ME 252 B

Computational Fluid Dynamics:

Wavelet transforms and their applications to turbulence

Marie Farge1 & Kai Schneider2

Winter 2004

University of California, Santa Barbara

1 2
LMD-CNRS, Ecole Normale Supérieure CMI, Université de Provence
24 rue Lhomond 39 rue Joliot-Curie
75231 Paris Cedex 05, France 13453 Marseille Cedex 13, France
Email : farge@lmd.ens.fr Email : kschneid@cmi.univ-mrs.fr
 Classification of signals (1d)

s : t → s(t) with t ∈ IR, s(t) ∈ IR or CI

1) Continuous ←→ discrete

s(t), t ∈ IR s(tn ), n ∈ ZZ

s(t) s(t n )

t ∆t tn

ME 252
ME 252 B,
B, Wavelet
Wavelet transforms
transforms and
and their
their applications
applications to
to turbulence,
turbulence, Marie
Marie Farge
Farge &
& Kai
Kai Schneider,
Schneider, UCSB,
UCSB, Winter
Winter 2004
2004
2) Nonperiodic ←→ periodic

• continuous

Period T , s(t) = s(t + nT ), n ∈ ZZ

• discrete

Period T , s(tn ) = s(tn + nT ),

n ∈ ZZ with T = N ∆T
ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
3) Compact support
s(t) 6= 0 for t ∈ [A, B] and s(t) = 0 else

Α Β

4) Signals with finite energy

• continuous, nonperiodic
Z ∞
E= |s(t)|2 dt < ∞
−∞

• continuous, periodic
Z T
E= |s(t)|2 dt < ∞
0
 • discrete, nonperiodic
X
∞
E= |s(tn )|2 < ∞
n=−∞

• discrete, periodic
NX
−1
E= |s(tn )|2 < ∞
n=0

For mathematicians: spaces of square-integrable functions

(norm + scalar product)

s(t) ∈ L2 (IR), L2 (TT ), l2 (IR), l2 (TT ) where TT = IR/ZZ

ME
ME 252
252 B,
B, Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationsto
toturbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
5) Absolutely integrable signals
Z ∞
S= |s(t)|dt < ∞ s(t) ∈ L1 (IR)
−∞

Classification of signals (higher dimensions)

2d −→ images s(~
x) = s(x, y), x, y ∈ IR or s(m, n), n, m ∈ ZZ

3d, n-d
scalar-valued ←→ vector-valued signals

- temperature - velocity
- pressure - RGB signal

=⇒ similar classification possible.

ME 252 B, Wavelet transforms and their applications to turbulence, Marie Farge & Kai Schneider, UCSB, Winter 2004
ME 252 B, Wavelet transforms and their applications to turbulence, Marie Farge & Kai Schneider, UCSB, Winter 2004
 The Fourier transform
Motivation

• representation of signals with sine and cosine functions

• transformation of signals into the frequency plane

• fast algorithms (FFT), N log2 N complexity

• correlation and convolution can be efficiently computed in the

frequency domain

• system theory:
sine and cosine are eigenfunctions of linear time-shift invariant
systems
ME 252
ME 252 B,
B, Wavelet
Wavelet transforms
transforms and
and their
their applications
applications to
to turbulence,
turbulence, Marie
Marie Farge
Farge &
& Kai
Kai Schneider,
Schneider, UCSB,
UCSB, Winter
Winter 2004
2004
 linear time-shift
u(t) 

ariant system y(t)

u(t) = sin 2πf t y(t) = a sin(2πf t + φ)

cos 2πf t a cos(2πf t + ψ)

For simplification one uses complex exponentials:

eit = cos t + i sin t
Recall complex numbers: I z = x + iy = reiθ
z ∈ C,
x = <z, y = =z

r 2 = x2 + y 2, θ = arctan y/x
Recall trigonometric polynomials:

X
s(t) = ak cos 2πkt + bk sin 2πkt
k≥0
ME
ME252
252B,
B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationsto
toturbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
 Fourier transforms

