Atomic Clocks as Quantum Estimators

A passive atomic clock is a sequential quantum estimator wrapped around an oscillator. During each cycle, an ensemble of atoms compares the oscillator phase with a reproducible transition. A measurement converts the accumulated relative phase into data, an estimator updates the inferred oscillator state, and a controller applies a correction. The continuously available output is the disciplined oscillator, not a stream of autonomous atomic ticks.

A narrow Ramsey fringe or a small state-preparation variance is only one ingredient. The oscillator continues to drift between atomic measurements, and the controller must identify the correct fringe branch while correcting that drift. This page develops the statistical consequences of repeated interrogation; AMO Atomic Clocks develops physical clock architectures, systematic shifts, and the sensitivity-function treatment.

Required background. Parameter estimation and Fisher information provide likelihoods and local bounds; the standard quantum limit and Ramsey interferometry supply the independent-atom readout.

Atomic measurements and oscillator feedback

A transition with unperturbed angular frequency $\omega_0$ provides the quantum reference. The microwave or optical oscillator carries phase continuously, while each interrogation compares its phase with the atoms. An estimator converts the outcome into an inferred frequency error, and the controller uses that error to steer the oscillator.

The atoms are excellent long-term references but are observed intermittently. The local oscillator supplies continuity but wanders. The clock is therefore a hybrid quantum–classical control system whose state is only partially observed.

An in-loop error signal can be small because feedback suppresses it. That does not independently demonstrate that the output is correct. A trustworthy clock requires a second measurement path, a comparison with another reference, or a validated model showing how the in-loop record predicts out-of-loop behavior.

The Per-Cycle Phase Channel

Let the local oscillator have instantaneous fractional frequency offset

$$y_{\mathrm{LO}}(t) = \frac{\omega_{\mathrm{LO}}(t)-\omega_0}{\omega_0}.$$

During cycle $j$, the interrogation sensitivity function $g_j(t)$ weights oscillator fluctuations. A useful relative-phase model is

$$\phi_j = \omega_0 \int_{t_j}^{t_j+T_c} g_j(t)y_{\mathrm{LO}}(t)\,dt + \phi_{\mathrm{sys},j} + \phi_{\mathrm{ctrl},j}.$$

Here $T_c$ is the full cycle time, $\phi_{\mathrm{sys},j}$ collects atomic and apparatus shifts, and $\phi_{\mathrm{ctrl},j}$ is the known analysis or servo phase. For ideal Ramsey free evolution of duration $T$, $g_j(t)=1$ during the dark interval and zero elsewhere, so a constant detuning gives

$$\phi_j=\delta\omega_jT+\phi_{\mathrm{sys},j} +\phi_{\mathrm{ctrl},j}.$$

Conditional on the phase history, $N$ independent two-level atoms with state-resolved population detection have a binomial excited count:

$$K_j\sim\operatorname{Binomial} \!\left(N,p_e(\phi_j)\right),$$

with an effective Ramsey fringe

$$p_e(\phi_j) = \frac{1+C\sin\phi_j}{2}$$

when the analysis phase places the clock at quadrature. Here $C$ describes the conditional single-atom contrast, including independent dephasing or preparation imperfections. An unresolved oscillator phase shared by the atoms must instead be marginalized over the whole binomial likelihood; it correlates their outcomes. Replacing it by a reduced marginal contrast does not preserve independent-binomial statistics. Detection errors likewise require a response model.

The atomic outcome is not a direct observation of instantaneous frequency. It is a noisy, periodic observation of an integrated phase. This distinction is central to clock design.

