Fused Lasso Nearly-Isotonic Signal Approximation in General Dimensions

Abstract In this paper we introduce and study fused lasso nearly-isotonic signal approximation, which is a combination of fused lasso and generalized nearly-isotonic regression. We show how these three estimators relate to each other and derive solution to a general problem. Our estimator is computationally feasible and provides a trade-off between monotonicity, block sparsity and goodness-of-fit. Next, we prove that fusion and near-isotonisation in one dimensional case can be applied interchangably, and this step-wise procedure gives the solution to the original optimization problem. This property of the estimator is very important, because it provides a direct way to construct path solution when one of the penalization parameters is fixed. Also, we derive unbiased estimator of degrees of freedom of the estimator.