probabilistic perspectives on climate dynamics · probabilistic perspectives on climate dynamics...

Probabilistic Perspectives on ClimateDynamics

Adam Monahanmonahana@uvic.ca

School of Earth and Ocean SciencesUniversity of Victoria

Probabilistic Perspectives on Climate Dynamics – p. 1/63

Probability and Climate

“Climate is what you expect,

“Climate is what you expect, weather is what you get.”- Robert A. Heinlein

⇒ “expectation” lies at heart of notion of climate

⇒ this is a fundamentally probabilistic perspective

The ultimate goal of climate physics is the “measure” of the climatesystem, with emphasis on

characterisation

physical understanding

predictability(evolution of trajectory and response of measure to forcing)

characterisation

Toward this end, a natural language of climate physics is

dynamical systems

stochastic processes

random dynamical systems (where they meet)

Triple goals of probabilistic climate dynamics:

1. characterise distributions of climate states

2. understand how these distributions arise from the underlyingphysics

3. use pdfs to improve forecasts/predictions

dynamical systems

Overview

Why is climate complex?

Origins of stochasticity in the climate system.

Case Study I: Stochastic dynamics of sea-surface winds.

Case Study II: Stochastic dynamics of El Nino/Southern Oscillation.

Overview

Why is Climate Complex? Coupling Across Scales

From von Storch and Zwiers, 1999Probabilistic Perspectives on Climate Dynamics – p. 6/63

From von Storch and Zwiers, 1999

Generic dynamical equation for climate state

dt= Lz + N(z, z) + F

Decompose state into “slow” and “fast” variables x and y

⇒ coupled dynamics

dt= Lxxx + Lxyy +N (x)

xx (x,x) +N (x)xy (x,y) +N (x)

yy (y,y) + Fx

dt= Lyxx + Lyyy +N (y)

xx (x,x) +N (y)xy (x,y) +N (y)

yy (y,y) + Fy

Averaging to project on “slow” dynamics retains “upscale” influenceof eddies on resolved flow (“closure problem”)

dt= Lxxx +N (x)

xx (x,x) +N(x)yy (y,y) + Fx

yy (y,y) + Fx

yy (y,y) + Fy

dt= Lxxx +N (x)

yy (y,y) + Fx

yy (y,y) + Fy

dt= Lxxx +N (x)

yy (y,y) + Fx

yy (y,y) + Fy

dt= Lxxx +N (x)

yy (y,y) + Fx

yy (y,y) + Fy

dt= Lxxx +N (x)

yy (y,y) + Fx

yy (y,y) + Fy

dt= Lxxx +N (x)

yy (y,y) + Fx

yy (y,y) + Fy

dt= Lxxx +N (x)

Why is Climate Complex? Coupling Across Systems

From IPCC Third Assessment Report

Why is Climate Complex? Feedback Loops

From http://eesc.columbia.edu/courses/ees/slides/climate/

Why is Climate Complex? Non-Stationarity

Why is Climate Complex? Data Surfeit & Paucity

1880 1900 1920 1940 1960 1980 2000

4 (° C

From ERA-40 Project Report Series 19

Origins of Stochasticity: Multiscale Dynamics

Climate system displays variability over broad range of space andtime scales

From Saltzman, 2002Probabilistic Perspectives on Climate Dynamics – p. 13/63

When modelling slower components of system, don’t want (need?)to explicitly simulate faster components

⇒ “subgrid-scale parameterisations” (closure)

Analogy with statistical mechanics; can talk about “pressure” and“temperature” of gas without accounting for state of each molecule

Can also consider separation of scales in space; parameterisationproblem arises because of finite gridsize of operational models

Nonlinear dynamics⇒ fast dynamics has upscale effect on slow

Coarse-graining results not only in unresolved scales, but alsounresolved processes (e.g. internal gravity waves, convection, cloudmircophysics)

State of unresolved processes/scales conditioned on resolved scalesnot unique: adopt distributional perspective on subgrid-scale

Formally decompose climate state z into slow and fast variables x

and y (resp. “climate” and “weather”):

dt= f(x,y, t)

εg(x,y, t)

Obtain effective stochastic dynamics for “climate” as ε→ 0

In many (most?) climate applications, ε is not small

Fast/slow decomposition not unique;“one person’s noise is another person’s signal”

dt= f(x,y, t)

εg(x,y, t)

dt= f(x,y, t)

εg(x,y, t)

dt= f(x,y, t)

