessian is purely local (the first term on the right- shown to reproduce the redistribution of probability towards
hand side (RHS) of Eq. (2)). All the nonlocal effects of pres- enstrophy production, as well as displaying non-vanishing
sure Hessian enter through the anisotropic part (or deviatoric action, even when Q ¼ 0.
part corresponding to the second term in the RHS of Eq. (2)). The second part of the paper studies to what degree spa-
Hence, in this view, the RE approximation can be understood tial locality is important in determining these properties of
as the neglect of all the nonlocal effects implied by the the anisotropic part of the pressure Hessian. As can be seen
incompressibility condition (or pressure field). from the expression for the pressure Hessian (Eq. (2)), the
In order to quantify the precise action of pressure in nu- anisotropic part of the pressure Hessian is also the part that
merical turbulent flows, it was proposed, in Ref. 8, to study is spatially nonlocal, i.e., the part that requires knowledge of
the probability current associated with pressure in the plane tr(A2) at positions y = x. Arguably, the more non-local
spanned by the two highly relevant invariants of A, R and Q effects are important, the more challenging it is to formulate
(the so-called RQ-plane). One of these invariants, defined as closures in terms of local quantities. In order to examine the
degree of locality, in the second part of this paper, we
1 1 1 decompose the space integration in Eq. (2) into two parts: a
Q ¼  trðA2 Þ ¼ jxj2  trðS2 Þ; (4)
2 4 2 local part given by the integration over a small ball of radius
given by the Kolmogorov length scale gK and the remainder
quantifies the net balance, or competition, between enstrophy being the “nonlocal” portion. We will show that neglecting
and dissipation. The other important invariant, defined as the second non-local contribution leads to an expression that
1 1 1 models the anisotropic part of the pressure Hessian in terms
R ¼  trðA3 Þ ¼  xi Sij xj  trðS3 Þ; (5) of the Hessian of the invariant Q. Using DNS data, this
3 4 3
expression is compared with the true pressure Hessian.
quantifies the competition between enstrophy production and
strain skewness (i.e., dissipation production). As it will be II. PRESSURE HESSIAN FROM DNS AND SYNTHETIC
recalled in the following, in terms of the velocity gradient VELOCITY FIELDS
evolution in statistically stationary turbulence, pressure has
two important roles. First of all, pressure counteracts the de- A. Probability current in the RQ-plane
velopment of the singularity implied by the self-stretching 1. Definition of the probability current
term. This feature is found to be well reproduced by existing
closures.8 The other important pressure action is the redis- We follow the approach used in Refs. 8 and 19, based on
tribution of probabilities towards enstrophy production a Fokker-Planck equation for the dynamics of R and Q. To
dominated regions (i.e., towards R < 0). This is not repro- summarize the approach, we remark that it can be shown that
duced well by existing closures. As discussed in Refs. 8 and the time evolution along a Lagrangian trajectory of the non-
14, a related deficiency of the closures is that they all predict dimensional invariants R* ¼ R=r3 and Q* ¼ Q=r2 is given by
that the pressure Hessian is proportional to Q. In the R  Q
dQ 1 1
plane dynamics, this implies that when Q ¼ 0, the effect of ¼ 3R  3 Aik Hkip  3 Aik Hki ; (6)
pressure Hessian also vanishes. For real turbulence, there is dt r r
no such vanishing of pressure Hessian effects when Q ¼ 0.8
In this article, we investigate whether these particular dR 2  2 1 1
¼ ðQ Þ  4 Aik Akl Hlip  4 Aik Akl Hli ; (7)
features of pressure (redistribution of probabilities towards dt 3 r r
enstrophy production, and non-vanishing action even when   
Q ¼ 0) are inherent to true Navier-Stokes turbulence or can where r2 ¼ tr S2 is the strain variance and t* ¼ rt is the
also be observed in various approximations, namely syn- non-dimensional time. Also, Hp stands for (minus) the devia-
thetic turbulent velocity fields. Various types of synthetic toric part of the pressure Hessian, i.e.,
fields are considered. The first type of synthetic field consid-  2 
ered is Gaussian field obtained by superposing random- p @ p dij @ 2 p
Hij ¼   ; (8)
phase Fourier modes with prescribed spectra. The second @xi @xj 3 @xk @xk
type is called “multifractal”15 and consists of a Gaussian
field whose vorticity field is amplified by means of the “fluid and H ¼ r2 A is the viscous term (recall that in the RE
deformation closure” and made consistent with multifractal- approximation, Hp ¼ H ¼ 0). The Fokker-Planck equation
ity’s long-range correlations in physical space.16 The third describing the time evolution of the joint density PðQ ; R Þ
type of synthetic field is generated using the Lagrangian may be written as
mapping technique.17,18 It also relies on random-phase 0
@
1
Gaussian fields but then applies a multi-scale deformation @P B @Q C
of fluid particles using a simple Lagrangian mapping. For þ@ A:W ¼ 0; (9)
@t @
each of these synthetic velocity fields, a pressure field is
@R
obtained numerically by means of the pressure Poisson
equation. As will be seen, unlike the local closures dis- where the divergence of the probability current W controls
cussed above, these synthetic fields reproduce many correct the time variations of the joint probability density P.

