Optimal Reduced Isotonic Regression

Abstract: Isotonic regression is a shape-constrained nonparametric regression in which the regression is an increasing step function. For $n$ data points, the number of steps in the isotonic regression may be as large as $n$. As a result, standard isotonic regression has been criticized as overfitting the data or making the representation too complicated. So-called "reduced" isotonic regression constrains the outcome to be a specified number of steps $b$, $b \leq n$. However, because the previous algorithms for finding the reduced $L_2$ regression took $\Theta(n+bm^2)$ time, where $m$ is the number of steps of the unconstrained isotonic regression, researchers felt that the algorithms were too slow and instead used approximations. Other researchers had results that were approximations because they used a greedy top-down approach. Here we give an algorithm to find an exact solution in $\Theta(n+bm)$ time, and a simpler algorithm taking $\Theta(n+b m \log m)$ time. These algorithms also determine optimal $k$-means clustering of weighted 1-dimensional data.