εg(x,y, t)

dt= f(x,y, t)

εg(x,y, t)

Origins of Stochasticity: Model Uncertainty

All models of climate (or subsystems) contain both

model error, and

poorly constrained (sometimes unphysical) parameters

Ideally: given pdf of parameters and model structure, obtain pdf ofclimate state

Reality:

full climate state pdf cannot be computed; must adoptMonte-Carlo or ensemble forecast approach in which pdf issampled; “curse of dimensionality”

parameter pdfs generally not well known a priori

Building large ensembles computationally expensive

model error, and

Reality:

model error, and

Reality:

model error, and

Reality:

model error, and

Reality:

model error, and

Reality:

model error, and

Reality:

model error, and

Reality:

Parameter Uncertainty: Ensemble Prediction

climateprediction.net uses idle private CPUs to integrateensembles with different parameter settings

http://www.climateprediction.net

Initial Condition Uncertainty: Ensemble Forecasting

Model uncertainties can also include initial conditions

http://chrs.web.uci.edu/images/ensemble_large_atmo.jpg

Stochastic Climate Models: Case Studies

Two distinct end-member approaches to modelling pdfs of climatevariables:

running fully complex general circulation models

considering physically-motivated idealised models

First approach has benefit of being more realistic, but is also muchmore complex; mechanisms are not always clear

Second approach not always quantitatively accurate, but importantfor developing understanding and elucidating mechanism

Will now consider two “idealised” stochastic models for:

stochastic dynamics of sea-surface winds

stochastic dynamics of El Niño-Southern Oscillation

Sea Surface Winds: Why Should We Care?

Air/Sea Exchange

ocean and atmosphere interact through respective boundarylayers, exchanging momentum, energy, freshwater, and gases

fluxes depend on surface winds, in general nonlinearly

ocean currents largely driven by surface winds

Sea State

sea state important for shipping, recreation

determined by both local and remote winds

Power Generation

wind power potentially significant source of energy

generation rate scales as cube of wind speed; extreme eventsimportant

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Air/Sea Exchange

Sea State

Power Generation

Vector Wind Moments

mean(U) (m/s)

0 50 100 150 200 250 300 350−60

std(U) (m/s)

0 50 100 150 200 250 300 350−60

skew(U)

0 50 100 150 200 250 300 350−60

Zonal Wind

mean(V) (m/s)

0 50 100 150 200 250 300 350−60

std(V) (m/s)

0 50 100 150 200 250 300 350−60

skew(V)

0 50 100 150 200 250 300 350−60

Meridional Wind

Mean and Skewness of Vector Wind

Joint pdfs of mean and skew for zonal and meridional winds

−10 −5 0 5 10

mean(U) (ms−1)

0.010 0.022 0.046 0.100

−5 0 5−1.5

−0.5

mean(V) (ms−1)

0.010 0.022 0.046 0.100 0.215 0.464

(note logarithmic contour scale)

Wind Speed Moments

mean(w) (ms−1)

0 50 100 150 200 250 300 350

std(w) (ms−1)

0 50 100 150 200 250 300 350

skew(w)

0 50 100 150 200 250 300 350

−0.5

Wind Speed pdf: Weibull distribution

The pdf of wind speed w has traditionally (and empirically) beenrepresented by 2-parameter Weibull distribution:

0 5 10 15 20 250

b=2b=3.5b=7

p(w) =b

)b−1exp

[−(wa

a is the scale parameter (pdf centre)

b is the shape parameter (pdf tilt)

pw(w) is unimodal

0 5 10 15 20 250

b=2b=3.5b=7

p(w) =b

)b−1exp

[−(wa

pw(w) is unimodal

0 5 10 15 20 250

b=2b=3.5b=7

p(w) =b

)b−1exp

[−(wa

pw(w) is unimodal

0 5 10 15 20 250

b=2b=3.5b=7

p(w) =b

)b−1exp

[−(wa

pw(w) is unimodal

Wind Speed pdfs: Observed

Observed speed moments fall around Weibull curve

mean(w)/std(w)

1.5 2 2.5 3 3.5 4 4.5 5 5.5

−0.8

−0.6

−0.4

−0.2

10.000

Boundary Layer Dynamics

Horizontal momentum equations:

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Momentum tendency due to:

advection (transport by flow; secondary importance on dailytimescales )

pressure gradient force

Coriolis force

turbulent momentum flux (in vertical)

Integrated momentum budget for slab of thickness h:

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

∂t+ u · ∇u = −1

ρ∇p− f k× u− 1

∂(ρu′u′3)

Coriolis force

dt= −1

ρ∇p− f k× u +

(u′u′3(0)− u′u′3(h)

Surface Wind Stress

Surface wind stress is turbulent momentum flux across air/seainterface:

τs = ρau′u′3(0)

u = along-mean wind component

v = cross-mean wind component

u = (u, v)

u3 = vertical wind component

Flux parameterised in terms of u by bulk drag formula:

τs = ρacdwu

where w =‖ u ‖ is the wind speed.