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-3 Local and nonlocal pressure Hessian effects Phys. Fluids 23, 095108 (2011)

The probability current can be written in terms of conditional power supply. It provides, in the units of the simulation, a
averages as constant energy injection rate hi ¼ 0:001. The kinematic
0 1 viscosity of the fluid is  ¼ 0.000285. The Kolmogorov’s
* dQ  + scale is gK ¼ ð 3 =hiÞ1=4 ¼ 0:0123 so that dx=gK  1 since

B dt C   dx ¼ 2p=512.
W ¼ @  AQ ; R PðQ ; R Þ (10)
dR  We display in Fig. 1(a) the vector plot and streamlines
dt of the probability current W p associated with the pressure
Hessian (Eq. (11)) in the R*Q*-plane, as it was done in Ref.
and can be decomposed into W ¼ W RE þ W p þ W  , 8. Three main remarks can be made at this stage: (1) first, the
where the probability currents W RE , W p , and W  are asso- pressure Hessian counteracts the development of the finite
ciated, respectively, to the effects on the Lagrangian evolu- time singularity along the right tail of the Vieillefosse line
tion of the invariants Q and R (Eqs. (6) and (7)) of the implied by the RE term, (2) probabilities are found to be
restricted Euler term A2, of (minus) the pressure Hessian very low in the dissipation production dominated region
P and of diffusivity r2 A entering in Eq. (1). In this arti- (i.e., R* > 0 and above the Vieillefosse tail) meaning that
cle, we will focus on the probability current associated with pressure does not play there a significant role, and (3) pres-
the pressure Hessian W p . It can be written as sure redistributes the probabilities towards the enstrophy pro-
 
duction dominated region (i.e., the flux is directed to the left,
Aik Hkip =r3    towards R* < 0). As far as the restricted Euler term is con-
Wp ¼ Q ; R PðQ ; R Þ: (11)
Aik Akl Hlip =r4  cerned, as is well known,1,2 the deterministic W RE probabil-
ity current pushes probabilities toward the right tail of the
Vieillefosse line (data not shown). This result helps to create
More details are provided in Ref. 8.
a picture of the time evolution of velocity gradients along
Lagrangian trajectories in stationary flows: The RE term
2. DNS velocity fields
“pushes” the probabilities towards the right tail towards and
In the following, we will make extensive use of data along the Vieillefosse line, while the pressure regularizes the
from standard DNS of the Navier-Stokes equations, for a implied finite time singularity and redistributes the probabil-
Taylor-based Reynolds number of order Rk ¼ 145. DNS is ities towards the left part of the plane such that, in turn, the
based on a pseudo-spectral method with 2nd-order accurate RE term can act again, etc. To that picture should be added
Adams-Bashforth time stepping; the computation box is the viscous diffusion effects, W  pushes the probabilities to-
cubic (size 2p) with periodic boundary conditions in the ward vanishing R and Q not only along the Vieillefosse line
three directions and spatial resolution 5123. Statistical statio- but also everywhere else, and stochastic forcing, such that
narity is maintained by an isotropic external force acting at the full probability current W (Eq. (10)) is divergence free,
low wavenumbers in order to ensure a constant energy- in order to ensure stationary statistics (Eq. (9)).8

FIG. 1. Probability current W p associated with the pressure

3=2Hessian—Eq. (11)—for
  the DNS velocity field and the three synthetic velocity fields—Eqs. (12),
(13), and (15)—in the R*Q*-plane, where R ¼ R= Sij Sij and Q ¼ Q= Sij Sij . The streamlines and vector plots of W p are shown in (a,c,e,g). The iso-
probability contours of the magnitude of W p are shown in (b,d,f,h). Contours are logarithmically spaced by factors of 10, starting at 1 for the contour closest to
the origin. The thick lines represent the zero-discriminant (or Vieillefosse) line: 27 2 3
4 R þ Q ¼ 0.

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-4 Chevillard et al. Phys. Fluids 23, 095108 (2011)

