How far apart can L1 and L2 lines fit to the same data be?

112 Views Asked by At

Given $n$ points $(x_i, y_i)$ in the unit square with $x_i = {{i - 1} \over {n - 1}}$ uniformly spaced and $0 \leq y_i \leq 1$, consider the best-fit L1 line and the best-fit L2 line:

$ \qquad$ L1 aka Least absolute deviation line: minimize $\sum |y_i - a_1 \, x_i - b_1|$
$ \qquad$ L2 aka least squares or simple linear regression line: minimize $\sum (y_i - a_2 \, x_i - b_2)^2 $

How far apart can these two lines be, over all sets of $n$ points $(x_i, y_i)$ as above ?


The following patterns of $y_i \in \{0, 1\}$ have L1 lines constant 0, which simplifies life. They are local maxima, > nearby patterns, but that doesn't prove anything.

enter image description here