Phase and Frequency Information

For one atom with the quadrature likelihood above, the phase FI is $C^2$ at $\phi=0$. For $N$ independent atoms,

$$F_\phi^{\mathrm{CSS}} = NC^2,$$

where CSS denotes a coherent spin state. Since $\phi=\delta\omega T$, information transforms as

$$F_{\delta\omega}^{\mathrm{CSS}} = NC^2T^2.$$

If independent cycles of duration $T_c$ can be combined and the oscillator model does not introduce an additional limit, the information rate is

$$\dot F_{\delta\omega} = \frac{NC^2T^2}{T_c}.$$

Writing the duty factor as $D=T/T_c$ gives

$$\dot F_{\delta\omega} = NC^2TD.$$

This formula makes four levers visible: atom number, contrast, interrogation time, and duty factor. It also prevents preparation and readout time from disappearing when two protocols are compared.

For averaging time $\tau\gg T_c$, the corresponding atom-projection-noise benchmark for fractional frequency is

$$\sigma_y^{\mathrm{QPN}}(\tau) \simeq \frac{1}{\omega_0CT\sqrt N} \sqrt{\frac{T_c}{\tau}}.$$

The numerical prefactor depends on the modulation, line shape, estimator, and definition of one error sample. This is an independent-atom benchmark under a local, correctly locked model. It is not automatically the instability of a running clock.

The high carrier frequency of an optical transition improves fractional frequency sensitivity through the factor $1/\omega_0$. It does not by itself remove laser noise, systematic shifts, dead time, or transfer uncertainty.

The Clock as a State-Space Model

Successive clock cycles are generally not independent. Oscillator frequency noise has memory, the servo feeds previous outcomes into future controls, and slow systematic variables drift. A state-space model makes these dependencies explicit. Let $x_j$ contain the oscillator offset, drift, and selected nuisance parameters; let $a_j$ denote the known interrogation settings. After observing $K_j$, the controller chooses the correction $u_j$ for the next transition:

$$K_j\sim p(K_j|x_j,a_j),\qquad x_{j+1}=f(x_j,u_j)+w_j.$$

Write $\pi_j^-(x)$ for the prior before measurement $j$ and $\pi_j^+(x)$ for its posterior. The two operations are

$$\begin{aligned} \pi_j^-(x) &=\int p(x|x',u_{j-1})\,\pi_{j-1}^+(x')\,dx',\\ \pi_j^+(x) &\propto p(K_j|x,a_j)\,\pi_j^-(x). \end{aligned}$$

The first line predicts the current state from the previous posterior. The second updates that same state with the current measurement. The controller then chooses $u_j$ from $\pi_j^+$. In a locally linear Gaussian model this becomes a Kalman filter; broad periodic likelihoods can require a circular, grid-based, or particle representation.

A scalar example shows why measurement precision and oscillator quality must be optimized together. Suppose the latent cycle-to-cycle frequency follows a random walk,

$$x_{j+1}=x_j+w_j, \qquad \operatorname{Var}(w_j)=q,$$

and a locally linear atomic readout is

$$r_j=x_j+v_j, \qquad \operatorname{Var}(v_j)=R.$$

If $P$ is the steady-state posterior variance, prediction gives $P+q$ and measurement update gives

$$P = \frac{R(P+q)}{P+q+R}.$$

The nonnegative solution is

$$P = \frac{ \sqrt{q^2+4qR}-q }{2}.$$

Reducing atomic measurement noise $R$ helps, but the gain depends on process noise $q$. When $q$ is large, oscillator unpredictability replenishes uncertainty between observations. When $q$ is very small, the filter can average many weak measurements statistically. The prediction variance before the next observation is $P+q$, even if the posterior $P$ is small. A continuously available clock output is affected by this between-measurement uncertainty, not just the uncertainty immediately after an update.

Local Oscillator Noise Is Part of the Task

The local oscillator is not merely a technical imperfection external to quantum metrology. Its phase is precisely what the atoms are asked to estimate and stabilize. Oscillator noise can

• broaden the phase prior before each measurement;
• reduce observed contrast when untracked;
• drive the state outside the discriminator’s local range;
• correlate outcomes from different cycles;
• enter through pulse phases and optical-path delivery;
• be aliased by dead time; and
• set the benefit obtainable from lower atomic readout noise.