We display also in Fig. 1(b) the amplitude of W p , using above the Viellefosse tail) than in the DNS. The pressure
logarithmic spacing of iso-probability lines. It can be seen obtained from the Gaussian field is only realistic in the ens-
that indeed no current is discernible in the right part of the trophy production dominated region (i.e., R* < 0), where the
plane. We remark that it would be interesting to quantify behavior shows indeed a trend to push the probability density
whether the effect of anisotropic pressure Hessian is orthogo- towards this region. Also, the streamlines cross the Q ¼ 0
nal to the RQ-plane in this region, when the RQ plane is line meaning that the pressure field from the Gaussian veloc-
extended into three-dimensions, as proposed and studied in ity field does produce non-zero effect even when Q ¼ 0.
Ref. 20. In Fig. 1(d), W p -amplitude is shown. It can be seen that
the amplitude iso-values are symmetric with respect to the
R* ¼ 0 line. We are thus led to the conclusion that a Gaussian
3. Incompressible Gaussian stochastic velocity field velocity field, and its associated pressure field, do not make
with K41 correlation structure
difference between dissipation production dominated regions
Let us write a Gaussian homogeneous, isotropic, and (R* > 0) and enstrophy production dominated regions
incompressible vectorial field u(x),21,22 having a correlation (R* < 0).
structure consistent with K41 scalings. It reads, in d -
dimensions, 4. Incompressible multifractal stochastic velocity field
ð with KO62 statistics
xy
u ðxÞ ¼ uL ðx  yÞ d 2
^ dWðyÞ; (12) Based on the recent fluid deformation imposed by the
Rd þ
jx  yj2 3 Euler flow7 and further heuristic introduction of the multi-
fractal structure of turbulence as observed from extensive
where dW(y) ¼ (dW1(y), dW2(y),…,dWd(y)) is a Gaussian empirical data (see, e.g., Ref. 16), Ref. 15 proposed the fol-
vectorial white noise and uL is a large-scale cut-off which lowing 3D random vectorial field:
involves the integral length scale L. The deterministic kernel ð
xy
entering into Eq. (12) is regularized over the small length u ðxÞ ¼ uL ðx  yÞ 3 2
^ eS ðyÞ dWðyÞ; (13)
scale e, namely j:j ¼ h j:j (* stands
 for theÐconvolution prod-
R 3 þ
jx  yj2 3
uct), with a mollifier h ðxÞ ¼ 1d h x and hðxÞdx ¼ 1. It is
shown in Ref. 21 that the velocity u ðxÞ has a well-defined where S is a tensorial Gaussian log-correlated noise of the
limit when  ! 0, denoted by u(x), and such that form,
hjuðx þ ‘eÞ  uðxÞjq i  Cq ð‘=LÞq=3 when ‘ ! 0, with Cq a rffiffiffiffiffiffi ð "
constant independent on the vector e. 5 ðy  rÞ  ½ðy  rÞ ^ dWðrÞ
S ðyÞ ¼ k
A Gaussian vectorial field, such as from Eq. (12), is a 4p jy  rjL jy  rj7=2
poor representation of turbulence since it does not reproduce #
several important features such as a mean energy transfer ½ðy  rÞ ^ dWðrÞ  ðy  rÞ
þ ; (14)
towards small scales (i.e., the skewness phenomenon), the jy  rj7=2
non-Gaussianity of velocity increments (i.e., the intermit-
tency phenomenon), and the alignment of vorticity with the with  denoting the tensorial product. The form of the sym-
intermediate eigenvector of the strain rate tensor.1,3,4 Never- metric matrix S is inspired by the recent fluid deformation
theless, it is useful to consider it in the analysis of the statisti- closure experienced by the fluid over short times,15 and the
cal quantities in which we are interested, such as the exponent 7=2 has been selected such that the components of
probability current W p associated with the pressure Hessian S are correlated logarithmically in space. A free parameter k
(Eq. (11)). In particular, we look at which part can be attrib- enters this construction and governs the level of intermit-
uted to Gaussian statistics and which part is really linked to tency of the field. We will take in the sequel k2 ¼ 0.025 in
turbulence. order to be consistent with empirical findings.23
The Gaussian velocity field (Eq. (12)) is computed in a Generation of the vectorial field u ðxÞ can be done accu-
periodic box in d ¼ 3 space dimensions, using N ¼ 10243 col- rately and efficiently in periodic boxes using up to 10243 col-
location points. The regularizing parameter  is chosen as locations points, in a similar way as done for the Gaussian
 ¼ 6dx, where the spatial resolution is dx ¼ 1=N. For the velocity field (Eq. (12)). The cost of the computation of the
mollifier h and the large scale cut-off u, we take Gaussian matrix exponential is the limiting numerical step. It is esti-
functions. See Ref. 15 for further numerical details. The mated at each point of space using a Padé approximant with
Pressure p is defined via the Poisson equation Dp ¼ tr(A2), scaling and squaring (see Ref. 15 for details).
where A is the (Gaussian) velocity gradient tensor. It has been shown numerically (Ref. 15) that the multi-
Figure 1(c) shows the vector plot and streamlines of the fractal velocity field—Eq. (13)—gives a realistic representa-
probability current W p obtained from the Gaussian velocity tion of instantaneous realizations of velocity fields in fully
field—Eq. (12). It can be seen that the pressure from the developed turbulence in the inertial range in regard to the
Gaussian field does not counteract directly the singularity following properties: (1) Longitudinal d‘u and transverse ve-
along the direction of the right tail of the Vieillefosse line, as locity increments are intermittent, k being the intermittency
is the case in the DNS. Also, there is a significantly higher coefficient, (2) the third-order moment hðd‘ uÞ3 i is negative
probability current in the right side region (i.e., R* > 0 and and proportional to the scale ‘. The fact that there is negative

