The Spectral Theorem for Symmetric Matrices

Multiply a vector by most matrices and it gets stretched and rotated — the output points in some new, unpredictable direction. But for a special and enormously common class of matrices — symmetric ones, which is what every covariance matrix and every "similarity" or "distance" matrix in finance and statistics happens to be — something remarkable is guaranteed: there always exists a set of directions the matrix does not rotate at all, only stretches, and those directions are always mutually perpendicular. That guarantee is the spectral theorem, and it's the reason principal component analysis, risk decomposition, and half of multivariate statistics work at all.

The analogy: kneading dough along fixed grain lines

Picture a block of dough with grain running in two perpendicular directions, like woven fabric. Now knead it — push and pull — but only ever along those grain lines, never twisting them. The dough stretches unevenly (more along one grain than the other) but the grain lines themselves stay put; they just get longer or shorter. Any other line drawn on the dough, diagonal to the grain, does rotate as you knead. A symmetric matrix is a kneading operation with this special property built in: there are always some perpendicular "grain lines" — the eigenvectors — that survive the kneading without rotating, only stretching by a factor called the eigenvalue.

Writing it down

Let $A$ be a real, symmetric matrix, meaning $A^\top = A$ — the entry in row $i$, column $j$ equals the entry in row $j$, column $i$. A vector $v$ is an eigenvector with eigenvalue $\lambda$ if

$$Av = \lambda v .$$

In words: multiplying $v$ by $A$ gives back the same direction $v$, just scaled by the number $\lambda$ — no rotation, pure stretch (or shrink, or flip, if $\lambda$ is negative).

The spectral theorem states that for any symmetric $A$ of size $n \times n$, there exist $n$ eigenvectors that are mutually orthogonal (perpendicular, zero dot product between any pair) and can be scaled to unit length, and their eigenvalues are always real numbers (never involving $\sqrt{-1}$, unlike general matrices). Collecting the eigenvectors as columns of a matrix $Q$ and the eigenvalues on the diagonal of a matrix $\Lambda$, the whole matrix factors as

$$A = Q\Lambda Q^\top .$$

In words: $A$ is equivalent to rotating into the grain-line coordinate system ($Q^\top$), stretching along each grain line by its eigenvalue ($\Lambda$), then rotating back ($Q$). Because $Q$'s columns are orthonormal, $Q^\top = Q^{-1}$ — rotating back is literally the reverse of rotating in, with no distortion, which is what makes this decomposition so clean compared to a general matrix.

This explorer draws the unit circle transformed by a 2x2 matrix, with its eigenvectors overlaid. For a symmetric matrix, the two eigenvector arrows sit at exactly 90 degrees to each other, and they point straight along the axes of the resulting ellipse — that alignment, guaranteed by the spectral theorem, is what you should watch for as you adjust the matrix entries.

Worked example 1: a 2x2 covariance matrix, by hand

Take $A = \begin{pmatrix}5 & 2\\2 & 2\end{pmatrix}$ — symmetric, and a plausible (scaled) covariance matrix. Find eigenvalues by solving $\det(A-\lambda I)=0$:

$$(5-\lambda)(2-\lambda) - 4 = 0 \;\Rightarrow\; \lambda^2 - 7\lambda + 6 = 0 \;\Rightarrow\; (\lambda-6)(\lambda-1)=0 .$$

So $\lambda_1 = 6$, $\lambda_2 = 1$. For $\lambda_1=6$: solve $(A-6I)v=0$, i.e. $\begin{pmatrix}-1&2\\2&-4\end{pmatrix}v=0$, giving $v = (2,1)$ (unnormalised) — check: $-1(2)+2(1)=0$. Normalise: length $=\sqrt{5}$, so $v_1 = (2/\sqrt5, 1/\sqrt5) = (0.894, 0.447)$.

For $\lambda_2=1$: solve $(A-I)v=0$, i.e. $\begin{pmatrix}4&2\\2&1\end{pmatrix}v=0$, giving $v=(1,-2)$ (unnormalised) — check: $4(1)+2(-2)=0$. Normalise: length $=\sqrt5$, so $v_2=(1/\sqrt5,-2/\sqrt5)=(0.447,-0.894)$.

Sanity check the perpendicularity: $v_1\cdot v_2 = (0.894)(0.447)+(0.447)(-0.894) = 0.400-0.400=0$ — exactly orthogonal, as guaranteed. And $\lambda_1+\lambda_2 = 7 = A_{11}+A_{22}=5+2$ (the trace), a fast arithmetic check available on every 2x2 symmetric matrix.

Worked example 2: reconstructing $A$ and reading a risk decomposition

Verify $A=Q\Lambda Q^\top$ with $Q = \begin{pmatrix}0.894 & 0.447\\0.447&-0.894\end{pmatrix}$, $\Lambda=\begin{pmatrix}6&0\\0&1\end{pmatrix}$. Compute $Q\Lambda$ first: multiply each column of $Q$ by its eigenvalue — column 1 by 6, column 2 by 1: $Q\Lambda = \begin{pmatrix}5.364 & 0.447\\2.682&-0.894\end{pmatrix}$. Then $(Q\Lambda)Q^\top$: row 1 of $Q\Lambda$ dotted with each column of $Q^\top$ (which are the rows of $Q$) gives row 1 of the result: $5.364(0.894)+0.447(0.447) = 4.795+0.200=4.995\approx5.00$, and $5.364(0.447)+0.447(-0.894)=2.398-0.400=1.999\approx2.00$. That reproduces $A$'s first row $(5,2)$, confirming the factorisation.

Now read it as a risk decomposition: if $A$ is a covariance matrix, total variance (the trace, $5+2=7$) splits into $6$ units along direction $v_1$ and $1$ unit along $v_2$. In portfolio terms, $85.7\%$ of the total risk ($6/7$) sits along one specific combination of the two original variables — the direction a principal component analysis would call the first, dominant factor.

What this means in practice

• It's the mathematical basis of principal component analysis. PCA on a covariance matrix is literally applying the spectral theorem and reading off the eigenvector with the largest eigenvalue as "the direction of greatest variance."
• It guarantees real, orthogonal risk factors exist. Risk models decompose portfolio variance into independent contributions precisely because a covariance matrix is always symmetric, so the spectral theorem always applies — no special pleading needed.
• It underlies Cholesky's cousin decompositions. Where Cholesky gives a fast triangular square root for simulation, the spectral form gives an interpretable square root ($Q\Lambda^{1/2}$) used when you specifically want to see which directions carry the risk.