Atomic Clocks and Frequency Standards

Every clock is two machines bolted together: a flywheel that swings freely and a governor that checks whether it runs fast or slow. In a pendulum clock the bob is both, which is why no two quite agree — their periods depend on length, temperature, and local gravity. An atomic clock breaks the roles apart. The flywheel is an electromagnetic oscillator — quartz, microwave synthesizer, or ultrastable laser — supplying the countable ticks. The governor is a cloud of atoms whose transition frequency $\nu_0=(E_e-E_g)/h$ is fixed by quantum energy levels and constants alone. The atoms never make the ticks; they arbitrate them, nudging the oscillator back when it drifts off the line. Because that line is the same for every caesium atom in the universe, every such clock realizes the same second, independent of how it was built. This construction draws on three chapters at once: fine and hyperfine structure supplies a sharp reference transition; Doppler cooling, magneto-optical trapping, and the dipole potential hold the atoms nearly still so their motion does not smear it. What it adds is a discipline of measurement — two figures of merit, stability (reproducibility over an averaging time $\tau$ ) and accuracy (fidelity to the unperturbed atom), and one organizing number they both ride on, the line quality factor $Q=\nu_0/\Delta\nu$ . Push $Q$ up, by narrowing the line or raising $\nu_0$ from 9 GHz microwaves to $10^{14}$ -Hz optical transitions, and the clock becomes simultaneously more stable and more accurate. That lever, plus the optical frequency comb that made an optical tick countable, is the whole story of how frequency became the most precisely realized quantity in physics.

Key takeaways

An atomic clock locks a countable oscillator to a narrow atomic transition; the atoms are an ageless frequency reference immune to manufacturing tolerances, the oscillator supplies the ticks.
Since 1967 the SI second has been exactly 9,192,631,770 periods of the caesium-133 ground-state hyperfine transition, making time and frequency the most accurately realized of all physical quantities.
Two figures of merit govern a clock: stability (reproducibility, the Allan deviation $\sigma_y(\tau)$ ) and accuracy (agreement with the unperturbed atom, an uncertainty budget of systematic shifts).
Both scale inversely with the line quality factor $Q=\nu_0/\Delta\nu$ ; Ramsey's separated-field method makes the effective $\Delta\nu\approx1/(2T)$ depend only on the free-evolution time $T$ , not the pulse length.
Optical clocks in trapped ions or magic-wavelength lattices reach $\nu_0\sim10^{14}$ – $10^{15}$ Hz and uncertainties near $10^{-18}$ , about four orders beyond the best caesium fountains.
Key idea: a clock is a frequency divider; its fractional instability and inaccuracy both scale as $1/Q=\Delta\nu/\nu_0$ — why the field climbed from 9 GHz microwave to $10^{14}$ – $10^{15}$ Hz optical transitions.

The core insight

A clock is a frequency divider: lock a countable oscillator to an atomic transition, and both the fractional instability and the inaccuracy scale inversely with the line quality factor $Q=\nu_0/\Delta\nu$ — which is why the field moved from 9 GHz microwave transitions to $10^{14}$ – $10^{15}$ Hz optical ones.

Where this lives today

The living edge of the field is the optical lattice clock and its single-ion cousin. Thousands of $^{87}$ Sr or $^{171}$ Yb atoms are laser-cooled into a standing-wave lattice at the magic wavelength, so the trap leaves the clock transition untouched; a laser stabilized to a cryogenic high-finesse cavity then interrogates the ultranarrow line for seconds in the Lamb-Dicke regime, free of first-order Doppler and recoil. Single trapped ions — Al $^+$ , Yb $^+$ , Sr $^+$ — take the complementary route: one isolated atom, read out with quantum logic, reaching a few parts in $10^{18}$ . The bridge that made this usable is the optical frequency comb, whose self-referenced picket fence of evenly spaced modes divides an optical tick at $10^{14}$ Hz down to a countable radio frequency. The applications run in two directions. Downward into the Earth: transportable optical clocks do centimetre-level geodesy by comparing tick rates. Outward into fundamental physics: comparing clocks of different electronic structure over years constrains any drift of the fine-structure constant $\alpha$ , tests Lorentz invariance and the redshift, and feeds searches for ultralight dark matter. A redefinition of the SI second on an optical transition — retiring caesium after half a century — is now under discussion at the international metrology bodies.