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-5 Local and nonlocal pressure Hessian effects Phys. Fluids 23, 095108 (2011)

skewness S ¼ hðd‘ uÞ3 i=hðd‘ uÞ2 i3=2 means that u exhibits a now with the field decomposed at a smaller filtering scale. In
non-vanishing mean energy transfer towards the small this way, the effects are superposed and accumulated over a
scales, and (3) vorticity gets preferentially aligned with the range of spatial scales. Further details and characteristics of
eigenvector of the strain-rate tensor corresponding to the in- these synthetic velocity fields can be found in Refs. 17 and
termediate eigenvalue. 18. The present MTLM velocity field was generated in a per-
We show in Fig. 1(e) the vector plot and streamlines of iodic box, using 5123 collocation points, with M ¼ 6 genera-
the probability current W p obtained from the multifractal tions in the hierarchy, and an energy spectrum corresponding
velocity field. It can be seen that, in a similar fashion to the to Rk  250.
Gaussian case, streamlines are roughly symmetric with The results for the probability current W p obtained
respect to the R* ¼ 0 line. This is not consistent with DNS in from the MTLM velocity field (Eq. (15)) are shown in Figs.
the R* > 0 region, but it is still realistic in the R* < 0 region. 1(e) and 1(f). When compared with the DNS results (Figs.
The difference with the Gaussian field is the fact that now 1(a) and 1(b)), we can see a close agreement of both magni-
the joint density of R* and Q* is not symmetric with respect tude and direction of W p . In particular, the MTLM velocity
to the R* ¼ 0 line, showing thus the predominance of the field reproduces, for pressure-related part of the probability
regions for enstrophy-enstrophy production (upper-left quad- current, the void of probability in the R* > 0 region. We can
rant, which in turbulent flows is correlated with vortex conclude that, of the three cases studied, the MTLM velocity
stretching) and dissipation-dissipation production (lower- field gives the most realistic synthetic turbulence, as far as
right quadrant, connected with biaxial straining in turbulent anisotropic pressure Hessian effects are concerned. But some
flows). This is also the case for the probability current ampli- small difference can be observed in the R* > 0 region, close
tude, as shown in Fig. 1(f). We can see, therefore, that both to the origin, where the MTLM fields seem to predict a circu-
the Gaussian and multifractal velocity fields do not repro- lar motion that is not present in DNS.
duce the void in probability in the R* > 0 region as observed
in DNS, but they do reproduce accurately the probability B. Mean pressure Hessian norm conditioned on Q
current evolutions in the R* < 0 regions and the presence of
probability flux at Q ¼ 0. As previously noticed in Ref. 8, current closures for the
pressure Hessian5,7,14 are proportional to the invariant Q. This
5. MTLM velocity field is in particular the case for the closure provided in Ref. 7,
namely
We consider a third case of a synthetic velocity field.
The minimal turnover Lagrangian map (MTLM) velocity C1
sg
field is obtained by distorting an initially random solenoidal P¼ trðA2 Þ; (16)
trðC1
sg Þ
vector field, u0(x), over a hierarchy of spatial scales
f‘n ¼ 2nL, n ¼ 1,…,Mg, where L is of the order of the inte-
where Csg is the statistically stationary Cauchy-Green tensor
gral scale and the smallest scale, ‘M, is in the order of Kol-
at the Kolmogorov time scale sg (Refs. 7 and 8) and
mogorov scale. This generates the multiscale recursive
tr(A2) ¼ 2Q. Indeed, it is tempting to close the anisotropic
sequence,
part of P as a symmetric tensor proportional to Q since the
un ðxÞ ¼ T½un  1 ðxÞ; ‘n  n ¼ 1; …; M; (15) isotropic part is itself proportional to Q as seen on the Pois-
son equation Dp ¼ 2Q. It would imply in particular that the
whose final step, uM(x), is the synthetic velocity field. Here, probability current W p vanishes on the Q ¼ 0 line. This is
T[ ] stands for the distortion operations applied. At each level not observed on DNS (see Figs. 1(a) and 1(b)). To quantify
n in the sequence, the velocity is filtered at scale ‘n and more precisely the behavior of P in the neighborhood of van-
decomposed into low-pass and high-pass filtered parts: u< ishing Q, we proposed in Ref. 8 to estimate the average pres-
and u>, respectively. The u< part is deformed by mapping sure Hessian (square) norm jPj2 ¼ tr(P2) conditioned on the
the velocity vectors from their collocation points, x, to new    Q. The corresponding conditional average
invariant
positions that fluid particles moving at constant velocity in tr P2 jQ was shown for fields obtained by DNS as well as
Lagrangian coordinates would reach: u<(X(t), t) ¼ u<(x, 0), by applying the closure (Eq. (16)). It was observed that for
with X(t) ¼ x þ tu<(x). The parameter t is taken equal to the the DNS case, such conditional average at Q ¼ 0 does not
eddy-turnover time-scale corresponding to the spatial scale vanish and, furthermore, it behaves nearly linearly with Q,
‘n, computed using standard Kolmogorov scaling. New ve- whereas the closure Eq. (16) predicts a vanishing conditional
locity values at the collocation points are obtained by inter- average for Q ¼ 0 and a quadratic behavior with Q.   
polation over nearby velocities that have come into a We display in Fig. 2 the conditional average tr P2 jQ
neighborhood of radius ‘n around x after the mapping. This as a function of Q for the four different velocity fields: the
deformed field u< is made solenoidal again by projection in DNS, Gaussian (Eq. (12)), multifractal (Eq. (13)), and
Fourier space, and the amplitudes of its Fourier modes are MTLM (Eq. (15)) velocity fields. As previously observed,
scaled to conform to the target energy spectrum. Finally, u< for the DNS case (solid line), the conditional average does
is recombined with the u> part, which at this stage still not vanish at Q ¼ 0 and behaves linearly with Q in the neigh-
remains as a Gaussian field. borhood around Q ¼ 0. For all the three remaining synthetic
The next generation in the hierarchy will take the com- velocity fields, the conditional average does not vanish at
plete field u and will apply the same distortion operations, Q ¼ 0 and, hence, they perform better than the deterministic

