Ucl 2004

Local and Global Aspects of Mixing
Jean-Luc Thiffeault
Department of Mathematics
Imperial College London
with
Steve Childress
Courant Institute of Mathematical Sciences
New York University
http://www.ma.imperial.ac.uk/jeanluc
Local and Global Aspects of Mixing p.1/30
Experiment of Rothstein et al.: Persistent Pattern
Disordered array of
magnets with oscillatory current drive a
thin layer of electrolytic solution.
periods 2, 20, 50, 50.5
[Rothstein, Henry, and Gollub,
Nature 401, 770 (1999)]
Evolution of Pattern
Striations
Smoothed by diffusion
Eventually settles into pattern (eigenfunction)

Local vs Global Regimes of Mixing

Local theory:
Based on distribution of Lyapunov exponents.

Local theory:
[Antonsen et al., Phys. Fluids (1996)]
[Balkovsky and Fouxon, PRE (1999)]

[Son, PRE (1999)]
Average over angles

Statistical model
Statistical model

Local theory:

[Son, PRE (1999)]
Average over angles

Statistical model
Statistical model
Global theory:
Eigenfunction of advectiondiffusion operator.

Local theory:

[Son, PRE (1999)]
Average over angles

Statistical model
Statistical model
Global theory:
[Pierrehumbert, Chaos Sol. Frac. (1994)]
Strange eigenmode
[Fereday et al., Wonhas and Vassilicos, PRE (2002)] Bakers map
[Sukhatme and Pierrehumbert, PRE (2002)]
[Fereday and Haynes (2003)]
Unified description

Local theory:

[Son, PRE (1999)]
Average over angles

Statistical model
Statistical model
Global theory:
So far, local theories are Lagrangian and global theories are

Eulerian.

Local theory:

[Son, PRE (1999)]
Average over angles

Statistical model
Statistical model
Global theory:

Eulerian.
Today: Try to connect the two pictures.

Local theory:

[Son, PRE (1999)]
Average over angles

Statistical model
Statistical model
Global theory:

Eulerian.
Today: Try to connect the two pictures.
Cannot often do this! Map allows (mostly) analytical results.

A Bit of History
Eulerian (spatial) coordinates are due to. . .
A Bit of History
Eulerian (spatial) coordinates are due to. . .
dAlembert
A Bit of History
. . . and Lagrangian (material) coordinates to. . .
dAlembert
Euler
The people responsible for the confusion. . .
The people responsible for the confusion. . .
Lagrange
Dirichlet
(See footnote in Truesdell, The Kinematics of Vorticity.)

The Map
We consider a diffeomorphism of the 2-torus T2 = [0, 1]2 ,
M(x) = M x + (x),
where
M=
2 1
;
1 1
(x) =
2
sin 2x1
sin 2x1
M x is the Arnold cat map.
The map M is area-preserving and chaotic.
For = 0 the stretching of fluid elements is homogeneous in

space.
For small the system is still uniformly hyperbolic.
Advection and Diffusion: Eulerian Viewpoint

Iterate the map and apply the heat operator to a scalar field (which
we call temperature for concreteness) distribution (i1) (x),
(i) (x) = H (i1) (M1 (x))
where is the diffusivity, with the heat operator H and kernel h

Z
H (x) :=
h (x y)(y) dy;
T2
X
h (x) =
exp(2ik x k2 ).
k
In other words: advect instantaneously and then diffuse for one

unit of time.
Transfer Matrix
Fourier expand (i) (x),
(i) (x) =
(i)
k e2ikx .
The effect of advection and diffusion becomes

X
(i)
(i1)
Tkq q ,
k (x) =
q
with the transfer matrix,

Z

2
Tkq :=
exp 2i (q x k M(x)) q dx,
=e
T2
q 2
0,Q2 iQ1 JQ1 ((k1 + k2 ) ) ,
Q := k M q,
where the JQ are the Bessel functions of the first kind.

Variance: A Measure of Mixing

In the absence of diffusion ( = 0) the variance (i)
Z
X (i)
(i) 2
(i) 2
(i)
(i)
(x) dx =
:=
k ,
k :=
k
T2
is preserved. (We assume the spatial mean of is zero.)

For > 0 the variance decays.
We consider the case 1, of greatest practical interest.