The physical picture

Start with the division of labour. The oscillator runs on its own — a laser locked to a mirror cavity holds its frequency for a while, then drifts. Periodically its light is sent through the atoms and the absorbed fraction measured: on the line the response is symmetric, but drift high or low and the atoms report an asymmetry, an error signal fed back to pull the oscillator onto resonance. The narrower the line, the more sharply the atoms tell "too fast" from "too slow", and the more tightly the flywheel is disciplined. This single servo loop is the architecture of every atomic clock, from disciplined quartz to a lattice clock.

Figure

A clock is a servo loop — a local oscillator at frequency ν is interrogated against the invariant atomic transition ν0 = (E_e − E_g)/h, the error δ = ν − ν0 is fed back to steer the oscillator, and a comb counts the disciplined output: Cs microwave ν0 = 9 192 631 770 Hz, Sr optical ν0 = 429 228 004 229 873 Hz

The precision of that arbitration is set by how many cycles you watch coherently. A transition at $\nu_0$ observed for a time $T$ winds up a phase $2\pi\nu_0 T$ , and the fractional resolution is roughly $1/(\nu_0 T)$ , so there are two ways to sharpen the reading: raise $\nu_0$ — why the field abandoned 9 GHz microwaves for $10^{14}$ -Hz optical lines — or lengthen $T$ , why laser cooling matters, since micro-kelvin atoms in a fountain drift through the cavity for half a second. Ramsey supplied the trick that makes long $T$ practical: split the interaction into two brief phase-coherent $\pi/2$ pulses separated by a dark period $T$ . The superposition precesses freely in the dark, accumulating the atom-oscillator phase; the second pulse converts it into a population that cosine-oscillates with detuning, a fringe whose width is set by $T$ , not the pulse duration — a two-slit interferometer run in time (derived below).

Figure

Ramsey fringes P_e = ½(1 + cos δT) versus detuning ν − ν0 for a caesium fountain with dark time T = 0.5 s — the central fringe half-width is Δν ≈ 1/(2T) = 1 Hz, so the quality factor is Q = ν0/Δν ≈ 9.2×10⁹, and the fringe width depends on T only, never on the pulse length

Accuracy is a different discipline — bookkeeping rather than interferometry. Every effect that shifts the observed line from its bare value — Zeeman, blackbody Stark, collisions, second-order Doppler, the lab's gravitational potential — is measured, subtracted, and assigned a residual; the total inaccuracy is the quadrature sum. A clock can be exquisitely stable yet inaccurate if one shift is mis-evaluated.

The core equations

The chain begins with a definition, not a derivation. In 1967 the international community fixed the second by counting cycles of one atomic transition, retiring the astronomical second tied to Earth's rotation — the anchor of the entire SI, through which the metre, the volt, and much else are realized.

1\ \mathrm{s} \equiv 9{,}192{,}631{,}770\ \text{periods of the Cs-133 hyperfine transition} \qquad (1)

where: the integer $9\,192\,631\,770$ is exact by definition — the number of cycles of the caesium-133 $(F=4)\leftrightarrow(F=3)$ hyperfine transition in one SI second. Equivalently, this transition is exactly $9\,192\,631\,770$ Hz with zero uncertainty, because the second is defined to make it so. Any clock transition is set by the energy gap between two atomic states, chosen long-lived and weakly coupled so the line is narrow and the frequency well defined.

\nu_0=\frac{E_e-E_g}{h} \qquad (2)

where: $\nu_0$ — unperturbed transition frequency, the reference the oscillator locks to; $E_e,E_g$ — energies of the upper and lower clock states; $h$ — Planck constant. Because $E_e-E_g$ is fixed by atomic structure, $\nu_0$ is a property of the species, identical for every atom of the isotope everywhere. The single most important derived number is the line quality factor, the transition frequency divided by its width — how finely the oscillator can be split against the line, and the lever the whole field pulls on.

