Difference between revisions of "Equivalent conditions for a linear map between two normed spaces to be continuous everywhere"

From Maths
Jump to: navigation, search
m (Linking to proof subpages)
m (Proof: Removing todo notice)
 
Line 22: Line 22:
 
{{Begin Inline Proof}}{{:Equivalent conditions for a linear map between two normed spaces to be continuous everywhere/4 implies 1}}
 
{{Begin Inline Proof}}{{:Equivalent conditions for a linear map between two normed spaces to be continuous everywhere/4 implies 1}}
 
{{End Proof}}{{End Theorem}}
 
{{End Proof}}{{End Theorem}}
<br/><br/>
 
{{Requires proof|See page 154 in<ref name="APIKM"/>}}
 
  
 
==Notes==
 
==Notes==

Latest revision as of 01:35, 28 February 2016

Statement

Given two normed spaces [ilmath](X,\Vert\cdot\Vert_X)[/ilmath] and [ilmath](Y,\Vert\cdot\Vert_Y)[/ilmath] and a linear map [ilmath]L:X\rightarrow Y[/ilmath] between them, then the following are equivalent (meaning if you have 1 you have all the others):

  1. If we have a sequence [ilmath](x_n)_{n=1}^\infty\subseteq X[/ilmath] with [ilmath]x_n\rightarrow 0[/ilmath] then [ilmath](\Vert L(x_n)\Vert_Y)_{n=1}^\infty[/ilmath] is a bounded sequence[1][Note 1]
  2. [ilmath]L[/ilmath] is continuous at a particular point in [ilmath]X[/ilmath][1][2][Note 2]
  3. [ilmath]L[/ilmath] is a bounded linear map, that is [ilmath]\exists A>0\ \forall x\in X[\Vert L(x)\Vert_Y\le A\Vert x\Vert_X][/ilmath][1][2]
  4. [ilmath]L[/ilmath] is continuous everywhere[1][2]

Proof

[ilmath]1)\implies 2)[/ilmath]: That [ilmath]L[/ilmath] maps all null sequences to bounded sequences [ilmath]\implies[/ilmath] [ilmath]L[/ilmath] is continuous at [ilmath]p\in X[/ilmath]


This is a proof by contrapositive. That is we will show that if [ilmath]L[/ilmath] is not continuous at [ilmath]p[/ilmath] [ilmath]\implies[/ilmath] [ilmath]L[/ilmath] takes a null sequence to one that isn't bounded (an unbounded one).

  • Let the normed spaces [ilmath]X[/ilmath] and [ilmath]Y[/ilmath] be given, as well as a linear map [ilmath]L:X\rightarrow Y[/ilmath]
    • Suppose that [ilmath]L[/ilmath] is not continuous at [ilmath]p[/ilmath], this means:
  • [ilmath]\exists (x_n)_{n=1}^\infty\rightarrow p[/ilmath] such that [math]\lim_{n\rightarrow\infty}\left(L(x_n)\right)\ne L(p)[/math]

Recall that continuity states that:

  • [ilmath]L[/ilmath] is continuous at [ilmath]p[/ilmath] [ilmath]\iff[/ilmath] [math]\forall(x_n)_{n=1}^\infty\subseteq X \left[\left(\lim_{n\rightarrow\infty}(x_n)=p\right)\implies\left(\lim_{n\rightarrow\infty}\left(L(x_n)\right)=L\left(\lim_{n\rightarrow\infty}(x_n)\right)=L(p)\right)\right][/math]

So it follows that to not be continuous at [ilmath]p[/ilmath]:

  • [math]\exists (x_n)_{n=1}^\infty\subseteq X\left[\left(\lim_{n\rightarrow\infty}(x_n)=p\right)\wedge\neg\left(\lim_{n\rightarrow\infty}(L(x_n))=L(p)\right)\right][/math], by negation of implies. Additionally we may negate the [ilmath]=[/ilmath] and thus we see this is the same as:
    • [math]\exists (x_n)_{n=1}^\infty\subseteq X\left[\left(\lim_{n\rightarrow\infty}(x_n)=p\right)\wedge\left(\lim_{n\rightarrow\infty}(L(x_n))\ne L(p)\right)\right][/math]