Z
X (i)
(i) 2
(i) 2
(i)
(i)
(x) dx =
:=
k ,
k :=
k
T2

Three phases:
The variance is initially constant;

Z
X (i)
(i) 2
(i) 2
(i)
(i)
(x) dx =
:=
k ,
k :=
k
T2

Three phases:
It then undergoes a rapid superexponential decay;

Z
X (i)
(i) 2
(i) 2
(i)
(i)
(x) dx =
:=
k ,
k :=
k
T2

Three phases:
It then undergoes a rapid superexponential decay;
(i)
settles into an eigenfunction of the AD operator that sets

the exponential decay rate.
ag replacements
Decay of Variance
10
= 103
variance
10
10
20
e15.2i
40
10
10
60
= 10
10
0.5
10
12
14
iteration
Variance: 5 iterations for = 0.3 and = 103

0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Eigenfunction for = 0.3 and = 103

(Renormalised by decay rate)
0.9
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.1
0.2
0.3
0.4
0.5
0.6
i = 25
0.7
0.8
0.9
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
i = 30
Decay Rate
For small , the dominant Bessel function is J1 , so the decay
factor 2 for the variance is given by

2
1
= T(0 1),(0 1) = e J1 () = 2 + O , .
Hence, for small the decay rate is limited by the (0 1) mode.

The decay rate is independent of for 0.
Decay Rate

2
1
= T(0 1),(0 1) = e J1 () = 2 + O , .

This is an analogous result to the bakers map [Fereday et al.,

Wonhas and Vassilicos, PRE (2002)]. Here the agreement with
numerical results is good for quite close to unity.
Decay Rate

2
1
= T(0 1),(0 1) = e J1 () = 2 + O , .

This is an analogous result to the bakers map [Fereday et al.,

Wonhas and Vassilicos, PRE (2002)]. Here the agreement with
numerical results is good for quite close to unity.
In the bakers map the discontinuity implies a slow convergence
of the Fourier modes. However, it is a one-dimensional problem.
Decay Rate as 0
0
2
4
log 2
ag replacements
10
12
14
16
18
20 4
10
10
10
10
10
10
Lagrangian Viewpoint
Puzzle: Superexponential decay in Lagrangian coordinates.
Fix this by averaging over initial conditions: local argument

(Antonsen et al., 1996). No pattern possible.

How to reconcile? Try to do analytically as far as feasible, for

our map with small .


Discover what large-scale eigenfunction looks like in

Lagrangian coordinates (hint: they are not eigenfunctions!).



Why do this? The two viewpoints are a priori unrelated,

because they for these highly-chaotic systems they are
connected by an extremely convoluted (i.e., inaccessible)
transformation!



Why do this? The two viewpoints are a priori unrelated,

because they for these highly-chaotic systems they are
connected by an extremely convoluted (i.e., inaccessible)
transformation!
But must give same answer for a scalar quantity like the
decay rate.
Advection and Diffusion: Eulerian to Lagrangian

Advection-diffusion (AD) equation:
t + v x =
e x2 .

t + v x =
e x2 .
We define Lagrangian coordinates X by

x = v(x, t),
x(0) = X.

t + v x =
e x2 .
We define Lagrangian coordinates X by

x = v(x, t),
x(0) = X.
Transform AD equation to Lagrangian coordinates,

= X (D X ).
Anisotropic diffusion tensor, in terms of metric or CauchyGreen

strain tensor:
X xi xi
D :=
e g 1 ;
gpq :=
.
p
q
X X
i
From Flow to Map

Velocity field doesnt enter the Lagrangian equation directly:
regard the time dependence in D as given by map rather than flow.
The solution of the AD equation in Fourier space is then
X
(i1)
(i)
(i)
exp G k` `
,
k =
`
where i denotes the ith iterate of the map, and

Z
(i)
Gk` = 4 2 T
(k D(i) `) e2i(k`)X d2 X .
T2
From Flow to Map

Velocity field doesnt enter the Lagrangian equation directly:
regard the time dependence in D as given by map rather than flow.
The solution of the AD equation in Fourier space is then
X
(i1)
(i)
(i)
exp G k` `
,
k =
`
where i denotes the ith iterate of the map, and

Z
(i)
Gk` = 4 2 T
(k D(i) `) e2i(k`)X d2 X .
T2
This is an exact result, but the great difficulty lies in calculating

the exponential of G (i) . We shall accomplish this perturbatively.
Back to the Beginning
M=
M(x) = M x + (x),
!
!
sin 2x1
2 1
;
(x) =
;
2 sin 2x1
1 1
The eigenvalues of M are
1
u = = 2 (3 + 5) = cot2 ,
s =
1
2 (3
5) = tan2
and the corresponding eigenvectors,

(
u s) =
cos sin
sin cos
!
Back to the Beginning
M=
M(x) = M x + (x),
!
!
sin 2x1
2 1
;
(x) =
;
2 sin 2x1
1 1
The eigenvalues of M are
1
u = = 2 (3 + 5) = cot2 ,
s =
PSfrag
replacements
and the corresponding
eigenvectors,
(
u s) =
cos sin
sin cos
1
2 (3
5) = tan2
Contract
Stretch
Coefficients of Expansion: Perturbation Theory