If an untracked Gaussian phase has variance $\sigma_{\mathrm{LO}}^2(T)$, its average coherence factor is

$$C_{\mathrm{LO}}(T) = \exp\!\left[ -\frac{\sigma_{\mathrm{LO}}^2(T)}{2} \right].$$

Substituting this factor into a fringe model captures contrast loss but not every consequence of colored oscillator noise. The same noise also determines the phase prior, cycle-to-cycle correlations, and slip probability. A full clock model therefore uses a phase-noise or fractional-frequency-noise spectrum and the actual sensitivity function. Noise Spectra develops spectral conventions; the AMO Atomic Clocks page derives the clock-specific Dick-effect sum.

Dead time and aliasing

Preparation and detection create intervals during which the atoms do not observe the oscillator. Periodic sampling mixes local-oscillator noise near harmonics of $1/T_c$ into the clock record: the Dick effect. Increasing $N$ or squeezing the atomic readout does not suppress that aliased oscillator noise. Indeed, reducing projection noise can reveal a Dick-noise floor that was previously hidden.

Important mitigation strategies have different resource costs:

• faster preparation and readout raise duty factor;
• interleaved ensembles can provide nearly continuous sensitivity;
• a quieter oscillator reduces phase diffusion but may require a larger cryogenic, optical, or vibration-isolated subsystem;
• synchronous comparison rejects shared oscillator noise but changes the claim from independent stability to differential stability;
• nondestructive or weak measurements can track phase while adding measurement backaction and control complexity.

Achieving nearly continuous atomic sensitivity can require an extra ensemble and coordinated interrogation. Its additional controls and optical paths affect both practical cost and frequency uncertainty.

Phase Wraps and Clock Reliability

A two-outcome Ramsey fringe is periodic. Near a chosen operating point the error signal is approximately linear, but outside its capture interval the same measurement can indicate the wrong branch. For a zero-mean Gaussian predicted phase with standard deviation $\sigma_\phi$, the probability of leaving the nominal capture interval $|\phi|<\phi_{\mathrm{cap}}$ is

$$P_{\mathrm{out}} \simeq \operatorname{erfc}\!\left( \frac{\phi_{\mathrm{cap}}}{\sqrt2\sigma_\phi} \right).$$

The actual slip probability also depends on measurement noise, the controller, and recovery logic. Identifying $P_{\mathrm{slip}}$ with this tail probability is a simplified model in which every excursion causes a slip.

Even a small per-cycle probability can matter in a long data set. If there are $M$ nearly independent opportunities to slip, then

$$P(\text{at least one slip}) = 1-(1-P_{\mathrm{slip}})^M \simeq MP_{\mathrm{slip}}$$

when $MP_{\mathrm{slip}}\ll1$. Without independence, the union bound still gives $P(\text{at least one slip})\le MP_{\mathrm{slip}}$ for equal marginal probabilities; the exact product formula need not hold.

Longer interrogation increases local frequency information as $T^2$ but also widens the oscillator phase distribution and narrows the unambiguous frequency range. Multi-ensemble protocols can use short or dual-quadrature interrogations to identify the branch for a longer interrogation. Adaptive protocols can carry a full phase posterior and choose $T$ or analysis phase from its width. Every auxiliary ensemble and short interrogation counts toward the global resource comparison.

A clock that occasionally makes undetected fringe hops may have an impressive conditional Allan deviation after bad runs are removed. Reliability claims must report the detection rule, removed intervals, false-alarm rate, recovery time, and effect on availability.

