A matrix is a rectangular array of numbers. The order of a matrix with n rows and k columns is . The element in row i, column j of is denoted as , and the notation is sometimes used to refer to the twodimensional rowcolumn array
A vector is a onedimensional array of numbers. A column vector has a single column (). A row vector has a single row (). A scalar is a matrix of order —that is, a single number. A square matrix has the same row and column order, . A diagonal matrix is a square matrix where all offdiagonal elements are zero, if . The identity matrix is a diagonal matrix with for all i. The unit vector is a vector where all elements are 1. The unit matrix is a matrix of all 1s. Similarly, the elements of the null vector and the null matrix are all 0.
Basic matrix operations are as follows:
If and are of the same order, then is the matrix of elementwise sums,
If and are of the same order, then is the matrix of elementwise differences,
The dot product of two nvectors and is the sum of their elementwise products,
The dot product is also known as the inner product of and . Two vectors are said to be orthogonal if their dot product is zero.
Matrices and are said to be conformable for multiplication if the number of columns in equals the number of rows in . Suppose that is of order and that is of order . The product is then defined as the matrix of the dot products of the ith row of and the jth column of ,
The transpose of the matrix is denoted as and is obtained by interchanging the rows and columns,
A symmetric matrix is equal to its transpose, . The inner product of two column vectors and is .
The right inverse of a matrix is the matrix that yields the identity when is postmultiplied by it. Similarly, the left inverse of yields the identity if is premultiplied by it. is said to be invertible and is said to be the inverse of , if is its right and left inverse, . This requires to be square and nonsingular. The inverse of a matrix is commonly denoted as . The following results are useful in manipulating inverse matrices (assuming both and are invertible):










If is a diagonal matrix with nonzero entries on the diagonal—that is, —then . If is a blockdiagonal matrix whose blocks are invertible, then
In statistical applications the following two results are particularly important, because they can significantly reduce the computational burden in working with inverse matrices.
Suppose is a nonsingular matrix that is partitioned as
Then, provided that all the inverses exist, the inverse of is given by
where , , , and .
Suppose is nonsingular, is nonsingular, and and are and matrices, respectively. Then the inverse of is given by
This formula is particularly useful if and has a simple form that is easy to invert. This case arises, for example, in mixed models where might be a diagonal or blockdiagonal matrix, and .
Another situation where this formula plays a critical role is in the computation of regression diagnostics, such as in determining the effect of removing an observation from the analysis. Suppose that represents the crossproduct matrix in the linear model . If is the ith row of the matrix, then is the crossproduct matrix in the same model with the ith observation removed. Identifying , , and in the preceding inversion formula, you can obtain the expression for the inverse of the crossproduct matrix:
This expression for the inverse of the reduced data crossproduct matrix enables you to compute “leaveoneout” deletion diagnostics in linear models without refitting the model.
If is rectangular (not square) or singular, then it is not invertible and the matrix does not exist. Suppose you want to find a solution to simultaneous linear equations of the form
If is square and nonsingular, then the unique solution is . In statistical applications, the case where is rectangular is less important than the case where is a square matrix of rank less than k. For example, the normal equations in ordinary least squares (OLS) estimation in the model are
A generalized inverse matrix is a matrix such that is a solution to the linear system. In the OLS example, a solution can be found as , where is a generalized inverse of .
The following four conditions are often associated with generalized inverses. For the square or rectangular matrix there exist matrices that satisfy
The matrix that satisfies all four conditions is unique and is called the MoorePenrose inverse, after the first published work on generalized inverses by Moore (1920) and the subsequent definition by Penrose (1955). Only the first condition is required, however, to provide a solution to the linear system above.
Pringle and Rayner (1971) introduced a numbering system to distinguish between different types of generalized inverses. A matrix that satisfies only condition (i) is a inverse. The inverse satisfies conditions (i) and (ii). It is also called a reflexive generalized inverse. Matrices satisfying conditions (i)–(iii) or conditions (i), (ii), and (iv) are inverses. Note that a matrix that satisfies the first three conditions is a right generalized inverse, and a matrix that satisfies conditions (i), (ii), and (iv) is a left generalized inverse. For example, if is of rank k, then is a left generalized inverse of . The notation inverse for the MoorePenrose inverse, satisfying conditions (i)–(iv), is often used by extension, but note that Pringle and Rayner (1971) do not use it; rather, they call such a matrix “the” generalized inverse.
If the matrix is rankdeficient—that is, —then the system of equations
does not have a unique solution. A particular solution depends on the choice of the generalized inverse. However, some aspects of the statistical inference are invariant to the choice of the generalized inverse. If is a generalized inverse of , then is invariant to the choice of . This result comes into play, for example, when you are computing predictions in an OLS model with a rankdeficient matrix, since it implies that the predicted values
are invariant to the choice of .
Taking the derivative of expressions involving matrices is a frequent task in statistical estimation. Objective functions that are to be minimized or maximized are usually written in terms of model matrices and/or vectors whose elements depend on the unknowns of the estimation problem. Suppose that and are real matrices whose elements depend on the scalar quantities and —that is, , and similarly for .
The following are useful results in finding the derivative of elements of a matrix and of functions involving a matrix. For more indepth discussion of matrix differentiation and matrix calculus, see, for example, Magnus and Neudecker (1999) and Harville (1997).
The derivative of with respect to is denoted and is the matrix of the first derivatives of the elements of :
Similarly, the second derivative of with respect to and is the matrix of the second derivatives
The following are some basic results involving sums, products, and traces of matrices:












