Eigenvalues and Diagonalization

How eigenvalues reveal invariant directions of a linear map and why diagonalization turns repeated matrix actions into simple scalar dynamics.

Modified

April 26, 2026

Keywords

eigenvalues, eigenvectors, diagonalization, spectral methods, matrix powers, symmetric matrices

1 Role

This page is the first real entry point into spectral thinking.

It explains how some directions are special for a linear map, why those directions are encoded by eigenvectors, and how diagonalization turns a hard matrix action into independent scalar actions.

2 First-Pass Promise

Read this page after Orthogonality and Least Squares.

If you stop here, you should still understand:

what an eigenvector and eigenvalue mean geometrically
what diagonalization is saying
why diagonalization helps with powers of a matrix
why spectral structure matters for dynamics, graphs, and later SVD

3 Why It Matters

This topic matters because eigenvalues and eigenvectors tell you what a linear map is really doing along its special directions:

in repeated dynamics, they tell you which modes grow, decay, or stay fixed
in graph methods, they describe frequency-like structure and connectivity patterns
in numerical methods, they control conditioning, convergence, and stability behavior
in SVD, they reappear through the eigenstructure of \(A^\top A\)

So this page is where matrix algebra starts turning into modes, spectra, and long-term behavior.

4 Prerequisite Recall

a matrix represents a linear map once bases are chosen
if \(A = S\Lambda S^{-1}\), then \(A\) and \(\Lambda\) represent the same map in different coordinates
orthogonal changes of coordinates preserve Euclidean geometry

5 Intuition

Most vectors change direction when a matrix acts on them.

But some directions are special: the matrix only stretches, shrinks, or flips them. Those are eigenvectors.

So if

\[ Av = \lambda v, \]

then the vector \(v\) does not get turned into a new direction. It only gets scaled by \(\lambda\).

That matters because if you can find enough such directions, the whole matrix becomes simpler. In the right coordinates, the matrix stops mixing directions and becomes diagonal.

Diagonalization is exactly that change of viewpoint:

find the coordinates in which the map acts independently on each special direction.

Once that happens, powers of the matrix become easy, because diagonal matrices are easy to power.

6 Formal Core

Definition 1 (Definition) Let \(A \in \mathbb{R}^{n \times n}\).

A nonzero vector \(v \in \mathbb{R}^n\) is an eigenvector of \(A\) with eigenvalue \(\lambda \in \mathbb{R}\) if

\[ Av = \lambda v. \]

The condition \(v \neq 0\) is essential.

Proposition 1 (Key Statement) If a matrix \(A \in \mathbb{R}^{n \times n}\) has \(n\) linearly independent eigenvectors \(v_1,\dots,v_n\) with eigenvalues \(\lambda_1,\dots,\lambda_n\), then with

\[ S = \begin{bmatrix} \vert & & \vert \\ v_1 & \cdots & v_n \\ \vert & & \vert \end{bmatrix}, \qquad \Lambda = \operatorname{diag}(\lambda_1,\dots,\lambda_n), \]

we have

\[ A = S\Lambda S^{-1}. \]

Equivalently,

\[ S^{-1}AS = \Lambda. \]

This is called diagonalization.

Proposition 2 (Powers Become Easy) If \(A = S\Lambda S^{-1}\), then for every positive integer \(k\),

\[ A^k = S\Lambda^k S^{-1}. \]

So repeated application of \(A\) is controlled by repeated powers of its eigenvalues.

Proposition 3 (Symmetric Preview) If \(A\) is symmetric, then its eigenvectors can be chosen orthonormal, so the diagonalization takes the especially clean form

\[ A = Q\Lambda Q^\top. \]

This is the first glimpse of the spectral theorem and is one reason symmetric matrices are so important.

7 Worked Example

Consider the matrix

\[ P = \begin{bmatrix} 0.8 & 0.2 \\ 0.2 & 0.8 \end{bmatrix}. \]

You can think of this as a two-state mixing matrix: each step keeps most of the current value and exchanges a smaller part with the other coordinate.

First find eigenvectors by inspection.

The vector

\[ v_1 = \begin{bmatrix} 1 \\ 1 \end{bmatrix} \]

satisfies

\[ Pv_1 = \begin{bmatrix} 1 \\ 1 \end{bmatrix} = 1 \cdot v_1, \]

so \(v_1\) is an eigenvector with eigenvalue \(1\).

The vector

\[ v_2 = \begin{bmatrix} 1 \\ -1 \end{bmatrix} \]

satisfies

\[ Pv_2 = \begin{bmatrix} 0.6 \\ -0.6 \end{bmatrix} = 0.6 \, v_2, \]

so \(v_2\) is an eigenvector with eigenvalue \(0.6\).

