2
$\begingroup$

Let $H$ be a Hilbert space. Let $L: H \rightarrow H$ be a linear bounded positive operator (i.e. $\langle L(u), u \rangle \geq 0$ for all $u \in H$).

A) Prove that $I+aL$ is bijective for every $a > 0$ (I is the identity map)

B) Show that $\lim\limits_{a \rightarrow \infty}(I+aL)^{-1}u = Pu$ (where P is the orthogonal projection onto $Ker(L)$


For A, injectivity seems straightforward (if we use positivity). For surjectivity I defined:

$$S = S_a = \sum\limits_{n=0}^{\infty} (-aL)^n$$

Then, after observing that $(I+aL)(S(v)) = v$, I concluded that $S(v)$ is a pre-image of $(I+aL)$, which makes $(I+aL)$ surjective. But is S a valid operator?

For B, I decomposed $v = v_1 + v_2$ and $S(v) = u_1 + u_2$ (according to the orthogonal projection theorem) and then used uniqueness to show that:

$$L(u_2) = \frac{v_2-u_2}{a}$$

From there, I think that if $a \rightarrow \infty$ then $L(u_2) \rightarrow 0$ (and thus $u_2 \rightarrow 0$ and the proof follows). Is this correct? The fact that $u_2$ also shows up on the RHS makes me a bit uneasy.

$\endgroup$
  • $\begingroup$ you mean $\ker L$ instead of $T$? $\endgroup$ – el_tenedor Apr 6 '17 at 17:40
  • $\begingroup$ I find it hard to believe that the operator $S$ is well defined for all $a > 0$ how do we know it converges. I understand you probably tried 1/(1+aL) and used the series expansion. However, that is valid on if $\|aL\| < 1$. $\endgroup$ – Gregory Apr 6 '17 at 17:47
  • $\begingroup$ Why should the series $S_a$ converges? $\endgroup$ – user251257 Apr 6 '17 at 17:48
  • $\begingroup$ I do mean $Ker(L)$ thanks. @Gregory That was one of main concerns too. But how do we prove it's surjective then? It doesn't seem obvious to me. $\endgroup$ – Pellenthor Apr 6 '17 at 17:50
3
$\begingroup$

For A) you can use Lax-Milgram:

$$ \langle aLu + u, u \rangle = \langle aLu, u \rangle + \langle u, u \rangle \geq \|u\|^2. $$

This shows, that $I + aL$ is invertible with $\|(I + aL)^{-1}\| \leq 1$, for all $a > 0$.

For B): If $u \in \ker L$ we have $Lu = 0$. This implies $$ (aL + I) u = u, \text{ for all } a > 0 $$ and thus $$ (aL + I)^{-1}u = u, \text{ for all } a > 0. $$ Taking the limit $a \to \infty$ shows, that $(aL + I)^{-1}$ works as the identity on $\ker L$.

Note that since $L$ is positive it is in particular self adjoint. This gives you $$ (\ker L)^\perp = \overline{\operatorname{im} L^*} = \overline{\operatorname{im} L}. $$

Now if $u \in \operatorname{im} L$ there is some $v$ such that $u = Lv$. We can calculate $$ (aL + I)^{-1} Lv = (aL + I)^{-1} \frac{1}{a} (aL + I - I) v =\frac{1}{a} \|v\|- \frac{1}{a} (aL + I)^{-1} v, $$ which gives $$ \|(aL + I)^{-1} Lv\| \leq \frac{1}{a} v + \frac{1}{a} \|(aL + I)^{-1}\| \|v\| \leq \frac{2}{a} \|v\| \to 0, \text{ for } a \to \infty. $$

We just proved that $$\lim_{a \to \infty}(aL + I)^{-1}w$$ for all $w \in \operatorname{im}L$. But since the net $(aL + I)^{-1}$ is uniformly bounded in operator norm this lets us extend the convergence to the closure $\overline{\operatorname{im}L}$:

Let $w \in \overline{\operatorname{im} L}$. Let $\varepsilon > 0$ be given. Then, there exists $Lv \in \operatorname{im} L$ such that $$\|w - Lv\| \leq \frac{\varepsilon}{2}.$$ Furthermore, we have some $\tilde a > 0$ such that for all $a \geq \tilde a$ the inequality $$\|(aL + I)^{-1} Lv\| < \frac{\varepsilon}{2}$$ holds. Thus for all $a \geq \tilde a$ we have $$ \|(aL + I)^{-1} w\| = \|(aL + I)^{-1} (w - Lv + Lv) \| \leq \|w - Lv\| + \|(aL + I)^{-1} Lv\| \leq \varepsilon. $$ This proves $$\lim_{a \to \infty} (aL + I)^{-1} w = 0.$$

$\endgroup$
  • 1
    $\begingroup$ Thank you. So the full conclusion of Lax-Milgram (with all the details) would be something like this: ...and thus there is a unique $v \in H$ such that $\langle aLv + v, u\rangle = \langle u, u\rangle$ for all $u$, which means that $(I+aL)(v) = u$. This shows that $I +aL$ is invertible, On top of that $\Vert v \Vert = \Vert (I+aL)^{-1} \Vert \leq (1/c)*M = 1$ (where $c=1$ is the inequality constant and $M=1$ is the (trivial) operator norm. Is that correct? $\endgroup$ – Pellenthor Apr 6 '17 at 18:37
  • $\begingroup$ @Pellenthor: Sounds good! $\endgroup$ – el_tenedor Apr 6 '17 at 18:41
  • 1
    $\begingroup$ Isn't it the case that: $u \in Ker(L^*) \Leftrightarrow L^*(u) = 0 \Leftrightarrow \langle v, L^*(u) \rangle = 0 (\forall v \in H) \Leftrightarrow \langle L(v), u \rangle = 0 (\forall v \in H) \Leftrightarrow$ $u \in Im(L)^{\perp}$ and thus $Ker(L) = Ker(L^*) =Im(L)^{\perp}$? (so we don't have to worry about the closure) $\endgroup$ – Pellenthor Apr 6 '17 at 18:45
  • 2
    $\begingroup$ $\|(I+aL)^{-1}u \| \le \|(I+aL)^{-1}u_n \| + \|u-u_n\|$ where $u_n\in im(L)$. $\endgroup$ – user251257 Apr 6 '17 at 18:45
  • 2
    $\begingroup$ @Pellenthor: $((A^\perp)^\perp = \overline A$ $\endgroup$ – el_tenedor Apr 6 '17 at 19:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.