Surface Wind Stress

u = (u, v)

τs = ρacdwu

Surface Wind Stress

u = (u, v)

τs = ρacdwu

Drag Coefficient

Drag coefficient influenced by

surface roughness z0

Obukhov length L (buoyancy flux, stratification)

cd = k2

)− ψm

)]−2

ψm is an empirical function

za is the anemometer height, typically 10 m

Surface winds modify surface wave field

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Drag Coefficient

cd = k2

)− ψm

)]−2

⇒ z0 depends on w

Neutral Drag Coefficient: Observations

2 52 01 51 0500.5

Large & Pond 1982

Smith 1980

Smith 1988

Yelland et al. 1998

Smith 1992

u10n (m/s)

Taylor et al. (2000) WMO/TD No. 1036Probabilistic Perspectives on Climate Dynamics – p. 29/63

Surface Momentum Budget

To close momentum budget, need parameterisation of turbulentmomentum flux at z = h

Use “finite-differenced” eddy viscosity:

u′u′3(h) =K

h(U− u)

⇒ Surface layer momentum budget

dt= −1

ρ∇p− f k× u− cd

h2(U− u)

= Π− cdhwu− K

whereΠ = −1

ρ∇p− f k× u +

u′u′3(h) =K

h(U− u)

dt= −1

h2(U− u)

= Π− cdhwu− K

whereΠ = −1

ρ∇p− f k× u +

u′u′3(h) =K

h(U− u)

dt= −1

h2(U− u)

= Π− cdhwu− K

whereΠ = −1

ρ∇p− f k× u +

u′u′3(h) =K

h(U− u)

dt= −1

h2(U− u)

= Π− cdhwu− K

whereΠ = −1

ρ∇p− f k× u +

u′u′3(h) =K

h(U− u)

dt= −1

h2(U− u)

= Π− cdhwu− K

whereΠ = −1

ρ∇p− f k× u +

u′u′3(h) =K

h(U− u)

dt= −1

h2(U− u)

= Π− cdhwu− K

whereΠ = −1

ρ∇p− f k× u +

Mechanistic Model: SDE

Decomposing Π into mean and fluctuations:

Πu(t) = 〈Πu〉+ σW1(t)

Πv(t) = σW2(t)

where Wi is Gaussian white noise⟨Wi(t1)Wj(t2)

⟩= δijδ(t1 − t2)

we obtain stochastic differential equation

dt= 〈Πu〉 −

cdhwu− K

h2u+ σW1

dt= −cd

hwv − K

h2v + σW2

Πv(t) = σW2(t)

⟩= δijδ(t1 − t2)

dt= 〈Πu〉 −

cdhwu− K

h2u+ σW1

dt= −cd

hwv − K

h2v + σW2

Πv(t) = σW2(t)

⟩= δijδ(t1 − t2)

dt= 〈Πu〉 −

cdhwu− K

h2u+ σW1

dt= −cd

hwv − K

h2v + σW2

Mechanistic Model: pdf

Solution of associated Fokker-Planck equation for stationary pdf:

puv(u, v) = N1 exp

{〈Πu〉u−

2h2(u2 + v2)

∫ √u2+v2

′)w′2 dw′})

Changing to polar coordinates and integrating over angle gives windspeed pdf:

pw(w) = NwI0

(2 〈Πu〉wσ2

(− 2

2h2w2 +

′)w′2 dw′})

Mechanistic Model: pdf

Solution of associated Fokker-Planck equation for stationary pdf:

puv(u, v) = N1 exp

{〈Πu〉u−

2h2(u2 + v2)

∫ √u2+v2

′)w′2 dw′})

Changing to polar coordinates and integrating over angle gives windspeed pdf:

pw(w) = NwI0

(2 〈Πu〉wσ2

(− 2

2h2w2 +

′)w′2 dw′})

