6. Process or Product Monitoring and Control
6.5. Tutorials

6.5.5.

Principal Components

Dimension reduction tool

A Multivariate Analysis problem could start out with a substantial number of correlated variables. Principal Component Analysis is a dimension-reduction tool that can be used advantageously in such situations. Principal component analysis aims at reducing a large set of variables to a small set that still contains most of the information in the large set.

Principal factors

The technique of principal component analysis enables us to create and use a reduced set of variables, which are called principal factors. A reduced set is much easier to analyze and interpret. To study a data set that results in the estimation of roughly 500 parameters may be difficult, but if we could reduce these to 5 it would certainly make our day. We will show in what follows how to achieve substantial dimension reduction.

Inverse transformaion not possible

While these principal factors represent or replace one or more of the original variables, it should be noted that they are not just a one-to-one transformation, so inverse transformations are not possible.

Original data matrix

To shed a light on the structure of principal components analysis, let us consider a multivariate data matrix X, with n rows and p columns. The p elements of each row are scores or measurements on a subject such as height, weight and age.

Linear function that maximizes variance

Next, standardize the X matrix so that each column mean is 0 and each column variance is 1. Call this matrix Z. Each column is a vector variable, z_i, i = 1, . . . , p. The main idea behind principal component analysis is to derive a linear function y for each of the vector variables z_i. This linear function possesses an extremely important property; namely, its variance is maximized.

Linear function is component of z

This linear function is referred to as a component of z. To illustrate the computation of a single element for the jth y vector, consider the product y = z v' where v' is a column vector of V and V is a p x p coefficient matrix that carries the p-element variable z into the derived n-element variable y. V is known as the eigen vector matrix. The dimension of z is 1 x p, the dimension of v' is p x 1. The scalar algebra for the component score for the ith individual of y_j, j = 1, ...p is:

y_ji = v'₁z_1i + v'₂z_2i + ... + v'_pz_pi

This becomes in matrix notation for all of the y:

Y = ZV

Mean and dispersion matrix of y

The mean of y is m_y = V'm_z = 0, because m_z = 0.

The dispersion matrix of y is

D_y = V'D_zV = V'RV

R is correlation matrix

Now, it can be shown that the dispersion matrix D_z of a standardized variable is a correlation matrix. Thus R is the correlation matrix for z.

Number of parameters to estimate increases rapidly as p increases

At this juncture you may be tempted to say: "so what?". To answer this let us look at the intercorrelations among the elements of a vector variable. The number of parameters to be estimated for a p-element variable is

• p means
• p variances
• (p² - p)/2 covariances
• for a total of 2p + (p²-p)/2 parameters.

So

• If p = 2, there are 5 parameters
• If p = 10, there are 65 parameters
• If p = 30, there are 495 parameters

Uncorrelated variables require no covariance estimation

All these parameters must be estimated and interpreted. That is a herculean task, to say the least. Now, if we could transform the data so that we obtain a vector of uncorrelated variables, life becomes much more bearable, since there are no covariances.