Choosing the Interrogation Time

If independent atomic coherence decays as

$$C(T)=e^{-\Gamma T},$$

then the ideal CSS information rate is

$$\dot F_{\delta\omega}(T) = \frac{NT^2e^{-2\Gamma T}}{T+t_d},$$

where $t_d=T_c-T$ is fixed dead time. The optimal $T$ solves

$$\frac{2}{T} - 2\Gamma - \frac{1}{T+t_d} =0.$$

With zero dead time, $T_*=1/(2\Gamma)$. Positive dead time makes longer interrogations relatively more valuable because each cycle carries fixed overhead. Oscillator phase wraps or non-Markovian noise can impose a shorter optimum than this contrast-only model predicts.

This optimization illustrates a recurring rule: maximize information or task utility per wall-clock resource, not information per successful interrogation in isolation.

Entangled Probes in a Clock

A CSS has frequency QFI $NT^2$; an ideal $N$-atom GHZ state has $N^2T^2$. The latter also has a phase period reduced by $N$, faster decoherence under independent noise, and potentially slower preparation. Heisenberg Scaling derives these generator and range comparisons.

For independent Markovian dephasing with single-atom contrast $e^{-\Gamma T}$ and no overhead, the information rates are

$$\dot F_{\mathrm{CSS}}=NT e^{-2\Gamma T},\qquad \dot F_{\mathrm{GHZ}}=N^2T e^{-2N\Gamma T}.$$

Their optimal times are $1/(2\Gamma)$ and $1/(2N\Gamma)$, respectively, and both attain $N/(2e\Gamma)$. This is a comparison under a specific noise and timing model, not a prohibition on entanglement-enhanced clocks. Other states or noise structures can give different improvements. The derivation is developed in Noise and Decoherence in Metrology.

Spin squeezing

A spin-squeezed state is often more practical than a GHZ state. Its Wineland parameter

$$\xi_R^2 = \frac{N(\Delta J_\perp)^2}{|\langle\mathbf J\rangle|^2}$$

already compares noise with signal response. In a locally linear Ramsey measurement, an idealized phase variance is

$$(\Delta\phi)^2 \simeq \frac{\xi_R^2}{N},$$

For a locally Gaussian readout with parameter-independent variance this corresponds to $F_\phi=N/\xi_R^2$; more generally these first and second moments give a lower bound on the readout FI. Spin Squeezing develops the covariance, entanglement witnesses, preparation, and readout.

For the running clock, compare the phase variance after the actual interrogation, preparation time, failed cycles, oscillator noise, and readout. The worked example below shows explicitly how the contrast of the reference and the longer cycle time change the expected gain.

Synchronous and Independent Comparisons

Two clocks interrogated synchronously with equal sensitivity to a shared local oscillator can reject its phase. Choosing the difference sign so the atomic contribution is $\omega_A-\omega_B$, a simplified phase is

$$d_j = (\omega_A-\omega_B)T + n_{A,j}-n_{B,j} + \phi_{\mathrm{link},j},$$

where the common oscillator term has canceled. This can reveal atomic projection noise or a frequency ratio beneath the free-running oscillator noise. For different carrier frequencies or unequal sensitivity functions, frequency-transfer factors and appropriate weighting are required for this cancellation. It does not directly show how either clock would perform with an independent oscillator.

For two fractional-frequency estimates with covariance, the differential variance is

$$\operatorname{Var}(y_A-y_B) = \operatorname{Var}(y_A) + \operatorname{Var}(y_B) - 2\operatorname{Cov}(y_A,y_B).$$

Ignoring the covariance can either exaggerate or hide noise. Shared lasers, environmental monitors, atom preparation, transfer links, and data processing all create correlations. Conversely, an independent comparison includes both clocks and the link; attributing the result to one clock requires an additional model or a third reference.

The same distinction applies to entanglement-enhanced comparisons. A differential gain with a common oscillator establishes a valuable network or ratio result under that architecture. A claim about stand-alone timekeeping must include the oscillator and control resources needed at each node.

