Skip to main content

Subsection 4.6.2 Summary

The LLS problem can be states as: Given \(A \in \Cmxn \) and \(b \in \Cm \) find \(\widehat x \in \Cn \) such that

\begin{equation*} \| b - A \widehat x \|_2 = \min_{x \in \Cn} \| b - A x \|_2 . \end{equation*}

Given \(A \in \Cmxn \text{,}\)

  • The column space, \(\Col( A ) \text{,}\) which is equal to the set of all vectors that are linear combinations of the columns of \(A \)

    \begin{equation*} \{ y ~\vert~ y = A x \}. \end{equation*}
  • The null space, \(\Null( A ) \text{,}\) which is equal to the set of all vectors that are mapped to the zero vector by \(A \)

    \begin{equation*} \{ x ~\vert~ A x = 0 \}. \end{equation*}
  • The row space, \(\Rowspace( A ) \text{,}\) which is equal to the set

    \begin{equation*} \{ y ~\vert~ y^H = x^H A \}. \end{equation*}

    Notice that \(\Rowspace( A ) = \Col( A^H ) \text{.}\)

  • The left null space, which is equal to the set of all vectors

    \begin{equation*} \{ x ~\vert~ x^H A = 0 \}. \end{equation*}

    Notice that this set is equal to \(\Null( A^H ) \text{.}\)

  • If \(A x = b \) then there exist \(x_r \in \Rowspace( A ) \) and \(x = x_r + x_n \) where \(x_r \in \Rowspace( A ) \) and \(x_n \in \Null( A ) \text{.}\)

These insights are summarized in the following picture, which also captures the orthogonality of the spaces.

If \(A \) has linearly independent columns, then the solution of LLS, \(\widehat x \text{,}\) equals the solution of the normal equations

\begin{equation*} ( A^H A) \widehat x = A^H b. \end{equation*}

as summarized in

The (left) pseudo inverse of \(A \) is given by \(A^{\dagger} = ( A^H A )^{-1} A^H \) so that the solution of LLS is given by \(\widehat x = A^\dagger b \text{.}\)

Definition 4.6.2.1. Condition number of matrix with linearly independent columns.

Let \(A \in \Cmxn \) have linearly independent columns (and hence \(n \leq m \)). Then its condition number (with respect to the 2-norm) is defined by

\begin{equation*} \kappa_2( A ) = \| A \|_2 \| A^\dagger \|_2 = \frac{\sigma_0}{\sigma_{n-1}}. \end{equation*}

Assuming \(A \) has linearly independent columns, let \(\widehat b = A \widehat x \) where \(\widehat b \) is the projection of \(b \) onto the column space of \(A \) (in other words, \(\widehat x \) solves the LLS problem), \(\cos( \theta ) = \| \widehat b \|_2 / \| b \|_2 \text{,}\) and \(\widehat b + \delta\! \widehat b = A ( \widehat x + \delta\! \widehat x ) \text{,}\) where \(\delta\! \widehat b \) equals the projection of \(\delta\!b \) onto the column space of \(A \text{.}\) Then

\begin{equation*} \frac{\| \delta\! \widehat x\|_2}{\| \widehat x \|_2} \leq \frac{1}{\cos( \theta )} \frac{\sigma_0}{\sigma_{n-1}} \frac{\| \delta\! b \|_2}{\| b \|_2} \end{equation*}

captures the sensitivity of the LLS problem to changes in the right-hand side.

If \(A \) has linearly independent columns and \(A = U_L \Sigma_{TL} V_L^H \) is its Reduced SVD, then

\begin{equation*} \widehat x = V_L \Sigma_{TL}^{-1} U_L^H b \end{equation*}

solves LLS.

Given \(A \in \C^{m \times n}\text{,}\) let \(A = U_L \Sigma_{TL} V_L^H \) equal its Reduced SVD and \(A = \left( \begin{array}{c | c} U_L \amp U_R \end{array} \right) \left( \begin{array}{c | c} \Sigma_{TL} \amp 0 \\ \hline 0 \amp 0 \end{array} \right) \left( \begin{array}{c | c} V_L \amp V_R \end{array} \right)^H \) its SVD. Then

\begin{equation*} \widehat x = V_L \Sigma_{TL} U_L^H b + V_R z_b, \end{equation*}

is the general solution to LLS, where \(z_b \) is any vector in \(\C^{n-r} \text{.}\)

Solving LLS via Gram-Schmidt QR factorization for \(A \in \Cmxn \text{:}\)

  • Compute QR factorization via (Classical or Modified) Gram-Schmidt: approximately \(2 m n^2 \) flops.

  • Compute \(y = Q^H b \text{:}\) approximately \(2 m n^2 \) flops.

  • Solve \(R \widehat x = y \text{:}\) approximately \(n^2 \) flops.

Solving LLS via Householder QR factorization for \(A \in \Cmxn \text{:}\)

  • Householder QR factorization: approximately \(2 m n^2 - \frac{2}{3} n^3 \) flops.

  • Compute \(y_T = Q^H bnn \) by applying Householder transformations: approximately \(4 m n - 2 n^2 \) flops.

  • Solve \(R_{TL} \widehat x = y_T \text{:}\) approximately \(n^2 \) flops.