Topic :: Multivariate regression

1. Basic Matrix Operation

Basic Matrix Operation

  • Let’s briefly review the basic matrix operation and results for vectors and matrices.

  • If we denote $L, R, Q$ be the lower-triangular, upper-triangular, and orthogonal matrix, then
    • LR decomposition: $A = LR$
    • Cholesky decomposition: $A = LL^T$
    • QR decomposition: $A = QR$
  • If $A$ and $B$ are $(J \times J)$ matrices, then

  • $ \lvert Q \rvert = 1 $

  • If and A, B are both square and non-singular, then

  • The rank of $A$, denoted $r(A)$, is the size of the largest submatrix of $A$ that has a nonzero determinant.

    Note that $r(AB) = r(A)$ if $\lvert B \rvert \ne 0$, and, in general, $r(AB) \le \text{min} \left( r(A), r(B) \right) $.

    If $A$ is square, $(J \times J)$, and nonsingular, then a unique $(J \times J)$ inverse matrix $A^{-1}$ exists such that $AA^{-1}=I_J$.

    If $A$ is orthogonal, then $A^{-1}=A^T$. Note that $(AB)^{-1}=B^{-1}A^{-1}$, and $\lvert A^{-1} \rvert = \lvert A \rvert ^{-1}$.

  • where, $A$ and $D$ are $(J \times J)$ and $(K \times K)$ nonsingular matrices, respectively.

    In special case of this result, if $A$ is $(J \times J)$ and u and v are $J$-vectors, then

  • If $A$ and $D$ are symmetric matrices and $A$ is nonsingular, then,

    where $E = D-B^T A^{-1} B$ is nonsingular and $F = A^{-1} B$.

Vectoring and Kronecker Products

  • vec(A) denotes the $(JK \times 1)$-column vector formed by placing the columns of a $(J\times K)$-matrix A under one another successively.

  • If a $(J \times K)$-matrix $A$ is such that the $jk$th element $A_{jk}$ is itself a submatrix, then $A$ is termed a block matrix. The Kronecker product of a $(J \times K)$-matrix $A$ and an $(L \times M)$-matrix B is the $(JL \times KM)$ block matrix

    This is commonly known as the left Kronecker product.

  • The following operation hold for Kronecker products:

    If $A$ is $(J \times J)$ and $B$ is $(K \times K)$, then

    If $A$ is $(J \times K)$ and $B$ is $(L \times M)$, then

    If $A$ and $B$ are square and nonsingular, then,

    One of the most useful results that combines vectoring with Kronecker products is that

Eigenanalysis for Square Matrices

  • If $A$ is a $(J \times J)$-matrix, then $\lvert A - \lambda I_J \rvert $ is a polynomial of order $J$ in $\lambda$. The equation will have $J$ (possibly complex-valued) roots denoted by $\lambda_j = \lambda_j (A), ~ j=1, 2, \cdots, J$.

    The root $\lambda_j$ is called the eigenvalue (characteristic root, latent root) of $A$, and the set ${\lambda_j }$ is called the spectrum of $A$.

Associated with $\lambda_j$, there is a $J$-vector $v_j=v_j(A)$ (not all of whose entries of zero) such that

The vector $\text{v}_j$ is called eigenvector (characteristic vector, latent vector) associated with $\lambda_j$.

  • Eigenvectors $\text{v}_j$ and $\text{v}_k$ associated with distinct eigenvalues $(\lambda_j \ne \lambda_k)$ are orthogonal. If $V = (v_1, v_2, \cdots, v_J)$, then

    where $\Lambda = \text{diag} ( \lambda_1, \lambda_2, \cdots, \lambda_J ) $ is a matrix with the eigenvalues along the diagonal and zeros elsewhere, and $V^T V = I_J$.

  • The “outer product” of a $J$-vector $\text{v}$ with itself is the $(J\times J)$-matrix $\text{v} \text{v}^T$, which has rank 1.

    The spectrum theorem expresses the $(J\times J)$-matrix $A$ as a weighted average of rank-1 matrices,

    where $I_J = \sum_{j=1} ^J v_j v_j ^T$, and where the weights, $\lambda_1, \cdots, \lambda_J$, are the eigenvalues of $A$.

  • The rank of $A$ is the number of nonzero eigenvalues, the trace is

    and the determinant is

Functions of Matrices

  • If $A$ is a symmetric $(J\times J)$-matrix and $\phi : R^J \rightarrow R^J$ is a function, then where $\lambda_j$ and $v_j$ are the $j$th eigenvalue and corresponding eigenvector, respectively, of $A$.

  • Examples include the following:

    Hence, $\lambda_j (\phi (A)) = \phi ( \lambda_j (A) )$ and $v_j (\phi (A)) = v_j (A)$.

2. Multivariate regression

  • Consider the multivariate regression framework: where $Y$ is an $n \times q$ matrix of responses, $X$ is an $n \times p$ design matrix, $B$ is a $p \times q$ regression coefficients matrix to be estimated, and $W$ is an matrix of errors.

  • Least squares figure out the regression coefficients that minimize the following objective function: $\parallel \cdot \parallel_F$ is the Frobenius norm.

  • Let us assume that the responses are correlated with each other. Then, since the above objective function do not include the correlation among responses, we will use the Mahalanobis distance instead of Eucleadian distance of Frobenius norm.

Category statistics, topic