Q=\frac{\nu_0}{\Delta\nu} \qquad (3)

where: $Q$ — dimensionless quality factor of the reference resonance; $\Delta\nu$ — full linewidth or Ramsey fringe width. A caesium fountain ( $\nu_0\approx9.2\times10^9$ Hz, $\Delta\nu\sim1$ Hz) reaches $Q\sim10^{10}$ ; a strontium optical clock ( $\nu_0\approx4.3\times10^{14}$ Hz, $\Delta\nu\sim1$ Hz) reaches $Q\sim10^{14}$ – $10^{15}$ . Stability — how the clock's random frequency error averages down with time — is quantified by the Allan deviation. In the white-noise regime it falls as the square root of averaging time, its prefactor set by $Q$ and the single-shot signal-to-noise ratio.

\sigma_y(\tau)=\frac{1}{\pi Q}\,\frac{1}{(S/N)}\,\sqrt{\frac{T_c}{\tau}} \qquad (4)

where: $\sigma_y(\tau)$ — Allan deviation, the standard fractional-frequency stability at averaging time $\tau$ ; $S/N$ — signal-to-noise ratio of one interrogation (shot-noise limited, $S/N=\sqrt{N}$ for $N$ independent atoms); $T_c$ — clock cycle time; $\tau$ — total averaging time. Larger $Q$ , more atoms, and longer averaging all lower $\sigma_y$ . The Ramsey lineshape fixes $\Delta\nu$ in a real clock: after two phase-coherent $\pi/2$ pulses bracketing a dark time $T$ , the excited-state probability is a cosine of the accumulated phase $\delta T$ .

P_e(\delta)=\tfrac{1}{2}\big[1+\cos(\delta\,T)\big],\qquad \delta=2\pi(\nu-\nu_0) \qquad (5)

where: $P_e$ — probability of finding the atom in the excited clock state after the two pulses; $\delta$ — angular detuning of the drive from resonance; $T$ — free-evolution dark time. The fringes have period $2\pi/T$ in $\delta$ , hence a central width $\Delta\nu\approx1/(2T)$ — independent of pulse length, the whole point of the method. Now the systematics. First-order Doppler is linear in velocity and geometry-cancellable; second-order Doppler is the relativistic time dilation of a moving atom, always a redshift, removable only by cooling.

\frac{\Delta\nu}{\nu_0}=\frac{v}{c}\cos\theta-\frac{1}{2}\frac{v^2}{c^2} \qquad (6)

where: $v$ — atomic speed; $\theta$ — angle to the light; $c$ — speed of light. The first-order term averages to zero for a symmetric velocity distribution or counter-propagating interrogation (as in a fountain); the second-order term shrinks only when $v^2$ shrinks, which is why micro-kelvin cooling is essential at the $10^{-16}$ level. The dominant systematic for room-temperature clocks is the blackbody-radiation Stark shift: the chamber's thermal photon bath shifts the levels through their differential polarizability, with a mean-square field growing as $T^4$ .

\frac{\Delta\nu_{\mathrm{BBR}}}{\nu_0}=-\frac{1}{2}\frac{\Delta\alpha\,\langle E^2\rangle}{h\nu_0},\qquad \langle E^2\rangle\propto T^4 \qquad (7)

where: $\Delta\alpha$ — differential static polarizability of the clock states; $\langle E^2\rangle$ — mean-square thermal field, $\propto T^4$ (Stefan-Boltzmann); $T$ — environment temperature. For strontium this is a few parts in $10^{15}$ at room temperature, so it must be evaluated — or cryogenically suppressed — to reach $10^{-18}$ . Because a clock's rate depends on gravitational potential, two clocks at different heights tick at measurably different rates — not an error to remove but a signal that turns clocks into altimeters.

