# Segment 48. Principal Component Analysis (PCA)

(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

## Contents

#### Watch this segment

(Don't worry, what you see statically below is not the beginning of the segment. Press the play button to start at the beginning.)

{{#widget:Iframe |url=http://www.youtube.com/v/frWqIUpIxLg&hd=1 |width=800 |height=625 |border=0 }}

The direct YouTube link is http://youtu.be/frWqIUpIxLg

Links to the slides: PDF file or PowerPoint file

### Problems

#### To Compute

1. Suppose that only one principal component is large (that is, there is a single dominant value $\displaystyle s_i$ ). In terms of the matrix $\displaystyle \mathbf V$ (and anything else relevant), what are the constants $\displaystyle a_j$ and $\displaystyle b_j$ that make a one-dimensional model of the data? This would be a model where $\displaystyle x_{ij} \approx a_j \lambda_i + b_j$ with each of the data points (rows) having its own value of an independent variable $\displaystyle \lambda_i$ and each of the responses (columns) having it's own constants $\displaystyle a_j,b_j$ .

2. The file dataforpca.txt has 1000 data points (rows) each with 3 responses (columns). Make three scatter plots, each showing a pair of responses (in all 3 possible ways). Do the responses seem to be correlated?

3. Find the principal components of the data and make three new scatter plots, each showing a pair of principal coordinates of the data. What is the distribution (histogram) of the data along the largest principal component? What is a one-dimensional model of the data (as in problem 1 above)?

#### To Think About

1. Although PCA doesn't require that the data be multivariate normal, it is most meaningful in that case, because the data is then completely defined by its principal components (i.e., covariance matrix) and means. Can you design a test statistic that measures "quality of approximation of a data set by a multivariate normal" in some quantitative way? Try to make your statistic approximately independent of $\displaystyle N$ , the number of data points.