Asymptotic Minimax Risk for the White Noise Model on the Sphere (Q4956047)

The regression model \(dX(x)=f(x)dx+\varepsilon dW(x)\) is considered on the unit sphere \(S_d\) in d+1-dimensional space, where \(W\) is the Gaussian white noise on \(S_d\) and \(f\) is an unknown function. The minimax risk for \(f\) estimation in \(L_2\) (with respect to Lebesgue measure on \(S_d\)) is obtained for the case of Hölder \(f\) and for \(f\) in Sobolev-space balls. E.g., if \(B(s,M)\) is a ball in \(s\)-order Sobolev norm on \(S_d\) and the risk of the estimator \(\hat f\) is defined as \(\sup_{f\in B(s,M)}E_{\varepsilon, f}\|\hat f-f\|^2_{L_2},\) then \[ \underset_{\hat f} \inf R_\varepsilon(\hat f)\sim \varepsilon^{4s/(2s+d)}M^{d/(2s+d)}P_{s,d},\quad \text{as} {s} 0, \] \[ \text{with}\quad P_{s,d}=\left(2s/(s+d)(d-1)!\right)^{2s/(2s+d)}(2s+d)^{d/(2s+d)}/d \] ..