\frac{\Delta\nu}{\nu_0}=\frac{\Delta\Phi}{c^2}=\frac{g\,\Delta h}{c^2} \qquad (8)

where: $\Delta\Phi$ — difference in Newtonian gravitational potential; $g\approx9.81\ \mathrm{m\,s^{-2}}$ ; $\Delta h$ — height difference. Numerically $g/c^2\approx1.09\times10^{-16}\ \mathrm{m^{-1}}$ : a clock raised by 1 cm ticks faster by about $1\times10^{-18}$ , within reach of the best optical clocks. Finally, for lattice clocks the trap itself would shift the transition through the AC Stark effect — unless the trap laser is tuned to the magic wavelength, where the two clock states shift equally and the transition is untouched.

\lambda_{\mathrm{magic}}:\quad \alpha_g(\lambda)=\alpha_e(\lambda) \qquad (9)

where: $\alpha_g,\alpha_e$ — dynamic (AC) polarizabilities of the ground and excited clock states; $\lambda_{\mathrm{magic}}$ — the wavelength where they are equal, so the light shifts cancel and thousands of trapped atoms feel no net trap shift. For $^{87}$ Sr this falls near 813 nm.

Derivation ARamsey fringes and the width that beats transit-time broadening

The goal is to derive the Ramsey lineshape and its central width from the two-pulse geometry, and see why the resolution is governed by the long dark time $T$ , not the pulse duration. We model the clock atom as a two-level system on the Bloch sphere, apply two phase-coherent $\pi/2$ pulses separated by a free-evolution time $T\gg\tau_p$ , and read the accumulated phase back out as a population.

Step 1. The first $\pi/2$ pulse rotates the Bloch vector to the equator, the equal superposition $\tfrac{1}{\sqrt2}(|g\rangle+|e\rangle)$ . Why: a resonant $\pi/2$ pulse is the atomic beam-splitter, putting the atom into both clock states at once.

Step 2. During the dark time $T$ the state evolves to $\tfrac{1}{\sqrt2}(|g\rangle+e^{-i\delta T}|e\rangle)$ , the relative phase $\delta T$ set by the detuning $\delta=2\pi(\nu-\nu_0)$ . Why: in the oscillator's rotating frame an off-resonant atom precesses at $\delta$ , so $\delta T$ records how far it drifted over the dark time.

Step 3. The second $\pi/2$ pulse, phase-locked to the first, interferes the components and maps the phase onto population: $P_e=\tfrac12[1+\cos(\delta T)]$ . Why: the recombining beam-splitter — constructive interference when $\delta T$ is a multiple of $2\pi$ , turning an invisible phase into a countable population.

Step 4. Adjacent fringe maxima are separated by $\delta T=2\pi$ ; the central fringe falls to half-height near $\delta T\approx\pi$ , so $\Delta\nu\approx1/(2T)$ , independent of $\tau_p$ . Feeding $Q\approx2\nu_0 T$ into equation (4) with shot-noise $S/N=\sqrt N$ gives $\sigma_y(\tau)\approx\dfrac{1}{2\pi\nu_0 T}\dfrac{1}{\sqrt N}\sqrt{\dfrac{T_c}{\tau}}$ . Why: resolving power bought entirely with dark time, so a fountain with $T\sim0.5$ s reaches $\Delta\nu\sim1$ Hz though each pulse alone is far too short. *As a check: $T=0.5$ s gives $\Delta\nu=1$ Hz and $Q=9.2\times10^9$ , consistent with the performance of operating caesium fountains.*

Go deeper · Derivation B · 1 minFrom line quality factor to the Allan-deviation floor

Derivation BFrom line quality factor to the Allan-deviation floor

The goal is to show why the stability prefactor in equation (4) is fixed by $Q$ and the atom number, and to expose the shot-noise floor no passive clock beats without entangling its atoms. We reason from the fringe slope, feed in quantum projection noise, and convert the single-shot uncertainty into the averaging that defines the Allan deviation.

Step 1. Locking the oscillator means sitting on the steep side of the central fringe, where the slope $dP_e/d\nu\sim Q/\nu_0$ fractionally. Why: the fringe is the error signal, so a steeper fringe (larger $Q$ ) turns a given frequency wobble into a larger, more detectable population swing.