Mechanistic Model: Predictions

s−3/2

<Πu> (ms−2× 10−3)

mean(u) (ms−1)

0 1 2 3 40.05

<Πu> (ms−2× 10−3)

std(u) (ms−1)

0 1 2 3 40.05

−0.4

−0.3

−0.35

−0.3

−0.2

−0.25

−0.2

−0.1

.1−0

<Πu> (ms−2× 10−3)

skew(u)

0 1 2 3 40.05

<Πu> (ms−2× 10−3)

s−3/2

mean(w)

0 2 40.05

1.522.5

<Πu> (ms−2 × 10−3)

std(w)

0 2 40.05

−0.2

−0.1

0.10.2

<Πu> (ms−2) × 10−3

skew(w)

0 2 40.05

Mechanistic Model: Predictions

s−3/2

<Πu> (ms−2× 10−3)

mean(u) (ms−1)

0 1 2 3 40.05

<Πu> (ms−2× 10−3)

std(u) (ms−1)

0 1 2 3 40.05

−0.4

−0.3

−0.35

−0.3

−0.2

−0.25

−0.2

−0.1

.1−0

<Πu> (ms−2× 10−3)

skew(u)

0 1 2 3 40.05

<Πu> (ms−2× 10−3)

s−3/2

mean(w)

0 2 40.05

1.522.5

<Πu> (ms−2 × 10−3)

std(w)

0 2 40.05

−0.2

−0.1

0.10.2

<Πu> (ms−2) × 10−3

skew(w)

0 2 40.05

Mechanistic Model: Comparison with Observations

mean(w)/std(w)

1.5 2 2.5 3 3.5 4 4.5 5 5.5

−0.8

−0.6

−0.4

−0.2

10.000ObservedWeibullModel

Case Study I: Conclusions

Sea surface wind pdfs characterised by relationships betweenmoments

These moment relationships reflect physical processes producingdistributions

Idealised stochastic models can be constructed from basic physicalprinciples to (qualitatively) explain physical origin of pdf structure

More accurate quantitative simulation requires a more sophisticatedmodel; qualitative utility of relatively simple model suggests itcaptures essential physics

El Niño - Southern Oscillation (ENSO)

ENSO is the dominant mode of climate variability on interannualtimescales, involving coupled interactions between the ocean and theatmosphere

Dynamics primarily contained in equatorial Pacific; impacts feltglobally

Skillful ENSO forecasts are believed to be primary potential sourceof skill for seasonal climate forecasting

Tropical Pacific: Mean State

From www.pmel.noaa.gov/tao/elnino/nino-home.html

Tropical Pacific: El Niño State

Tropical Pacific: La Niña State

ENSO Indices: SOI and East Pacific SST

El Niño Growth: Bjerknes’ Hypothesis

E. PacificThermoclineDepth

Temperature ofUpwelled Water

E. Pacific SST

E. Pacific SLP

(Rising Air, Precip.)

Strength ofTrade Winds

SEC andUpwelling

Bjerknes’ Hypothesis

(+)(+)

ENSO Cycles: Delayed Oscillator Mechanism

From Chang and Battisti, Physics World, 1998.

ENSO Irregularity: Stochastic Oscillator Mechanism

El Niño Impacts: Northern Hemisphere Winter

From iri.columbia.edu/climate/ENSO/

ENSO Impacts: Changes in Mean Climate

Adapted from Sardeshmukh et al., J. Clim (2000)

ENSO Impacts: Changes in Climate Variability

From Sardeshmukh et al., J. Clim (2000)

ENSO Impacts: Extreme Events

From Bove et al., J. Clim (1998)Probabilistic Perspectives on Climate Dynamics – p. 47/63

ENSO Dynamics

As a first approximation, ENSO dynamics modelled as stable lineardynamical system driven by noise:

dt= Ax +B(x) ◦ W

wherex is the state vector (e.g. sea surface temperature field)

A is the linearised dynamical operator

B(x) is the noise amplitude matrix

W is a vector of independent white noise processes

For simplicity, will assume that B is state-independent (additivenoise)

ENSO Dynamics

dt= Ax +B(x) ◦ W

ENSO Dynamics

dt= Ax +B(x) ◦ W

ENSO Dynamics

Analytic solution to SDE:

x(t) = P (t)x(0) +

0P (t− t′)BW(t′)dt′

whereP (t) = exp(At)

Evolution of moments:

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

Assuming max(Re(eig(A)))< 0, stationary covariance

C =⟨xxT

satisfies fluctuation-dissipation relationship

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

P (t) = exp(At)

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

ENSO Dynamics

x(t) = P (t)x(0) +

〈x(t)〉 = P (t)x(0) ;⟨x(t+ τ)x(t)T

⟩= P (τ) < x(t)x(t)T >

C =⟨xxT

AC + CAT = −BBT

Optimal Perturbations

Linearised operator A will generally be non-normal, i.e.