The coefficients of expansion and characteristic directions for the
linear cat map are uniform in space. Perturb off this.
To leading order in , the coefficient of expansion is written as
(i)
= i (1 + (i) )
where is the coefficient of expansion for the unperturbed cat

map; the perturbed eigenvectors are similarly written
(i)
+
=u
(i)
s ,
(i)
s
.
= s (i) u
Coefficients of Expansion: Perturbation Theory

The coefficients of expansion and characteristic directions for the
linear cat map are uniform in space. Perturb off this.
To leading order in , the coefficient of expansion is written as
(i)
= i (1 + (i) )
where is the coefficient of expansion for the unperturbed cat

map; the perturbed eigenvectors are similarly written
(i)
+
=u
(i)
s ,
(i)
s
.
= s (i) u
Simple application of matrix perturbation theory to Jacobian

matrix of the map. The symmetrised Jacobian is the metric:
(i)
g
(i) 2 (i) (i)

u
[ ] u
(i) 2 (i) (i)

+ [ ] s s .
Perturbation Results
(i)
= i (1 + (i) ),
(i)
1
2
sin 2
i1
X
j=0
(i)
(i)
1
2 (cos 2
(i)
= u
+ (i) s ,
u
j
1)
j=0
cos 2(M X)1 ;
1
(i)
(i)
= 2i
(
+
+
),
2i

i1
X
skip
2(ij)
cos 2(M X)1 .
Observe that the perturbation to the eigenvectors converges

exponentially, as required.
Perturbed Metric Tensor
(i)
(i) 1
[g ] ;
(i) 1
[g ]
(i) 2 (i) (i)

[ ] s s
skip
(i) 2 (i) (i)

u
.
+ [ ] u
To leading order in , we have

(i)
u
+ 2 (i) (2i s s 2i u
u
)
[g ]1 = 2i s s + 2i u

(i)
2i
2i
),

(
u s + s u
where the only functions of X are (i) and (i) .
Perturbed Metric Tensor
(i)
(i) 1
[g ] ;
(i) 1
[g ]
(i) 2 (i) (i)

[ ] s s
skip
(i) 2 (i) (i)

u
.
+ [ ] u
To leading order in , we have

(i)
u
+ 2 (i) (2i s s 2i u
u
)
[g ]1 = 2i s s + 2i u

(i)
2i
2i
),

(
u s + s u
where the only functions of X are (i) and (i) .
Recall the solution to the AD equation:

X
(i1)
(i)
(i)
exp G k` `
.
k =
`
The Exponent G (i)

(i)
Gk` = 4 2 T
=
skip
(k D(i) `) e2i(k`)X d2 X
T2
(i)
(i)
Ak` + Bk`
The Exponent G (i)

(i)
Gk` = 4 2 T
=
skip
(k D(i) `) e2i(k`)X d2 X
T2
(i)
(i)
Ak` + Bk`
where
(i)
Ak`
(i)
Bk`
=

2i
2i
ks2
= 2 ks `s
2i
2i
ku2
k` ,
(i)
:= 4 2
eT
ku ù k`
(i)
(ku `s + ks ù ) (+
), ks := (k s).
with ku := (k u
(i)
k` +
k` )
The Exponent G (i) = A(i) + B (i) (contd)
skip
The diagonal part, A(i) , inexorably leads to superexponential

decay of variance, because it grows exponentially.
Upon making use of the Fourier-transformed (i) and (i) , we find
(i)
Bk`
ij
Bk`
12
2i
i1
X
j=0
ij
Bk`
= sin 2 ks `s
k,`+e1 Mj + k,è1 Mj
2i
ku ` u

+ (ku `s + ks ù ) 2(ij) sin2 2(ij) cos2 .
So B (i) is not diagonal (it couples different modes to each other).

= Dispersive in Fourier space.
But can we Compute the Exponential, exp(G (i) )?

To leading order in , for A diagonal, we have G (i) = A(i) + B (i) ,
(i)
(i)
(i)
[exp(A +B )]k` = e
(i)
Akk
(i)
k` +Ek` ;
(i)
(i)
Ek` = Bk`
eAkk
(i)
Akk
(i)
eA``
(i)
A``
From Eulerian considerations, we know we must avoid

superexponential decay of (i) for long times.

(i)
(i)
(i)
[exp(A +B )]k` = e
(i)
Akk
(i)
k` +Ek` ;
(i)
(i)
Ek` = Bk`
eAkk
(i)
Akk
(i)
eA``
(i)
A``

However, the 2i term in Akk precludes any optimism about

the situation: it dooms us to a grim superexponential death.
(i)

(i)
(i)
(i)
[exp(A +B )]k` = e
(i)
Akk
(i)
k` +Ek` ;
(i)
(i)
Ek` = Bk`
eAkk
(i)
Akk
(i)
eA``
(i)
A``

However, the 2i term in Akk precludes any optimism about

the situation: it dooms us to a grim superexponential death.
For = 0, this is indeed what happens. But for a finite value

of , the E term breaks the diagonality of G , so that given
some initial set of wavevectors, the variance contained in
those modes can be transferred elsewhere.
(i)
A Few Words about Numerics
skip
Impractical to take the matrix exponential for large matrices.
skip
Perturbative expansion sidesteps this problem.
skip
However, still need to go to extremely high wavenumber

. . . impossible to use mesh, since would have to refine
exponentially fast.
skip

exponentially fast.
So keep track of only the required wavevectors: their number

should grow exponentially . . . but it doesnt!
skip

exponentially fast.