Step 2. The population is uncertain by quantum projection noise: measuring $N$ independent atoms gives fractional population noise $1/\sqrt N$ . Dividing by the fringe slope converts it into a single-shot frequency uncertainty $\sim\dfrac{1}{\pi Q}\dfrac{1}{\sqrt N}$ . Why: the quantum floor — $N$ independent atoms read the frequency only $\sqrt N$ times better than one, the statistics of coin flips.

Step 3. Repeating the cycle $T_c$ over a total time $\tau$ gives $\tau/T_c$ independent measurements, so the uncertainty averages down as $\sqrt{T_c/\tau}$ , yielding $\sigma_y(\tau)=\dfrac{1}{\pi Q}\dfrac{1}{\sqrt N}\sqrt{\dfrac{T_c}{\tau}}$ . Why: each cycle is independent white noise, so the error averages down as $\tau^{-1/2}$ , and the whole curve slides down when $Q$ or $N$ rises.

Step 4. The lesson: to reach a target $\sigma_y$ faster, raise $Q$ or $N$ ; below the shot-noise floor lies only spin-squeezing. Why: it explains why optical clocks, with $Q$ larger by $10^4$ – $10^5$ , reach a given stability in seconds where a fountain needs hours. *As a check: for $Q=9.2\times10^9$ , $N=10^6$ , $T_c=\tau=1$ s, equation (4) gives $\sigma_y(1\,\mathrm{s})\approx3.5\times10^{-14}$ , averaging to $\sim1\times10^{-16}$ after about $1.2\times10^5$ s (roughly 1.4 days), as observed in caesium fountains.*

Go deeper · Derivation C · 1 minWhy climbing to optical frequencies buys accuracy

Derivation CWhy climbing to optical frequencies buys accuracy

The goal is to make quantitative the central slogan — that raising $\nu_0$ improves not just stability but accuracy — and to be honest about the caveat it hides. We compare a microwave and an optical clock under the same absolute perturbation, form the fractional shift, take the ratio, and see where the clean scaling breaks down.

Step 1. Accuracy is quoted as a fractional shift $\delta\nu/\nu_0$ , the dimensionless error in the realized second, so a perturbation of absolute size $\delta\nu$ matters less as $\nu_0$ grows. Why: the same 1 mHz wobble is a fractional error of $10^{-13}$ on a 9 GHz line but only $2\times10^{-18}$ on a $4\times10^{14}$ Hz line — the denominator does the work.

Step 2. The optical-to-microwave frequency ratio is $\dfrac{\nu_0^{\mathrm{opt}}}{\nu_0^{\mathrm{mw}}}\approx\dfrac{4.3\times10^{14}}{9.2\times10^{9}}\approx4.7\times10^{4}$ , so a common absolute shift is fractionally $\sim5\times10^4$ times smaller optically. Why: this factor is the whole case for optical clocks — five orders of head-room from the choice of transition.

Step 3. The same ratio raises $Q$ , so the optical clock also averages down $\sim5\times10^4$ times faster (Derivation 2). This is why optical clocks now reach $\sim10^{-18}$ against $\sim10^{-16}$ for fountains. Why: stability and accuracy improve in lockstep because both ride the same $1/Q=\Delta\nu/\nu_0$ .

Step 4. The caveat: the argument assumes absolute shifts are comparable across species, which they are not — strontium's blackbody sensitivity is large, and each species has its own systematics to evaluate to target. Why: raising $\nu_0$ makes accuracy possible, not automatic — realized only when every shift can be pinned down at $10^{-18}$ . *As a check: using $^{87}$ Sr $\nu_0=429{,}228{,}004{,}229{,}873$ Hz (698 nm) and Cs $\nu_0=9{,}192{,}631{,}770$ Hz gives $\nu_0^{\mathrm{opt}}/\nu_0^{\mathrm{mw}}=4.67\times10^4$ .*

