The Covariance Matrix

The covariance matrix Σ is the single object every risk calculation in this course leans on — portfolio variance, the optimizer’s solution, the Sharpe ratio, all of it routes through Σ. So before we can optimize, we have to really understand this table: how it’s assembled, the structural rules it must obey (symmetry, positive semidefiniteness), how monstrously large it grows as you add assets, and the property — conditioning — that quietly determines whether your optimizer produces sane weights or noise-driven nonsense. Get Σ right and the rest is arithmetic; get it wrong and no amount of clever optimization can rescue you.

Building Σ: variances on the diagonal, covariances off it

Analogy. Picture a spreadsheet with the same list of assets down the rows and across the columns. Each cell answers: “how do the row asset and the column asset move together?” The cells on the main diagonal compare an asset with itself — that’s just its own variance — and every off-diagonal cell measures a genuine pair.

Definition. For $N$ assets with returns $r_1, \dots, r_N$, the covariance matrix $\Sigma$ is the $N \times N$ table with entries $\Sigma_{ij} = \mathrm{Cov}(r_i, r_j) = \mathbb{E}\big[(r_i - \mu_i)(r_j - \mu_j)\big].$ On the diagonal ($i = j$) this is the variance $\Sigma_{ii} = \sigma_i^2$. Off the diagonal it factors into volatilities and correlation: $\Sigma_{ij} = \sigma_i\,\sigma_j\,\rho_{ij},$ where $\rho_{ij} \in [-1, 1]$ is the correlation. The correlation matrix is the same table with the volatilities stripped out — all ones on the diagonal — so it’s the pure co-movement skeleton, and $\Sigma = D R D$ where $D$ is the diagonal matrix of volatilities and $R$ the correlation matrix.

Worked example (3 assets). Suppose $\sigma = (20\%,\ 10\%,\ 15\%)$ and the correlations are $\rho_{12} = 0.3$, $\rho_{13} = -0.2$, $\rho_{23} = 0.5$. Then each covariance is volatility × volatility × correlation:

• $\Sigma_{12} = 0.20 \times 0.10 \times 0.3 = 0.006$.
• $\Sigma_{13} = 0.20 \times 0.15 \times (-0.2) = -0.006$.
• $\Sigma_{23} = 0.10 \times 0.15 \times 0.5 = 0.0075$.

So the full matrix is $\Sigma = \begin{pmatrix} 0.0400 & 0.0060 & -0.0060 \\ 0.0060 & 0.0100 & 0.0075 \\ -0.0060 & 0.0075 & 0.0225 \end{pmatrix}.$ Read it: asset 1 is the most volatile (diagonal 0.04 = 0.20²), assets 1 and 3 are a diversifying pair (negative covariance), and assets 2 and 3 move together (positive).

Positive semidefiniteness: the rule Σ can’t break

Analogy. A portfolio’s variance is a real, physical quantity — the spread of its returns. It can be zero (a perfectly hedged book) but it can never be negative: nothing is “less risky than perfectly certain.” A valid covariance matrix is exactly one that guarantees this for every possible portfolio you could build from the assets.

Definition. A symmetric matrix $\Sigma$ is positive semidefinite (PSD) if for every weight vector $w$, $w^\top \Sigma\, w \ge 0.$ Since $w^\top \Sigma w$ is the portfolio variance, PSD is just the statement “no portfolio can have negative variance.” Equivalently, all of $\Sigma$‘s eigenvalues are $\ge 0$. A matrix is positive definite (PD) if the inequality is strict ($> 0$) for all nonzero $w$ — i.e. no nonzero portfolio has exactly zero variance, which is what you need for $\Sigma$ to be invertible.

Why a matrix can fail to be PSD. If you estimate correlations carelessly — say from data where different pairs were measured over different time windows, or you hand-tweak a few cells — you can produce a symmetric matrix with a negative eigenvalue. That implies some portfolio has negative variance, which is nonsense, and the optimizer will exploit it ruthlessly: it’ll pour weight into that “free risk reduction” and hand you an absurd answer.

Worked example — spotting a broken matrix. Consider the correlation matrix claiming three assets are each strongly positively correlated except one pair is strongly negative: $R = \begin{pmatrix} 1 & 0.9 & 0.9 \\ 0.9 & 1 & -0.9 \\ 0.9 & -0.9 & 1 \end{pmatrix}.$ This is logically impossible: if asset 1 is 90% correlated with both 2 and 3, then 2 and 3 cannot be strongly negatively correlated — they’re both shadowing asset 1. Plug it in and you’d find a negative eigenvalue, flagging it as not PSD. The transitivity of correlation isn’t a suggestion; it’s enforced by the PSD constraint.

