. . A statistical mechanics approach to stochastic parametrizations J. Wouters V. Lucarini eoretical Meteorology, University of Hamburg October 31, 2013 J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 1 / 19
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
.
......
A statistical mechanics approach to stochasticparametrizations
J. Wouters V. Lucarini
eoretical Meteorology, University of Hamburg
October 31, 2013
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 1 / 19
What is parametrization?
Parametrization = deriving reduced models
e climate system has variability over a large range of space and timescales¹
10
102
103
104
km
hour day week month season year
convectionthunderstorms
meso-scalevariability
MCC
synoptic-scalevariabilitytraveling
cyclones
low-frequencyvariability (LFV)*
ENSO
blockingpersistentanomalies
horiz
onta
l sca
le
time scale
¹ECMWF 2005 workshop on stochastic-dynamic models, based on Fraedrich (1978) JASJ. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 2 / 19
What is stochastic parametrization?
Climate models have a coarse resolution (∼ 300 km)
Unresolved processes (clouds, convection, precip.) are effected by and effectresolved dynamics.
We need to represent unresolved processes in resolved dynamics
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 3 / 19
Similarities to statistical mechanics
Reduction of degrees of freedom is a central task of statistical mechanics,e.g. Brownian motion..Hasselman (1976), “Stochastic climate models”, Tellus..
......
“e essential feature of stochastic climate models is that the non-averaged“weather” components are also retained. ey appear formally as randomforcing terms. e climate system, acting as an integrator of thisshort-period excitation, exhibits the same random-walk responsecharacteristics as large particles interacting with an ensemble of muchsmaller particles in the analogous Brownian motion problem.”
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 4 / 19
Multiscale methodsApplies to dynamical systems with a large time scale separation..Averaging ∼ LLN..
......
{x = f(x, y)
y = 1ϵ g(x, y) +
1√ϵβ(x, y)V
For ϵ << 1 and t ∈ [0, T], x(t) → sol. of X = F(X), F(X) = ⟨f(X, y)⟩ρy|X
.Homogenisation ∼ CLT
..
......
{x = 1
ϵ f0(x, y) + f1(x, y) + α(x, y)U
y = 1ϵ2g(x, y) + 1
ϵβ(x, y)V
For ϵ << 1 and t ∈ [0, T], weak convergence of x to the solutions of areduced SDE with Gauss. white noise.
Khasminskii, Kurtz, Papanicolaou,… (60s, 70s), see also Pavliotis & Stuart (2008)
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 5 / 19
Stochastic averaging: applications
Several low dimensional models:Monahan & Culina(2011) J. of Clim.From triad to barotropic flows:Majda, Timofeyev, Vanden-Eijnden, Franzke(1999) PNAS, (2001) Comm. Pure App. Math., (2003) J. Atmos. Sci.,(2005) J. Atmos. Sci.2D zonal jets:Bouchet, Nardini, Tangarife(2013) J Stat PhysEnergy conserving multi-scale toy model of the atmosphere:Frank & Gowald(2013) Physica D
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 6 / 19
Mori-Zwanzig: formal derivationNon-linear ODE:
z = f(z) z = (x, y)
Linear PDE (Liouville equation):
u(t, z) = Lu(t, z) u(0, z) = A(z) L = f.∇
u(t, z) = LetLu(0, z) = etL(P+Q)LA(z)
P a projector on a space of x-observables, Q = 1− P
Using a Duhamel-Dyson decomposition:
etL = etQL +
∫ t
0dτe(t−τ)LPLeτQL
xi(t, x) = etLPLxi + etQLQLxi +∫ t
0dτe(t−τ)LPLeτQLQLxi
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 7 / 19
Mori-Zwanzig projection operators
xi(t, x) = etLPLxi + etQLQLxi +∫ t
0dτe(t−τ)LPLeτQLQLxi
term 1: Markovian partterm 2: fluctuating part: obeys orthogonal dynamics QLterm 3: memory term
Note:
Formal rewriting of the equationsDemonstrates the presence of correlations and memory in reducedsystemsFurther approximation is needed for application
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 8 / 19
Mori-Zwanzig: aproximations
Short memory: white noise, no memoryt-model: etQL → etL in the memory term²
Applications:
Burgers equation: Bernstein (2007) Mult. Mod. Sim.Euler equation: Hald & Stinis (2007) PNASKuramoto-Sivashinski equation: Stinis (2004) Mult. Mod. Sim.
