4.3-exam2-2008-edelman¶
Quiz 2, 2008 (edelman) · Section 4.3: Least Squares Approximations Source: exam2_1.pdf, pp. 177, 157
Problem¶
4 (15 pts.) Suppose we have obtained from measurements \(n\) data points \((t_i, b_i)\), and you are asked to find a best least squares fit function of the form \(y = C + Dt + E(1 - t)\). Are \(C, D\), and \(E\) uniquely determined? Write down a solvable system of equations that gives a solution to the least squares problem.
Solution
We have that \[ C + Dt + E(1 - t) = (C + E) + (D - E)t. \] Thus we see that \(E\) is a free variable: it is not uniquely determined, and in fact can take any value. Given this, just write down the usual least squares equations but treat \(C + E\) and \(D - E\) as your variables: the matrix \(A\) has two columns: the first consists of \(n\) 1's, the second is the vector \((t_i)\). Then solve \[ A^TA \begin{bmatrix} C + E \\ D - E \end{bmatrix} = A^Tb. \]Keywords: least squares, free variable, normal equations