1) Continuous signals

We consider an absolutely integrable signal s(t) ∈ L1 (IR)(∩L2 (IR)),

t, s ∈ IR

The Fourier transform is defined as:

Z ∞
b )=
S(f s(t)e−i2πf t dt
−∞
Z ∞ Z ∞
= s(t) cos 2πf tdt + i s(t) sin 2πf tdt
−∞ −∞

b ) ∈ C.
Note that in general S(f I

b )| and phase φ = arctan =S(f

Define modulus |S(f b )/<S(f
b )

ME
ME252
252B,
B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
The inverse Fourier transform is defined as:

Z ∞
s(t) = b )ei2πf t df
S(f
−∞

ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
Example:

(
1 for − T ≤ t ≤ T,
s(t) = (1)
0 elsewhere

b )= sin 2πf T
S(f
πf

s(t)

-T T t

ME 252 B, Wavelet transforms and their applications to turbulence, Marie Farge & Kai Schneider, UCSB, Winter 2004
 2) Properties

a) scaling
1 b f
s(at) ⇐⇒ S( ) a ∈ IR, a 6= 0
|a| a

b 1 t
S(af ) ⇐⇒ s( )
|a| a

b) time-shift

s(t − t0 ) ⇐⇒ b )
exp(−i2πf t0 )S(f t0 ∈ IR

c) frequency-shift
b
S(f − f0 ) ⇐⇒ exp(i2πf0 t)s(t) f0 ∈ IR

ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge& &Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
d) differentiation (with respect to time)
If s(t) is n-times continuously differentiable and s(n)(t) ∈ L1 (IR),
then
s(n)(t) ⇐⇒ (i2πf )n S(f
b )

e) differentiation (with respect to frequency)

If tm s(t) ∈ L1 (IR) for m = 0, 1, ..., M , then Sb(m)(f ) exists and

(−i2πt)m s(t) ⇐⇒ Sb(m) (f )

f) multiple application of the Fourier transform

Z ∞
F {s(t)}(f ) = s(t)e−i2πf t dt = S(f
b )
−∞
Z ∞
F 2 {{s(t)}(f )}(t) = F {S(f
b )}(t) = b )e−i2πf t df = s(−t)
S(f
−∞
−→ F 2 corresponds to time inversion
and hence F 4 = Identity
MEME
252 B, B,
252 Wavelet transforms
Wavelet and
transforms their
and applications
their to to
applications turbulence, Marie
turbulence, Farge
Marie & Kai
Farge Schneider,
& Kai UCSB,
Schneider, Winter
UCSB, 2004
Winter 2004
 F 3 = F −1 = F ? (inverse Fourier transform)
Remark: The Fourier transform is a cyclic operator of 4th degree.

g) convolution
given s1(t) and s2(t) with s1(t) ∈ L2 (IR) and s2(t) ∈ L∞ (IR).
Z ∞
s1(t) ? s2(t) = s1(τ )s2 (t − τ )dτ
−∞
? commutes, i.e. s1 ? s2 = s2 ? s1
? is associative, i.e. s1 ? s2 ? s3 = s1 ? (s2 ? s3 ) = (s1 ? s2) ? s3

s1 (t) ⇐⇒ Sb1(f ) and s2 (t) ⇐⇒ Sb2 (f )

s1(t) ? s2(t) ⇐⇒ Sb1(f )Sb2 (f )

h) correlation
- cross-correlation: s1 (t), s2(t) ∈ L2 (IR)
Z ∞
φ12(t) = s1(τ )s2 (t + τ )dτ = s1(t) ? s2 (−t)
−∞
ME252
ME 252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
 Z ∞
φ21(t) = s1(t + τ )s2 (τ )dτ = s1(−t) ? s2(t)
−∞
If Sb1(f ) and Sb2(f ) exist, then
c (f ) = F {φ (t)}(f ) = F {s (t) ? s (−t)}(f )
Φ 12 12 1 2

= F {s1(t) ? F 2{s2 (t)}}(f ) = F {s1(t)}(f )F 3 {s2 (t)}}(f )