Which is exactly "there exists a sequence in [ilmath]X[/ilmath] whose limit is [ilmath]p[/ilmath] and where [ilmath]\lim_{n\rightarrow\infty}(L(x_n))\ne L(p)[/ilmath]"

  • Let us now take [ilmath]L(x_n)\not\rightarrow L(p)[/ilmath] and subtract [ilmath]L(p)[/ilmath] from both sides. We see:
    • [ilmath]L(x_n)-L(p)\not\rightarrow L(p)-L(p)[/ilmath], using the fact that [ilmath]L[/ilmath] is linear we see that:
      • [ilmath]L(x_n-p)\not\rightarrow L(0)[/ilmath] and [ilmath]L(0)=0\in Y[/ilmath] so:
    • [ilmath]L(x_n-p)\not\rightarrow 0[/ilmath]
  • Thus [ilmath]\Vert L(x_n-p)\Vert_Y\not\rightarrow 0[/ilmath] (as [ilmath]\Vert0\Vert_Y=0[/ilmath] by definition)
  • So [math]\exists C>0\ \forall N\in\mathbb{N}\ \exists n\in\mathbb{N}[n>N\wedge\Vert L(x_n-p)\Vert_Y>\epsilon][/math]

If a sequence converges to [ilmath]0[/ilmath] then we have:

  • [ilmath]\forall\epsilon>0\ \exists N\in\mathbb{N}\ \forall n\in\mathbb{N}[n>N\implies d(x_n,x)<\epsilon][/ilmath] we know we don't have this, so we negate it:
  • [ilmath]\exists\epsilon>0\ \forall N\in\mathbb{N}\ \exists n\in\mathbb{N}[n>N\wedge d(x_n,x)>\epsilon][/ilmath].

That is there exists an [ilmath]\epsilon>0[/ilmath] such that for all [ilmath]N[/ilmath] there exists a bigger [ilmath]n[/ilmath] such that [ilmath]d(x_n,x)>\epsilon[/ilmath], we shall later call such an [ilmath]\epsilon[/ilmath] [ilmath]C[/ilmath] and construct a subsequence out of the [ilmath]n[/ilmath]s

  • Thus it is possible to construct a subsequence, [ilmath](\Vert L(x_{n_k}-p)\Vert_Y)_{k=1}^\infty[/ilmath] of the image [ilmath](x_n)[/ilmath] where for every [ilmath]k[/ilmath] we have:
  • [ilmath]\Vert L(x_{n_k}-p)\Vert_Y>C[/ilmath]

By the negation of convergent sequence we see that there exists a [ilmath]C[/ilmath] such that for all [ilmath]N\in\mathbb{N} [/ilmath] there exists another natural number, [ilmath]n[/ilmath] such that [ilmath]\Vert L(x_n-p)\Vert_Y>C[/ilmath], we construct a sequence, [ilmath](n_k)_{k=1}^\infty[/ilmath] of such [ilmath]n[/ilmath]-ns. That is we know there exists a [ilmath]C[/ilmath], so we pick [ilmath]N=1[/ilmath] and get the [ilmath]n[/ilmath] that works, this is our first term, we then set [ilmath]N=2[/ilmath] and the [ilmath]n[/ilmath] that works is our second term, and so forth.

Some of these [ilmath]n[/ilmath]s may be the same, but that doesn't matter.

[ilmath](x_{n_k})[/ilmath] is the subsequence of [ilmath](x_n)[/ilmath] which contains only the terms that satisfy [ilmath]\Vert L(x_n-p)\Vert_Y> C[/ilmath]

It would have been better if I used [ilmath]n[/ilmath] and [ilmath]m[/ilmath] as terms, to make which [ilmath]n[/ilmath] I am talking about clearer, but a reader able to attempt this proof should follow.

