E3101 2002 SVD Fun 1

Applications of the SVD

Marc Spiegelman

Detail from Durer’s Melancolia, dated 1514., 359x371 image

E3101 2002 SVD Fun 2

Image Compression
Given an original image (here 359 × 371 pixels)
Detail from Durer’s Melancolia, dated 1514., 359x371 image

We can write it as a 359 × 371 matrix A which can then be decomposed via the
singular value decomposition as

A = U ΣV T

where U is 359 × 359, Σ is 359 × 371 and V is 371 × 371.

E3101 2002 SVD Fun 3

The matrix A however can also be written as a sum of rank 1 matrices

A = σ1 u1 vT1 + σ2 u2 vT2 + . . . + σn un vTn

where each rank 1 matrix u i vT

i is the size of the original matrix. Each one of
these matrices is a mode.

Because the singular values σ i are ordered σ1 ≥ σ2 ≥ . . . ≥ σn ≥ 0,

however, significant compression of the image is possible if the spectrum of
singular values has only a few very strong entries.
E3101 2002 SVD Fun 4

Spectrum of Singular values for A

Singular Values
5
10

4
10

3
10

2
10

1
10

0
10
0 50 100 150 200 250 300 350 400

Here the spectrum is contained principally in the first 100–200 modes (max).
E3101 2002 SVD Fun 5

We can therefore reconstruct the image from just a subset of modes. For example
in matlabese we can write just the first mode as

[U,S,V]=svd(A)
B=U(:,1)*S(1,1)*V(:,1)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 1 modes
E3101 2002 SVD Fun 6

Or as a sum of the first 10 modes as

B=U(:,1:10)*S(1:10,1:10)*V(:,1:10)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 10 modes

which only uses 5% of the storage (10 × 359 + 10 × 371 + 10 = 7310 pixels
vs 359 × 371 = 133189 pixels.
E3101 2002 SVD Fun 7

Adding modes, just adds resolution

B=U(:,1:20)*S(1:20,1:20)*V(:,1:20)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 20 modes
E3101 2002 SVD Fun 8

Adding modes, just adds resolution

B=U(:,1:30)*S(1:30,1:30)*V(:,1:30)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 30 modes
E3101 2002 SVD Fun 9

Adding modes, just adds resolution

B=U(:,1:40)*S(1:40,1:40)*V(:,1:40)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 40 modes
E3101 2002 SVD Fun 10

Adding modes, just adds resolution

B=U(:,1:50)*S(1:50,1:50)*V(:,1:50)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 50 modes
E3101 2002 SVD Fun 11

Adding modes, just adds resolution

B=U(:,1:100)*S(1:100,1:100)*V(:,1:100)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 100 modes
E3101 2002 SVD Fun 12

Adding modes, just adds resolution

B=U(:,1:200)*S(1:200,1:200)*V(:,1:200)’
Detail from Durer’s Melancolia, dated 1514., 359x371 image EOF reconstruction with 200 modes
E3101 2002 SVD Fun 13

Application 2: EOF analysis

Pattern extraction—Mid-ocean ridge topography

Here we consider a real research use of the SVD by Chris Small (LDEO)

A Global Analysis Of Midocean Ridge Axial Topography

GEOPHYSICAL JOURNAL INTERNATIONAL 116 (1): 64-84 JAN 1994
E3101 2002 SVD Fun 14

120˚ 150˚ 180˚ 210˚ 240˚ 270˚ 300˚ 330˚ 0˚ 30˚ 60˚ 90˚
80˚ 80˚

60˚ 60˚

40˚ 40˚

20˚ 20˚

0˚ 0˚

-20˚ -20˚

-40˚ -40˚

-60˚ -60˚

120˚ 150˚ 180˚ 210˚ 240˚ 270˚ 300˚ 330˚ 0˚ 30˚ 60˚ 90˚
E3101 2002 SVD Fun 15

The data: cross axis topography profiles from

different spreading rates
Cross axis topography of mid−ocean ridges
1500
slow
medium
fast
1000
relative Elevation (m)

500

−500

−1000

−1500
0 10 20 30 40 50 60 70 80
Km across axis
E3101 2002 SVD Fun 16

Form a matrix A (179 × 80) of elevation vs. distance across the ridge
Cross axis topography of mid−ocean ridges

2000

20
1500

40
1000
spreading rate 60

500
80

0
100

−500
120

−1000
140

20 40 60
Km across axis

and again take the SVD A= U ΣV T . Here U is the same size as A and Σ and
V are both square 80 × 80 matrices.
E3101 2002 SVD Fun 17

Now the rows of V T form an orthonormal basis for the row space of A, i.e. each
profile (row of A) can be written as a linear combination of the rows of V T or

A = CV T

which by inspection of the SVD shows that C = U Σ. Here, the rows of V T are
known as Empirical Orthogonal Functions or EOFs.
E3101 2002 SVD Fun 18

Again, if the spectrum of Singular values contains a few large values and a long
tail of very small values, it may be possible to reconstruct the rows of A with only
a small number of EOFs. The spectrum for this data looks like
Singular Values
5
10

4
10

3
10

2
10

1
10
0 10 20 30 40 50 60 70 80

which suggests that you only need about 4 EOF’s to explain most of the data.
E3101 2002 SVD Fun 19

The first 4 EOFs

First 4 EOFs
0.4
EOF1
EOF2
0.3 EOF3
EOF4

0.2

0.1

−0.1

−0.2

−0.3
0 10 20 30 40 50 60 70 80
Km across axis
E3101 2002 SVD Fun 20

And we can reconstruct individual profiles as combinations of the first 4 EOF’s.

For example here is one for a slow spreading rate
EOF reconstruction, sample 10=(−2994.2,−3475.3,−2403.3,−148.1)
1500

1000
relative Elevation (m)

500

−500

−1000

−1500
0 10 20 30 40 50 60 70 80
Km across axis
E3101 2002 SVD Fun 21

Intermediate spreading rate

EOF reconstruction, sample 60=(−557.5,699.8,−299.1,676.9)
700

600

500

400
relative Elevation (m)

300

200

100

−100

−200

−300
0 10 20 30 40 50 60 70 80
Km across axis
E3101 2002 SVD Fun 22

Fast spreading rate

EOF reconstruction, sample 100=(−617.3,1316.3,−557.2,135.8)
500

400
relative Elevation (m)

300

200

100

−100
0 10 20 30 40 50 60 70 80
Km across axis