= Sb1(f )Sb2? (f )
and analogously
Φ b? (f )S
c (f ) = S b (f )
21 1 2

i) autocorrelation
s1(t) ∈ L2(IR)

Z ∞
φ11(t) = s1(τ )s1 (t + τ )dτ = s1(t) ? s1 (−t)
−∞

ME252
ME 252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
 and with s1(t) ⇐⇒ Sb1(f )
we obtain in frequency space
c (f ) = F {φ (t)}(f ) = S
Φ b (f )S b (f )|2
b? (f ) = |S
11 11 1 1 1

j) multiplication
Z ∞
s1(t)s2 (t) ⇐⇒ Sb1(f ) ? Sb2(f ) = Sb1(ξ)Sb2 (f − ξ)dξ
−∞

k) Parseval’s identity
Z ∞ Z ∞
s1(t)s2 (t)dt = Sb1(f )Sb2 (−f )df
−∞ −∞
Z ∞ Z ∞
−→ s1(t)s?2 (t)dt = Sb1(f )Sb2? (f )df
−∞ −∞
and in particular for s1 = s2 = s ⇐⇒ b ) we have
S(f
Z ∞ Z ∞
|s(t)|2 dt = b )|2 df
|S(f
−∞ −∞
MEME
252252
B, Wavelet transforms
B, Wavelet andand
transforms their applications
their to turbulence,
applications Marie
to turbulence, Farge
Marie & Kai
Farge Schneider,
& Kai UCSB,
Schneider, Winter
UCSB, 2004
Winter 2004
 l) energy spectrum
b 2 R∞
E(f ) = |S(f )| and E = 0 E(f )df
E(f ) is called spectral energy density, or energy spectrum.

m) symmetries
s(t) = seven(t) + sodd (t)
1 (s(t) + s(−t)) and s
with seven(t) = 2 1
odd (t) = 2 (s(t) − s(−t))
Decomposing the corresponding Fourier transform into real and
imaginary part we obtain:
b )=S
S(f b (f ) + iS
b (f )
r i
where Sbr (f ) = <S(f
b ) and S
b (f ) = =S(f
i
b )

seven(t) ⇐⇒ Sbr (f )

sodd (t) ⇐⇒ Sbi(f )

and additionally, we have that Sbr (f ) is even (cosine-transform)
and Sbi(f ) is odd (sine-transform).
ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge& &Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
n) real valued signals
b
If s(t) is real valued, then we have S(−f ) = Sb? (f )

o) regularity
If sn (t) ∈ L1 (IR) then limf →±∞ |(i2πf )n S(f
b )| = 0, i.e.

b ) = O(|f |−n−)
S(f

ME252
ME 252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
Bandwidth of signals and Heisenberg’s uncertainty principle
Z ∞
θ2 = (t − t0 )2|s(t)|2 dt
−∞
Z ∞
B2 = (f − f0 )2|S(f
b )|2 df
−∞
R∞ R∞ b
2
where −∞ |s(t)| dt = −∞ |S(f )|2 df = 1
and t0 and f0 are the center of gravity in the t/f plane, respectively:
Z ∞ Z ∞
t0 = t|s(t)|2 dt f0 = b )|2 df
f |S(f
−∞ −∞
Heisenberg’s uncertainty principle yields:
1
θB ≥
4π

Proof:

w.l.o.g. let t0 = f0 = 0
Using Schwarz inequality
Z b Z b Z b
| g1(t)g2(t)dt|2 ≤ |g1(t)|2 dt |g2(t)|2 dt
−a −a −a
ME 252
ME 252 B,
B, Wavelet
Wavelet transforms
transforms and
and their
their applications
applications to
to turbulence,
turbulence, Marie
Marie Farge
Farge && Kai
Kai Schneider,
Schneider, UCSB,
UCSB, Winter
Winter 2004
2004
for a, b → ∞ and with
g1(t) = ts(t) and g2(t) = ds/dt we have