AAT 6= ATA

so eigenvectors of A are not orthogonal, with consequences:

eigenvectors of covariance C (EOFs) do not coincide witheigenvectors of A (dynamical modes)

perturbation norm (with metric M )

N(t) = x(t)TMx(t) = x(0)TP (t)TMP (t)x(0)

may grow (by potentially large amount) over finite times eventhough asymptotically stable:

limt→∞

N(t) = 0

AAT 6= ATA

limt→∞

N(t) = 0

AAT 6= ATA

limt→∞

N(t) = 0

AAT 6= ATA

limt→∞

N(t) = 0

AAT 6= ATA

limt→∞

N(t) = 0

AAT 6= ATA

limt→∞

N(t) = 0

Defining amplification factor

n(t) =x(0)TP (t)TMP (t)x(0)

x(0)TMx(0)

optimal perturbation e maximises n(t) subject to constraintx(0)TMx(0) = 1⇒ generalised eigenvalue problem:

P (t)TMP (t)e = λMe

Response to fluctuating forcing:

var(X(t)) = BT

(∫ t

0P (s)TMP (s)ds

⇒ importance of projection of noise structure on “average” optimals formaintaining variance;

“stochastic optimals”

x(0)TMx(0)

var(X(t)) = BT

(∫ t

0P (s)TMP (s)ds

x(0)TMx(0)

var(X(t)) = BT

(∫ t

0P (s)TMP (s)ds

x(0)TMx(0)

var(X(t)) = BT

(∫ t

0P (s)TMP (s)ds

x(0)TMx(0)

var(X(t)) = BT

(∫ t

0P (s)TMP (s)ds

x(0)TMx(0)

var(X(t)) = BT

(∫ t

0P (s)TMP (s)ds

“stochastic optimals” Probabilistic Perspectives on Climate Dynamics – p. 51/63

Stochastic ENSO Models

How are A and B determined?

1. Empirical: Linear Inverse Modelling

estimate covariances from observations and compute

τln(⟨

x(t+ τ)x(t)T⟩ ⟨

x(t)x(t)T⟩−1)

Compute B from Lyapunov equation

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

Issues:i. if x(t) not truly Markov, estimates will depend on lag τ

ii. must enforce positive-definiteness of BBT

τln(⟨

x(t)x(t)T⟩−1)

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

τln(⟨

x(t)x(t)T⟩−1)

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

τln(⟨

x(t)x(t)T⟩−1)

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

τln(⟨

x(t)x(t)T⟩−1)

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

Issues:

i. if x(t) not truly Markov, estimates will depend on lag τii. must enforce positive-definiteness of BBT

τln(⟨

x(t)x(t)T⟩−1)

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

τln(⟨

x(t)x(t)T⟩−1)

A⟨xxT

⟩+⟨xxT

⟩AT = −BBT

LIM: Optimal Perturbations from Observations

Optimal perturbation e

Perturbation P (t)e at t = 7 months

From Penland and Sardeshmukh (1995)Probabilistic Perspectives on Climate Dynamics – p. 53/63

LIM: Response to Stochastic Forcing

From Penland and Sardeshmukh (1995)Probabilistic Perspectives on Climate Dynamics – p. 54/63

2. Mechanistic: Stochastic Reduction of Tangent Linear Model

Partition dynamics into “slow” and “fast” variables x and y:

dt= f(x,y)

εg(x,y)

Reduce coupled system to effective stochastic dynamics for x:

dt= Lx +N(x,x) + S(x) ◦ W

(many ways of doing this; some formal, some ad hoc)

Linearise model around appropriate state (e.g. climatologicalmean)

dt= f(x,y)

εg(x,y)

dt= f(x,y)

εg(x,y)

dt= f(x,y)

εg(x,y)

dt= f(x,y)

εg(x,y)

Mechanistic Model: 6-Month Stochastic Optimals

From Kleeman and Moore, J. Atmos. Sci., 1997.

