Skip to content

6.3-exam3-2013-strang

Quiz 3, 2013 (strang) · Section 6.3: Symmetric Positive Definite Matrices Source: exam3_2.pdf, pp. 72, 76-77

Problem

(3) (30 pts.)

(a) For a really large number \(N\), will this matrix be positive definite? Show why or why not.

\[ A = \begin{bmatrix} 2 & 4 & 3 \\ 4 & N & 1 \\ 3 & 1 & 4 \end{bmatrix}. \]

(b) Suppose: \(A\) is positive definite symmetric, \(Q\) is orthogonal (same size as \(A\)), \(B\) is \(Q^TAQ = Q^{-1}AQ\). Show that: 1. \(B\) is also symmetric. 2. \(B\) is also positive definite.

(c) If the SVD of \(A\) is \(U\Sigma V^T\), how do you find the orthogonal \(V\) and the diagonal \(\Sigma\) from the matrix \(A\)?

Solution (a) The easiest test to use here is to check whether the upper-left determinants are positive. 1 \(\times\) 1: This is 2, which is always greater than 0. 2 \(\times\) 2: This is \(2N - 16\), which is greater than 0 if \(N\) is really large (in particular if \(N > 8\)). 3 \(\times\) 3: Use the method of your choice to compute the determinant of \(A\), in terms of \(N\). By the big formula, \[ \det A = 8N + 12 + 12 - 2 - 64 - 9N = -42 - N. \] This is going to be very negative if \(N\) is really large. So the matrix will not be positive definite. (b) First we show that \(B\) is symmetric. This means we need to check \(B^T = B\). Using what we're told, \[ B^T = (Q^TAQ)^T = Q^TA^T(Q^T)^T = Q^TAQ = B, \] where in the next-to-last step we used the fact that \(A\) itself is symmetric (\(A^T = A\)). For positive definiteness, one way is to use the energy test. If \(x\) is any nonzero vector, then \[ x^TBx = x^T(Q^TAQ)x = (Qx)^TA(Qx) = y^TAy, \] where \(y = Qx\). We know that \(y\) is nonzero, because \(Q\) is orthogonal and therefore has no nullspace. Hence \(x^TBx > 0\). Another approach is via eigenvalues. We know that \(B = Q^{-1}AQ\), so \(B\) is similar to \(A\). That means that they have the same eigenvalues. Since \(A\) is positive definite, its eigenvalues are all positive, so those of \(B\) are as well. A third approach: \(A\) is positive definite, so \(A = R^TR\) for some \(R\) with independent columns. Then \(B = Q^TR^TRQ = (RQ)^T(RQ)\). \(RQ\) is a matrix with independent columns, since \(Q\) is orthogonal. So \(B\) is positive definite. (c) First compute the matrix \(A^TA\). Find the eigenvalues and eigenvectors. The first \(r\) columns of \(V\) should be length 1 eigenvectors of \(A^TA\), corresponding to nonzero eigenvalues, arranged in order of decreasing eigenvalue. (A small complication: if there is a repeated eigenvalue, make sure to pick orthogonal eigenvectors for that eigenvalue.) The diagonal entries of \(\Sigma\) should be square roots of the eigenvalues of \(A^TA\), again in decreasing order. The remaining columns of \(V\) should be an orthonormal basis for the nullspace of \(A^TA\) (which is the same thing as the nullspace of \(A\)). This will give enough columns for \(V\) to be a square matrix. The other diagonal entries of \(\Sigma\) should be 0.

Keywords: positive definite, determinant, similar matrices, singular value decomposition, symmetric matrix