# A name for operator-dependent cross-product

Suppose that we have a $\\n\times p$ matrix $\mathbf{M}$. Different transformations using different column-wise operators can lead to a new $\\p\times p$ symmetric matrix $\mathbf{S}$. For example, the covariance matrix $\mathbf{C}$ can be computed using the dot product operator, whereby each value of the covariance matrix is the dot product of two columns of the original matrix $\mathbf{M}$ (divided by $n-1$):

$\mathbf{C} = { 1 \over {n-1} } \mathbf{M}^{T} \cdot \mathbf{M}$

Similarly, the correlation matrix $\mathbf{P}$ can be defined by $\mathbf{P}_{ij} = \mathrm{corr}(C_i,C_j)$ where $\mathbf{C}_i$ and $\mathbf{C}_j$ are columns of $\mathbf{C}$, and $\mathrm{corr}$ is a measure of correlation like the Pearson product-moment coefficient. In this case, the operator is the bivariate correlation coefficient.

Does this operator-dependent transformation have a name?

• In this answer I've outlined how all these matrices are the modifications of the SSCP X'X matrix. For further scholastic interest: Zegers, ten Berge (Psychometrica, 1985) unite by the single general formula 4 coefficients: identity coef, additivity coef (based on covariance coef), cosine, Pearson correlation. – ttnphns Mar 29 '16 at 13:14
• I would not recommend to believe the Wikipedia article on "Gramian matrix". Actually, "Gramian" means different things. For example, within factor analysis literature this word is long (since 1930s) used as a synonym to "positive semidefinite square matrix". In many sources, "Gram matrix" is not the same as "Gramian matrix", and so on. – ttnphns Mar 29 '16 at 13:18
• Some vector-oriented computing systems, such as APL and Mathematica, have built-in generalizations of this type. See reference.wolfram.com/language/ref/Inner.html for instance. – whuber Mar 29 '16 at 14:00
• Ismael: If you need a label for X'X matrix - there are many synonymous or equivalent, coming from people of different fields and backgrounds. Multivariate statistical data analysis is one of the oldest branch. There it is called Sums-of-Squares-and-Crossproducts matrix or simply Crossproduct matrix. I don't recommend much using words Gramian or Gram due to reasons pointed in my comment above. – ttnphns Mar 29 '16 at 18:04
• @ttnphns I think Ismael's question is more broad, he wants a name for any $f(X_i, X_j)$ matrix where $f()$ can be any (perhaps only symmetric?) function (and $X_i$ are columns of $X$); e.g. a $p\times p$ matrix of mutual information between variables or a $p\times p$ matrix of distance covariances between them, of whatever. Ismael, did I understand correctly? If so, then I don't think there is a general name for that. – amoeba Mar 29 '16 at 21:58

My understanding of the question is that it asks for a name for any $p\times p$ matrix $\mathbf F$ with elements $F_{ij} = f(\mathbf X_i, \mathbf X_j)$, where $\mathbf X_k$ are columns of a $n\times p$ data matrix $\mathbf X$ and $f(\cdot, \cdot)$ is some arbitrary function.
This can be seen as a generalization of covariance matrix (if $f$ is covariance), correlation matrix (if $f$ is correlation), sum-of-squares-and-cross-products matrix (if $f$ is dot product), etc.
Note that even if $f$ is symmetric and "sensible", the resulting matrix $\mathbf F$ can fail to be positive-definite. This is the case e.g. when $f$ is mutual information (as stated in the title of this paper).
I doubt that there is a generic term for $\mathbf F$. If you really have to have a name for it, I think you should invent one. If your $f$ is supposed to measure some relationship between variables $i$ and $j$, then perhaps $\mathbf F$ can be called "cross-relationship matrix" or simply "relationship matrix"?
• I think you can use the language of reproducing kernel Hilbert spaces and consider $\mathbf F$ to be the Gram matrix associated with the kernel $f$, provided $f$ satisfies the standard conditions. – cardinal Mar 30 '16 at 21:02
• @cardinal I was thinking whether I should mention this in my answer. But usually kernel matrix (Gram matrix) is understood to be $n\times n$, i.e. it measures the "relationships" (kernel distances) between the samples, whereas here OP considers a $p \times p$ matrix measuring the "relationships" between the variables. So I figured that to call it kernel/Gram matrix might be confusing; what do you think? (Also, one can consider functions that do not satisfy kernel conditions, e.g. apparently mutual information.) – amoeba Mar 30 '16 at 21:08