Prediction Utility

Can assess “utility” of ensemble prediction p(x) by comparingforecast pdf with “climatological” pdf q(x), using relative entropy

∫p(x) ln

Assuming p and q are both Gaussian, can decompose R:

R = Signal + Dispersion

whereSignal =

2(µp − µq)TΣ−2

q (µp − µq)−n

Disp =1

(det(Σ2

det(Σ2p)

)+ tr(Σ2

pΣ−2q )

Prediction Utility

∫p(x) ln

whereSignal =

q (µp − µq)−n

Disp =1

(det(Σ2

det(Σ2p)

)+ tr(Σ2

pΣ−2q )

Prediction Utility

∫p(x) ln

whereSignal =

q (µp − µq)−n

Disp =1

(det(Σ2

det(Σ2p)

)+ tr(Σ2

pΣ−2q )

Prediction Utility

∫p(x) ln

whereSignal =

q (µp − µq)−n

Disp =1

(det(Σ2

det(Σ2p)

)+ tr(Σ2

pΣ−2q )

Prediction Utility

∫p(x) ln

Signal =1

q (µp − µq)−n

Disp =1

(det(Σ2

det(Σ2p)

)+ tr(Σ2

pΣ−2q )

Prediction Utility

∫p(x) ln

whereSignal =

q (µp − µq)−n

Disp =1

(det(Σ2

det(Σ2p)

)+ tr(Σ2

pΣ−2q )

“Observed” Prediction Utility

From Tang, Kleeman, & Moore, J. Atmos. Sci., 2005.Probabilistic Perspectives on Climate Dynamics – p. 58/63

Individual Forecast Contributions to Correlation

From Tang, Kleeman, & Moore, J. Atmos. Sci., 2005.

Predictive utility relates well to “skillful” forecasts

ENSO forecast skill derives mostly from initial conditions

Individual Forecast Contributions to Correlation

From Tang, Kleeman, & Moore, J. Atmos. Sci., 2005.

Predictive utility relates well to “skillful” forecasts

ENSO forecast skill derives mostly from initial conditions

Case Study II: Conclusions

State of Tropical Pacific has influence on “weather” and “climate”pdfs globally

Linearised stochastic dynamics can be constructed giving insight to:

maintenance of ENSO variability

sources of predictability

importance of “non-normal” dynamics

Other studies have relaxed some of the above assumptions, allowingfor e.g. multiplicative noise and nonlinearity

Overall Conclusions

Climate science is fundamentally probabilistic

Need for probabilistic approach arises from essential complexity ofsystem; some of these complexities can be addressed by bettermodels or observations, but some are irreducible

Fundamental challenges:

characterise climate state pdfs

understand physical origin of pdfs

use pdfs to improve forecasts/predictions

Individual processes/phenomena have been investigated withconsiderable success, but much remains to be done

Overall Conclusions

References

1. Imkeller, P. and J.-S. von Storch (eds), 2001: Stochastic ClimateModels. Birkhäuser, 432 pp.

2. Imkeller, P. and A. Monahan, 2002: Conceptual stochastic climatemodels. Stoch. and Dynamics, 2, 311-326.

3. Palmer, T. et al., 2005: Representing model uncertainty in weatherand climate prediction. Ann. Rev. Earth Planet. Sci., 33, 163-93.

4. Palmer, T. and P. Williams: Stochastic Physics and ClimateModelling. Phil. Trans. Roy. Soc., to appear in 2008.

5. Penland, C. 2003: Noise out of chaos and why it won’t go away.Bull. Amer. Met. Soc. 84, 921-925.

Tropical Multiscale Convective Systems

probabilistic perspectives on climate dynamics · probabilistic perspectives on climate dynamics...

Documents

a. biographical information personal15th. 2016 mozuraitis,...

stochastic dynamics of sea surface...

probabilistic reliability engineering -...

analytic and probabilistic perspectives on the...

probabilistic design introduction an example motivation...

probabilistic damage

overview of astec integral code status and perspectives 2019...

probabilistic graphical...

modelling and verification of probabilistic...

probabilistic thinking

€¦ · web viewby what “term” would you classify...

perturbation dynamics of a planktonic...

probabilistic programming and stan · probabilistic...

probabilistic language acquisition 1 running head...

modelling interactions between weather and climatemodelling...

scalable probabilistic pca - ww2.amstat.org › meetings ›...

probabilistic prediction

probabilistic programming: bayesian modellingmade...

battle for your decision:deterministic vs. probabilistic...

volcanic activity: frontiers and challenges in forecasting...