Graphics Reference
In-Depth Information
′
and they must satisfy Σw
. herefore, w
is obtained byfinding
the eigenvector associated with the leading eigenvalue α
.Forthesecondprincipal
component,welookforaunitvector w
whichisorthogonaltow
andmaximizesthe
variance of the projection of X along w
.hatis,intermsofaLagrangianproblem,
we solve for α
, w
and β in the following optimization formula:
max
α
,β,w
=
α
w
,andw
w
=
w
′
w
′
w
′
Σw
−
α
(
w
−
)−
β
(
w
)
.
(
.
)
Using a similar procedure, we are able to find the leading principal components se-
quentially.
Assume for simplicity that the data
are already centered on their
mean, and so the sample covariance matrix is given by Σ
n
x
,...,x
n
n
j
=
x
j
x
′
n.Byap-
plying the above sequential procedure to the sample covariance Σ
n
,wecanobtain
the empirical principal components.
For KPCA using the feature representation (
.
), the data mapped in the feature
space
=
j
H
κ
are
γ
,...,γ
n
. he sample covariance (which is also known as a covari-
ance operator in
H
κ
)isgivenby
n
j
=
n
C
n
=
(
γ
j
−
γ
)(
γ
j
−
γ
)
(
.
)
where f
κ
.
Applying similar arguments to before, we aim to find the leading eigencomponents
of C
n
.hatis,wesolveforh in the following optimization problem:
g is a linear operator defined by
(
f
g
)(
h
)=
g, h
H
κ
f for f , g, h
H
max
h
H
κ
h, C
n
h
H
κ
subject to
h
=
.
(
.
)
H
κ
n
j
=
β
j
γ
j
It can be shown that the solution to this is of the form h
=
H
κ
,whereβ
j
's
are scalars. As
n
i, j
=
′
n
n
β
′
h, C
n
h
=
β
i
β
j
γ
i
, C
n
γ
j
=
K
I
n
−
Kβ
n ,
H
κ
H
κ
n
H
κ
β
′
and
h
=
Kβ,whereK
=
κ
(
x
i
, x
j
)
denotes the n
n kernel data matrix, the
optimization problem can be reformulated as
′
n
n
β
′
n subject to β
′
max
β
K
I
n
−
Kβ
Kβ
=
.
(
.
)
n
R
n
he Lagrangian of the above optimization problem is
′
n
n
′
′
max
α
R
n
β
K
I
n
−
Kβ
n
−
α
(
β
Kβ
−
)
,
n
R,β
where α is the Lagrange multiplier. Taking derivatives with respect to the β's and
setting them to zero, we get
′
n
′
n
n
n
K
I
n
−
Kβ
n
=
αKβ ,or
I
n
−
Kβ
=
nαβ.
(
.
)
n
n