Local Operations, Instruments, and AQFT Measurement

A localized measurement is not an arbitrary state-update rule attached to a region. It is a normal completely positive instrument whose outcomes sum to a channel and whose nonselective action is the identity on the causal complement. Individual postselected branches may change conditional expectations in a remote correlated system; averaging over outcomes must not transmit a signal.

Required background. Normal channels and operations provide complete positivity and preduals, while Haag–Kastler locality supplies causal complements. Helpful background. Split inclusions clarify independent ancillas and Reeh–Schlieder warns against particle-localization intuition. See also local measurement instruments, selective operations, and localization and measurement cost.

Instruments and algebraic localization

For a finite outcome set $X$, a Heisenberg instrument on a von Neumann algebra $\mathcal M$ is a family of normal completely positive maps

$$\mathcal I_x:\mathcal M\longrightarrow\mathcal M, \qquad \sum_{x\in X}\mathcal I_x(1)=1.$$

The effect for outcome $x$ is $E_x=\mathcal I_x(1)$. In a normal state $\omega$, its probability is $p_x=\omega(E_x)$ and, when $p_x>0$, the conditional state is

$$\omega_x(A)=\frac{\omega(\mathcal I_x(A))}{p_x}.$$

The nonselective channel is $\mathcal I=\sum_x\mathcal I_x$. For general outcome spaces, countable additivity is imposed ultraweakly on an operation-valued measure. This is the Davies–Lewis notion of an instrument Davies and Lewis 1970, §§2–3, pp. 242–250.

The dilation supplies the proof mechanism. Prepare a normal probe state $\sigma$, apply a local normal $*$-automorphism generated by the coupling, and evaluate a probe effect with a normal slice map. Automorphisms and slice maps are completely positive and normal, so every outcome operation has those properties; summing a normalized probe observable gives a unital map. Fewster and Verch establish this construction for algebraic QFT on globally hyperbolic spacetimes Fewster and Verch 2020, §§3.1–3.3, pp. 859–870. A list of desired probabilities without such an operation-valued map is not yet an instrument.

Let $O$ be the coupling region of a local net. A sufficient localization condition is

$$\mathcal I(B)=B, \qquad B\in\mathcal A(O'),$$

together with localization of the effects in $\mathcal A(O)$ and compatibility with the net representation. The condition belongs to the sum of branches. Requiring $\omega_x(B)=\omega(B)$ for every branch would forbid ordinary conditioning on pre-existing Bell correlations and is not a locality principle.

A two-outcome Weyl instrument

Let $\phi$ be the free real scalar field and choose a real test function $f\in C_0^\infty(O)$. Its Weyl unitary is $W(f)=e^{i\phi(f)}\in\mathcal A(O)$. Couple a two-level probe by the controlled unitary

$$U=|0\rangle\langle0|\otimes1 +|1\rangle\langle1|\otimes W(f),$$

prepare the probe in $|+\rangle=(|0\rangle+|1\rangle)/\sqrt2$, and measure it in the $X$ basis. The system Kraus operators are

$$M_+=\frac{1+W(f)}2, \qquad M_-=\frac{1-W(f)}2,$$

and the instrument is $\mathcal I_\pm(A)=M_\pm^*AM_\pm$. Direct multiplication gives

$$M_+^*M_++M_-^*M_-=1,$$

so the two branches are normal completely positive and their sum is unital. Their effects encode the real part of the Weyl observable: $E_\pm=(2\pm W(f)\pm W(f)^*)/4$.

These effects are generally not projections. Indeed, $E_+^2=E_+$ would impose a special spectral restriction on $W(f)$ that a generic Weyl unitary does not satisfy. The probe therefore realizes an unsharp two-outcome system measurement even though its own $X$ measurement is sharp. This is expected: slicing out a prepared probe can make the induced system effect less sharp. Complete positivity, rather than projectivity of every effect, is the stable structural requirement.

If $B\in\mathcal A(O')$, locality gives $[B,W(f)]=0$. Therefore

$$\sum_{s=\pm}\mathcal I_s(B) =\frac12\bigl(B+W(f)^*BW(f)\bigr)=B.$$

This constructs the stated compactly smeared two-level probe measurement and proves that its nonselective action commutes with every causal-complement observable. It is the field-theoretic realization developed further on the local-instrument application page.

An independent normalization check uses probabilities: for every normal $\omega$,

$$p_++p_-=\omega(E_++E_-)=1,$$

while $|p_+-p_-|=|\operatorname{Re}\omega(W(f))|\leq1$ by unitarity. Thus each $p_\pm$ lies in $[0,1]$ without choosing a density matrix for $\mathcal A(O)$.

The tail test and selective conditioning

Replace $f$ by a profile that is not compactly supported and has nonzero causal symplectic pairing $E(f,g)$ with some $g$ supported in $O'$. The Weyl relations give

$$W(f)W(g)=e^{-iE(f,g)}W(g)W(f).$$

Unless the phase is trivial, the nonselective action sends $W(g)$ to

$$\mathcal I(W(g)) =\tfrac12\bigl(1+e^{iE(f,g)}\bigr)W(g),$$

not to $W(g)$. The alleged localization fails exactly where the compact-support hypothesis was dropped. Rapid decay is not a substitute for causal support.

Adversarial test. Choose $E(f,g)=\pi$. Then the factor multiplying $W(g)$ is zero, so the purportedly “local” nonselective operation erases the remote Weyl expectation completely. The failure is order one even if the tail producing the pairing is pointwise small; the relevant datum is causal symplectic support, not a visual estimate of the profile.

Even with compact $f$, a selected outcome can alter $\omega_x(B)$ for remote $B$ when the initial state correlates $\mathcal A(O)$ and $\mathcal A(O')$. No observer in $O'$ can choose or learn $x$ without a classical signal, and the average remains unchanged. This distinction prevents a false inference from conditional state change to superluminal control.

Exercises

1. Compute the effects. Derive $E_\pm$ and show $0\leq E_\pm\leq1$.

Solution

$E_\pm=M_\pm^*M_\pm=(2\pm W\pm W^*)/4$. Positivity follows from the form $M_\pm^*M_\pm$. Since $E_++E_-=1$, each positive effect is bounded above by one.

2. Remote expectation. Let $B$ commute with both Kraus operators. Show that its nonselective expectation is unchanged, but explain why a branch expectation need not be.

Solution

Commutation gives $\sum_s\omega(M_s^*BM_s)=\omega(B\sum_sM_s^*M_s)=\omega(B)$. A branch divides $\omega(BE_s)$ by $\omega(E_s)$, which equals $\omega(B)$ only when the state has the relevant factorization or zero covariance. Correlation, not causal influence, produces the difference.

References

• Davies, Edward B., and John T. Lewis. “An Operational Approach to Quantum Probability.” Communications in Mathematical Physics 17 (1970): 239–260. DOI.
• Fewster, Christopher J., and Rainer Verch. “Quantum Fields and Local Measurements.” Communications in Mathematical Physics 378 (2020): 851–889. DOI; Open PDF.