On Additive Approximate Submodularity
This addresses a theoretical gap in machine learning for handling functions with additive errors in submodularity, though it is incremental relative to known results for modularity and convexity.
The paper tackles the problem of quantifying how close approximately submodular functions are to truly submodular ones, showing an upper bound of O(n^2) and a lower bound of Ω(√n) on the distance, and providing an algorithmic tool for adapting submodular optimization algorithms.
A real-valued set function is (additively) approximately submodular if it satisfies the submodularity conditions with an additive error. Approximate submodularity arises in many settings, especially in machine learning, where the function evaluation might not be exact. In this paper we study how close such approximately submodular functions are to truly submodular functions. We show that an approximately submodular function defined on a ground set of $n$ elements is $O(n^2)$ pointwise-close to a submodular function. This result also provides an algorithmic tool that can be used to adapt existing submodular optimization algorithms to approximately submodular functions. To complement, we show an $Ω(\sqrt{n})$ lower bound on the distance to submodularity. These results stand in contrast to the case of approximate modularity, where the distance to modularity is a constant, and approximate convexity, where the distance to convexity is logarithmic.