Isotonic Regression by Dynamic Programming

Abstract

For a given sequence of numbers, we want to find a monotonically increasing sequence of the same length that best approximates it in the sense of minimizing the weighted sum of absolute values of the differences. A conceptually easy dynamic programming approach leads to an algorithm with running time O(n logn). While other algorithms with the same running time are known, our algorithm is very simple. The only auxiliary data structure that it requires is a priority queue. The approach extends to other error measures. 2012 ACM Subject Classification Human-centered computing → Graph drawings. Mathematics of computing → Graph algorithms