129
23
v1v2 (latest)

Isotonic Regression in Multi-Dimensional Spaces and Graphs

Abstract

In this paper we study minimax and adaptation rates in general isotonic regression. For uniform deterministic and random designs in [0,1]d[0,1]^d with d2d\ge 2 and N(0,1)N(0,1) noise, the minimax rate for the 2\ell_2 risk is known to be bounded from below by n1/dn^{-1/d} when the unknown mean function ff is nondecreasing and its range is bounded by a constant, while the least squares estimator (LSE) is known to nearly achieve the minimax rate up to a factor (logn)γ(\log n)^\gamma where nn is sample size, γ=4\gamma = 4 in the lattice design and γ=max{9/2,(d2+d+1)/2}\gamma = \max\{9/2, (d^2+d+1)/2 \} in the random design. Moreover, the LSE is known to achieve the adaptation rate (K/n)2/d{1log(n/K)}2γ(K/n)^{-2/d}\{1\vee \log(n/K)\}^{2\gamma} when ff is piecewise constant on KK hyperrectangles in a partition of [0,1]d[0,1]^d. Due to the minimax theorem, the LSE is identical on every design point to both the max-min and min-max estimators over all upper and lower sets containing the design point. This motivates our consideration of estimators which lie in-between the max-min and min-max estimators over possibly smaller classes of upper and lower sets, including a subclass of block estimators. Under a qq-th moment condition on the noise, we develop q\ell_q risk bounds for such general estimators for isotonic regression on graphs. For uniform deterministic and random designs in [0,1]d[0,1]^d with d3d\ge 3, our 2\ell_2 risk bound for the block estimator matches the minimax rate n1/dn^{-1/d} when the range of ff is bounded and achieves the near parametric adaptation rate (K/n){1log(n/K)}d(K/n)\{1\vee\log(n/K)\}^{d} when ff is KK-piecewise constant. Furthermore, the block estimator possesses the following oracle property in variable selection: When ff depends on only a subset SS of variables, the 2\ell_2 risk of the block estimator automatically achieves up to a poly-logarithmic factor the minimax rate based on the oracular knowledge of SS.

View on arXiv
Comments on this paper