The Matrix Square Root

For an ordinary number, a square root undoes squaring: the square root of 9 is 3 because 3 times 3 is 9. A matrix square root does the same job for a whole matrix — it's a matrix $B$ such that $B \times B$ reproduces the original matrix $A$. This matters in finance because a covariance matrix describes how a basket of assets moves together, but simulating correlated returns needs a way to turn independent random numbers into correlated ones, and the matrix square root is exactly the tool that does that conversion.

For a symmetric, positive-definite matrix like a covariance matrix, the square root is built from its eigendecomposition: write $A$ as $Q \Lambda Q^T$, where $Q$ holds the eigenvectors (the directions the matrix stretches space along) and $\Lambda$ is a diagonal matrix of eigenvalues (how much stretching happens along each direction). Taking the ordinary square root of each eigenvalue and reassembling gives $B = Q \Lambda^{1/2} Q^T$, and $B \times B$ recovers $A$ exactly. As a small check, a matrix with eigenvalues 4 and 9 along two independent directions has a square root with eigenvalues 2 and 3 along those same directions.

In practice, a Cholesky decomposition is usually preferred for simulation because it's cheaper to compute, but the symmetric matrix square root is the version used when the result itself needs to stay symmetric, such as in some risk-parity and portfolio-construction calculations.