Eigenvalues and eigenvectors reveal directions in space that remain unchanged (except for scaling) when a linear transformation is applied. This property is remarkably useful for uncovering the underlying structure within data, forming the foundation of Principal Component Analysis (PCA).

PCA is a cornerstone technique in machine learning and data analysis, primarily used for dimensionality reduction. The main objective is to identify a new set of coordinate axes, called principal components, for the data. These new axes are chosen such that the maximum possible variance in the data is captured along the first axis, the maximum remaining variance along the second axis (orthogonal to the first), and so on. This allows us to represent the data using fewer dimensions while minimizing information loss.

The Covariance Matrix: Capturing Data Spread

To find these directions of maximum variance, we first need to quantify how the different features in our dataset vary together. This is precisely what the covariance matrix ( $C$ ) does. For a dataset with $p$ features, the covariance matrix is a $p \times p$ symmetric matrix where:

The diagonal element $C_{ii}$ is the variance of the $i$ -th feature.
The off-diagonal element $C_{ij}$ is the covariance between the $i$ -th feature and the $j$ -th feature.

This matrix summarizes the spread and linear relationships within the data.

Eigenvectors as Directions of Maximum Variance

Here's the critical connection: The eigenvectors of the data's covariance matrix $C$ point along the directions of maximum variance in the data. These directions are the principal components. Furthermore, the eigenvalue $\lambda$ associated with each eigenvector quantifies the amount of variance along that specific direction.

Think back to the definition $Ax = \lambda x$ . In the context of PCA, our matrix $A$ is the covariance matrix $C$ . The eigenvectors $x$ (often denoted as $u$ or $v$ in PCA literature) are the principal components. When we apply the transformation represented by $C$ to one of its eigenvectors $u$ , the result is simply the same vector scaled by its eigenvalue $\lambda$ : $Cu = \lambda u$ .

Why is this important? It turns out that finding the direction $u$ (a unit vector) that maximizes the variance of the data projected onto it ( $u^T C u$ ) leads directly to the eigenvector equation. The direction $u$ that maximizes this variance is the eigenvector corresponding to the largest eigenvalue of $C$ . The maximum variance achieved is equal to that largest eigenvalue $\lambda_{max}$ .

Principal Components: The New Axes

So, the principal components of the data are simply the eigenvectors of its covariance matrix $C$ , typically ordered from highest eigenvalue to lowest.

First Principal Component (PC1): The eigenvector associated with the largest eigenvalue ( $\lambda_1$ ). This direction captures the most variance in the dataset.
Second Principal Component (PC2): The eigenvector associated with the second-largest eigenvalue ( $\lambda_2$ ). This direction is orthogonal (perpendicular) to PC1 and captures the maximum remaining variance.
Subsequent Components: Each subsequent principal component is orthogonal to all preceding components and captures the maximum variance possible from what's left.

Visualizing Principal Components

Imagine a scatter plot of 2D data points that form an elliptical cloud. The eigenvectors of the covariance matrix of this data will align with the axes of this ellipse.

Scatter plot showing 2D data points. PC1 (red line) indicates the direction of maximum variance, determined by the eigenvector with the largest eigenvalue. PC2 (orange dotted line), orthogonal to PC1, captures the next largest variance.

How PCA Reduces Dimensions

The eigenvalues tell us the amount of variance captured by each principal component (eigenvector). By ranking the eigenvectors according to their corresponding eigenvalues in descending order, we know which directions are most "important" for describing the data's spread.

To reduce data from $p$ dimensions to a smaller number $k$ (where $k < p$ ), PCA involves these steps:

Calculate the covariance matrix $C$ of the original data.
Compute the eigenvalues and eigenvectors of $C$ .
Select the $k$ eigenvectors that correspond to the $k$ largest eigenvalues. These form the new basis for our reduced-dimension subspace.
Transform the original data points into this new subspace by projecting them onto the selected $k$ eigenvectors.

The result is a $k$ -dimensional dataset that retains the most significant variance present in the original $p$ -dimensional data. The proportion of total variance retained by the first $k$ components can be calculated by summing the top $k$ eigenvalues and dividing by the sum of all eigenvalues.

Summary

Eigenvalues and eigenvectors are fundamental to Principal Component Analysis because:

Eigenvectors of the covariance matrix define the principal components: the orthogonal axes along which the data exhibits maximum variance.
Eigenvalues measure the magnitude of variance along these principal component directions.

This relationship allows PCA to systematically identify the most informative directions in the data, making it possible to reduce dimensionality effectively while preserving essential data structure. Understanding this connection is significant for anyone applying or interpreting PCA results in machine learning. You will learn how to perform these calculations using libraries like NumPy in the upcoming practical exercises.