HDS

Exercise 14.3: Sharper Rates via Entropy Integrals

chapter 14 entropy integral

Note that f1ni=1nεf(xi) is a sub-Gaussian process with respect to the metric d(f,g)=fgn. Therefore, (1)E[supfF1ni=1nεf(xi)]1n0DlogN(u,F,n)du where (2)F={fθ(x)=ϕ(x),θ:θR3,f1} and D=supf,gF=fgn2. Note that θϕ(X)θ is an isomorphism between R3 and col(ϕ(X)) with the norm ϕ(X)θn=fn. Hence (3)N(u,F,n)N(u,{θR3:ϕ(),θ1},ϕ(X)n)(4)N(u,{θR3:ϕ(X)θ1},ϕ(X)n)(5)N(u,{θR3:ϕ(X)θn1},ϕ(X)n)(6)3log(1+2/u), so the entropy integral is finite and independent of n.

Published on 9 April 2021.