Scribd
Upload
48 views

Appendix II: Derivation of Principal Component Analysis

This document derives the principal component analysis (PCA) technique of using the eigenvectors of the covariance matrix to minimize the mean-square error when approximating a random vector with a linear combination of orthogonal basis vectors. It shows that the eigenvectors of the covariance matrix satisfy the necessary condition of minimizing the expectation of the residual error. The PCA problem is thus equivalent to computing the eigendecomposition of the covariance matrix to obtain the principal components that maximize the retained variance when reconstructing the data.

Uploaded by

cthunder_1
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
48 views

Appendix II: Derivation of Principal Component Analysis

This document derives the principal component analysis (PCA) technique of using the eigenvectors of the covariance matrix to minimize the mean-square error when approximating a random vector with a linear combination of orthogonal basis vectors. It shows that the eigenvectors of the covariance matrix satisfy the necessary condition of minimizing the expectation of the residual error. The PCA problem is thus equivalent to computing the eigendecomposition of the covariance matrix to obtain the principal components that maximize the retained variance when reconstructing the data.

Uploaded by

cthunder_1
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

Appendix II: Derivation of

Principal Component
Analysis

2.1

Eigenvectors of the Covariance Matrix Derivation

2.1.1

Principal Component Feature Extraction

Consider a random vector x

for which a basis u is to be determined such that an approxi-

mation of x can be obtained by a linear combination of m orthogonal basis vectors:

x = y 1 u 1 + y 2 u 2 + + y m u m

[154]

where m < n and y j is the weighting coefficient of basis vector u j formed by taking the inner
product of x with u j :
T

yj = x uj .

[155]

By forming a k n observation matrix X , the rows of which are the observations x k , we can
express Equation 154 and Equation 155 respectively in the following form:
T

= YU T
X
m

[156]

Y m = XU m

[157]

where U m is an n m matrix whose columns are an uncorrelated basis for X , and Y m is a

k m matrix whose columns are the coefficients for each column vector in U m . By the orthog157

Eigenvectors of the Covariance Matrix Derivation

onality of U m and by the imposition of an additional constraint that the columns of U m have
unit norm, i.e. U j = 1 , it follows that:
T

Um Um = Im

[158]

where I m is a m m diagonal matrix with unit entries. We refer to Equation 156 as the PCAfeature re-synthesis equation and Equation 157 as the PCA-feature projection equation where the
columns of Y m correspond to the estimated features and the columns of U m are the projection
vectors which perform the linear transformation of the input to the new uncorrelated basis.
The problem, then, for deriving a PCA is to obtain the matrix U m such that the residual error in
approximating x with x is minimized:
n

= x x =

y ju j

[159]

j = m+1

that is, the expansion of all the unused features results in a minimal signal which is the residual
error . A suitable criteria is the minimization of the expectation of the mean-square residual
error:
2

= E [ ] = E [ x x ]

[160]

where the expectation of an arbitrary function of x , say ( x ) , is defined as the element-wise


operation:

E[(x)]

( x ) px ( x ) dx

[161]

where p x ( x ) is the probability density function of the random variable x . Since the expectation
operator is linear and due to the condition that the columns of U m are orthonormal it follows that:

158

Eigenvectors of the Covariance Matrix Derivation

T
= E [ ] = E y i u i y j u j
i = m + 1
j = m + 1

E[y j ]

[162]

j = m+1

from which it follows that:


2

E[y j ]

= E [ ( u j x ) ( x u j ) ] = u j E [ xx ]u j = u j Ru j
T

[163]

where R is the correlation matrix for x . Now by substitution of Equation 163 into Equation 162
we arrive at the quantity to be minimized:
n

u j Ru j

[164]

j = m+1

We can express the minimization as a multi-variate differential equation by using a set of


Lagrangian multipliers j and setting the derivative of the expectation of the mean-square error
with respect to the basis components u j to zero. This is a necessary condition for the minimum
and gives the characteristic equation:

' = 2 ( Ru j ju j ) = 0, j = m + 1, , n
u j

[165]

It is well known that the solutions to this equation constitute the eigenvectors of the correlation
matrix R . It is also worth noting that the correlation matrix is related to the covariance matrix by
the following expression:
T

Q x = R x mm .

[166]

Thus, for zero-mean or centered data, the problem is equivalent to finding the eigenvectors of the
covariance matrix Q x . Since the columns of U are now determined to be the eigenvectors, we
can re-express the residual error as the sum of the eigenvalues of the unused portion of the basis:
n

[167]

j = m+1

159

Eigenvectors of the Covariance Matrix Derivation

and the solution to the minimization reduces to ordering the basis vectors u j such that the columns with the smallest eigenvalues occur in the unused portion of the basis which also implies that
the m columns of U m which are used for reconstruction should comprise the eigenvectors with
the m largest eigenvalues.
Now that we have arrived at the form of the solution for optimal orthonormal basis reconstruction
(in the square-error sense) we must find a general form for representing the solution. Since the
eigenvalues form the diagonal elements of the covariance of the transformed data with all other
elements equal to zero we can express the solution to the eigenvalue decomposition as a diagonalization of the input covariance matrix Q x .

160

You might also like