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-6 Chevillard et al. Phys. Fluids 23, 095108 (2011)

cated synthetic velocity fields, such as Eqs. (13) and (15), can
reproduce the motion of the probability current W p associated
with the pressure Hessian in the R* < 0 region. Additionally,
taking into account the spatial distribution of the velocity field
also leads to a non-vanishing conditional mean pressure Hes-
sian norm for Q ¼ 0. In this section, we study to what degree
spatial locality is important in determining these properties of
the anisotropic part of the pressure Hessian.
The exact expression (2) for the pressure Hessian is very
useful since it allows interpreting its isotropic part as being
local, whereas the anisotropic part is governed by tr(A2), or
equivalently Q at different locations, i.e., it contains non-local
contributions from the spatial variations of Q. In this section,
we will work with an equivalent form of Eq. (2) that under-
   
FIG. 2. Conditional expectation tr P2 jQ of pressure Hessian norm on lines the role played by the Hessian Q of the invariant Q.
Q*: DNS (solid line), Gaussian (dotted-dashed line), multifractal (dashed Indeed, taking two spatial derivatives of the Poisson equation
line), and MTLM (dotted line) velocity fields.
that commute 2 with the Laplacian, one obtains DP ¼ 2Q,
closure (Eq. (16)) on this issue. For the Gaussian case, how- where Qij ¼ @x@ i @x
Q
j
. It is then easily seen that a similar relation
ever (dash-dotted line), the tails of the conditional average exists between the deviatoric parts of P and Q, namely
are not realistic, being far below the curves corresponding to DPd ¼ 2Qd, where the superscript d denotes the deviatoric
the DNS case. Numerical simulations of the Gaussian fields part, i.e., for example, Pd ¼ P  13trðPÞI, with I being the
at lower resolutions, i.e., N ¼ 2563 or N ¼ 5123 (data not identity matrix. We finally reach a relation, equivalent to Eq.
shown), showed no difference with the N ¼ 10243 case. (2), between Pd and Qd
Interestingly, in this regard, the multifractal field (dashed ð
1 1
line) performed much better, exhibiting conditional average d
P ðxÞ ¼  Qd ðyÞdy: (17)
tails very close to the DNS result. Some discrepancy is found 2p jx  yj
for negative Q s, where the tail in the DNS case has a steeper
Relation (17) is exact. In the following, we will truncate the
slope. The MTLM velocity field (dotted line) also performs
integral present in Eq. (17) over a ball, centered at x, and of
well against DNS data, although its tails are found to be quite
radius g, namely
symmetric, at odds with DNS.
ð
Overall, the behavior of the pressure Hessian obtained 1 1
from the three synthetic velocity fields is reasonably satisfac- Pd ðxÞ   Qd ðyÞdy: (18)
2p jx  yjg jx  yj
tory when compared against DNS data. The probability cur-
rent is well reproduced in the R < 0 region and the It is easily seen that from Eq. (18), we recover Eq. (17) by
conditional average shown in Fig. 2 does not vanish for van- taking g ! þ1. We now make the strong assumption that g
ishing Q. Only the MTLM velocity field can reproduce addi- is of order of the Kolmogorov length scale gK. In this case,
tionally the void in probability observed in DNS over the we can Taylor-expand the Hessian of Q at the position y
R* > 0 region (explaining, or at least giving an interpretation around its value at the location x, take out Qd(x) from the in-
of the lack of action of the pressure Hessian in the dissipa- tegral, and perform the remaining integration in spherical
tion production dominated region R* > 0 remains, however, coordinates. We get an expression for the deviatoric part of
an open problem). Furthermore, we have shown that syn- the pressure Hessian,
thetic velocity fields do predict the pressure Hessian square
norm as being closer to linearly proportional to the invariant  d  d
@2p 2 @2Q
Q, rather than proportional to Q2 as is the case in existing  g : (19)
@xi @xj @xi @xj
closures, in particular Eq. (16).
At this stage, one could reach the conclusion that some This now expresses the anisotropic part of the pressure Hes-
approximate surrogates of an actual turbulent field, even sian in terms of local properties of the invariant Q, although
when obtained with the simplest Gaussian approximation, the latter’s spatial derivatives are needed. These derivatives
contain a better prediction of the behavior of the pressure are unknown in the closure and Lagrangian models of, e.g.,
Hessian, in connection with its dependence on Q, than the Ref. 7, and so this does not constitute a practical closure yet.
deterministic closure given by Eq. (16). We will see in the To ensure that this expression yields the same norm as the
following that the anisotropic part of the pressure Hessian true pressure Hessian, we define the ball’s radius according to
can, in fact, be accurately closed by the local spatial varia-
tions of the invariant Q. sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
htrðP2 Þi
g2 ¼ : (20)
htrðQ2 Þi
III. LOCALITY OF THE PRESSURE HESSIAN
We have seen in the first part of this work that a simple The expression (19) and the choice of the length scale g by
Gaussian approximation, given by Eq. (12), or more sophisti- Eq. (20) are consistent only if (1) eigenvalues and