²Chorin, Hald, Kupferman (2002) Phys. DChorin & Hald “Stochastic Tools in Mathematics and Science” (2013) Springer
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 9 / 19
Weakly coupled systems and response
Dynamical system
{x = fx(x) + ϵψx(y)
y = fy(y) + ϵψy(x)(1)
Linear responsez = f(z) + ϵδf(z)
δρ(1)(A) =∂⟨A⟩ρϵ∂ϵ
=
∫ ∞
0dτ⟨δf(x)∇eτLA⟩ρ0 =
∫ ∞
0dτRA,δf(τ)
Can one find a dyn. syst. x such that
⟨Ax⟩ρ = ⟨Ax⟩ρϵ + O(ϵn)
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 10 / 19
Weak coupling: 1st order response
Calculate response δρ(1)(Ax) to coupling Ψ at first order
δ(1)ρ =
∫ ∞
0dτ ⟨⟨Ψx(y)⟩ρy∇eτLAx⟩ρx
=
∫ ∞
0dτ ⟨M∇eτLAx⟩ρx
Consider a reduced system˙x = fx(x) + ϵM
then ρ(Ax) = ρ(Ax) + O(ϵ2)J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 11 / 19
Weak coupling: 2nd order (1/2)
Second order response (1/2):
[δρ(2)
]1∼ ⟨δΨx(y)δΨx(fs(y))⟩ρy = CΨx;Ψx(s)
parametrized by correlated noise
˙x =fx(x) + ϵM+ ϵσ(t)
⟨σ(t)⟩ = 0
⟨σ(t)σ(t+ s)⟩ = CΨx;Ψx(s)
Higher order moments are determined by higher order repsonseJ. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 12 / 19
Weak coupling: 2nd order (2/2)
Second order response (2/2):
[δρ(2)
]2∼ ⟨Ψy(x).∇Ψx(fs(y))⟩ρY = RΨx,Ψy(x)(s)
parametrized by a memory term
˙x =fx(x) + ϵM+ ϵσ(t) + ϵ2∫ ∞
0dτ RΨy,Ψx(x(t−τ))(s)
then ρ(Ax) = ρ(Ax) + O(ϵ3)
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 13 / 19
Numerical experiment (with S. Dolaptchiev & U. Achatz)
Stochastically perturbed additive triadFast Ornstein-Uhlenbeck process (y) coupled to a slow variable (x) bynonlinear coupling
dxdt = B0y1y2dy1dt = B1y2x− γ1
ϵ y1 +σ1√ϵU
dy2dt = B2xy1 − γ2
ϵ y2 +σ2√ϵV
→
dxdτ = ϵB0y1y2dy1dτ = ϵB1y2x− γ1y1 + σ1Udy2dτ = ϵB2xy1 − γ2y2 + σ2V
τ = tϵ
Homogenisation results in an Ornstein-Uhlenbeck process
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 14 / 19
Numerical experiment: weak coupling
e parametrization can be explicitly computed
CΨx;Ψx(s) = ⟨Ψx(y(t))Ψx(y(t+ s))⟩= (ϵB(0))2⟨y1(t)y2(t)y1(t+ s)y2(t+ s)⟩∼ e−(γ1+γ2)s
RΨx;Ψy(s) = ⟨Ψy(x(t), y(t))∇yΨx(y(t+ s))⟩= ⟨ϵB(1)y2x∂y1(ϵB(0)y1(t+ s)y2(t+ s))⟩+ (1 ↔ 2)
∼ e−(γ1+γ2)s
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 15 / 19
Numerical experiment: results
Autocorrelation of x for ϵ = 0.125
0 2 4 6 8 10 12 14 160
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1normalized ACF, eps=0.125
full
weak
hom
t
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 16 / 19
Numerical experiment: results
Autocorrelation of x for ϵ = 0.25
0 1 2 3 4 5 6 7 8�0.2
0
0.2
0.4
0.6
0.8
1
1.2normalized ACF, eps=0.25
full
weak
hom
t
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 17 / 19
Numerical experiment: results
Autocorrelation of x for ϵ = 0.5
0 0.5 1 1.5 2 2.5 3 3.5 4�0.2
0
0.2
0.4
0.6
0.8
1
1.2normalized ACF, eps=0.5
full
weak
hom
t
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 18 / 19
Conclusions
In reduced models, correlated noise and memory appearCorrelation function and memory kernel can be determined in theuncoupled Y system.Preliminary numerical experiments show encouraging results.
Response: Wouters, Lucarini (2012) J. Stat. Mech.Mori-Zwanzig: Wouters, Lucarini (2013) J. Stat. Phys.
ank you!
J. Wouters, V. Lucarini Stochastic parametrizations October 31, 2013 19 / 19
Related Documents
