Nach den Kommentaren in meiner anderen Antwort (und dem Titel der OP-Frage!) Folgt hier eine nicht sehr strenge theoretische Untersuchung des Problems.
Wir wollen , um zu bestimmen , ob Bias B ( θ n ) = E ( θ n ) - θ unterschiedliche Konvergenzrate als die Quadratwurzel der Varianz aufweisen kann,B(θ^n)=E(θ^n)−θ
B ( θ n ) = O ( 1 / n δ ) ,√Var ( θ n ) =O(1/n γ ),γ ≠ δ? ? ?
B(θ^n)=O(1/nδ),Var(θ^n)−−−−−−−√=O(1/nγ),γ≠δ???
Wir haben
B ( θ n ) = O ( 1 / n δ )⟹lim n δ E ( θ n ) < K⟹lim n 2 δ [ E ( θ n ) ] 2 < K '
B(θ^n)=O(1/nδ)⟹limnδE(θ^n)<K⟹limn2δ[E(θ^n)]2<K′
⟹[ E ( θ n ) ] 2 = O ( 1 / n 2 δ )
⟹[E(θ^n)]2=O(1/n2δ)(1)
während
√Var(ˆθn)=O(1/nγ)⟹limnγ√E(ˆθ2n)−[E(ˆθn)]2<M
Var(θ^n)−−−−−−−√=O(1/nγ)⟹limnγE(θ^2n)−[E(θ^n)]2−−−−−−−−−−−−−√<M
⟹lim√n2γE(ˆθ2n)−n2γ[E(ˆθn)]2<M
⟹limn2γE(θ^2n)−n2γ[E(θ^n)]2−−−−−−−−−−−−−−−−−−√<M
⟹limn2γE(ˆθ2n)−limn2γ[E(ˆθn)]2<M′
⟹limn2γE(θ^2n)−limn2γ[E(θ^n)]2<M′(2)
We see that (2)(2) may hold happen if
A) both components are O(1/n2γ)O(1/n2γ), in which case we can only have γ=δγ=δ.
B) But it may also hold if
limn2γ[E(ˆθn)]2→0⟹[E(ˆθn)]2=o(1/n2γ)
limn2γ[E(θ^n)]2→0⟹[E(θ^n)]2=o(1/n2γ)(3)
For (3)(3) to be compatible with (1)(1), we must have
n2γ<n2δ⟹δ>γ
n2γ<n2δ⟹δ>γ(4)
So it appears that in principle it is possible to have the Bias converging at a faster rate than the square root of the variance. But we cannot have the square root of the variance converging at a faster rate than the Bias.