Dominated convergence: $$f_n(\omega)\to f(\omega)\quad(\omega\in\Omega)\implies f_n(E)\to f(E)$$ (This gives a tool for analysis of operators.)


Given a Borel space $\Omega$ and a Hilbert space $\mathcal{H}$.

Consider a spectral measure: $$E:\mathcal{B}(\Omega)\to\mathcal{B}(\mathcal{H})$$

And its Borel measures: $$\nu_\varphi(A):=\|E(A)\varphi\|^2$$ $$\mu_{\varphi\psi}(A):=\langle E(A)\varphi,\psi\rangle$$

Regard its measurable calculus: $$\mathcal{D}f(E):=\left\{\varphi\in\mathcal{H}:\int_\Omega|f|^2\mathrm{d}\nu_\varphi\right\}:\quad\langle f(E)\varphi,\psi\rangle:=\int_\Omega f\mathrm{d}\mu_{\varphi\psi}$$

Suppose pointwise convervence: $$f_n(\omega)\to f(\omega)$$

Denote their common domain by: $$\mathcal{D}:=\bigcap_{n\in\mathbb{N}}\mathcal{D}f_n(E)$$

Define their limiting operator by: $$\mathcal{D}T:=\{\varphi\in\mathcal{D}:f_n(E)\varphi\to\psi\}:\quad T\varphi:=\psi$$ (That is the one taken in Stone's theorem.)

Operator Domain

Can it happen that this domain shrinks?

Domain Criterion

Does a common domain suffice: $$\mathcal{D}f_n(E)\supseteq\mathcal{D}f(E):\quad T\supseteq f(E)$$ In that case one obtains: $$\mathcal{D}f_n(E)\supseteq\mathcal{D}f(E)\iff |f_n|^2\leq C_n(|f|^2+1)$$ So one may miss a dominant for: $$\|f(E)\varphi-f_n(E)\varphi\|^2=\|(f-f_n)(E)\varphi\|^2=\int|f-f_n|^2\mathrm{d}\nu_\varphi$$

But what is an explicit nonexample?

Dominant Criterion

Assuming a dominant one gets: $$|f_n|^2\leq C(|f|^2+1):\quad T=f(E)$$

Clearly one has: $$\mathcal{D}f(E)\subseteq\mathcal{D}f_n(E)$$

So the above shows: $$\|f(E)\varphi-f_n(E)\varphi\|^2=\ldots=\int_\Omega|f-f_n|^2\mathrm{d}\nu_\varphi\to0$$

How to prove the converse inclusion?

Operator Core

Does one have as core: $$\mathcal{D}f_n(E)\supseteq\mathcal{D}f(E):\quad\overline{T}=f(E)$$ (This solves the failure above!)

  • $\begingroup$ what is "finite measure space"? Finite set? $\endgroup$ – Yurii Savchuk Dec 10 '14 at 8:12
  • $\begingroup$ what is $\mathcal D$ ? $\endgroup$ – Yurii Savchuk Dec 10 '14 at 8:13
  • $\begingroup$ @YuriiSavchuk: Oh, it actually it must be only a measurABLE space; for instance a topological space together with its Borel algebra. $\endgroup$ – C-Star-W-Star Dec 10 '14 at 13:15
  • $\begingroup$ @YuriiSavchuk: The domain of an operator: $\mathcal{D}(T)$ $\endgroup$ – C-Star-W-Star Dec 10 '14 at 13:18

Ok, I think I finally got it...

(However, if anybody finds bugs, typos, loopholes etc. then please let me know. Thanks!)


Given the natural numbers $\Omega:=\mathbb{N}$ and the Hilbert space $\mathcal{H}:=\ell^2(\mathbb{N})$.

Choose the canonical basis by: $e_n:=\chi_n$

Consider the spectral measure $E(\{n\}):=P_n$.

Operator Domain

Regard the series of measurable functions $f_n:=\sum_{k=1}^nk\chi_{\{k\}}$.

Denote their sum by: $f:=\sum_{k=1}^\infty k\chi_{\{k\}}$

It approaches pointwise its sum: $f_n\to f$

It has a dominant: $|f_n|^2\leq1(|f|^2+1)$

However, not every element converges: $$\varphi:=\sum_{k=1}^\infty\frac{1}{k}e_k:\quad f_n(E)\varphi=\sum_{k=1}^n1e_k\nrightarrow\psi$$

So the domain may shrink: $\mathcal{D}T\subsetneq\mathcal{D}$


Regard the sequence of measurable functions $f_n:=n\chi_{\{n\}}$.

It vanishes pointwise: $f_n\to0$

Now, it fulfils the domain criterion: $$\mathcal{D}f_n(E)=\mathcal{H}\supseteq\mathcal{H}=\mathcal{D}f(E)$$ However, its measurable calculus does not converge: $$\varphi:=\sum_{k=1}^\infty\frac{1}{k}e_k:\quad f_n(E)\varphi=e_n\nrightarrow0=f(E)\varphi$$

So a common domain is not sufficient: $T\subsetneq f(E)$


Suppose $\varphi\in\mathcal{D}T$ with $T\varphi=\psi$.

By Fatou applied to the measurable calculus one has: $$0\leq\int|f-f_n|^2\mathrm{d}\nu_\varphi\leq\liminf_m\int|f_m-f_n|^2\mathrm{d}\nu_\varphi=\liminf_m\|f_m(E)\varphi-f_n(E)\varphi\|^2\to0$$ But that implies especially: $\varphi\in\mathcal{D}f(E)$

Moreover, one has then: $f(E)\varphi=\lim_nf_n(E)\varphi=\psi=T\varphi$

Concluding that the operators agree: $f(E)=T$

Operator Core

Suppose $\varphi\in\mathcal{D}f(E)$ with $f(E)\varphi=\psi$.

By Egorov's theorem one finds: $$\nu_\varphi(\Omega_N)\to\nu_\varphi(\Omega):\quad\|f-f_n\|_{\Omega_N}\stackrel{n\to\infty}{\to}0$$

Set the approximation to be: $$\varphi_N:=1_N(E)\varphi:=E(\Omega_N)\varphi$$

So on the hand one has: $\varphi_N\to\varphi$

Also note that: $f_n(E)1_N(E)=(f_n1_N)(E)$

Now by uniform convergence: $$|f_n1_N|\leq R_N(|f1_N|+1)\quad|f1_N|\leq R_N(|f_n1_N|+1)$$ So one has a common domain: $\mathcal{D}(f1_N)(E)=\mathcal{D}(f_n1_N)(E)$

But also one has trivially: $$|f1_N|\leq1(|f|+1)$$ Thus they all include: $\mathcal{D}f(E)\subseteq\mathcal{D}(f1_N)(E)$

Hence one can compute: $$\left\|(f1_N)(E)\varphi-(f_n1_N)(E)\varphi\right\|^2=\int_{\Omega_N}|f-f_n|^2\mathrm{d}\nu_\varphi\leq\|f-f_n\|_{\Omega_N}\|\varphi\|^2\stackrel{n\to\infty}{\to}0$$ That is $\varphi_N\in\mathcal{D}T$ with $T\varphi_N=(f1_N)(E)\varphi$.

But by the preceding dominant-criterion: $$|f1_N|\leq1(|f|+1):\quad(f1_N)(E)\varphi\to f(E)\varphi$$ So as was desired: $T\varphi_N\to f(E)\varphi$

Concluding that it is a core: $\overline{T}=f(E)$


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.