Limiting cases

Case	Result	Meaning
Master figure	$Q=\dfrac{\nu_0}{\Delta\nu}$	Both stability and fractional accuracy ride on $Q$ . Cs microwave $Q\sim10^{10}$ ; Sr optical $Q\sim10^{14}$ – $10^{15}$ at the same $\sim1$ Hz width.
White-noise averaging	$\sigma_y(\tau)\propto\dfrac{1}{Q\sqrt N}\dfrac{1}{\sqrt\tau}$	Shot-noise-limited Allan deviation falls as $\tau^{-1/2}$ ; larger $Q$ and $N$ lower the whole curve.
Quantum projection floor	$\sigma_y(\tau)\approx\dfrac{1}{2\pi\nu_0 T}\dfrac{1}{\sqrt N}\sqrt{\dfrac{T_c}{\tau}}$	The floor from counting statistics of $N$ independent atoms; only spin-squeezing beats the $1/\sqrt N$ .
Ramsey vs. Rabi width	$\Delta\nu_{\mathrm{Ramsey}}\approx\dfrac{1}{2T},\ \ \Delta\nu_{\mathrm{Rabi}}\approx\dfrac{0.8}{\tau_p}$	Set by the long dark time $T$ , not the short pulse $\tau_p$ ; hence far narrower for a given apparatus.
Cold vs. thermal atoms	second-order Doppler $\propto v^2$	300 K beam $\sim3\times10^{-13}$ ; 1 µK cold $\sim10^{-21}$ .

Worked examples

1 · A caesium fountain, from linewidth to averaging time (confidence builder)

A caesium fountain launches $N=10^{6}$ laser-cooled atoms through a microwave cavity with a Ramsey dark time $T=0.5$ s and a cycle time $T_c=1$ s. Find the fringe width, the quality factor, the single-shot stability, and how long it must average to reach $1\times10^{-16}$ .

Step 1. Fringe width from the dark time: $\Delta\nu\approx1/(2T)=1/(2\times0.5)=\mathbf{1\ Hz}$ — the Ramsey central-fringe half-width, set entirely by $T$ .

Step 2. Quality factor: $Q=\nu_0/\Delta\nu=(9.192631770\times10^{9})/1=\mathbf{9.19\times10^{9}}$ — a ten-billion-fold splitting, the highest $Q$ before optical clocks.

Step 3. Single-shot stability from equation (4) with $S/N=\sqrt N=\sqrt{10^6}=10^3$ : $\sigma_y(1\,\mathrm{s})=\dfrac{1}{\pi Q}\dfrac{1}{10^3}\sqrt{\dfrac{1}{1}}=\dfrac{1}{\pi\times9.19\times10^{9}\times10^{3}}=\mathbf{3.5\times10^{-14}}$ .

Step 4. Averaging time to $1\times10^{-16}$ : since $\sigma_y\propto\tau^{-1/2}$ , $\tau=(3.5\times10^{-14}/1\times10^{-16})^{2}\approx(346)^2\approx\mathbf{1.2\times10^{5}\ s\approx1.4\ days}$ . A day-plus of integration per point — which is why higher- $Q$ optical clocks, reaching the same stability in seconds, were worth the effort.

2 · Relativistic geodesy: a clock raised by one centimetre

By what fractional rate does an optical clock speed up when lifted 1 cm, and is that resolvable?

Step 1. Use the gravitational redshift, equation (8): $\Delta\nu/\nu_0=g\,\Delta h/c^2$ with $g=9.81\ \mathrm{m\,s^{-2}}$ and $\Delta h=0.01$ m.

Step 2. Evaluate the sensitivity: $g/c^2=9.81/(2.998\times10^8)^2=9.81/(8.988\times10^{16})=\mathbf{1.09\times10^{-16}\ m^{-1}}$ .

Step 3. Multiply by the height: $\Delta\nu/\nu_0=1.09\times10^{-16}\times0.01=\mathbf{1.09\times10^{-18}}$ . Optical clocks resolving a few parts in $10^{19}$ measure a 1 cm lift comfortably — and conversely, to define a frequency at $10^{-18}$ you must know the height above the geoid to a centimetre. Chronometric levelling: the clock is an altimeter.

3 · Why cold atoms matter: second-order Doppler

Compare the second-order (time-dilation) Doppler shift for a thermal caesium beam at 300 K and for laser-cooled caesium at 1 µK.