  • We now have a sequence [ilmath](x_{n_k})[/ilmath] such that [ilmath]\Vert L(x_{n_k}-p)\Vert_Y>C[/ilmath]
  • Define a new sequence [math]b_k:=\frac{1}{\sqrt{\Vert x_{n_k}-p\Vert} }[/math]
    • It is easy to see that [ilmath]b_k\rightarrow +\infty[/ilmath] (as [ilmath](x_{n_k}-p)\rightarrow 0[/ilmath])

TODO: Prove that this tends to [ilmath]+\infty[/ilmath]


  • Define a new sequence [math]d_k:=b_k(x_{n_k}-p)[/math]
  • Clearly [math]d_k\rightarrow 0[/math]

If [ilmath](d_k)_{k=1}^\infty\rightarrow 0[/ilmath] then [ilmath]\forall\epsilon>0\exists N\in\mathbb{N}\forall n\in\mathbb{N}[n>N\implies\Vert d_k-0\Vert_X<\epsilon][/ilmath].

  • Notice [math]\Vert d_k\Vert_X=b_k\Vert x_{n_k}-p\Vert_X=\frac{\Vert x_{n_k}-p\Vert_X}{\sqrt{\Vert x_{n_k}-p\Vert_X} }[/math] [math]=\sqrt{\Vert x_{n_k}-p\Vert_X}[/math]

TODO: Finish this proof


  • But [math]\Vert L(d_k)\Vert_Y=\Vert L(b_k(x_{n_k}-p))\Vert_Y=b_k\Vert L(x_{n_k}-p)\Vert_Y\ge Cb_k\rightarrow +\infty[/math]
  • Thus we have shown if [ilmath]L[/ilmath] is not continuous at [ilmath]p[/ilmath] that the mapping of a null sequence is unbounded, the contrapositive of what we set out to claim

[ilmath]2)\implies 3)[/ilmath]: [ilmath]L[/ilmath] is continuous at [ilmath]p\in X[/ilmath] [ilmath]\implies[/ilmath] [ilmath]\exists A\ge 0\ \forall x\in X[\Vert L(x)\Vert_Y \le A\Vert x\Vert_X][/ilmath] (That is that [ilmath]L[/ilmath] is a bounded linear map)


The key to this proof is exploiting the linearity of [ilmath]L[/ilmath]. As will be explained in the blue box.

  • Suppose that [ilmath]L:X\rightarrow Y[/ilmath] is continuous at [ilmath]p\in X[/ilmath]. Then:
    • [ilmath]\forall\epsilon>0\ \exists\delta>0[\Vert x-p\Vert_X<\delta\implies\Vert L(x-p)\Vert_Y<\epsilon][/ilmath] (note that [ilmath]L(x-p)=L(x)-L(p)[/ilmath] due to linearity of [ilmath]L[/ilmath])
  • Define [ilmath]u:=x-p[/ilmath], then:
    • [ilmath]\forall\epsilon>0\ \exists\delta>0[\Vert u\Vert_X<\delta\implies\Vert Lu\Vert_Y<\epsilon][/ilmath] (writing [ilmath]Lu:=L(u)[/ilmath] as is common for linear maps)

We now know we may assume that [ilmath]\forall\epsilon>0\ \exists\delta>0[\Vert u\Vert_X<\delta\implies\Vert Lu\Vert_Y<\epsilon][/ilmath] - see this box for a guide on how to use this information and how I constructed the rest of the proof. The remainder of the proof follows this box.

The key to this proof is in the norm structure and the linearity of [ilmath]L[/ilmath].

  • Pick (fix) some [ilmath]\epsilon>0[/ilmath] we now know:
    • [ilmath]\exists\delta>0[/ilmath] such that if we have [ilmath]\Vert x\Vert_X<\delta[/ilmath] then we have [ilmath]\Vert Lx\Vert<\epsilon[/ilmath]

In the proof we will have to show at some point that: [ilmath]\forall x\in X[\Vert Lx\Vert_Y\le A\Vert x\Vert_X][/ilmath] this means:

  • At some point we'll be given an arbitrary [ilmath]x\in X[/ilmath]

However we have a [ilmath]\delta[/ilmath] such that [ilmath]\Vert u\Vert_X<\delta\implies\Vert Lu\Vert_Y<\epsilon[/ilmath]

