The Schur Complement

Inverting a large matrix is expensive, but often a matrix has natural blocks — say, a joint covariance matrix split into "assets I already hold" and "a new asset I'm considering adding." Recomputing the whole inverse from scratch every time the second block changes wastes the work already done on the first. The Schur complement is the algebraic tool that lets you eliminate one block, solve a smaller problem involving only the other block, and reuse that structure — it's exactly what happens when you compute a conditional variance or conditional distribution from a joint one.

An analogy: solving two equations by substitution

In basic algebra, to solve two equations in two unknowns you often solve one equation for one variable and substitute it into the other, collapsing the problem to a single equation in a single unknown. The Schur complement is the matrix-block generalization of exactly that move: given a block system, solve the first block's equation for its variable in terms of the second, substitute it into the second block's equation, and what's left — a smaller matrix expression — is the Schur complement. It captures "the second block, after accounting for everything explainable by the first."

The math, one symbol at a time

Partition a matrix into blocks:

$$M = \begin{pmatrix} A & B \\ C & D \end{pmatrix}.$$

The Schur complement of $A$ in $M$ is

$$S = D - C A^{-1} B.$$

In words: take the $D$ block, and subtract off the part of it that's "explained" by its interaction with $A$ through $B$ and $C$ — what remains, $S$, is what $D$ contributes on its own, once $A$'s influence is removed. This single object controls whether $M$ as a whole is invertible (it is, exactly when both $A$ and $S$ are invertible) and gives a formula for $M^{-1}$ built from smaller pieces, avoiding a full inversion of $M$ at once.

Worked example 1: a numeric 2x2 block system

Let $M = \begin{pmatrix}2&0&1\\0&2&1\\1&1&3\end{pmatrix}$, blocked as $A = \begin{pmatrix}2&0\\0&2\end{pmatrix}$ (top-left $2\times2$), $B = C^\top = \begin{pmatrix}1\\1\end{pmatrix}$, $D = 3$. Then $A^{-1} = \begin{pmatrix}0.5&0\\0&0.5\end{pmatrix}$, so $CA^{-1}B = \begin{pmatrix}1&1\end{pmatrix}\begin{pmatrix}0.5&0\\0&0.5\end{pmatrix}\begin{pmatrix}1\\1\end{pmatrix} = 0.5+0.5 = 1$. The Schur complement is $S = 3 - 1 = 2$. Since both $A$ (invertible, determinant 4) and $S=2$ are nonzero, $M$ is guaranteed invertible — established without ever inverting the full $3\times3$ matrix.

Worked example 2: conditional variance from a joint covariance matrix

Two assets have joint covariance $\Sigma = \begin{pmatrix}\sigma_1^2 & \sigma_{12}\\\sigma_{12} & \sigma_2^2\end{pmatrix} = \begin{pmatrix}4 & 2\\2 & 9\end{pmatrix}$. The variance of asset 2 conditional on knowing asset 1's value is exactly the Schur complement of the top-left block: $\sigma_2^2 - \sigma_{12}(\sigma_1^2)^{-1}\sigma_{12} = 9 - 2(0.25)(2) = 9 - 1 = 8$. Asset 2's unconditional variance is 9, but once you know asset 1 (which explains part of asset 2's movement through their covariance of 2), the leftover uncertainty drops to 8 — a smaller number, exactly as intuition demands, and the Schur complement is the formula that produces it directly from the joint covariance matrix.

What this means in practice

The Schur complement shows up wherever a problem naturally splits into blocks: conditional distributions of multivariate normals (risk factor models, Kalman filters), efficient updates to a covariance matrix when adding one new asset without refactoring the whole matrix, and the interior-point methods used inside portfolio optimizers. It's the computational shortcut behind the The Sherman-Morrison-Woodbury Formula, which handles low-rank updates to an inverse using the same block-elimination idea.