Quantum Sensors in the Field

A field sensor must estimate a physical quantity under variable environmental conditions with a usable uncertainty, update rate, and operating burden. Intrinsic sensitivity is only one part of that task. A quantum sensor can exploit quantized transitions or coherent interference without using entanglement; a claim of entanglement-enhanced performance requires an additional matched comparison.

Helpful background. Measurement as estimation defines the inference problem; noise and decoherence in metrology explains how probe coherence limits it. This page concerns the translation from an isolated probe to a deployed measurement system.

Information per wall-clock time

Suppose one interrogation produces Fisher information $F_1$ about parameter $\theta$, takes sensing time $t$, and requires dead time $t_d$ for reset and readout. For independent repeated shots over duration $T$,

$$F_{\mathrm{tot}}\approx \frac{T}{t+t_d}F_1, \qquad \operatorname{Var}(\widehat\theta) \ge\frac{t+t_d}{TF_1},$$

for locally unbiased estimators under the model’s regularity assumptions, with many complete cycles in $T$. Here $F_1$ is the classical information in the actual measurement outcomes; the inequality is a lower bound, not a claim that an estimator achieves it. Optimizing $F_1$ alone can select a poor instrument if it increases dead time or reduces duty cycle.

In a simplified Ramsey sensor estimating angular frequency $\theta=\omega$, the accumulated phase is $\omega t$. With $N$ independent probes, exponential visibility $e^{-t/T_2}$, and operation at the most informative fringe phase,

$$F_1=N t^2e^{-2t/T_2}.$$

For ordinary frequency $\nu=\omega/(2\pi)$ in hertz, the Fisher information instead contains $(2\pi t)^2$ in place of $t^2$. With zero dead time, maximizing $F_1/t$ gives $t=T_2/2$. With dead time, the optimum satisfies

$$\frac2t-\frac2{T_2}-\frac1{t+t_d}=0.$$

The instrument optimum differs from the isolated-shot optimum, which is $t=T_2$. As dead time grows, it becomes worthwhile to extract more information from each preparation before resetting; the optimal sensing time moves from $T_2/2$ toward $T_2$ in this model. Colored noise, local-oscillator drift, adaptive tracking, or probe interactions require a different likelihood. A long interrogation also narrows the unambiguous frequency range, so this local calculation presumes a known operating fringe or successful coarse tracking.

Degen, Reinhard, and Cappellaro provide the sensing framework behind such sensitivity calculations. Its field use requires the full acquisition and inference protocol to be specified.

Nuisance parameters consume information

Suppose the data depend on the desired parameter $\theta$ and nuisance parameters $\boldsymbol\zeta$, such as temperature, tilt, laser frequency, or a background field. Partition the classical Fisher matrix of the actual joint measurement model:

$$F= \begin{pmatrix} F_{\theta\theta}&F_{\theta\zeta}\\ F_{\zeta\theta}&F_{\zeta\zeta} \end{pmatrix}.$$

When the nuisance block is invertible, jointly estimating it leaves effective information

$$F_{\mathrm{eff}}= F_{\theta\theta} -F_{\theta\zeta}F_{\zeta\zeta}^{-1}F_{\zeta\theta}.$$

The subtraction quantifies confusion between signal and nuisance effects. If a magnetic field and temperature shift a single measured resonance in indistinguishable ways, more repetitions of that same measurement cannot separate them. Additional transitions, reference channels, modulation, or independent calibration can change identifiability.

The expression assumes a correctly specified regular statistical model. Substituting a quantum Fisher matrix does not generally make its multiparameter precision bound jointly attainable: optimal measurements for different parameters may be incompatible. Unknown systematic effects are not repaired by treating every fitted covariance matrix as complete uncertainty.