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-7 Local and nonlocal pressure Hessian effects Phys. Fluids 23, 095108 (2011)

eigenvectors of Pd and Qd are well correlated and (2) g is est eigenvalues are well correlated. In Fig. 3(d), we display
indeed of the order of the Kolmogorov length scale gK since the probability of the cosine of the angle between the eigen-
we assume that in the neighborhood of x, Qd(y)  Qd(x). vectors of the left and right terms of Eq. (19). It is found that
Both Hessian tensors, for the pressure and for Q, are com- in all the cases, corresponding to the three different eigenval-
puted in Fourier space, for the periodic DNS flows. The ues, the maximum of probability is reached when the eigen-
expression (19) requires the computation of the second deriva- vectors of Pd and g2Qd are aligned. These results show
tive of Q, which is already the square of a first spatial deriva- then that Pd and g2Qd are correlated both in amplitude and
tive. This is a reason for our use of the highly resolved DNS eigendirections. We have seen that, in tensorial structure, the
under study, in which dx=gK  1 (see Sec. II A 2). Using true pressure Hessian Pd and the local expression in terms of
gK ¼ ð 3 =hiÞ1=4 ¼ 0:0123, we find that for the current DNS, velocity gradients g2Qd are quite similar. We may now
our choice of the length scale g as Eq. (20) implies that wonder if the local expression is able to reproduce the proba-
g ¼ 1.83 gK. Hence, g is of the order of gK, as required. Let us bility current associated to pressure effects as seen in Figs.
stress that a well resolved DNS is necessary since some noise 1(a) and 1(b). To this purpose, we show in Figs. 4(a) and
can be introduced by the computation of third order deriva- 4(b) the probability current W p obtained from DNS when
tives, leading to a bad estimation of the parameter g. More the true pressure Hessian is replaced by the local expression
work is needed to clarify this point and to assess the depend- (19). We see in Fig. 1(a) that, indeed, the local expression
ence of g on resolution effects and Reynolds numbers. This reproduces the counteractive action of the pressure along the
could be done, for example, performing specifically designed right tail of the Vieillefosse line. Furthermore, a void in
DNS aimed at quantifying accurately high-order velocity probability is found in dissipation production dominated
derivatives, as was proposed in Refs. 24 and 25. regions (i.e., R* > 0), as can be clearly seen in Fig. 1(b).
We display in Fig. 3 the joint probability densities of the Finally, the local expression also reproduces some of the
three eigenvalues of the deviatoric part of the true pressure probability redistribution in the enstrophy production domi-
Hessian (denoted by kp) and the corresponding eigenvalues nated region (R* < 0), as observed in DNS (Figs. 1(a) and
of the expression g2Qd (denoted by kq). The smallest (Fig. 1(b)), although the direction of the probability flux stream-
3(a)) and largest eigenvalues (Fig. 3(c)) are found to be well lines for R* < 0 is seen to be more vertical than the left-
correlated with joint density contours that are elongated wards directions seen in the DNS. Also, the streamlines are
along the perfect correlation line (i.e., kp ¼ kq) and correla- found more curved for the local expression than for P, and
tion coefficients of q ¼ 0.714 and q ¼ 0.703, respectively. the probability current amplitude jW p j is found to decrease
As far as the intermediate eigenvalue is concerned, faster at high values of R and Q than in the DNS case. This
weaker correlation is found (q ¼ 0.275), with isolines being could be due to limitations of the localized expression in
close to circles. We can conclude that the smallest and larg- reproducing very high turbulent fluctuations.
The overall behavior of the local expression in the RQ-
plane is on the whole quite satisfactory when compared
against DNS. Some differences appear: (1) the streamlines
of the probability current are found more curved for the local
closure than for P, very much in opposition to the stream-
lines imposed by the RE approximation (see Ref. 8) and (2)
the probability current amplitude jW p j is found to decrease
faster at high values of R and Q than in the DNS case. This
could be due to limitations of the localized expression in
reproducing very high turbulent fluctuations. On the whole,
however, the trends provided by the local expression Eq.
(19) agree quite well compared to the DNS results in terms
of the probability fluxes in the RQ plane.
We also show in Fig. 4(c) the conditional expectation of
the pressure Hessian square norm, based on Q. For the sake
of clarity, we show again the conditional average as obtained
in DNS (solid line) (it was already shown in Fig. 2). We use
open symbols (*) to show the conditional expectation
obtained from DNS using the local approximation (19). It
can be seen that the conditional expectation obtained from
the local approximation is in very good agreement with the
DNS, with some discrepancies appearing for negative Q*.
Interestingly, we see that the conditional expectation does
FIG. 3. In (a,b,c), we show the joint PDF of eigenvalues kp of the pressure not vanish for Q ¼ 0. This is consistent with former remarks
Hessian and eigenvalues kq of the proposed local expression (Eq. (19)). Iso- made about the requirement that a realistic model of P can-
probability lines correspond to e4, e3, e2, e1, 1. In (d) is shown the PDF
of the cosine of the angle between the eigenvectors of P and those of the
not be simply proportional to Q.
local closure (Eq. (19)): eigenvectors associated to the smallest (dashed), in- Finally, to quantify the agreement between individual
termediate (solid), and biggest (dot-dashed) eigenvalues. tensor elements, we show in Fig. 4(d) the joint Probability