Z ∞ Z ∞ Z ∞
| ts(t)ds/dtdt|2 ≤ |ts(t)|2 dt |ds/dt|2 dt
−∞ −∞ −∞
As s ∈ L2 (IR), limt→±∞ |s(t)| ≤ C √1
t
Z ∞
1
ts(t)ds/dtdt = −
−∞ 2
and
Z ∞ Z ∞
|ds/dt|2 dt = b )|2 df
|2πf S(f
−∞ −∞
Z Z
1 2
∞
2 2
∞
≤ 4π t |s(t)| dt f 2 |S(f
b )|2 df
4 −∞ −∞

ME252
ME 252B,
B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationsto
toturbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
Distributions

(
∞ for t = t0 ,
δ(t − t0 ) = (2)
0 elsewhere
Z ∞
δ(t − t0 )dt = 1
−∞

Properties:

Z ∞
s(t)δ(t − t0)dt = s(t0 )
−∞
r
n
exp(−nt2 ) −→ δ(t) for n −→ ∞
π
r
n −π 2f 2
exp(−nt2) ⇐⇒ exp( )
π n
ME
ME252
252B,
B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
 δ(t) ⇐⇒ 1

1 ⇐⇒ δ(t)

δ (k)(t) ⇐⇒ (i2πf )k

(−i2πt)k ⇐⇒ δ (k)(f )

Scaling:

1
δ (k)(at) ⇐⇒ (i2πf a)k
|a|

1 (k) t
δ ( ) ⇐⇒ (i2πf a)k
|a| a
and for k = 0
1 t
δ( ) ⇐⇒ 1
|a| a
ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
Shift:
δ (k)(t − t0 ) ⇐⇒ exp(−i2πf t0 )(i2πf )k

exp(i2πf0 t)(−i2πt)k ⇐⇒ δ (k)(f − f0 )

and for k = 0
δ(t − t0) ⇐⇒ exp(−i2πf t0 )

exp(i2πf0t) ⇐⇒ δ(f − f0 )

1
sin(2πf0 t) ⇐⇒ (δ(f − f0 ) − δ(f + f0 ))
2i
1
cos(2πf0 t) ⇐⇒ (δ(f − f0 ) + δ(f + f0 ))
2

ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
Convolution:
s(t) ⇐⇒ b )
S(f

δ (k)(t) ? s(t) ⇐⇒ (i2πf a)k S(f

b )

and for k = 0

δ(t) ? s(t) = s(t)

δ(t − t0) ? s(t) = s(t − t0 )

Sampling theorem: Let s(t) ∈ L1(IR) with S(f

b ) = 0 for |f | > f .
c
Then we have
∞
X sin π(t − nT )/T 1
s(t) = s(nT ) for T ≤
n=0 π(t − nT )/T 2fc

ME
ME252 B,B,
252 Wavelet transforms
Wavelet and
transforms their
and applications
their to to
applications turbulence, Marie
turbulence, Farge
Marie &&
Farge Kai
KaiSchneider, UCSB,
Schneider, Winter
UCSB, 2004
Winter 2004
 2) Periodic signals

Periodic signal s(t) = s(t + nT ), t ∈ IR, n ∈ ZZ with period T ,

b b 1 RT
Discrete Fourier coefficients Sk , k ∈ ZZ with Sk = T 0 s(t)e−i2πkt dt

P b i2πkt/T
and s(t) = k∈ZZ Sk e

3) Discrete signals

Discrete signal sn, n ∈ ZZ

Periodic Fourier transform

X
b )=
S(f sn e−i2πnf
n∈ZZ

4) Discrete periodic signals

Discrete periodic signal sn , 0 ≤ n ≤ N − 1 with sn = sn+mN , m ∈ ZZ

MEME
252 B, B,
252 Wavelet transforms
Wavelet and
transforms their
and applications
their to to
applications turbulence, Marie
turbulence, Farge
Marie & Kai
Farge Schneider,
& Kai UCSB,
Schneider, Winter
UCSB, 2004
Winter 2004
Periodic discrete Fourier transform
NX
−1
1
Sbk = sne−i2πkn/N , 0 ≤ k ≤ N − 1
N n=0