  • All we have to do is make [ilmath]\Vert x\Vert_X[/ilmath] so small that it is less than [ilmath]\delta[/ilmath] - we can do this using scalar multiplication.
  • Note that if [ilmath]x=0[/ilmath] the result is trivial, so assume that arbitrary [ilmath]x[/ilmath] is [ilmath]\ne 0[/ilmath]
  • With this in mind the task is clear: we need to multiply [ilmath]x[/ilmath] by something such that the actual vector part has [ilmath]\Vert\cdot\Vert_X<\delta[/ilmath]
    • Then we can say (supposing [ilmath]x=\alpha p[/ilmath] for some positive [ilmath]\alpha[/ilmath] and [ilmath]\Vert p\Vert_X<\delta[/ilmath]) that [ilmath]\Vert L(\alpha p)\Vert_Y=\alpha\Vert Lp\Vert_Y[/ilmath]
      • But [ilmath]\Vert p\Vert_X<\delta\implies\Vert Lp\Vert_Y<\epsilon[/ilmath]
    • So [ilmath]\Vert L(\alpha p)\Vert_Y=\alpha\Vert Lp\Vert_Y<\alpha\epsilon[/ilmath], if we can get a [ilmath]\Vert x\Vert_X[/ilmath] involved in [ilmath]\alpha[/ilmath] the result will follow.

Getting an arbitrary [ilmath]\Vert x\Vert_X[/ilmath] to have magnitude [ilmath]<\delta[/ilmath]

  1. Lets normalise [ilmath]x[/ilmath] and then multiply it by the magnitude again (so as to do nothing)
    • So notice [math]x=\overbrace{\Vert x\Vert_X}^\text{scalar part}\cdot\overbrace{\frac{1}{\Vert x\Vert_X}x}^\text{vector part}[/math],
  2. Now the vector part has magnitude [ilmath]1[/ilmath] we can get it to within [ilmath]\delta[/ilmath] by multiplying it by [math]\frac{\delta}{2}[/math]
    • So [math]x=\overbrace{\Vert x\Vert_X\cdot\frac{2}{\delta} }^\text{scalar part}\cdot\overbrace{\frac{\delta}{2\Vert x\Vert_X}x}^\text{vector part}[/math].

Now we have [math]\left\Vert\frac{\delta}{2\Vert x\Vert_X}x\right\Vert_Y<\delta[/math] which [math]\implies \left\Vert L\left(\frac{\delta}{2\Vert x\Vert_X}x\right)\right\Vert_Y<\epsilon[/math]

  • Thus [math]\Vert Lx\Vert_Y=\frac{2\Vert x\Vert_X}{\delta}\cdot\left\Vert L\left(\frac{\delta}{2\Vert x\Vert_X}x\right)\right\Vert_Y<\frac{2\Vert x\Vert_X}{\delta}\cdot\epsilon=\frac{2\epsilon}{\delta}\Vert x\Vert_X[/math]

But [ilmath]\epsilon[/ilmath] was fixed, and so was the [ilmath]\delta[/ilmath] we know to exist based off of this, so set [ilmath]A=\frac{2\epsilon}{\delta}[/ilmath] after fixing them and the result will follow!

