Why the Spectral Theorem Guarantees Orthogonal Eigenvectors

This article provides a complete overview of why the Spectral Theorem guarantees that every real symmetric matrix has orthogonal eigenvectors. By examining the algebraic relationship between matrix transposition and the dot product, this guide breaks down how distinct eigenvalues naturally force eigenvectors to be perpendicular, and how invariant subspaces ensure orthogonality even when eigenvalues repeat.

The Secret Engine: Symmetry and the Dot Product

To understand the Spectral Theorem, you must first understand how a real symmetric matrix interacts with the dot product.

A matrix \(A\) is symmetric if it is equal to its own transpose (\(A = A^T\)), meaning its entries are mirrored across the main diagonal. In linear algebra, the dot product of two vectors \(u\) and \(v\) can be written using matrix multiplication as \(u^T v\).

When we apply a symmetric matrix \(A\) to a vector inside a dot product, something special happens:

\[(Au) \cdot v = (Au)^T v = u^T A^T v\]

Because \(A = A^T\), we can substitute \(A\) back in:

\[u^T A^T v = u^T (Av) = u \cdot (Av)\]

This gives us the fundamental identity of symmetric matrices: \((Au) \cdot v = u \cdot (Av)\). You can move a symmetric matrix from one side of a dot product to the other without changing the result. This single property is the engine that drives the Spectral Theorem.

Case 1: Distinct Eigenvalues Force Orthogonality

Let us assume a real symmetric matrix \(A\) has two different eigenvalues, \(\lambda_1\) and \(\lambda_2\), with corresponding eigenvectors \(v_1\) and \(v_2\). By the definition of eigenvalues and eigenvectors:

  • \(Av_1 = \lambda_1 v_1\)
  • \(Av_2 = \lambda_2 v_2\)

We want to prove that \(v_1\) and \(v_2\) are orthogonal, which means their dot product must equal zero (\(v_1 \cdot v_2 = 0\)).

Let us evaluate the expression \((Av_1) \cdot v_2\) in two different ways. First, by simply substituting \(Av_1 = \lambda_1 v_1\):

\[(Av_1) \cdot v_2 = (\lambda_1 v_1) \cdot v_2 = \lambda_1 (v_1 \cdot v_2)\]

Second, by using our symmetric matrix identity to move \(A\) to the other side of the dot product:

\[(Av_1) \cdot v_2 = v_1 \cdot (Av_2) = v_1 \cdot (\lambda_2 v_2) = \lambda_2 (v_1 \cdot v_2)\]

Since both expressions equal \((Av_1) \cdot v_2\), they must equal each other:

\[\lambda_1 (v_1 \cdot v_2) = \lambda_2 (v_1 \cdot v_2)\]

Subtracting the right side from the left gives:

\[(\lambda_1 - \lambda_2)(v_1 \cdot v_2) = 0\]

Because we started with the assumption that the eigenvalues are distinct (\(\lambda_1 \neq \lambda_2\)), the term \((\lambda_1 - \lambda_2)\) cannot be zero. Therefore, for the equation to hold true, \(v_1 \cdot v_2\) must equal zero.

This proves that eigenvectors corresponding to distinct eigenvalues of a symmetric matrix are strictly guaranteed to be orthogonal.

Case 2: Repeated Eigenvalues and Invariant Subspaces

A common question arises when a symmetric matrix has repeated eigenvalues (where \(\lambda_1 = \lambda_2\)). In the equation \((\lambda_1 - \lambda_2)(v_1 \cdot v_2) = 0\), the first term becomes zero, which means \(v_1 \cdot v_2\) doesn't have to be zero. Why doesn't the matrix become defective and run out of orthogonal eigenvectors?

The guarantee holds because of invariant subspaces.

Every real symmetric matrix has at least one real eigenvalue and one corresponding eigenvector, \(v_1\). Now, imagine the entire space of vectors that are completely orthogonal to \(v_1\). Let us call this perpendicular space \(S\).

If we take any vector \(w\) inside this perpendicular space \(S\) (meaning \(w \cdot v_1 = 0\)) and multiply it by our symmetric matrix \(A\), where does the resulting vector \(Aw\) land? We can check if \(Aw\) is still perpendicular to \(v_1\) by taking their dot product:

\[(Aw) \cdot v_1 = w \cdot (Av_1) = w \cdot (\lambda_1 v_1) = \lambda_1 (w \cdot v_1)\]

Since \(w \cdot v_1 = 0\), the entire expression equals zero. Therefore, \((Aw) \cdot v_1 = 0\).

This proves that multiplying any vector in \(S\) by \(A\) keeps that vector inside \(S\). Because \(A\) traps vectors within this perpendicular space, we can treat \(S\) as a brand-new, smaller symmetric system and find a new eigenvector \(v_2\) inside of it. Since \(v_2\) lives inside \(S\), it is automatically orthogonal to \(v_1\), regardless of whether its eigenvalue is the same as \(\lambda_1\) or different.

By repeating this process—peeling off one eigenvector at a time and looking into the remaining perpendicular space—we are guaranteed to find a full set of \(n\) mutually orthogonal eigenvectors for an \(n \times n\) matrix. If multiple eigenvectors share the same eigenvalue, they form a flat plane or hyperplane (an eigenspace) where every vector is an eigenvector, allowing us to freely pick perpendicular axes within that space using the Gram-Schmidt process.

The Geometric Intuition

Algebra aside, the Spectral Theorem makes intuitive sense when you view matrices as geometric transformations.

A general matrix can stretch, compress, shear, and rotate space. Rotations are the reason some matrices fail to have real eigenvectors—if a matrix rotates every vector in a plane, no vector stays on its original span.

However, the condition \(A = A^T\) algebraically cancels out any rotational or shearing asymmetry. If row \(i\) influences column \(j\) in the exact same way column \(i\) influences row \(j\), all twisting forces are balanced. Consequently, a real symmetric matrix represents a pure deformation—a simple stretching or compressing of space along a set of perpendicular axes. Those natural, perpendicular axes of stretch are the orthogonal eigenvectors guaranteed by the Spectral Theorem.