Stability, Accuracy, and Estimator Risk

Define fractional frequency deviation

$$y(t)=\frac{\nu(t)-\nu_0}{\nu_0}.$$

For adjacent averages $\bar y_k(\tau)$, Allan variance is

$$\sigma_y^2(\tau) = \frac12 \mathbb E\!\left[ (\bar y_{k+1}-\bar y_k)^2 \right].$$

It is a stability statistic for a specified record and averaging time. It is not automatically a frequentist confidence interval for the mean, a Bayesian posterior width, or a systematic uncertainty. Drift, flicker floors, gaps, servo transients, and selection can all invalidate a casual $1/\sqrt\tau$ extrapolation.

Systematic accuracy is a separate estimation problem. A schematic clock measurement equation is

$$y_{\mathrm{corr}} = y_{\mathrm{meas}} - \sum_r \beta_r x_r,$$

where $x_r$ are environmental or operating variables and $\beta_r$ are sensitivities. Uncertainty in both, together with covariance, propagates to the corrected result. Quantum Fisher information for the Ramsey phase does not bound an unmodeled bias. Detailed shift evaluation belongs to the AMO clock pages.

A mature result therefore keeps at least four objects separate:

• single-cycle phase resolution;
• locked-clock instability versus averaging time;
• systematic uncertainty of the realized frequency;
• task loss, such as time error, ratio uncertainty, availability, or anomaly-detection probability.

Optimizing one does not guarantee improvement in the others.

Worked comparison with reduced contrast and longer cycles

Consider an optical clock with

$$\nu_0=4.3\times10^{14}\,\mathrm{Hz}, \quad N=10^4, \quad T=0.50\,\mathrm{s},$$

$C=0.80$, and $T_c=1.0\,\mathrm{s}$. The white-frequency-noise model has the one-second coefficient

$$\begin{aligned} \sigma_y^{\mathrm{QPN}}(1\,\mathrm{s}) &\simeq \frac{1}{ 2\pi(4.3\times10^{14})(0.80)(0.50)(100) } \\ &\simeq 9.3\times10^{-18}. \end{aligned}$$

Suppose a squeezed protocol has metrological parameter $\xi_R^2=10^{-5/10}\simeq0.316$ but increases cycle time to $1.20\,\mathrm{s}$. If all other clock noise is absent and $\xi_R^2$ already includes the response reduction, its phase variance is $\xi_R^2/N$. The actual CSS reference has phase variance $1/(NC^2)$, so for averaging times long compared with both cycles the instability ratio is

$$\frac{\sigma_y^{\mathrm{sq}}}{\sigma_y^{\mathrm{CSS}}} = C\sqrt{\xi_R^2} \sqrt{\frac{1.20}{1.00}} \simeq 0.493.$$

The idealized improvement relative to this contrast-$0.80$ reference is about $1/0.493\simeq2.03$. The $5$ dB Wineland value was defined relative to a unit-contrast coherent spin, so it is not itself the variance ratio to the degraded reference used here. If an experiment instead reports a variance ratio of $0.316$ directly relative to its actual CSS protocol, the cycle-time penalty would give $\sqrt{0.316\times1.20}\simeq0.616$. State which reference the reported gain uses. If Dick noise or oscillator phase slips already exceed the CSS projection-noise term, the total improvement is smaller and may vanish.

Exercises

1. CSS and GHZ frequency information

For $N$ two-level atoms interrogated for time $T$, calculate the ideal frequency QFI of a CSS and a GHZ state. What additional issue appears in the GHZ likelihood?

Solution

Frequency is encoded by $e^{-i\delta\omega TJ_z}$, so

$$F_{\delta\omega} = 4T^2\operatorname{Var}(J_z).$$

For a CSS, $\operatorname{Var}(J_z)=N/4$, giving

$$F_{\delta\omega}^{\mathrm{CSS}}=NT^2.$$

For a GHZ state, $\operatorname{Var}(J_z)=N^2/4$, giving

$$F_{\delta\omega}^{\mathrm{GHZ}}=N^2T^2.$$

The GHZ signal depends on $N\delta\omega T$, so its phase period is reduced by a factor $N$. A single sine readout also has a reflection ambiguity within that period. The local QFI does not resolve those branches.