  • Fix some arbitrary [ilmath]\epsilon>0[/ilmath] (it doesn't matter what)
    • We know there [ilmath]\exists\delta>0[\Vert x\Vert_X<\delta\implies\Vert Lu\Vert_Y<\epsilon][/ilmath] - take such a [ilmath]\delta[/ilmath] (which we know to exist by hypothesis) and fix it also.
  • Define [ilmath]A:=\frac{2\epsilon}{\delta}[/ilmath] (if you are unsure of where this came from, see the blue box)
  • Let [ilmath]x\in A[/ilmath] be given (this is the [ilmath]\forall x\in X[/ilmath] part of our proof, we have just claimed an [ilmath]A[/ilmath] exists on the above line)
    • If [ilmath]x=0[/ilmath] then
      • Trivially the result is true, as [ilmath]L(0_X)=0_Y[/ilmath], and [ilmath]\Vert 0_Y\Vert_Y=0[/ilmath] by definition, [ilmath]A\Vert x\Vert_X=0[/ilmath] as [ilmath]\Vert 0_X\Vert_X=0[/ilmath] so we have [ilmath]0\le 0[/ilmath] which is true. (This is more workings than the line is worth)
    • Otherwise ([ilmath]x\ne 0[/ilmath])
      • Notice that [math]x=\frac{2\Vert x\Vert_X}{\delta}\cdot\frac{\delta}{2\Vert x\Vert_X}x[/math] and [math]\left\Vert \frac{\delta}{2\Vert x\Vert_X}x\right\Vert_X<\delta[/math]
        • and [math]\left\Vert \frac{\delta}{2\Vert x\Vert_X}x\right\Vert_X<\delta\implies\left\Vert L\left(\frac{\delta}{2\Vert x\Vert_X}x\right)\right\Vert_Y<\epsilon[/math]
    • Thus we see [math]\Vert Lx\Vert_Y=\left\Vert L\left(\frac{2\Vert x\Vert_X}{\delta}\cdot\frac{\delta}{2\Vert x\Vert_X}x\right)\right\Vert_Y=\frac{2\Vert x\Vert_X}{\delta}\cdot\overbrace{\left\Vert L\left(\frac{\delta}{2\Vert x\Vert_X}x\right)\right\Vert_Y}^{\text{remember this is }<\epsilon}[/math] [math]<\frac{2\Vert x\Vert_X}{\delta}\cdot\epsilon=\frac{2\epsilon}{\delta}\Vert x\Vert_X=A\Vert x\Vert_X[/math]
      • Shortening the workings this states that: [math]\Vert Lx\Vert_Y< A\Vert x\Vert_X[/math]
  • So if [ilmath]x=0[/ilmath] we have equality, otherwise [math]\Vert Lx\Vert_Y< A\Vert x\Vert_X[/math]

In either case, it is true that [math]\Vert Lx\Vert_Y\le A\Vert x\Vert_X[/math]
This completes the proof

[ilmath]3)\implies 4)[/ilmath]: [ilmath]\exists A\ge 0\ \forall x\in X[\Vert L(x)\Vert_Y \le A\Vert x\Vert_X][/ilmath] (That is that [ilmath]L[/ilmath] is a bounded linear map) [ilmath]\implies[/ilmath] [ilmath]L[/ilmath] is continuous at every point


This is a straight forward implication proof. We suppose by hypothesis that [ilmath]\exists A\ge 0\ \forall x\in X[\Vert L(x)\Vert_Y \le A\Vert x\Vert_X][/ilmath] is true.
We wish to show that [ilmath]L[/ilmath] is continuous everywhere, this means [ilmath]\forall x\in X[L\ \text{is continuous at }x][/ilmath], I will use sequential continuity in this proof (that is: [ilmath]\forall(x_n)_{n=1}^\infty\rightarrow x[\lim_{n\rightarrow\infty}(L(x_n))=L(x)][/ilmath], or the image of the sequence under [ilmath]L[/ilmath] converges to the image of [ilmath]x[/ilmath] under [ilmath]L[/ilmath])

  • Let [ilmath]x\in X[/ilmath] be given (arbitrary point)
  • Let [ilmath](x_n)_{n=1}^\infty\rightarrow x[/ilmath] be given (an arbitrary sequence that converges to [ilmath]x[/ilmath])

Remember that to converge means [ilmath]\forall\epsilon>0\exists N\in\mathbb{N}\forall n\in\mathbb{N}[n>N\implies d(x_n,x)<\epsilon][/ilmath] and in our normed space:

  • [ilmath]d(x_1,x_2):=\Vert x_1-x_2\Vert_X[/ilmath] (see metric induced by norm), so what we're really saying is:
  • [ilmath]\forall\epsilon>0\exists N\in\mathbb{N}\forall n\in\mathbb{N}[n>N\implies \Vert x_n-x\Vert_X<\epsilon][/ilmath]
  • Let [ilmath]\epsilon >0[/ilmath] be given (we will show that [math]\left\Vert L(x_n)-L(x)\right\Vert_Y<\epsilon[/math])
  • Choose [ilmath]N[/ilmath] such that [math]\forall n\in\mathbb{N}\left[n>N\implies\Vert x_n-x\Vert_X<\frac{\epsilon}{A}\right] [/math] (Where [ilmath]A[/ilmath] is the [ilmath]A[/ilmath] in the hypothesis of [ilmath]L[/ilmath] being a bounded linear map)