This is because as i increases,

most modes are damped as

exp 2i ks2 + 2i ku2 , except for those that have very
.
small ks = (k s), i.e., those that are aligned with u
skip

exponentially fast.



.
Just let computer take care of pruning via underflow!
skip

exponentially fast.



.
Just let computer take care of pruning via underflow!
The surviving modes need to become more and more aligned

as time goes on.
with u
Comparison: Eulerian and Lagrangian Views
10
= 104
50
variance
ag replacements
10
10
100
=0.1
=0.01
=0.001
=0.0001
Lagrangian
10
iteration
15
20
Convergence
|Eulerian Lagrangian|
10
ag replacements
10
10
skip
iteration = 4
= 0.01
2
10
10
10
10
10
10
10
10
10
Rescaled Pattern for i = 6, . . . , 12
log10 amplitude (rescaled)
10
15
ag replacements
= 104
= 0.1
20
25
0
0.1
0.2
0.3
0.4
0.5
, scaled by i )
k (aligned with u
Conclusions
In the Eulerian view, large-scale eigenmode dominates

exponential phase, as for bakers map.
Conclusions

Global structure matters!
Conclusions

It is not possible to simply transform the Eulerian result to

Lagrangian coordinates, since orbits are chaotic . . . must
solve Lagrangian problem from the start.
Conclusions


There exists a kind of pattern in Lagrangian coordinates (not

eigenfunction) that is cascading to large wavenumbers.
Conclusions



Pattern confined to dominant mode in Eulerian coordinates,

but dispersed in Lagrangian space.
Conclusions




Could the numerical economy be scaled to more complex

problems?
Conclusions




Could the numerical economy be scaled to more complex

problems?
Still some kinks to iron out!


Ucl 2004

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Ucl 2004

Загружено:

Авторское право:

Доступные форматы

Local and Global Aspects of Mixing

Local and Global Aspects of Mixing p.1/30

Experiment of Rothstein et al.: Persistent Pattern

Local and Global Aspects of Mixing p.2/30

Eventually settles into pattern (eigenfunction)

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

Local and Global Aspects of Mixing p.4/30

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

[Antonsen et al., Phys. Fluids (1996)]

[Balkovsky and Fouxon, PRE (1999)]

Average over angles

Local and Global Aspects of Mixing p.4/30

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

[Antonsen et al., Phys. Fluids (1996)]

[Balkovsky and Fouxon, PRE (1999)]

Average over angles

Eigenfunction of advectiondiffusion operator.

Local and Global Aspects of Mixing p.4/30

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

[Antonsen et al., Phys. Fluids (1996)]

[Balkovsky and Fouxon, PRE (1999)]

Average over angles

Eigenfunction of advectiondiffusion operator.

[Pierrehumbert, Chaos Sol. Frac. (1994)]

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

[Antonsen et al., Phys. Fluids (1996)]

[Balkovsky and Fouxon, PRE (1999)]

Average over angles

Eigenfunction of advectiondiffusion operator.

So far, local theories are Lagrangian and global theories are

Local and Global Aspects of Mixing p.4/30

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

[Antonsen et al., Phys. Fluids (1996)]

[Balkovsky and Fouxon, PRE (1999)]

Average over angles

Eigenfunction of advectiondiffusion operator.

So far, local theories are Lagrangian and global theories are

Today: Try to connect the two pictures.

Local and Global Aspects of Mixing p.4/30

Local vs Global Regimes of Mixing

Based on distribution of Lyapunov exponents.

[Antonsen et al., Phys. Fluids (1996)]

[Balkovsky and Fouxon, PRE (1999)]

Average over angles

Eigenfunction of advectiondiffusion operator.

So far, local theories are Lagrangian and global theories are

Today: Try to connect the two pictures.

Cannot often do this! Map allows (mostly) analytical results.

Local and Global Aspects of Mixing p.5/30

Local and Global Aspects of Mixing p.5/30

Local and Global Aspects of Mixing p.5/30

The people responsible for the confusion. . .

Local and Global Aspects of Mixing p.6/30

The people responsible for the confusion. . .

(See footnote in Truesdell, The Kinematics of Vorticity.)

M x is the Arnold cat map.

The map M is area-preserving and chaotic.