User:IssaRice/Extreme value theorem: Difference between revisions

From Machinelearning
No edit summary
No edit summary
Line 9: Line 9:
So suppose <math>f(c) < M</math>. We want to find <math>M' < M</math> such that <math>f(t) < M'</math> for all <math>t \in [a,c]</math>. That would mean that <math>\sup V_c \leq M' < M</math>.
So suppose <math>f(c) < M</math>. We want to find <math>M' < M</math> such that <math>f(t) < M'</math> for all <math>t \in [a,c]</math>. That would mean that <math>\sup V_c \leq M' < M</math>.


* Consider <math>t = c - \delta/2</math>. Then since <math>t < c</math>, there exists <math>x \in X</math> such that <math>t < x</math>. So <math>\sup V_t \leq \sup V_c < M</math> so <math>f(t) \leq \sup V_t < M</math>.
* Consider <math>t = c - \delta/2</math>. Then since <math>t < c</math>, there exists <math>x \in X</math> such that <math>t < x</math>. So <math>\sup V_t \leq \sup V_c < M</math> so <math>f(t) \leq \sup V_t < M</math>. This also means that for all <math>x \in [a,c-\delta/2]</math> we have <math>f(x) \leq \sup V_t < M</math>.
* If <math>t = c</math>, then
* But now if <math>c-\delta/2 \leq t \leq c</math>, then by continuity at <math>c</math>, We can choose <math>\epsilon > 0</math> with <math>\epsilon < M - f(c)</math>.<ref group="note">it is important here that <math>\epsilon</math> does not equal <math>M - f(c)</math>; choosing this <math>\epsilon</math> would be too weak and we would not be able to conclude <math>\sup V_c < M</math>, rather only that <math>\sup V_c \leq M</math>.</ref> By continuity at <math>c</math>, there exists a <math>\delta > 0</math> such that <math>|t-c|<\delta</math> implies <math>|f(t)-f(c)|<\epsilon</math>. This means <math>f(t) < f(c) + \epsilon < M</math>.


Now we can choose <math>M' = \max\{\sup V_{c-\delta/2}, f(c) + \epsilon\}</math>.


We can choose <math>\epsilon > 0</math> with <math>\epsilon < M - f(c)</math>.<ref group="note">it is important here that <math>\epsilon</math> does not equal <math>M - f(c)</math>; choosing this <math>\epsilon</math> would be too weak and we would not be able to conclude <math>\sup V_c < M</math>, rather only that <math>\sup V_c \leq M</math>.</ref> By continuity at <math>c</math>, there exists a <math>\delta > 0</math> such that <math>|t-c|<\delta</math> implies <math>|f(t)-f(c)|<\epsilon</math>. This means <math>f(t) < f(c) + \epsilon < M</math>. If <math>t \leq c-\delta/2</math> then there exists some <math>x \in X</math> such that <math>t < x</math>. This means <math>\sup V_t \leq \sup V_x < M</math> so <math>t \in X</math>. Otherwise if <math>c-\delta/2 \leq t \leq c</math> then <math>|t-c|<\delta</math> so <math>f(t) < f(c) + \epsilon < M</math>. So now what can we say about <math>\sup V_c</math>? We want to say <math>\sup V_c < M</math>. We can do this by showing that there exists a number <math>M' < M</math> such that <math>f(t) < M'</math> for all <math>t \in [a,c]</math>. That way, <math>\sup V_c \leq M' < M</math>. But <math>M' = \sup V_c</math> works.
 
If <math>t \leq c-\delta/2</math> then there exists some <math>x \in X</math> such that <math>t < x</math>. This means <math>\sup V_t \leq \sup V_x < M</math> so <math>t \in X</math>. Otherwise if <math>c-\delta/2 \leq t \leq c</math> then <math>|t-c|<\delta</math> so <math>f(t) < f(c) + \epsilon < M</math>. So now what can we say about <math>\sup V_c</math>? We want to say <math>\sup V_c < M</math>. We can do this by showing that there exists a number <math>M' < M</math> such that <math>f(t) < M'</math> for all <math>t \in [a,c]</math>. That way, <math>\sup V_c \leq M' < M</math>. But <math>M' = \sup V_c</math> works.


<math>\sup V_{c-\delta/2} < M</math> so let <math>M' = (M+\sup V_{c-\delta/2})/2</math>. Then <math>\sup V_{c-\delta/2} < M' < M</math> and if <math>v \in V_{c-\delta/2}</math> we have <math>v \leq \sup V_{c-\delta/2} < M'</math>.
<math>\sup V_{c-\delta/2} < M</math> so let <math>M' = (M+\sup V_{c-\delta/2})/2</math>. Then <math>\sup V_{c-\delta/2} < M' < M</math> and if <math>v \in V_{c-\delta/2}</math> we have <math>v \leq \sup V_{c-\delta/2} < M'</math>.

Revision as of 22:56, 1 June 2019

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):atx} 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 aX. 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 supVcM<M.

  • Consider t=cδ/2. Then since t<c, there exists xX such that t<x. So supVtsupVc<M so f(t)supVt<M. This also means that for all x[a,cδ/2] we have f(x)supVt<M.
  • But now if cδ/2tc, then by continuity at c, We can choose ϵ>0 with ϵ<Mf(c).[note 1] By continuity at c, there exists a δ>0 such that |tc|<δ implies |f(t)f(c)|<ϵ. This means f(t)<f(c)+ϵ<M.

Now we can choose M=max{supVcδ/2,f(c)+ϵ}.


If tcδ/2 then there exists some xX such that t<x. This means supVtsupVx<M so tX. Otherwise if cδ/2tc then |tc|<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, supVcM<M. But M=supVc works.

supVcδ/2<M so let M=(M+supVcδ/2)/2. Then supVcδ/2<M<M and if vVcδ/2 we have vsupVcδ/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 Mf(c); choosing this ϵ would be too weak and we would not be able to conclude supVc<M, rather only that supVcM.