Step 1. The relevant speed is the rms speed $v=\sqrt{3k_BT/m}$ , with $m_{\mathrm{Cs}}=2.21\times10^{-25}$ kg and $k_B=1.38\times10^{-23}$ J/K. Thermal beam ( $T=300$ K): $v\approx\mathbf{237\ m/s}$ .

Step 2. The shift is $\tfrac12 v^2/c^2=\tfrac12(237)^2/(8.99\times10^{16})\approx\mathbf{3.1\times10^{-13}}$ — far too large for a $10^{-15}$ standard.

Step 3. Laser-cooled to $T=1$ µK gives $v\approx\mathbf{0.014\ m/s}$ , so the shift collapses to $\tfrac12(0.014)^2/(8.99\times10^{16})\approx\mathbf{1.0\times10^{-21}}$ — utterly negligible. Eight orders of magnitude in temperature cut the shift by eight orders in fraction, which is why the modern clock era waited for laser cooling.

Physics of the portrait

Atomic clocks are the child of a method more than a machine — interrogate an atom twice, briefly, and read its answer off the interference between the two visits. That method has one author.

Origins · 1 minNorman Ramsey and how this was discovered

Norman Ramsey

1915–2011 · separated oscillatory fields and the atomic clock

Norman Foster Ramsey spent a career turning atomic beams into rulers for frequency. Working in the molecular-beam tradition of his mentor I. I. Rabi, he hit a wall in the late 1940s: resolving a transition finely wanted a long, uniform oscillating field, but real fields were never uniform over a long region, and the atoms' transit smeared the line. His answer, published in 1950, was to stop fighting for uniformity and instead apply two short, phase-coherent pulses separated by a long field-free drift — the method of separated oscillatory fields. The resulting fringes were narrower than any single interaction could give, their width set by the drift time rather than a field the experimenter had to control, and tolerant of the very inhomogeneity that had defeated the older scheme. This unlocked the caesium standard: Essen and Parry's 1955 clock at the National Physical Laboratory used Ramsey interrogation, and by 1967 the SI second was redefined in caesium's terms. Ramsey went further, inventing the atomic hydrogen maser with Daniel Kleppner, and shared the 1989 Nobel Prize for the separated-oscillatory-fields method and the maser. Nearly every precision clock today, microwave or optical, still interrogates its atoms the way Ramsey taught it to.

History

1950

— Ramsey publishes the method of separated oscillatory fields

1955

— Essen and Parry operate the first caesium-beam frequency standard at the NPL

1967

— the SI second is redefined as 9,192,631,770 periods of the Cs-133 hyperfine transition

1999

— laser-cooled caesium fountains become the primary realization of the second

2000

— the self-referenced optical frequency comb makes optical clocks directly countable

Going further

Texts. Foot, Atomic Physics, Ch. 6 (precision measurement) and Ch. 9 (laser cooling) · Demtröder, Laser Spectroscopy, chapters on frequency standards and Ramsey spectroscopy · Ramsey, Molecular Beams (1956), Ch. V · BIPM, The International System of Units (SI Brochure, 9th ed., 2019) · Ludlow, Boyd, Ye, Peik and Schmidt, "Optical atomic clocks", Rev. Mod. Phys. 87, 637 (2015). The Cs-133 hyperfine transition is exactly 9,192,631,770 Hz and the $^{87}$ Sr clock transition is 429,228,004,229,873 Hz (698 nm). The Ramsey lineshape is $P_e=\tfrac12[1+\cos\delta T]$ with fringe width $\Delta\nu\approx1/(2T)$ ; the Allan-deviation scaling follows equation (4), and the Doppler, blackbody, and gravitational-redshift expressions are given in equations (6)–(8).

Prerequisites. Fine and Hyperfine Structure · Doppler Cooling and the Doppler Limit · Magneto-Optical Trap · Dipole Force and the Optical Dipole Potential. Leads to. Optical Molasses · Bose-Einstein Condensation in Dilute Gases.