2. Projection-noise benchmark

A Ramsey clock has $N=2500$, $C=0.75$, $T=0.20\,\mathrm{s}$, $T_c=0.50\,\mathrm{s}$, and $\nu_0=9.192631770\,\mathrm{GHz}$. Estimate its QPN-limited Allan deviation at $\tau=100\,\mathrm{s}$.

Solution

Use

$$\sigma_y^{\mathrm{QPN}}(\tau) = \frac{1}{2\pi\nu_0CT\sqrt N} \sqrt{\frac{T_c}{\tau}}.$$

Substitution gives

$$\begin{aligned} \sigma_y^{\mathrm{QPN}}(100\,\mathrm{s}) &= \frac{ \sqrt{0.50/100} }{ 2\pi(9.192631770\times10^9)(0.75)(0.20)(50) } \\ &\simeq 1.63\times10^{-13}. \end{aligned}$$

This is a projection-noise benchmark. It omits oscillator noise, Dick aliasing, detection noise, shifts, and servo dynamics.

3. Optimum interrogation with dephasing

Assume zero dead time and $C(T)=e^{-\Gamma T}$. Maximize the CSS frequency information rate and find its maximum.

Solution

With $T_c=T$,

$$\dot F(T)=NTe^{-2\Gamma T}.$$

Differentiating,

$$\frac{d\dot F}{dT} = Ne^{-2\Gamma T}(1-2\Gamma T).$$

Thus

$$T_*=\frac{1}{2\Gamma},$$

and

$$\dot F(T_*) = \frac{N}{2e\Gamma}.$$

The result assumes independent Markovian dephasing and no phase-wrap or oscillator constraint.

4. Fixed overhead shifts the optimum

For $\Gamma=1\,\mathrm{s}^{-1}$ and fixed dead time $t_d=0.20\,\mathrm{s}$, solve the interrogation-time condition

$$\frac2T-2\Gamma-\frac1{T+t_d}=0.$$

Solution

Multiplying by $T(T+t_d)$ gives

$$2(T+t_d)-2\Gamma T(T+t_d)-T=0.$$

With the stated values,

$$2T^2-0.6T-0.4=0.$$

The positive root is

$$T_* = \frac{0.6+\sqrt{0.36+3.2}}{4} \simeq 0.622\,\mathrm{s}.$$

This exceeds the zero-dead-time optimum $0.5\,\mathrm{s}$ because a longer interrogation amortizes the fixed overhead.

5. Steady-state tracking variance

For the scalar random-walk clock model, let process variance be $q=4\times10^{-4}$ and atomic measurement variance be $R=9\times10^{-4}$. Find the steady-state posterior variance $P$ and standard deviation.

Solution

Use

$$P = \frac{\sqrt{q^2+4qR}-q}{2}.$$

Then

$$\begin{aligned} P &= \frac{ \sqrt{(4\times10^{-4})^2 +4(4\times10^{-4})(9\times10^{-4})} -4\times10^{-4} }{2} \\ &\simeq 4.32\times10^{-4}. \end{aligned}$$

Therefore

$$\sqrt P\simeq2.08\times10^{-2}.$$

The posterior variance is neither $q$ nor $R$; it is set by their dynamical balance.

6. Long-run phase-slip risk

A clock has Gaussian predicted phase with $\sigma_\phi=0.25\,\mathrm{rad}$ and capture half-width $\phi_{\mathrm{cap}}=\pi/2$. Estimate the outside-capture probability. Under the additional approximation that each excursion causes a slip, find the probability of at least one slip in $10^9$ independent cycles.

