This demo will help you build intuition for the behavior of
eigenvectors and eigenvalues of a 2x2 symmetric real matrix. As you’ll
remember, an **eigenvector** $v$ of a matrix $M$ is any vector that
satisfies the following equation:

In other words, if you transform a vector and you end up with a scaled version of the vector, then it is an eigenvector. The amount by which the vector is scaled is known as the eigenvalue.

When the 2x2 matrix is symmetric, then there exist two eigenvectors that are orthogonal to each other. In that case, we can write the matrix as:

where $U$ is a matrix holding the eigenvectors and $\Sigma$ is a
diagonal matrix where the entries in the diagonal are the
eigenvalues. If you’ll remember from linear algebra, every time you
have a square matrix whose rows (or columns) are orthogonal to each
other, that is a **rotation** matrix and, in addition, rotation
matrices are such that their transposes are their inverses. So a good
way to think about this is that eigenvectors give you a **decomposition** of the matrix M into simpler matrices.

In other words, the operation of every symmetric matrix $M$ on a vector $v$ is $Mv = U \Sigma U^T v$, or:

- $U^T v$: transform the vector $v$ to the “eigenspace”: this is a rotation
- $\Sigma U^T v$: in the eigenspace, scale the vector’s coordinates by the eigenvalues
- $Mv = U \Sigma U^T v$: transform the scaled vector back to the original basis

In the interactive demo below, the unit-length eigenvectors are represented by the red dots.

$M_{00}$ | </input> | $M_{01}$ | </input> |

$M_{10}$ | 1 | $M_{11}$ | </input> |

$U_{00}$ | $U_{01}$ | ||

$U_{10}$ | $U_{11}$ |

$\lambda_0$ | $\lambda_1$ |

- Eigenvalue multiplicity (two eigenvectors not aligned with one another with equal eigenvalues) will mean that there are more than 4 unit-length eigenvectors, and my crappy power iteration algorithm stops working in that case.
- The rotation transition for $U$ and $U^T$ should be an actual rotation. For small rotations the linear interpolation looks fine, but for bigger ones it’s hard to see what’s going on.
- We should arbitrarily flip the eigenvector signs such that we get a smaller rotation on $U^T$ and $U$. Sometimes my crappy power iteration algorithm gives the “bad” eigenvector, and that makes it hard to see what’s going on.
- If one of the eigenvalues is zero, my crappy power iteration algorithm gives a bad eigenvector (notice a theme here?)

If you really want to understand eigenvectors and eigenvalues, the best thing to read continues to be chapter 5 of Shewchuk’s classic Introduction to the Conjugate Gradient Method Without the Agonizing Pain.

The particular presentation in this demo was inspired by Blinn’s also-classic Consider the lowly 2x2 matrix.