The next set of results is useful in finding the derivative of elements of and of functions of , if is a nonsingular matrix:












Now suppose that and are column vectors that depend on and/or and that is a vector of constants. The following results are useful for manipulating derivatives of linear and quadratic forms:








To decompose a matrix is to express it as a function—typically a product—of other matrices that have particular properties such as orthogonality, diagonality, triangularity. For example, the Cholesky decomposition of a symmetric positive definite matrix is , where is a lowertriangular matrix. The spectral decomposition of a symmetric matrix is , where is a diagonal matrix and is an orthogonal matrix.
Matrix decomposition play an important role in statistical theory as well as in statistical computations. Calculations in terms of decompositions can have greater numerical stability. Decompositions are often necessary to extract information about matrices, such as matrix rank, eigenvalues, or eigenvectors. Decompositions are also used to form special transformations of matrices, such as to form a “squareroot” matrix. This section briefly mentions several decompositions that are particularly prevalent and important.
Every square matrix , whether it is positive definite or not, can be expressed in the form , where is a unit lowertriangular matrix, is a diagonal matrix, and is a unit uppertriangular matrix. (The diagonal elements of a unit triangular matrix are 1.) Because of the arrangement of the matrices, the decomposition is called the LDU decomposition. Since you can absorb the diagonal matrix into the triangular matrices, the decomposition
is also referred to as the LU decomposition of .
If the matrix is positive definite, then the diagonal elements of are positive and the LDU decomposition is unique. Furthermore, we can add more specificity to this result in that for a symmetric, positive definite matrix, there is a unique decomposition , where is unit uppertriangular and is diagonal with positive elements. Absorbing the square root of into , , the decomposition is known as the Cholesky decomposition of a positivedefinite matrix:
where is upper triangular.
If is symmetric nonnegative definite of rank k, then we can extend the Cholesky decomposition as follows. Let denote the lowertriangular matrix such that
Then .
Suppose that is an symmetric matrix. Then there exists an orthogonal matrix and a diagonal matrix such that . Of particular importance is the case where the orthogonal matrix is also orthonormal—that is, its column vectors have unit norm. Denote this orthonormal matrix as . Then the corresponding diagonal matrix—, say—contains the eigenvalues of . The spectral decomposition of can be written as
where denotes the ith column vector of . The rightside expression decomposes into a sum of rank1 matrices, and the weight of each contribution is equal to the eigenvalue associated with the ith eigenvector. The sum furthermore emphasizes that the rank of is equal to the number of nonzero eigenvalues.
Harville (1997, p. 538) refers to the spectral decomposition of as the decomposition that takes the previous sum one step further and accumulates contributions associated with the distinct eigenvalues. If are the distinct eigenvalues and , where the sum is taken over the set of columns for which , then
You can employ the spectral decomposition of a nonnegative definite symmetric matrix to form a “squareroot” matrix of . Suppose that is the diagonal matrix containing the square roots of the . Then is a squareroot matrix of in the sense that , because
Generating the MoorePenrose inverse of a matrix based on the spectral decomposition is also simple. Denote as the diagonal matrix with typical element
Then the matrix is the MoorePenrose (generalized) inverse of .
The singularvalue decomposition is related to the spectral decomposition of a matrix, but it is more general. The singularvalue decomposition can be applied to any matrix. Let be an matrix of rank k. Then there exist orthogonal matrices P and Q of order and , respectively, and a diagonal matrix such that
where is a diagonal matrix of order k. The diagonal elements of are strictly positive. As with the spectral decomposition, this result can be written as a decomposition of into a weighted sum of rank1 matrices
The scalars are called the singular values of the matrix . They are the positive square roots of the nonzero eigenvalues of the matrix . If the singularvalue decomposition is applied to a symmetric, nonnegative definite matrix , then the singular values are the nonzero eigenvalues of and the singularvalue decomposition is the same as the spectral decomposition.
As with the spectral decomposition, you can use the results of the singularvalue decomposition to generate the MoorePenrose inverse of a matrix. If is with singularvalue decomposition , and if is a diagonal matrix with typical element
then is the generalized inverse of .