-
True functions, estimations, and the expected error
ym(x)=f(x)+ϵm(x).Ex[(ym(x)−f(x))2]=Ex[ϵm(x)2]
-
The average error made by M individual models
vs. Expected error of the ensemble
EAvg=M1m=1∑MEx[ϵm(x)2]
EEnsemble=Ex⎣⎢⎡(M1m=1∑Mym(x)−f(x))2⎦⎥⎤
=Ex⎣⎢⎡(M1m=1∑Mϵm(x))2⎦⎥⎤
EEnsemble=M1EAvg
-
In reality (errors are correlated), by the Cauchy's inequality
[m=1∑Mϵm(x)]2≤Mm=1∑Mϵm(x)2⇒[M1m=1∑Mϵm(x)]2≤M1m=1∑Mϵm(x)2
EEnsemble≤EAvg