2 Random Vectors and Matrices

Section 1 was conducted in subscripts. That will not scale: the designs ahead have several factors, and regression has an arbitrary number of predictors. This section rewrites the same ideas in matrix form, where a model is a single equation \(\underline {Y}=XB+\underline {\varepsilon }\) regardless of how many terms it contains.

Three tools are assembled. The algebra of matrices, ranks and inverses gives the language. The expectation and variance-covariance of a random vector give the first two moments in matrix form. Differentiating linear and quadratic forms gives the calculus needed to minimise a sum of squares in one step rather than one parameter at a time. The section closes with idempotent matrices and the distribution of quadratic forms — the result that every sum of squares in the course depends on.

Questions on this section

Stuck on something here? Ask below and it stays attached to this topic.