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-8 Chevillard et al. Phys. Fluids 23, 095108 (2011)

sian P when compared in DNS computations. These findings

show that the main contribution to P(x) is contained in the
local neighborhood around position x, in a ball centered at x,
and of radius of the order of gK. This raises the hope that
local closures involving a finite set of ordinary differential
equations may still be possible for studying the Lagrangian
dynamics of the velocity gradient tensor. To that end, it is
still necessary to express the Hessian of Q in terms of the
local values of A. Only then would we have a full closure.
Let us finally remark that if a tractable transport equa-
tion for the pressure Hessian is difficult to get, the Lagran-
gian derivatives of p and P can be related. We get the
following transport equation for the pressure Hessian:

dO P @p dp
¼ rruk þ rr ; (21)
dt @xk dt

where dO=dt stands for the upper convected time derivative

or Oldroyd derivative that relates the rate of change written
in the coordinate system rotating and stretching with the fluid
(see, for example, Ref. 26), i.e.,
FIG. 4. (a) and (b): Probability current W p associated with pressure Hessian dO P dP
(Eq. (11)) for the DNS velocity field when using the local
 closure
  (19). See ¼ þ A> P þ PA: (22)
Fig. 1 for further details. (c) Conditional expectation tr P2 jQ of pressure dt dt
Hessian norm on Q*: DNS (solid) and closure (*), in a similar fashion than
in Fig. 2. (d) Joint PDF of the components P12 and g2AQ12 in a logarithmic Then, if we neglect the right-hand side of Eq. (21) in the rate
scale. Isolines correspond to probability [103, 102, 101, 1, 10]. of change of the pressure Hessian, i.e., if we assume that the
Oldroyd derivative (22) vanishes, and we apply the recent
Density Function (PDF) of the component P12 and the com- fluid deformation (RFD) approximation (assuming A inde-
ponent g2Q12. This plot demonstrates the good level of cor- pendent on time), we get that for an early time
relation between these two tensors since the joint PDF is >
clearly elongated along the perfect correlation line (i.e., PðsÞ ¼ esA Pð0ÞesA : (23)
P12 ¼ g2Q12). The corresponding correlation coefficient is
q ¼ 0.55. Thus, also this statistical test confirms the good This remark justifies the use of matrix exponentials as clo-
agreement between the local approximation and the true sures of the pressure Hessian,7 as it was also noted for the
pressure Hessian. subgrid-scale stress tensor.27 If in addition we start from an
initial isotropic pressure Hessian, and include the Poisson
equation, we recover exactly the closure (16) proposed in Ref.
IV. CONCLUSIONS AND PERSPECTIVES 7. This represents useful perspectives for future investigations
This article focuses on the statistical nature of the pres- on the Lagrangian dynamics of the velocity gradient tensor.
sure Hessian that governs much of the Lagrangian dynamics
of the velocity gradient tensor in turbulence. In the first part,
we have seen that synthetic velocity fields reproduce many ACKNOWLEDGMENTS
properties of the pressure Hessian as they are seen in DNS We thank G. Eyink and B. Lüthi for fruitful discussions.
flows, such as the non-trivial behavior of the probability cur- We acknowledge the Leonardo Da Vinci Programme for
rent, and the conditional expectation of the pressure Hessian support in the internship of F.T. at the ENS, and the CNRS
norm on the invariant Q. Even the simplest Gaussian approx- for constant support. DNS and stochastic fields have been
imation for the velocity field (Eq. (12)) can represent some performed by using the local computing facilities (PSMN) at
non-trivial behaviors of the pressure that could not be pre- ENS Lyon under grant CPER-CIRA. This research was sup-
dicted by the closures in terms of A.5,7,14 Based on this ob- ported in part by the National Science Foundation under
servation and on an exact field description of the pressure Grant No. NSF PHY05-51164 during the stay of L.C. at the
Hessian by means of nonlocal integrals (Eqs. (2) and (17)), KITP (Santa Barbara). C.M. and H.Y. are supported by NSF
we formulate the hypothesis that considering only the inte- Grant No. CDI-0941530. C.R. acknowledges the support of
gration over a ball of radius gK and neglecting other contri- CONICYT under Project Fondecyt 11080025.
butions, the deviatoric part of P could be expressed in terms
of local properties of the velocity gradient tensor, but in 1
C. Meneveau, “Lagrangian dynamics and models of the velocity gradient
terms of higher-order derivatives. Specifically, the spatially tensor in turbulent flows,” Annu. Rev. Fluid Mech. 43, 219 (2011).
2
B. J. Cantwell, “Exact solution of a restricted Euler equation for the veloc-
local approximation is not expressed in terms of A, but in ity gradient tensor,” Phys. Fluids A 4, 782 (1992).
terms of second-order derivatives of Q. This approximation 3
A. Tsinober, An Informal Introduction to Turbulence (Kluwer Academic,
was found to be highly correlated with the true pressure Hes- Dordrecht, 2001).