Ask the tutor: want to walk through how a strontium lattice clock builds its full $10^{-18}$ uncertainty budget shift by shift, or how spin-squeezing pushes past the shot-noise floor?

What is the actual difference between a clock's stability and its accuracy?

Stability is reproducibility: how tightly the clock returns the same frequency over an averaging time $\tau$ , quantified by the Allan deviation $\sigma_y(\tau)$ , which for a good clock falls as $\tau^{-1/2}$ . Accuracy is truth: how close that number is to the unperturbed atom, after every systematic shift is measured and subtracted. The two are independent — a clock can be beautifully stable (the same reading daily to $10^{-17}$ ) yet inaccurate at $10^{-15}$ if its blackbody shift was mis-evaluated. Stability says the clock agrees with itself; only the uncertainty budget says it agrees with the atom.

Why is the second defined by caesium specifically, and not some other atom?

Caesium won for engineering reasons: its ground-state hyperfine transition near 9.2 GHz was accessible to 1950s microwave technology, the atom has a single stable isotope ( $^{133}$ Cs) with a convenient level structure, and forms an intense beam. When the second was redefined in 1967, caesium clocks were the best realized standard in existence, so the definition was fixed to that transition — 9,192,631,770 Hz — making the number exact. Nothing about caesium is uniquely fundamental. That is why a redefinition is now under discussion: optical clocks in strontium, ytterbium, or ions outperform it by orders of magnitude.

What actually limits the best optical clocks today?

At the $10^{-18}$ level the budget is dominated by a handful of systematics. The blackbody-radiation Stark shift is often the largest: room-temperature photons shift strontium by a few parts in $10^{15}$ , evaluated from the differential polarizability and chamber temperature — or suppressed cryogenically. Collisional shifts in a dense lattice, residual lattice light shifts if the magic wavelength is slightly off, second-order Zeeman shifts, and the gravitational redshift (needing the clock's height to a centimetre) all contribute. Above these sits a stability limit set by the interrogation laser: you can interrogate only as long as it stays coherent, so cavity-laser thermal noise, not the atoms, often caps the dark time.

Why does simply raising the clock a few centimetres change how fast it ticks?

Because time runs at a rate that depends on gravitational potential — a real prediction of general relativity, not an artefact. A clock higher in the field sits at higher potential and ticks faster; the fractional rate difference is $g\,\Delta h/c^2$ , about $1.1\times10^{-16}$ per metre, so 1 cm gives roughly $10^{-18}$ . Long too small to notice, it is now easily resolved by optical clocks: comparing two clocks measures their height difference (chronometric levelling), and quoting a frequency at $10^{-18}$ requires the height above the geoid. The same effect, far larger, is why GPS clocks must be corrected at all.

What is the optical frequency comb for, and why was it a breakthrough?

An optical clock ticks at $\sim10^{14}$ – $10^{15}$ Hz — far too fast for electronics to count directly. Before the comb this made optical clocks nearly useless as clocks: you could build a narrow reference but not count it or compare it to caesium. The comb, from a self-referenced femtosecond mode-locked laser, is a spectrum of hundreds of thousands of evenly spaced modes whose spacing and offset are countable radio frequencies. Lock one tooth to the optical clock and the comb becomes a precise gearbox, dividing the optical tick to a microwave output and linking clocks across the spectrum. It enabled the whole optical-clock era, and earned Hall and Hänsch a share of the 2005 Nobel Prize.

Can you beat the shot-noise limit, and how?

Yes, in principle, by entangling the atoms. The floor $\sigma_y\propto1/\sqrt N$ in equation (4) is the standard quantum limit, assuming the $N$ atoms are independent — like coin flips, they read the frequency only $\sqrt N$ times better than one. Prepare them in a spin-squeezed or entangled state and their projection noise correlates and partly cancels, approaching the Heisenberg limit $\propto1/N$ . Laboratory clocks have shown modest squeezing gains, a few decibels below shot noise. But entangled states are fragile and hard to hold over long interrogation times, so the workhorse clocks still run near the shot-noise floor and buy stability the old way — large $Q$ and large $N$ .