Detector and Instrument Validation

A detector or instrument model is credible only when separate tests challenge its response, normalization, complete positivity, spacetime support, energy balance, perturbative approximation, numerical convergence, and data lineage. Agreement with the same calculation used to design the model is calibration, not independent validation. This page turns those principles into a fully reproducible inertial-detector benchmark and shows how seeded faults test whether the checks have power.

Required background. Localized detector models defines the physical probe, and local measurement instruments defines the probability and update maps that must be tested.

Helpful background. Switching, smearing, and regularization supplies the regulator and limit-order checks.

Chapter map. Use the route from a local coupling to a field instrument, the chapter claim-validity table, and the three independent validity questions to place this construction within the full measurement protocol.

Freeze the protocol before testing it

Record the field theory and state, detector Hilbert space, trajectory or worldtube, switching $\chi$, smearing $F$, coupling, initial probe state, readout POVM, perturbative order, discretization, and analytic or software version. If a parameter is changed after viewing a result, the result belongs to calibration; reserve an untouched observable, parameter point, or implementation for validation.

The main checks answer different questions:

Check	Question	Representative failure
Null	Does a forbidden or switched-off response vanish?	Background offset
Positivity and normalization	Is this a valid probability law or instrument?	Negative branch probability
Analytic benchmark	Do conventions and numerical integrals agree with an independent formula?	Fourier-factor or gap error
Causal support	Does the nonselective operation act trivially on the causal complement?	Switching or smearing tail
Convergence	Are cutoffs and quadratures resolved independently?	Shared unresolved grid
Energy balance	Does control work equal field plus probe energy change?	Missing switching work
Perturbative scaling	Is the omitted order smaller than the claim tolerance?	Hidden $O(\lambda^4)$ term
Cross-model	Does a second realization expose implementation dependence?	Same POVM, different update

Passing one row does not license another. In particular, a nonnegative response does not prove causal support, and a small quadrature residual does not bound a perturbative remainder.

Analytic vacuum-response benchmark

Consider a two-level Unruh–DeWitt detector at rest in the four-dimensional massless-scalar vacuum. Set the monopole matrix element to one and use

$$\chi(t)=e^{-t^2/(2T^2)}, \qquad F(\mathbf x)=\frac{e^{-\mathbf x^2/(2L^2)}}{(2\pi L^2)^{3/2}}.$$

Fix Fourier conventions by

$$\widetilde\chi(\nu)=\int dt\,e^{-i\nu t}\chi(t) =\sqrt{2\pi}\,T e^{-T^2\nu^2/2},$$ $$\widetilde F(\mathbf k)=\int d^3x\,e^{-i\mathbf k\cdot\mathbf x}F(\mathbf x) =e^{-L^2\mathbf k^2/2}.$$

The leading excitation probability is $P_{g\to e}=\lambda^2\mathcal F(\Omega)+O(\lambda^4)$. The general double-integral response is given in Satz 2007, Eqs. (2.2)–(2.5). In the present inertial smeared model, the mode expansion gives the independent spectral representation

$$\mathcal F(\Omega) =\int\frac{d^3k}{2(2\pi)^3|\mathbf k|} \left|\widetilde F(\mathbf k)\right|^2 \left|\widetilde\chi(\Omega+|\mathbf k|)\right|^2.$$

Angular integration reduces this to one positive integral,

$$\mathcal F(\Omega) =\frac{T^2}{2\pi}\int_0^\infty dk\,k \exp\!\left[-L^2k^2-T^2(k+\Omega)^2\right].$$

Define

$$a=L^2+T^2, \qquad b=T^2\Omega, \qquad c=T^2\Omega^2.$$

Completing the square gives the analytic value

$$\mathcal F(\Omega) =\frac{T^2e^{-c}}{2\pi} \left[ \frac{1}{2a} -\frac{b\sqrt\pi}{2a\sqrt a} e^{b^2/a}\operatorname{erfc}\!\left(\frac{b}{\sqrt a}\right) \right].$$