Sensitivity can improve while accuracy does not

An illustrative mean-square-error budget is

$$\mathrm{MSE}(T)= \frac{\eta^2}{uT}+b^2+\sigma_{\mathrm{drift}}^2(T).$$

Define $\eta$ by white-noise variance $\eta^2/T_{\mathrm{active}}$ after active averaging time $T_{\mathrm{active}}$, and let $u$ be the usable duty fraction. The term $b$ is a fixed residual bias; the last term is the variance of the drift contribution to the final estimate. The additive form assumes zero-mean white-noise and drift contributions with negligible covariance. If $\eta$ is already measured per total elapsed time, do not divide by $u$ again. A spectral-density convention must also be specified before converting a quoted noise amplitude into this averaging coefficient.

Take sensor A with $\eta=1$ unit$/\sqrt{\mathrm{Hz}}$ and $u=0.1$, and sensor B with $\eta=2$ in the same units and $u=0.9$. Ignoring bias and drift, A’s variance coefficient is $10/T$, whereas B’s is $4.44/T$. The intrinsically noisier sensor produces the more precise result over fixed wall-clock time.

At long times a fixed bias stops averaging from helping. Allan deviation can diagnose characteristic temporal behavior, but a low Allan deviation does not by itself establish absolute accuracy: a constant bias cancels from successive differences.

A field demonstration is a constrained inverse problem

A gravity instrument measures acceleration or gradients; an underground object is inferred through a model of geometry and density. A magnetometer measures fields; source localization adds a spatial model and often prior information. Application performance must therefore include uncertainty in the forward model as well as the sensor noise.

Stray and collaborators demonstrated atom-interferometric gravity-gradient measurements in an outdoor survey. Their analysis included the geometry of an underground feature and uncertainties in the inference. This is evidence for that instrument and survey protocol, not a general guarantee that every subsurface object is identifiable.

A useful comparison matches spatial resolution, observation time, environment, initialization, recalibration, and rejected data. If a reference sensor supplies vibration correction or drift compensation, include its cost and performance in the quantum system boundary. Entanglement gain should be measured at matched particle number and operational constraints, not inferred merely from the presence of a nonclassical state.

The useful gain is the one that survives preparation, measurement, calibration and inference. The metrology research guide discusses open questions about finite data, nuisance parameters and practical precision.

Exercises

For a normalized two-parameter Fisher matrix with diagonal entries one and off-diagonal correlation $r$, determine the information remaining about the first parameter when the second is unknown.

Solution

The Schur complement gives $F_{\mathrm{eff}}=1-r^2$. At $r=0.99$, only $0.0199$ remains. The variance bound increases by about a factor of $50.25$ relative to a known nuisance parameter.

A sensor has white-noise variance $4/T$ and residual bias $0.1$ in the measured units. At what duration do the two MSE contributions become equal?

Solution

Set $4/T=0.1^2$, obtaining $T=400$ in the time unit used to define the sensitivity. Longer averaging reduces the statistical term but leaves the bias contribution $0.01$.

For the Ramsey model, take reset and readout dead time $t_d=T_2$. Find the sensing time that maximizes information per total cycle time.

Solution

Set $x=t/T_2$ in the stationary condition. It becomes $2/x-2-1/(x+1)=0$, or $2x^2+x-2=0$. The positive root gives

$$t=\frac{\sqrt{17}-1}{4}T_2\approx0.7808T_2.$$

This lies between the zero-dead-time optimum $T_2/2$ and the single-shot optimum $T_2$. The information rate vanishes as $t$ tends to zero or infinity, so the positive stationary point is the maximum. It optimizes the stated local statistical model, not robustness to an unknown fringe or drift.

References

• C. L. Degen, F. Reinhard, and P. Cappellaro, “Quantum sensing,” Reviews of Modern Physics 89, 035002 (2017). Author preprint.
• B. Stray et al., “Quantum sensing for gravity cartography,” Nature 602, 590–594 (2022). doi:10.1038/s41586-021-04315-3.

Further reading

Sensing case studies compares specific tasks; gravimetry and inertial sensing and magnetometry provide platform-specific signal models.