1
nEigenvectors (for a square m´m matrix S)
nHow many eigenvaluesare there at most?
only has a non-zero solution if
this is a m-th order equation in λ which can have at most m distinct solutions
(roots of the characteristic polynomial) – can be complex even though S is real.eigenvalue (right) eigenvector
Example
2
ú ú ú û ù ê ê
ê ë é
=
0 0 0
0 2 0
0 0 3
S has eigenvalues 3, 2, 0 with corresponding eigenvectors
÷ ÷
÷ ø ö ç ç ç è æ
= 0 0 1 v
1÷ ÷
÷ ø ö ç ç ç è æ
= 0 1 0 v
2÷ ÷
÷ ø ö ç ç ç è æ
= 1 0 0 v
3On each eigenvector, S acts as a multiple of the identity matrix: but as a different multiple on each.
Any vector (say x= ) can be viewed as a combination of the eigenvectors: x = 2v
÷ 1+ 4v
2+ 6v
3÷÷ ø ö çç ç è æ 6 4 2
3
§Thus a matrix-vector multiplication such as Sx (S, x as in the previous slide) can be rewritten in terms of the
eigenvalues/vectors:
§Even though x is an arbitrary vector, the action of S on x is determined by the eigenvalues/vectors.
§Suggestion: the effect of “small” eigenvalues is small.
3 3 2
2 1
1 3 2
1
3 2
1
6 4
2 6
4 2
) 6 4
2 (
v v
v Sv
Sv Sv
Sx
v v
v S Sx
l l
l + +
= +
+
=
+ +
=
4
0 and
,
1 2 1 2} 2 , 1 { } 2 , 1 { } 2 , 1
{
= v ¹ Þ v • v =
Sv l l l
For symmetric matrices, eigenvectors for distinct eigenvalues are orthogonal
 ΠÞ
=
=
- l l
l , if 0 and S S
Tcomplex
for S I
All eigenvalues of a real symmetric matrix are real.
0 v
Sv if then , 0
, ³ = Þ ³
 Î
" w
nw
TSw l l
All eigenvalues of a positive semidefinite matrix are non-negative
5
nLet
nThen
nThe eigenvalues are 1 and 3 (nonnegative, real).
nThe eigenvectors are orthogonal (and real):
ú û ê ù
ë
= é
2 1
1 S 2
. 0 1 ) 2 2 (
1
1
2
2= - -
ú Þ û ê ù
ë é
-
= -
- l
l l I l
S
÷÷ ø çç ö è æ
-1
1 ÷÷
ø çç ö è æ 1 1
Real, symmetric.
Plug in these values and solve for
eigenvectors. 6
§Let be a squarematrix with m linearly independent eigenvectors (a “non-defective” matrix)
§Theorem: Exists an eigen decomposition
§(cf. matrix diagonalization theorem)
§Columns of U are eigenvectorsof S
§Diagonal elements of are eigenvaluesof diagonal
Unique for distinct
eigen- values
7
ú ú ú û ù ê ê
ê ë é
= v v
nU
1...
Let U have the eigenvectors as columns:
ú ú ú û ù ê ê
ê ë é ú ú ú û ù ê ê
ê ë é
= ú ú ú û ù ê ê
ê ë é
= ú ú ú û ù ê ê
ê ë é
=
n n
n n
n
v v v v
v v
S SU
l l
l
l ... ... ...
...
1 1
1 1 1
Then, SU can be written
And S=ULU
–1. Thus SU=UL, or U
–1SU=L
8
Recall ; 1 , 3 . 2
1 1 2
2
1
= =
ú û ê ù
ë
= é l l
S
The eigenvectors and form
÷÷ ø çç ö è æ
-1
1 ÷÷ ø
çç ö è æ 1
1 ú
û ê ù
ë é
= -
1 1
1 U 1
Inverting, we have ú
û ê ù
ë
é -
=
-
2 / 1 2 / 1
2 / 1 2 /
1
1 U
Then, S=ULU
–1= ú
û ê ù
ë
é -
ú û ê ù
ë ú é û ê ù
ë é
- 1 / 2 1 / 2
2 / 1 2 / 1 3 0
0 1 1 1
1 1
Recall UU
–1=1.
9
Let s divide U (and multiply U
–1) by 2
ú û ê ù
ë
é -
ú û ê ù
ë ú é û ê ù
ë é
- 1 / 2 1 / 2
2 / 1 2 / 1 3 0
0 1 2 / 1 2 / 1
2 / 1 2 /
Then, S= 1
Q L (Q
-1= Q
T)
Why? Stay tuned …
10
n If is a symmetric matrix:
n Theorem: Exists a (unique) eigen decomposition
n where Q is orthogonal:
n Q
-1= Q
Tn Columns of Q are normalized eigenvectors
n Columns are orthogonal.
n (everything is real)
Q
TQ S = L
11
nExamine the symmetric eigen decomposition, if any, for each of the following matrices:
ú û ê ù
ë é
- 1 0 1
0 ú
û ê ù
ë é
0 1
1
0 ú
û ê ù
ë é
- 2 3 2
1 ú
û ê ù
ë é
4 2
2 2
12
V
TU A = S
m ´ m m ´ n V is n ´ n For an m ´ n matrix A of rank r there exists a
factorization (Singular Value Decomposition = SVD) as follows:
The columns of U are orthogonal eigenvectors of AA
T. The columns of V are orthogonal eigenvectors of A
TA.
i
i
l
s =
(
r)
diag s
1... s
=
S Singular values.
Eigenvalues l
1… l
rof AA
Tare the eigenvalues of A
TA.
13
nIllustration of SVD dimensions and sparseness
14
Let
ú ú ú û ù
ê ê ê ë
é -
=
0 1
1 0
1 1 A
Thus m=3, n=2. Its SVD is
ú û ê ù
ë é ú - ú ú û ù
ê ê ê ë é ú ú ú û ù
ê ê ê ë é
-
- 1 / 2 1 / 2
2 / 1 2 / 1 0 0
3 0
0 1
3 / 1 6
/ 1 2 / 1
3 / 1 6 / 1 2 / 1
3 / 1 6 / 2 0
Typically, the singular values arranged in decreasing order.
15
nSVD can be used to compute optimal low-rank approximations.
nApproximation problem: Find Akof rank k such that
Akand X are both
m ´ n matrices.
Typically, want k << r.
Frobenius norm k F
X rank X
k
A X
A = -
min
= ) ( :16
nSolution via SVD
set smallest r-k singular values to zero
T k
k
U V
A = diag ( s
1,..., s , 0 ,..., 0 )
column notation: sum of rank 1 matrices T
i i k
i i
k
u v
A = å=1s
k
17
§How good (bad) is this approximation?
§It’s the best possible, measured by the Frobenius norm of the error:
§ min
$:&'() $ *)𝐴 − 𝑋
F= 𝐴 − 𝐴
) F= ∑
(1*)34𝜎
12where the siare ordered such that si³ si+1.
Suggests why Frobenius error drops as k increased.
18
§Whereas the term-doc matrix A may have m=50000, n=10 million (and rank close to 50000)
§We can construct an approximation A100 with rank 100.
§ Of all rank 100 matrices, it would have the lowest Frobenius error.
§Great … but why would we?
§Answer: Latent Semantic Indexing
C. Eckart, G. Young, The approximation of a matrix by another of lower rank.
Psychometrika, 1, 211-218, 1936. 19