Take $T$ as the time unit, so numerical energies are reported in $T^{-1}$, and choose

$$\frac{L}{T}=1, \qquad \Omega T=1, \qquad \lambda=0.1,$$

the exact leading response and probability are

$$\mathcal F(1/T)=0.005039976195438958,$$ $$P_{g\to e}^{(2)} =\lambda^2\mathcal F(1) =0.0000503997619543896.$$

The switched-off null is exactly $P_{g\to e}^{(2)}(\lambda=0)=0$. A numerical implementation should reproduce the response without using the closed form. Composite Simpson integration on $0\le kT\le8$ gives:

Subintervals $N$	Numerical $\mathcal F(1/T)$	Absolute error from the analytic value
32	$0.005037848639437736$	$2.13\times10^{-6}$
64	$0.005039950696539459$	$2.55\times10^{-8}$
128	$0.005039975819240664$	$3.76\times10^{-10}$
256	$0.005039976189642097$	$5.80\times10^{-12}$

The integrand is positive, the radial measure contributes the factor $k$, and the $2\pi$ coefficient is fixed by the declared Fourier transform. These simple invariants catch many convention errors. Spatial-profile regularization and the pointlike limit are analyzed in Louko and Satz 2006, §§ 2–4; smooth switching and the sharp-switching limit are separated in Satz 2007, §§ 3–4.

Fault injection demonstrates test power

A validation suite should fail on deliberately corrupted inputs.

Gap bias. Replace the declared $\Omega=1$ by $\Omega=1.05$ in the integrand while leaving the report unchanged. The analytic formula gives

$$\mathcal F(1.05)=0.004349377658520056,$$

a relative shift of $-13.7024\%$. The analytic comparison catches this error even if the biased numerical quadrature is internally converged.

Perturbative-order error. Feed the analysis synthetic data

$$P_{\mathrm{seed}}(\lambda) =\lambda^2\mathcal F(1)(1+0.8\lambda^2)$$

while fitting a pure quadratic. The diagnostic ratio is

$$\frac{P_{\mathrm{seed}}(\lambda)}{\lambda^2\mathcal F(1)} =1+0.8\lambda^2,$$

which equals $1.008$, $1.032$, and $1.072$ at $\lambda=0.1,0.2,0.3$. A multi-coupling test therefore exposes the missing order even when a single small-$\lambda$ run looks accurate.

Support leak. Both Gaussian profiles in the benchmark have nonzero tails and therefore cannot validate an exact spacelike-support claim. For the causal test, replace them by a temporal bump $\chi_c\in C_0^\infty(\mathbb R)$ and a normalized spatial bump $F_c\in C_0^\infty(\mathbb R^3)$, and verify that the complete worldtube $K=\operatorname{supp}\chi_c\times\operatorname{supp}F_c$ is spacelike separated from the receiver. Then seed the temporal leak

$$\chi_{\mathrm{seed}}(t) =\chi_c(t)+10^{-3}\chi_c(t-t_{\mathrm{leak}}),$$

while retaining the compact spatial profile, with the translated temporal bump making the full leaked worldtube enter the receiver’s causal past. The geometry checker must now reject exact spacelike separation; geometry alone does not prove that one chosen marginal changes. To test the dynamics as well, choose a receiver observable and state for which the leaked coupling has a proved nonzero response, then require the implementation to detect that response. A complementary seed can translate $F_c$ so that its support enters the receiver’s causal past while leaving $\chi_c$ fixed. Without a demonstrated dynamical response, downgrade the result only to “exact no signaling not certified,” not to a claim that signaling was observed. Causal factorization for supported probe couplings is the independent structural check; see Fewster and Verch 2020, Theorem 3.5 and Eqs. (3.28)–(3.30).

If any seeded fault passes, the corresponding unseeded claim is not validated. This is especially important when two apparently independent calculations reuse the same code, regulator, or Fourier convention.