We will want to show at some point that [math]\lim_{n\rightarrow\infty}\left(L(x_n)\right)=L(x)[/math] this means we'll need to do something along the lines of:

  • [ilmath]\Vert L(x_n)-L(x)\Vert_Y=\Vert L(x_n-x)\Vert_Y\le A\Vert x_n-x\Vert_X[/ilmath]

If we pick [ilmath]N[/ilmath] such that [math]\Vert x_n-x\Vert_X<\frac{\epsilon}{A} [/math] then:

  • [ilmath]\Vert L(x_n)-L(x)\Vert_Y=\Vert L(x_n-x)\Vert_Y\le A\Vert x_n-x\Vert_X[/ilmath][math]<A\frac{\epsilon}{A}=\epsilon[/math]

and that'll complete the proof.
Note: we know such an [ilmath]N[/ilmath] exists as by definition [math]\lim_{n\rightarrow\infty}(x_n)=x[/math]

  • Let [ilmath]n\in\mathbb{N} [/ilmath] be given
  • If [ilmath]n\le N[/ilmath] then we're done. By the truth table of implies if the RHS ([ilmath]\Vert L(x_n)-L(x)\Vert_Y<\epsilon[/ilmath]) is true or false, the implication is true, so this literally doesn't matter.
  • If [ilmath]n>N[/ilmath] then:
  • [math]\Vert L(x_n)-L(x)\Vert_Y=\overbrace{\Vert L(x_n-x)\Vert_Y}^\text{by linearity}[/math][math]\le\overbrace{A\Vert x_n-x\Vert_X}^{\text{boundedness of }L}[/math][math]<A\frac{\epsilon}{A}=\epsilon[/math]
  • Thus we have shown [math]n>N\implies \Vert L(x_n)-L(x)\Vert_Y<\epsilon[/math] as required.

That completes the proof

[ilmath]4)\implies 1)[/ilmath]: [ilmath]L[/ilmath] is continuous at every point [ilmath]\implies[/ilmath] [ilmath]L[/ilmath] maps all null sequences to bounded sequences


There is actually a slightly stronger result to be had here, I shall prove that, to which the above statement is a corollary.

  • Let [ilmath](x_n)_{n=1}^\infty\rightarrow x[/ilmath] be a sequence that converges - we must show that the image of this sequence under [ilmath]L[/ilmath] is bounded
    • As [ilmath]L[/ilmath] is everywhere continuous we know that it is sequentially continuous at [ilmath]x[/ilmath]. This means that:
      • [math]\forall\left((x_n)_{n=1}^\infty\rightarrow x\right)\left[\big(L(x_n)\big)_{n=1}^\infty\rightarrow L(x)\right][/math]
    • So [ilmath]L[/ilmath] maps [ilmath](x_n)_{n=1}^\infty\rightarrow x[/ilmath] to [math]\left(L(x_n)\right)_{n=1}^\infty\rightarrow L(x)[/math]
    • Recall that if a sequence converges it is bounded, as [math]\left(L(x_n)\right)_{n=1}^\infty[/math] converges, it must therefore be bounded.


Corollary: the image of every null sequence under [ilmath]L[/ilmath] is bounded


This completes the proof.


Notes

  1. this may be better said (perhaps) as "[ilmath]L[/ilmath] maps all sequences convergent to [ilmath]0[/ilmath] to bounded sequences"
  2. Gamelin and Green's Introduction to Topology only demonstrates continuity at the point [ilmath]0\in X[/ilmath]

References

  1. 1.0 1.1 1.2 1.3 Analysis - Part 1: Elements - Krzysztof Maurin
  2. 2.0 2.1 2.2 Introduction to Topology - Theodore W. Gamelin & Robert Everist Greene