what is eof analysis? eof = empirical orthogonal function method of finding structures (or patterns)...

35
What is EOF analysis? • EOF = Empirical Orthogonal Function • Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time) dataset • Mathematically EOFs are eigenvectors of the covariance matrix of a dataset

Upload: willis-owens

Post on 22-Dec-2015

219 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

What is EOF analysis?

• EOF = Empirical Orthogonal Function

• Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time) dataset

• Mathematically EOFs are eigenvectors of the covariance matrix of a dataset

Page 2: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

• Any (space-time) dataset can be represented as a matrix:

X = M = Xij

N

Math

Page 3: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

• Define XT

XT = N = Xji

M

Math

Page 4: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Math

• And covariance matrix C=XXT

C = M N = M

N M

M

Page 5: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Math

• EOFs (ei) are the eigenvectors of C

C ei = ei

Page 6: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Math

• Principal components:

Fourier coefficients of the corresponding EOFs in the time expansion of the dataset

PCi (t) = (XT, ei)Too easy, huh?

Page 7: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Math• Why does 1st EOF maximize explained

variance?Answer: by construction.

(eTX,Xte) = ||eTX|| = max(eT,e) = 1

Or:(eT,Ce) = , ( C = XXT ), Ce = e

This maximizes on the eigenvector corresponding to the greatest eigenvalue.

Amazing!

Page 8: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

What do EOFs and PCs mean?

• EOF – a coherent orthogonal spatial pattern.

• First EOF explains most variance in a physical field

• PC – time behavior of the corresponding EOF (=spatial pattern)

Stunning! Let’s EOF everything!

Page 9: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

EOF interpretation

Direction of maximum variance

Page 10: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Example 1. El-Nino.

Page 11: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Example 2. Arctic Oscillation.

Page 12: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Tropospheric Winter Trends

Cohen et al, 2012, ERL

Page 13: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Northern Hemisphere Land Temperatures 1987-2010

Data: CRU temperatureAlexeev et al, 2012, Clim Change; Cohen et al, 2012, ERL

Page 14: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Major modes in the Northern Hemisphere

Page 15: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Major modes in the Northern Hemisphere

Page 16: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Major modes in the Northern Hemisphere

Page 17: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Why EOFs are not physical modes?

Your equations:

dx/dt + Ax = f

Physical modes: eigenvectors of A.(Solve Ay = y)Physical modes are not orthogonal(generally speaking)

Page 18: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Why EOFs are not physical modes?

Your equations:

dx/dt + Ax = f

EOFs – eigenvectors of a matrix derived from A AT

EOFs: orthogonal by construction

Page 19: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Other methods

• SVD = Singular Value Decomposition,

aka

MCA = Maximum Correlation Analysis

• Method is looking for correlated spatial patterns in two different fields

Page 20: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Math

• Correlation matrix CXY=XYT

CXY = M N = M

N L

L

Page 21: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Math

• SVD vectors of C: in U (X-field) and V

(Y-field) matrices

CXY = UVT

Page 22: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Other methods

• CCA = Canonical Correlation Analysis:

SVD over space of Fourier coefficients of EOFs

Page 23: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Other methods

• POP = Principal Oscillation Pattern Analysis

FDT over space of Fourier coefficients of EOFs

(FDT = Fluctuation-Dissipation Theorem)

Page 24: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

POP = Principal Oscillating Patterns

xn+1 = C xn + (C = ‘step forward’ operator)

Assume < x, > = 0< xn+1, xn > = C < xn, xn > + < x, >

We can approximate C from: C = C 0 C-1

1

Where C0 = < xn, xn > , C1 = < xn+1, xn >

Page 25: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Other methods

• Varimax, Quartimax, rotated EOF analysis

EOF modifications

Page 26: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Other methods

• MTM = Multi-Taper Method

Combination of EOF and Wavelet analyses

Page 27: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Other methods

• SSA = Singular Spectrum Analysis

• MSSA = Multi-channel SSA

• MTM-SVD

• EEOF = Extended EOF

• FDEOF = Frequency Domain EOF

• CEOF = Complex EOF

Page 28: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

When is EOF analysis useful?

• Analysis of repeating pronounced patterns over long time series

• Image/data compression

• Filtering

Not so fast….

Page 29: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

When EOF use is inappropriate?

• Short time series, lots of missing and/or inconsistent data

• Absence of a prominent signal

• Presence of a dominant trend in the data (e.g. seasonal cycle is dangerous!)

Page 30: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Why do people get so excited about EOFs?

• EOFs can be applied to any dataset

• Simplicity of the analysis is very appealing. Everyone does EOFs.

• Patterns are often tempting to analyze (because of method’s simplicity)

Page 31: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Do not overdo it with EOFs!

• “New” patterns sometimes turn out to be not so new.

• Artificial (mechanistic) data de-trending can lead to surprises (example: removal of seasonal cycle does not remove changing seasonal variability in most of the fields)

Page 32: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Do scientists have problems interpreting EOF results?

• Saying “I performed EOF analysis on my data” does not mean you explained any physics

• EOFs usually do not coincide with eigen-modes of the physical process you are trying to interpret/explain. POP analysis does not give you orthonormal modes, but it might approximate your physical modes

Page 33: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Are results of EOF analysis accurate?

• Statistical significance is always an issue.

• If something correlates (even very well) with something else (or appears to be systematically preceding/following), this does not mean one causes the other. They both can be caused by something else.

Page 34: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

What are EOF maniacs?

• People who eof (svd, cca …) everything with everything just for the sake of it

Page 35: What is EOF analysis? EOF = Empirical Orthogonal Function Method of finding structures (or patterns) that explain maximum variance in (e.g.) 2D (space-time)

Are there many EOF/SVD/CCA maniacs out there?

•Yes, there are!

(I am one of them)