Acceptance record and evidence boundary

A useful acceptance quantity for non-null runs is

$$R_n =\frac{|P_{n+1}-P_n|} {\max(P_{n+1},P_{\mathrm{scale}})},$$

where $P_{\mathrm{scale}}$ prevents meaningless relative errors near zero. Report time and momentum domains, grid sequences, tail estimates, perturbative fits, input units, source versions and hashes, random seeds where applicable, raw and transformed outputs, failed and passed null tests, and the rule used to accept or downgrade the claim. Preserve negative and seeded-failure results.

Scientific evidence boundary, checked 2026-08-26. The cited literature and the benchmark above validate the stated leading-order, smeared inertial model and the local system–probe composition framework. They do not validate a sharp detector, a nonperturbative response, exact support for Gaussian profiles, or every detector trajectory and field state. Each extension needs its own regulator, convergence, causal, and approximation evidence.

Common pitfalls

Using two quadratures with the same unresolved cutoff. Agreement is not independent if both methods share the same bias. Compare a closed form or a genuinely different representation.

Interpreting the Gaussian benchmark as compact support. Gaussian profiles are analytically convenient and never vanish exactly. Use a compact profile for exact causal separation or state a tail bound.

Combining numerical and perturbative error. Grid refinement controls integration, not omitted $O(\lambda^4)$ physics. Estimate the two remainders separately.

Exercises

Starting from the three-dimensional spectral integral, derive the one-dimensional coefficient $T^2/(2\pi)$.

Solution

Write $d^3k=4\pi k^2dk$. The field normalization contributes $[2(2\pi)^3k]^{-1}$, so the angular integral gives $k/(4\pi^2)$. Since $|\widetilde\chi|^2=2\pi T^2e^{-T^2(k+\Omega)^2}$, multiplication leaves $T^2k/(2\pi)$. The spatial factor is $|\widetilde F|^2=e^{-L^2k^2}$.

Evaluate the radial integral analytically by completing the square.

Solution

With $a,b,c$ as defined,

$$I=e^{-c}\int_0^\infty dk\,k e^{-ak^2-2bk}.$$

Let

$$J(b)=\int_0^\infty dk\,e^{-ak^2-2bk} =\frac{\sqrt\pi}{2\sqrt a}e^{b^2/a} \operatorname{erfc}\!\left(\frac b{\sqrt a}\right).$$

Because $I=-e^{-c}J'(b)/2$ and $J'(b)=2bJ/a-1/a$, one obtains

$$I=e^{-c}\left[ \frac1{2a}-\frac{b\sqrt\pi}{2a\sqrt a} e^{b^2/a}\operatorname{erfc}\!\left(\frac b{\sqrt a}\right) \right].$$

Multiplication by $T^2/(2\pi)$ gives the displayed response.

Why can the synthetic fourth-order seed evade a single run at $\lambda=0.1$ but not the three-point scaling test?

Solution

At $\lambda=0.1$, the correction multiplies the quadratic result by only $1.008$, so an $0.8\%$ discrepancy could be absorbed into calibration or a loose error bar. Dividing by $\lambda^2$ at three couplings exposes a systematic increase: the normalized ratios are $1.008$, $1.032$, and $1.072$. A true leading-order coefficient would be constant up to the declared remainder, so the curvature diagnoses the seeded $\lambda^4$ term.

References

• Fewster, C. J., and Verch, R. (2020). “Quantum Fields and Local Measurements.” Communications in Mathematical Physics 378, 851–889. DOI. Open PDF.
• Louko, J., and Satz, A. (2006). “How Often Does the Unruh–DeWitt Detector Click? Regularisation by a Spatial Profile.” Classical and Quantum Gravity 23, 6321–6344. DOI. Open PDF.
• Satz, A. (2007). “Then Again, How Often Does the Unruh–DeWitt Detector Click If We Switch It Carefully?” Classical and Quantum Gravity 24, 1719–1732. DOI. Open PDF.