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
095108-9 Local and nonlocal pressure Hessian effects Phys. Fluids 23, 095108 (2011)

4 16
J. M. Wallace, “Twenty years of experimental and direct numerical simu- U. Frisch, Turbulence (Cambridge University, Cambridge, 1995).
17
lation access to the velocity gradient tensor: What have we learned about C. Rosales and C. Meneveau, “A minimal multiscale Lagrangian map
turbulence?,” Phys. Fluids 21, 021301 (2009). approach to synthesize non-Gaussian turbulent vector fields,” Phys. Fluids
5
M. Chertkov, A. Pumir, and B. I. Shraiman, “Lagrangian tetrad dynamics 18, 075104 (2006).
18
and the phenomenology of turbulence,” Phys. Fluids 11, 2394 (1999). C. Rosales and C. Meneveau, “Anomalous scaling and intermittency in
6
E. Jeong and S. S. Girimaji, “Velocity-gradient dynamics in turbulence: three-dimensional synthetic turbulence,” Phys. Rev. E 78, 016313 (2008).
19
Effect of viscosity and forcing,” Theor. Comput. Fluid Dyn. 16, 421 F. van der Bos, B. Tao, C. Meneveau, and J. Katz, “Effects of small-scale
(2003). turbulent motions on the filtered velocity gradient tensor as deduced from
7
L. Chevillard and C. Meneveau, “Lagrangian dynamics and statistical geo- holographic PIV measurements,” Phys. Fluids 14, 2457 (2002).
20
metric structure of turbulence,” Phys. Rev. Lett. 97, 174501 (2006). B. Lüthi, M. Holzner, and A. Tsinober, “Expanding the QR space to three
8
L. Chevillard, C. Meneveau, L. Biferale, and F. Toschi, “Modeling the dimensions,” J. Fluid Mech. 641, 497 (2009).
21
pressure Hessian and viscous Laplacian in turbulence: Comparisons with R. Robert and V. Vargas, “Hydrodynamic turbulence and intermittent ran-
DNS and implications on velocity gradients dynamics,” Phys. Fluids 20, dom fields,” Commun. Math. Phys. 284, 649 (2008).
22
101504 (2008). P. D. Tafti and M. Unser, “Fractional Brownian vector fields,” Multiscale
9
K. Ohkitani, “Eigenvalue problems in three-dimensional Euler flows,” Model. Simul. 8, 1645 (2010).
23
Phys. Fluids A 5, 2570 (1993). L. Chevillard, B. Castaing, E. Lévêque, and A. Arneodo, “Unified multi-
10
P. Constantin, “Geometric statistics in turbulence,” SIAM Rev. 36, 73 fractal description of velocity increments statistics in turbulence: Intermit-
(1994). tency and skewness,” Physica D 218, 77 (2006).
11 24
K. Ohkitani and S. Kishiba, “Nonlocal nature of vortex stretching in an Z.-S. She, S. Chen, G. Doolen, R. H. Kraichnan, and S. A. Orszag,
inviscid fluid,” Phys. Fluids A 7, 411 (1995). “Reynolds number dependence of isotropic Navier-Stokes turbulence,”
12
A. Majda and A. Bertozzi, Vorticity and Incompressible Flow (Cambridge Phys. Rev. Lett. 70, 3251 (1993).
25
University, Cambridge, 2002). J. Schumacher, “Sub-Kolmogorov-scale fluctuations in fluid turbulence,”
13
K. Ohkitani, “A miscellany of basic issues on incompressible fluid equa- Europhys. Lett. 80, 54001 (2007).
26
tions,” Nonlinearity 21, T255 (2008). P. Haupt, Continuum Mechanics and Theory of Materials (Springer, Ber-
14
J. D. Gibbon and D. D. Holm, “Lagrangian particle paths and ortho- lin, 2000).
27
normal quaternion frames,” Nonlinearity 20, 1745 (2007). Y. Li, L. Chevillard, G. Eyink, and C. Meneveau, “Matrix exponential-
15
L. Chevillard, R. Robert, and V. Vargas, “A stochastic representation of based closures for the turbulent subgrid-scale stress tensor,” Phys. Rev. E
the local structure of turbulence,” Europhys. Lett. 89, 54002 (2010). 79, 016305 (2009).

Author complimentary copy. Redistribution subject to AIP license or copyright, see http://phf.aip.org/phf/copyright.jsp
View publication stats