The dimensionality explosion

Analogy. Adding one more asset to your universe doesn’t add one more number to estimate — it adds a whole new row and column of co-movements with everything already there. The bookkeeping grows like the area of the matrix, not its length.

The count. An $N \times N$ covariance matrix has $N^2$ cells, but symmetry means the distinct numbers are $\underbrace{N}_{\text{variances}} + \underbrace{\frac{N(N-1)}{2}}_{\text{covariances}} = \frac{N(N+1)}{2}.$

Worked examples.

For the S&P 500, you must pin down over 125,000 numbers. And here’s the killer: you typically have only a few years of daily data — maybe 750–1,250 observations per asset. You are estimating more parameters than you have independent observations. When the number of parameters approaches or exceeds the sample size, the estimated Σ becomes wildly noisy and, often, singular (non-invertible) — which torpedoes the optimizer. This curse of dimensionality is the deep reason shrinkage (lesson 5) exists.

Conditioning: why a near-singular Σ wrecks the optimizer

Analogy. Imagine balancing a pencil on its tip versus laying it flat. Flat, a small nudge barely moves it (well-conditioned). On its tip, the tiniest breath sends it flying (ill-conditioned). A covariance matrix is “balanced on its tip” when some assets are nearly redundant — highly correlated — so the matrix is almost singular.

Definition. The condition number of $\Sigma$ is the ratio of its largest to smallest eigenvalue, $\kappa(\Sigma) = \frac{\lambda_{\max}}{\lambda_{\min}}.$ A large $\kappa$ means $\Sigma$ is ill-conditioned: it’s close to singular, $\lambda_{\min}$ is tiny, and the inverse $\Sigma^{-1}$ (which the optimizer needs) has enormous entries scaled by $1/\lambda_{\min}$. Small errors in the inputs get magnified by roughly $\kappa$ in the output weights.

Why correlated assets cause it. If two assets have correlation near 1, they’re nearly the same asset — the matrix has a direction with almost no variance, so $\lambda_{\min} \approx 0$ and $\kappa$ blows up. The optimizer, seeing two near-identical assets with slightly different estimated returns, will take a giant long position in one and a giant short in the other to harvest the tiny apparent spread — a position that is pure noise and will reverse out of sample.

Worked intuition. Two assets, both 20% vol, correlation $\rho$. The eigenvalues of the correlation matrix are $1+\rho$ and $1-\rho$, so $\kappa = \frac{1+\rho}{1-\rho}$. At $\rho = 0$, $\kappa = 1$ (perfectly balanced). At $\rho = 0.9$, $\kappa = \frac{1.9}{0.1} = 19$. At $\rho = 0.99$, $\kappa = \frac{1.99}{0.01} = 199$. As $\rho \to 1$, $\kappa \to \infty$ and the matrix becomes singular — there’s no unique optimal split between two identical assets, so the optimizer’s answer is meaningless.

Putting it together

The covariance matrix Σ is built from volatilities and correlations ($\Sigma_{ij} = \sigma_i\sigma_j\rho_{ij}$), with variances on the diagonal and co-movements off it. It must be symmetric (covariance ignores order) and positive semidefinite ($w^\top\Sigma w \ge 0$, so no portfolio has negative variance — equivalently all eigenvalues $\ge 0$); a negative eigenvalue means the matrix is broken and the optimizer will abuse it. The number of distinct parameters grows like $N^2$ — over 125,000 for the S&P 500 — so you routinely estimate more numbers than you have data, producing a noisy, often singular sample estimate. And conditioning (the eigenvalue ratio $\kappa$) decides everything downstream: highly correlated, near-redundant assets make Σ nearly singular, so $\Sigma^{-1}$ has huge entries that amplify tiny input errors into wild, unstable weights. That fragility is the bridge to the next lesson — and the reason shrinkage and constraints exist.

Next — mean-variance mechanics: with Σ in hand, we actually solve for the weights. We’ll derive the minimum-variance and maximum-Sharpe (tangency) portfolios, see the Lagrangian and closed-form solution that invokes $\Sigma^{-1}$, and work a full numeric example by hand — turning the abstract optimization problem into concrete numbers.