Eigenvectors and eigenvalues | Chapter 14, Essence of linear algebra

This video provides a geometric intuition for eigenvectors and eigenvalues, explaining them as vectors that remain on their own span during a linear transformation. It demonstrates this concept using 2D matrices and extends it to finding the axis of rotation in 3D space. The segment then derives the characteristic equation $\det(A - \lambda I) = 0$ by showing that non-zero solutions require the transformation matrix to squash space into a lower dimension. This segment demonstrates how to compute eigenvalues and eigenvectors using the characteristic equation $\det(A - \lambda I) = 0$. It illustrates three distinct scenarios: a matrix with two independent real eigenvalues, a rotation matrix with complex eigenvalues indicating no real invariant lines, and a shear matrix where multiple vectors share a single eigenvalue. The lesson concludes by introducing the concept of an 'eigenbasis'—a basis formed entirely by eigenvectors—which allows a transformation to be represented as a diagonal matrix. This diagonalization simplifies calculating high powers of matrices.

Before you watch

• Linear transformations as matrices
• Determinants and area/volume scaling
• Solving linear systems of equations
• Matrix multiplication and determinants
• Concept of linear transformations
• Solving systems of linear equations
• Basic understanding of vector spaces

Chapters

Learning script

Generated from the video's visuals and explanation; not verbatim speech.

0:00The topic of eigenvectors and eigenvalues often confuses students because it relies heavily on prior concepts. To understand it intuitively, you must be comfortable viewing matrices as linear transformations, and possess a solid grasp of determinants, linear systems, and change of basis.

1:20Consider a 2D linear transformation represented by the matrix with columns [3, 0] and [1, 2]. Most vectors are knocked off their original line—called their span—when transformed. However, some special vectors remain on their own span, merely being stretched or squished. In this example, the vector along the x-axis is scaled by 3, while the diagonal vector [-1, 1] is scaled by 2. These special vectors are eigenvectors, and their scaling factors are eigenvalues. Negative eigenvalues simply indicate the vector flips direction but stays on its span.

4:03A proper rotation in three dimensions has a fixed axis. Nonzero vectors on that axis are eigenvectors with eigenvalue one; the axis is a line, not a single vector. For the identity rotation all directions are fixed. Eigenvalues belong to the linear map, while coordinate representations depend on the basis.

5:24To compute eigenvalues algebraically, we start with the definition Av = λv. By rewriting the right side as (λI)v and moving everything to one side, we get (A - λI)v = 0. We seek non-zero solutions for v. A matrix can only map a non-zero vector to zero if it squashes space into a lower dimension, which occurs exactly when its determinant is zero. Thus, we solve det(A - λI) = 0. For our initial 2x2 matrix, this yields the polynomial (3-λ)(2-λ) = 0, confirming the eigenvalues are 3 and 2.

10:00We begin by solving for the eigenvalues of the matrix [[3, 1], [0, 2]]. By setting the determinant of [[3-λ, 1], [0, 2-λ]] to zero, we obtain the quadratic equation $(3-\lambda)(2-\lambda)=0$, yielding roots $\lambda=2$ and $\lambda=3$.

10:14To find the specific eigenvector associated with $\lambda=2$, we substitute this value back into the modified matrix. We solve the system [[1, 1], [0, 0]][x, y]^T = [0, 0]^T. Geometrically, this reveals that all vectors lying on the line spanned by [-1, 1] are stretched by exactly a factor of 2 under the original transformation.

10:46Next, consider a 90-degree rotation matrix [[0, -1], [1, 0]]. Visually, every vector is rotated off its own span, suggesting no real eigenvectors exist. Algebraically, the characteristic polynomial becomes $\lambda^2 + 1 = 0$, which has only imaginary roots ($i$ and $-i$), confirming there are no real eigenvalues or eigenvectors.

11:35Now examine a shear transformation defined by [[1, 1], [0, 1]]. Here, all vectors along the x-axis remain fixed in place. These are eigenvectors with eigenvalue $\lambda=1$. Interestingly, while there is only one unique eigenvalue, there is an entire line's worth of corresponding eigenvectors.

13:03What happens if our standard basis vectors themselves are eigenvectors? If $\hat{i}$ scales by -1 and $\hat{j}$ scales by 2, the resulting matrix [[-1, 0], [0, 2]] is diagonal. In such matrices, the diagonal entries directly represent the scaling factors (eigenvalues) applied to each dimension independently.

14:35Finally, we discuss changing coordinates to utilize this property. Even if a matrix isn't initially diagonal, like our first example [[3, 1], [0, 2]], we can construct a 'change of basis' matrix using its eigenvectors as columns. Sandwiching the original matrix between this change-of-basis matrix and its inverse transforms it into a diagonal form containing just the eigenvalues. This makes computing powers like $A^{100}$ trivial. This step requires a full independent eigenbasis and cannot be applied to every matrix without that condition.

Knowledge cards