User:IssaRice/Extreme value theorem

From Machinelearning

Working through the proof in Pugh's book by filling in the parts he doesn't talk about.

For x∈[a,b], define Vx=f([a,x])={f(t):a≤t≤x} to be the image of f up to and including x.

Let M=sup{f(x):x∈[a,b]} and X={x∈[a,b]:supVx<M}.

Now suppose f(a)<M. Then a∈X. We already know that X is bounded above, for instance by the number b. We can thus take the least upper bound of X, say c=supX. We already know f(c)≤M, so if we can just eliminate the possibility that f(c)<M, we will be done.

So suppose f(c)<M. We want to find M′<M such that f(t)<M′ for all t∈[a,c]. That would mean that supVc≤M′<M.

  • Consider t=c−δ/2. Then since t<c, there exists x∈X such that t<x. So supVt≤supVc<M so f(t)≤supVt<M.
  • If t=c, then


We can choose ϵ>0 with ϵ<M−f(c).[note 1] By continuity at c, there exists a δ>0 such that |t−c|<δ implies |f(t)−f(c)|<ϵ. This means f(t)<f(c)+ϵ<M. If t≤c−δ/2 then there exists some x∈X such that t<x. This means supVt≤supVx<M so t∈X. Otherwise if c−δ/2≤t≤c then |t−c|<δ so f(t)<f(c)+ϵ<M. So now what can we say about supVc? We want to say supVc<M. We can do this by showing that there exists a number M′<M such that f(t)<M′ for all t∈[a,c]. That way, supVc≤M′<M. But M′=supVc works.

supVc−δ/2<M so let M′=(M+supVc−δ/2)/2. Then supVc−δ/2<M′<M and if v∈Vc−δ/2 we have v≤supVc−δ/2<M′.

Therefore, c=b, which implies that M=supVb=supVc<M, a contradiction. So the assumption that f(c)<M was false, and we conclude f(c)=M.

Notes

  1. ↑ it is important here that ϵ does not equal M−f(c); choosing this ϵ would be too weak and we would not be able to conclude supVc<M, rather only that supVc≤M.