Asymptotic Minimax Risk in the Uniform Norm for the White Noise Model on the Sphere.pdf

Asymptotic Minimax Risk in the Uniform Norm forthe White Noise Model on the Sphere Jussi KlemelaRolf Nevanlinna InstituteUniversity of HelsinkiAbstractEstimation of the signal function dened on the unit sphere of the Euclideanspace is considered. Gaussian continuous time white noise model is supposed.Uniform norm is chosen as a loss function and exact asymptotic minimax riskis derived extending the result of Korostelev (1993). The exact asymptoticminimax risk is given also for the L2-loss applying the result of Pinsker (1982).Mathematics Subject Classications: 62G07, 62C20, 62H11KeyWords and Phrases: asymptotic minimax risk, spherical data, uniform norm,white noise model.1 IntroductionLet Sd = fx 2 Rd+1 : kxk = 1g. Consider the observation of the formdX(x) = f(x)dx+ dW (x); x 2 Sdwhere  0 is the noise size and dW (x) is the Gaussian white noise in Sd, that isfor measurable sets B;B1; B2  Sd,This research was supported by Sonderforschungsbereich 373 \Quantikation undSimulation Okonomischer Prozesse. 1 (i) L RB dW (x)
= N(0; (B)) where  is the Lebesgue measure of Sd,(ii) if B1 \ B2 = ;, then RB1 dW (x) and RB2 dW (x) are independent.The Lebesgue measure of the sphere can be dened with the help of the Lebesguemeasure of Rd+1, see Rudin (1986, page 175). It is a Borel measure, invariant withrespect to rotations and (Sd) = 2(d+1)=2=((d+ 1)=2).Let (; L) be the class of Holder functions(; L) = f : M ! R  jf(x) f(y)j  L(x; y); x; y 2 Sd where 0  1 and  is the Riemannian metric of Sd, that is, (x; y) = arccos(x0y).We dene the discrepancy of an arbitrary estimator f?(x) and the true signal functionf(x) by the sup norm kf? fk1 = supx2Sd jf?(x) f(x)j:Let w(u), u  0, be a loss function, i. e. a continuous bounded monotone functionof u, w(0) = 0. Introduce the minimax riskr = r(w; ; L) = inff? supf2(;L)E;fw  1 f? f 1where  = (2 log 2)=(2+d).Theorem 1 Let A;d = lim!1 kd=2K;k2=(2+d)2 where K; : Sd ! R,K;(