where Sbk = Sbk+mN , m ∈ ZZ

ME
ME252 B,B,
252 Wavelet transforms
Wavelet and
transforms their
and applications
their to to
applications turbulence, Marie
turbulence, Farge
Marie &&
Farge Kai
KaiSchneider, UCSB,
Schneider, Winter
UCSB, 2004
Winter 2004
 5) Summary

b ), f ∈
• Continuous signal s(t), t ∈ IR ←→ continuous spectrum, S(f
IR

• Periodic signal s(t), t ∈ TT ←→ discrete spectrum, Sbk , k ∈ ZZ

b ), f ∈ TT
• Discrete signal sn , n ∈ ZZ ←→ periodic spectrum, S(f

• Discrete periodic signal sn, 0 ≤ n ≤ N − 1 with sn = sn+mN , m ∈ ZZ

←→ periodic discrete spectrum Sbk , 0 ≤ k ≤ N − 1 and with Sbk =
Sbk+mN , m ∈ ZZ

Extention to higher dimensions: tensor product ansatz

ME
ME 252
252 B,
B, Wavelet
Wavelet transforms
transforms and
and their
their applications
applications to
to turbulence,
turbulence, Marie
Marie Farge
Farge &
& Kai
Kai Schneider,
Schneider, UCSB,
UCSB, Winter
Winter 2004
2004
 Ad Fourier transform

Continuous signals: s(t), t ∈ IR

Z ∞ Z ∞
b )=
S(f s(t)e−i2πf t dt, f ∈ IR and s(t) = b )ei2πf t df
S(f
−∞ −∞

Periodic signals (continuous): s̃(t) = s̃(t + mT ), m ∈ ZZ

Z
1 T X
b
Sk = s(t)e−i2πkt/T and s(t) = Sbk ei2πkt/T
T 0 k∈ZZ
with
k
ei2πkt/T ⇐⇒ δ(f − )
T
X k
b )=
S(f b
Sk δ(f − )
k∈ZZ T

ME
ME252
252B,
B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004
Dirac pulse:

∞
X ∞
1 X
δ(t − kT ) = δ(f − k/T )
k=−∞ T k=−∞

Periodisation
∞
X
s̃(t) = s(t − nT )
n=−∞

s̃(t) = s̃(t + mT ), m ∈ ZZ
X k
s̃(t) ⇐⇒ b
Sk δ(f − )
k∈ZZ
T

with Sbk = T1 S(k/T

b )

ME
ME 252
252 B,
B, Wavelet
Wavelet transforms
transforms and
and their
their applications
applications to
to turbulence,
turbulence, Marie
Marie Farge
Farge && Kai
Kai Schneider,
Schneider, UCSB,
UCSB, Winter
Winter 2004
2004
Discrete signals: sn, n ∈ ZZ

X
b )=
S(f sn e−i2πnf
n∈ZZ

Discrete periodic signals: sn, n = 0, ..., N − 1

X
Sbk = sne−i2πkn/N
n∈ZZ

ME 252
ME 252 B,
B, Wavelet
Wavelet transforms
transforms and
and their
their applications
applications to
to turbulence,
turbulence, Marie
Marie Farge
Farge && Kai
Kai Schneider,
Schneider, UCSB,
UCSB, Winter
Winter 2004
2004
Sampled signals:

X
ssamp(t) = T s(kT )δ(t − kT )
k∈ZZ

1
fsamp = 2flimit =
T

∞
X
ssamp(t) ⇐⇒ Sbsamp(f ) = b
S(f − kfsamp )
k=−∞

ME
ME252
252B,B,Wavelet
Wavelettransforms
transformsand
andtheir
theirapplications
applicationstototurbulence,
turbulence,Marie
MarieFarge
Farge&&Kai
KaiSchneider,
Schneider,UCSB,
UCSB,Winter
Winter2004
2004