In the spectral theorem for bounded self-adjoint operators we get, for each self-adjoint (bounded) operator $A$ in a Hilbert space $\mathcal{H}$, and each bounded Borel-measurable function $f \in \mathcal{B}_0(\sigma(A))$ (where I denote by $\mathcal{B}_0(\sigma (A))$ the set of all Borel-measurable functions in the spectrum $\sigma(A)$ of $A$) a bounded operator $f(A)$ defined as

$$\langle \psi, f(A) \psi \rangle = \int_{\sigma(A)} f(\lambda) d\mu_{\psi}(\lambda) \ ,$$

where $\mu_{\psi}$ is a measure defined by a family of "projection-valued measures" $(P_B)_{B \in \mathcal{B}(\sigma(A))}$ which are just $\chi_B(A)$ in the sense of the functional calculus, for each Borel subset $B$ of $\sigma(A)$. Given all this information, the operator $A$ is denoted as

$$ A = \int_{\sigma(A)} \lambda dP_\lambda \ ,$$

and this decomposition is unique in the sense that any other projection-valued measure satisfying this equality should be equal to the first. My questions are: Is there some other reason for this notation for the operator $A$? How this integral should be interpreted, as it isn't actually constructed as a Lebesgue integral? The only way I know to obtain the operator in this case is to use the somewhat more general expression above, and use the polarization identity + Riesz lemma, as usual.

Sorry for the long question, my doubt is quite simple but there was a lot of context before I could express it.

  • $\begingroup$ Also because this integral can be seen as "a Lebegue-integral" w.r.t. the projection-valued measure $P$ defined on the intervals $[a,b]$ by $P_b-P_a$ or something alike. For details see e.g. Birman-Solomyak $\endgroup$ – Andrei Kh Apr 14 at 8:42

I'll answer the second question first, then the first question.

The interpretation of the integral comes the following result:

If $(P_B)_{B \in \Omega}$ is a projection-valued measure for $(X,\Omega,H)$ and $f:X\to\mathbb C$ is a bounded $\Omega$-measurable function, then there is a unique operator $T$ in $\mathcal B(H)$ such that if $\varepsilon>0$ and $\{B_1,\ldots,B_n\}$ is an $\Omega$-partition of $X$ with $\sup\{|f(x_0)-f(x_1)|:x_0,x_1\in B_k\}<\varepsilon$ for $1\leq k\leq n$, then for any $x_k\in B_k$, $$\|A-\sum_{k=1}^nf(x_k)P_{B_k}\|<\varepsilon.$$

For a proof of this result, see Proposition IX.1.10 of Conway's functional analysis book. Given the result, we define $T=\int f\ dE$. In the course of the proof, it is seen that the relation you state, namely $\langle \psi, T \psi \rangle = \int_{\sigma(A)} f(\lambda) d\mu_{\psi}(\lambda)$, is shown to hold.

The reason for the notation is because this is (part of) an infinite-dimensional analogue of the spectral theorem of linear algebra. In the finite-dimensional case, self-adjoint operators are sums of projections on eigenspaces. In infinite-dimensions, self-adjoint operators are integrals ("limits" of sums) of projections associated with the spectrum.

  • $\begingroup$ What I wonder is the reason for the condition imposed for the function: is that a more elaborate way to define the integral as a "Stieltjes integral over the 'vector-valued' functions $P_{B_j}$"? I am still trying to understand why the condition $\sup \{|f(x_0) - f(x_1)| : x_0, x_1 \in B_j \} < \epsilon$ for the partition $\{B_1, \cdots, B_n\}$ is natural... $\endgroup$ – B. Chinaski Feb 7 '18 at 2:07

For a continuous function $f$, you can treat the integral as a Riemann-Stieltjes type of vector integral $$ \int_a^b f(\lambda)dP(\lambda)x = \lim_{\|\mathscr{P}\|\rightarrow 0}\sum_{j=1}^{n}f(\lambda_j)P[\lambda_{j-1},\lambda_j)x. $$

The notation is very suggestive in that form. As for notation, if you think of $dP(\lambda)$ as an "infinitestimal" projection onto the eigenspace associated with $\lambda$, then $$ A dP(\lambda) = \lambda dP(\lambda). $$ The projections sum to the identity $$ I = \int dP(\lambda) $$ which is another way of saying that you have a "continuous" basis of eigenvectors which may include discrete components as well. Hence, $$ Ax = A\int dP(\lambda)x=\int AdP(\lambda)x=\int \lambda dP(\lambda)x. $$ And the eigenspaces are mututally orthogonal, represented symbolically as $$ dP(\lambda)dP(\mu) = 0,\;\; \lambda\ne \mu, \\ dP(\lambda)dP(\lambda)=dP(\lambda). $$ which makes integrals easy to multiply $$ \int f(\lambda)dP(\lambda)\int g(\mu)dP(\mu)x = \int f(\lambda)g(\lambda)dP(\lambda)x. $$ And there is a Parseval identity, $$ \|\int f(\lambda)dP(\lambda)x\|^2 = \int |f(\lambda)|^2 d\|P(\lambda)x\|^2, \\ \| x\|^2 = \|\int dP(\lambda)x\|^2 = \int d\|P(\lambda)x\|^2. $$ It makes sense to write $f(A)=\int f(\lambda)dP(\lambda)$ because $$ A^nx = \int \lambda^n dP(\lambda)x. $$ More generally, $(fg)(A)x=f(A)g(A)x$. It's effort to make all of this rigorous, but that was not your question.

This is John von Neumann's integral notation. He used $dE(\lambda)$ where $E$ and $\lambda$ were used to be suggestive to the Quantum Physicist.

  • $\begingroup$ It's weird to think of these integrals as generalized decompositions as in the finite-dimensional case as it would lead to think that we can always express an operator as a "general sum" of projectors in some eigenspaces. That's the analogy the physicists do, but in reality things are much more complicated... $\endgroup$ – B. Chinaski Feb 7 '18 at 2:15
  • $\begingroup$ @B.Chinaski : Spectral integrals are used for selfadjoint and normal operators only. The projections are commuting orthogonal projections, which are selfadjoint. So these integrals on the real line generalize the diagonalization of a selfadjoint matrix. If you allow for measures in the plane, they generalize the diagonalizations of normal matrices. It's just that the spectrum can fill a continuum, instead of being discrete, or in addition to being discrete. This generalization is the one that is rigorous for Math and Physics. $\endgroup$ – DisintegratingByParts Feb 7 '18 at 4:55

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.