LINEAR-ALGEBRA · Interactive Practice | Unit 24 · Video 1

The Spectral Theorem: Why Symmetric Matrices Are Special

IKey Formulas

Formula	Name	Description
$A = A^T$	Symmetric matrix	Equal to its own transpose
$A = Q \Lambda Q^T$	Spectral decomposition	$Q$ orthonormal, $\Lambda$ diagonal real
$Q^T Q = I$ , so $Q^{-1} = Q^T$	Orthogonal matrix	Inverse equals transpose
$A = \lambda_1 q_1 q_1^T + \lambda_2 q_2 q_2^T + \cdots$	Spectral sum	Weighted projections onto eigen-axes

IIStretching Along Perpendicular Axes

A symmetric matrix sends the unit circle to an ellipse whose axes lie along perpendicular eigenvectors.

Matrix

💡 Challenge: find an input direction where $Av$ is parallel to $v$ — those eigenvector directions stay perpendicular for every symmetric matrix.

IIIA Sum of Rank-One Projections

Every symmetric matrix is $\lambda_1 q_1 q_1^T + \lambda_2 q_2 q_2^T$ — eigenvalues scaling perpendicular projections.

λ₁ 4 λ₂ 2

💡 Challenge: set $\lambda_2 = 0$ — the ellipse collapses to a line and $A$ becomes rank-one, projecting the whole plane onto the $q_1$ axis.

IVSymmetric vs. Non-Symmetric Distortion

Adding an antisymmetric shear tilts the eigenvectors and injects rotation into a pure stretch.

Shear amount 0

Symmetric A = [3 1; 1 3]

💡 Symmetry kills rotation — only perpendicular stretching survives. Past a shear of $0.5$ the eigenvalues turn complex and the map becomes purely rotational.

VQuiz Questions

Question 1

Which of the following matrices is symmetric (i.e., satisfies $A = A^T$ )?

✅ Correct! The off-diagonal entries are both 3, so $A = A^T$ .

❌ Not quite. A matrix is symmetric when entry $(i,j)$ equals entry $(j,i)$ . The off-diagonals must match exactly.

Show solution

Solution:

A matrix is symmetric when $A_{ij} = A_{ji}$ for every off-diagonal pair. Check each option:

$\begin{pmatrix} 2 & 3 \\ 3 & 5 \end{pmatrix}$ : off-diagonals are $3$ and $3$ — symmetric.
$\begin{pmatrix} 2 & 3 \\ -3 & 5 \end{pmatrix}$ : $3 \neq -3$ — not symmetric (in fact, antisymmetric in the off-diagonal).
$\begin{pmatrix} 2 & 3 \\ 4 & 5 \end{pmatrix}$ : $3 \neq 4$ — not symmetric.
$\begin{pmatrix} 0 & 1 \\ -1 & 0 \end{pmatrix}$ : $1 \neq -1$ — not symmetric (it's a 90° rotation).

The first option is correct.

Question 2

Let $Q$ be a matrix whose columns are orthonormal (mutually perpendicular unit vectors). Which equation is always true?

✅ Correct! Orthonormal columns mean $Q^T Q = I$ , so the transpose is the inverse.

❌ Not quite. Compute $Q^T Q$ entry by entry: it is the dot products of columns.

Show solution

Solution:

The $(i, j)$ entry of $Q^T Q$ is the dot product of column $i$ of $Q$ with column $j$ of $Q$ .

Same column ( $i = j$ ): unit vector dotted with itself gives $1$ .
Different columns ( $i \neq j$ ): perpendicular vectors give $0$ .

So $Q^T Q = I$ . Multiplying both sides by $Q^{-1}$ on the right gives $Q^T = Q^{-1}$ .

This is the magical property of orthogonal matrices: the inverse is free — just transpose. It's why the spectral theorem $A = Q \Lambda Q^T$ never requires computing an inverse.

Question 3

Consider the symmetric matrix $A = \begin{pmatrix} 3 & 1 \\ 1 & 3 \end{pmatrix}$ .

Its characteristic equation is $(3 - \lambda)^2 - 1 = 0$ . What are the eigenvalues?

✅ Correct! Expanding gives $\lambda^2 - 6\lambda + 8 = 0$ , factoring as $(\lambda-4)(\lambda-2)$ .

❌ Not quite. Expand the characteristic polynomial first, then factor it. Real symmetric matrices never have complex eigenvalues.

Show solution

Solution:

Expand $(3 - \lambda)^2 - 1 = 0$ :

$\lambda^2 - 6\lambda + 9 - 1 = \lambda^2 - 6\lambda + 8 = 0$

Factor:

$(\lambda - 4)(\lambda - 2) = 0$

So $\lambda_1 = 4$ and $\lambda_2 = 2$ . Both real, as guaranteed for any real symmetric matrix.

The corresponding orthonormal eigenvectors are $q_1 = \tfrac{1}{\sqrt{2}}(1, 1)$ and $q_2 = \tfrac{1}{\sqrt{2}}(1, -1)$ , which are perpendicular ( $q_1 \cdot q_2 = 0$ ).

Question 4

True or False: Every real symmetric matrix is geometrically just a stretching along perpendicular axes — there is no rotation or shear, only pure scaling in the right rotated coordinate frame.

✅ Correct! This is the geometric content of the spectral theorem $A = Q \Lambda Q^T$ — rotate, scale, rotate back.

❌ Not quite. Re-examine the factorization $A = Q \Lambda Q^T$ : $Q$ is a rotation, $\Lambda$ is pure scaling, and $Q^T$ undoes the rotation. The eigenvectors are perpendicular axes.

Show solution

Solution:

True. The spectral theorem $A = Q \Lambda Q^T$ tells us exactly this geometric story:

$Q^T$ : rotate space so the eigenvectors become the standard axes.
$\Lambda$ : scale along each axis by the corresponding eigenvalue.
$Q$ : rotate back to the original frame.

There is no shear, no rotation in the net effect on the eigen-axes — just scaling. This is why the unit circle always maps to an ellipse whose principal axes are perpendicular (and aligned with the eigenvectors).

This insight is the foundation of:

PCA: covariance matrices are symmetric → principal components are perpendicular axes of variance.
Optimization: the Hessian is symmetric → its eigenvalues tell you curvatures along perpendicular directions.
Quantum mechanics: observables are symmetric (Hermitian) → measurable values are real eigenvalues.

Solved: 0 / 4