These vectors are independent, so \(P\) is diagonalizable. If we normalize them, we get the orthonormal matrix

\[ Q = \frac{1}{\sqrt{2}} \begin{bmatrix} 1 & 1 \\ 1 & -1 \end{bmatrix}, \qquad \Lambda = \begin{bmatrix} 1 & 0 \\ 0 & 0.6 \end{bmatrix}. \]

Then

\[ P = Q\Lambda Q^\top. \]

Now powers become easy:

\[ P^k = Q \begin{bmatrix} 1^k & 0 \\ 0 & 0.6^k \end{bmatrix} Q^\top. \]

If we start from

\[ x_0 = \begin{bmatrix} 1 \\ 0 \end{bmatrix}, \]

then

\[ x_k = P^k x_0 = \frac{1}{2} \begin{bmatrix} 1 \\ 1 \end{bmatrix} + \frac{0.6^k}{2} \begin{bmatrix} 1 \\ -1 \end{bmatrix}. \]

So as \(k\) grows, the second mode fades away and the state approaches

\[ \frac{1}{2} \begin{bmatrix} 1 \\ 1 \end{bmatrix}. \]

This example shows the main picture:

eigenvectors reveal invariant directions
eigenvalues tell how strongly each mode persists
diagonalization separates the dynamics into independent pieces

8 Computation Lens

In practice, eigenvalue problems usually appear in one of these forms:

Find special directions Solve \((A-\lambda I)v=0\) for nonzero \(v\).
Understand repeated action Use diagonalization to study \(A^k\), matrix exponentials, or iterative updates.
Measure dominant behavior The eigenvalues with largest magnitude often control long-term growth or decay.

For symmetric matrices, computation is especially well behaved because orthogonal diagonalization keeps geometry stable and interpretable.

That is why symmetric or positive semidefinite matrices appear constantly in optimization, covariance analysis, graph Laplacians, and later SVD.

9 Application Lens

In many engineering and ML settings, we repeatedly apply a linear update:

\[ x_{t+1} = Ax_t. \]

Diagonalization turns this into mode-by-mode evolution. Each eigenvector gives a mode, and each eigenvalue tells whether that mode grows, decays, oscillates, or stays fixed.

This viewpoint appears in:

consensus and diffusion on graphs
Markov-style mixing processes
stability analysis of linear dynamical systems
spectral graph methods, where Laplacian eigenvectors define meaningful graph modes

So eigenvalues are not just a textbook computation. They are the language of persistent structure under repeated transformation.

10 Stop Here For First Pass

If you can now explain:

what an eigenvector and eigenvalue mean
when a matrix is diagonalizable
why diagonalization helps with powers of a matrix
why dominant eigenvalues control long-term behavior

then this page has done its main job.

11 Go Deeper

If you want more after the main page:

Proof: Distinct Eigenvalues Give Independent Eigenvectors
Application: Spectral Modes in Consensus and Graphs
Visual intuition: Computation Lab: Matrix Powers and Spectral Modes
Practice: Exercises: Eigenvalues and Diagonalization

12 Optional Paper Bridge

13 Optional After First Pass

If you want more practice before moving on:

diagonalize a simple \(2 \times 2\) matrix with distinct eigenvalues
compare a diagonalizable matrix with one that has too few independent eigenvectors
study how powers of a matrix depend on whether \(|\lambda|<1\), \(|\lambda|=1\), or \(|\lambda|>1\)
continue to SVD and Low-Rank Approximation when you want spectral structure to become a matrix-approximation tool

14 Common Mistakes

forgetting that the zero vector is never an eigenvector
thinking every square matrix is diagonalizable
confusing eigenvalues with diagonal entries of the original matrix
forgetting that repeated eigenvalues do not automatically prevent diagonalization
assuming orthogonal diagonalization works for every matrix instead of just special classes such as symmetric matrices

15 Exercises

Show that

\[ A = \begin{bmatrix} 2 & 1 \\ 0 & 3 \end{bmatrix} \]

has eigenvalues \(2\) and \(3\), and find one eigenvector for each.
Let

\[ B = \begin{bmatrix} 2 & 0 \\ 0 & 0.5 \end{bmatrix}. \]

Describe what happens to \(B^k x\) as \(k \to \infty\) for a general vector \(x\).
Explain in words why having a basis of eigenvectors is exactly what makes diagonalization possible.

16 Sources and Further Reading

MIT 18.06: Lecture 21, Eigenvalues and Eigenvectors - First pass - a clean official introduction to what eigenvectors mean and how to compute them. Checked 2026-04-24.
MIT 18.06SC: Lecture 22, Diagonalization and Powers of A - First pass - official source for the diagonalization viewpoint and why it simplifies matrix powers. Checked 2026-04-24.
Hefferon, Linear Algebra - Second pass - strong self-study text with more exercises and a gradual treatment of eigenvalues and diagonalization. Checked 2026-04-24.
A Comprehensive Survey on Spectral Clustering with Graph Structure Learning - Second pass - current survey showing how spectral ideas scale into graph-based ML. Checked 2026-04-24.
Piecewise Constant Spectral Graph Neural Network - Paper bridge - a current example of eigenvalue-aware filtering inside graph learning. Checked 2026-04-24.

Sources checked online on 2026-04-24:

MIT 18.06 Lecture 21: Eigenvalues and Eigenvectors
MIT 18.06SC Lecture 22: Diagonalization and Powers of A
Hefferon, Linear Algebra
arXiv survey on spectral clustering with graph structure learning
TMLR 2025 paper on Piecewise Constant Spectral Graph Neural Network