# About matrix $R$, what is this called: $R^TR$? What is it for?

I am doing singular value decomposition on a matrix $R$. The first step is to compute such a matrix $R^TR$. What is this matrix? A reference told me this is cross product of matrix R. I use a programming function and verified it. But what is the relation between this cross product of matrices and cross product of vectors? How can I relate them?

Another reference told me $R^TR$ is called covariance matrix. I checked its definition but I didn't find how can it be equal to the matrix multiplication result of $R^TR$. So is this really covariance matrix? Why?

• I don't think there is a name for $R^TR$. – Git Gud Apr 6 '14 at 10:26
• What is the meaning of computing $R^TR$? Why doing this? – Tyler 十三将士归玉门 Apr 6 '14 at 10:29
• Are you asking why the algorithm for computing the SVD decomposition works? – Git Gud Apr 6 '14 at 10:29
• Yeah, this is one confusion I have. – Tyler 十三将士归玉门 Apr 6 '14 at 10:35
• Then you should read a proof of it in a book and come back asking what steps you don't understand, if any. – Git Gud Apr 6 '14 at 10:46

A better term for $R^TR$ would be dot-product matrix, which is what this book chapter calls it, cf. equation (6). The matrix entries of the product $R^T R$ are exactly dot products between columns of $R$. It would be pretty hard to make a connection with the cross-product of vectors in this construction.

The usefulness of $R^T R$ lies in its being symmetric and positive semi-definite for any real matrix $R$, not necessarily square. It can well be called a covariance matrix because of its application in statistics, but the construction also appears in least squares solutions of overdetermined linear systems.

Working with the (real) symmetric matrix $R^T R$ allows us the luxury of having a full basis of eigenvectors, and the corresponding eigenvalues will be nonnegative since $R^T R$ is positive semi-definite as well (indeed, positive definite if $R$ has rank equal to its number of columns).

This amounts to a way of defining the singular values of $R$ as (principal) square roots of the eigenvalues of $R^T R$. Another term for essential the same technique is principal components analysis or PCA, which may be more commonly used in statistics and signal processing.

• Thanks yes in the book you refered it is called dot-product matrix. But in the R (programming language) $A^TB$ is defined as the matrix crossproduct (stuff.mit.edu/afs/sipb/project/r-project/lib/R/library/base/…). Why the two names are so different? Also, in this SVD tuturial[ling.ohio-state.edu/~kbaker/pubs/…, it says "$AA^T$ is a matrix whose values are the dot product of all the terms, so it is a kind of dispersion matrix of terms throughout all the documents". What does this mean "a kind of dispersion matrix"? – Tyler 十三将士归玉门 Apr 6 '14 at 14:28
• Can the dispersion matrix be defined in many ways? Because in Wikipedia it is defined in this way: en.wikipedia.org/wiki/Covariance_matrix – Tyler 十三将士归玉门 Apr 6 '14 at 14:32
• I think in the book you referred to they call it dot product matrix maybe because $R^TR$ is the dot products of all $R$'s column vectors, and $RR^T$ is the dot products of all $R$'s row vectors. I guess? – Tyler 十三将士归玉门 Apr 6 '14 at 14:37
• They also call it cross-product matrix. See here: utdallas.edu/~herve/abdi-MatrixAlgebra2010-pretty.pdf. Maybe people also call it cross product because it seems $R^TR$ generalizes the cross product from vectors to matrices in the same form. – Tyler 十三将士归玉门 Apr 6 '14 at 14:58
• Note that the cross-product of two vectors being another vector is peculiar to $\mathbb{R}^3$. I don't doubt that the cross-product construction can be generalized, but if $R^T R$ is such a generalization, I'm not aware of it. – hardmath Apr 6 '14 at 15:18

There is some analogy between constructing $R^T R$ from the matrix $R$ and constructing $\overline{z} z = |z|^2$ from the complex number $z$.