Solution

The Gaussian outside-capture probability is

$$P_{\mathrm{slip}} \simeq \operatorname{erfc}\!\left( \frac{\pi/2}{\sqrt2(0.25)} \right) = \operatorname{erfc}(4.44).$$

Numerically this is approximately

$$P_{\mathrm{slip}} \simeq 3.32\times10^{-10}$$

per independent opportunity. Over $10^9$ cycles, the small-probability estimate gives an expected slip count of about $0.332$ and probability $1-e^{-0.332}\simeq0.282$ of at least one slip. Very small per-cycle risks can therefore matter in long operation.

7. A synchronous comparison

Two clocks have individual fractional-frequency variances $9$ and $16$ in arbitrary units and covariance $10$ because they share a local oscillator. Find the variance of their difference. What value would be inferred if the covariance were ignored?

Solution

The differential variance is

$$\operatorname{Var}(y_A-y_B) = 9+16-2(10) = 5.$$

Ignoring covariance would give $25$, five times too large. The shared oscillator can be strongly rejected in the difference. The low differential variance does not imply that either stand-alone output has variance $5$.

8. Why Allan variance does not measure a constant bias

For fractional frequency $y(t)=b+at$, compute the Allan variance using adjacent intervals of duration $\tau$. What happens for a constant offset?

Solution

The interval average is $\bar y_k=b+a(k+\tfrac12)\tau$. Adjacent averages differ by $a\tau$, giving

$$\sigma_y^2(\tau)=\frac{a^2\tau^2}{2}.$$

For $a=0$, the Allan variance is zero even when the constant offset $b$ is nonzero. A clock can be stable while systematically shifted. Conversely, linear drift produces Allan deviation proportional to $\tau$, so extending an independent-white-noise $\tau^{-1/2}$ law would give the wrong behavior.

References

• D. W. Allan, “Statistics of atomic frequency standards,” Proceedings of the IEEE 54, 221–230 (1966), doi:10.1109/PROC.1966.4634.
• S. F. Huelga, C. Macchiavello, T. Pellizzari, A. K. Ekert, M. B. Plenio, and J. I. Cirac, “Improvement of frequency standards with quantum entanglement,” Physical Review Letters 79, 3865–3868 (1997), doi:10.1103/PhysRevLett.79.3865.
• W. M. Itano, J. C. Bergquist, J. J. Bollinger, J. M. Gilligan, D. J. Heinzen, F. L. Moore, M. G. Raizen, and D. J. Wineland, “Quantum projection noise: Population fluctuations in two-level systems,” Physical Review A 47, 3554–3570 (1993), doi:10.1103/PhysRevA.47.3554.
• E. M. Kessler, P. Kómár, M. Bishof, L. Jiang, A. S. Sørensen, J. Ye, and M. D. Lukin, “Heisenberg-limited atom clocks based on entangled qubits,” Physical Review Letters 112, 190403 (2014), doi:10.1103/PhysRevLett.112.190403.
• A. D. Ludlow, M. M. Boyd, J. Ye, E. Peik, and P. O. Schmidt, “Optical atomic clocks,” Reviews of Modern Physics 87, 637–701 (2015), doi:10.1103/RevModPhys.87.637.
• G. Santarelli, C. Audoin, A. Makdissi, P. Laurent, G. J. Dick, and A. Clairon, “Frequency stability degradation of an oscillator slaved to a periodically interrogated atomic resonator,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control 45, 887–894 (1998), doi:10.1109/58.710548.
• M. Schulte, C. Lisdat, P. O. Schmidt, U. Sterr, and K. Hammerer, “Prospects and challenges for squeezing-enhanced optical atomic clocks,” Nature Communications 11, 5955 (2020), doi:10.1038/s41467-020-19403-7.
• D. J. Wineland, J. J. Bollinger, W. M. Itano, F. L. Moore, and D. J. Heinzen, “Spin squeezing and reduced quantum noise in spectroscopy,” Physical Review A 46, R6797–R6800 (1992), doi:10.1103/PhysRevA.46.R6797.