Experiment: The Cavendish Torsion Balance
Tests Equation (33.1). Assuming Phenomenon 33.4 and Equation (33.6).
In 1797–1798 Henry Cavendish measured, in an outbuilding at Clapham, the gravitational attraction between lead spheres on a laboratory bench [Cavendish:1798]. The declared aim was the mean density of the Earth; read retrospectively, it is the first laboratory determination of the constant \(G\) that fixes the strength of gravitation, and the archetype of every precision weak-force measurement since. Its role in this part is evidential: the inverse-square law of Central Forces and Statics rests observationally on celestial mechanics, and Cavendish supplies the laboratory half—the same law, with the same constant, between metre-scale masses. That universality is what the tests of Part V — General Relativity and Cosmology will probe to far greater precision.
Two threads run through the chapter and are treated with full honesty. The first is the theory of the torsion pendulum, which is what converts an observed angle into a force, and which does so absolutely: the instrument's force scale is fixed by a length, a mass distribution and a time, with no force standard anywhere in the chain. The second is the notorious mutual disagreement of the modern determinations. Six laboratory measurements quoting uncertainties of tens of parts per million scatter over some five hundred parts per million, and the recommended value therefore carries an uncertainty enlarged by a factor of nearly four to cover a spread nobody has explained [Tiesinga:2021] [Rothleitner:2017] [Mohr:2025]. \(G\) is, by a wide margin, the least precisely known of the fundamental constants, and this chapter says so plainly rather than quoting the recommended value as though it were settled.
Cavendish reports every length in feet and inches and every mass in pounds, and quotes his result as a pure number—the ratio of the mean density of the Earth to that of water. Editorial rule 3 of this treatise admits SI only, so every apparatus figure below has been converted, and the density ratio has been multiplied by \(10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) to become a density. The conversions are exact by definition (one inch is \(25.4\,\mathrm{mm}\), one pound is \(0.45359237\,\mathrm{kg}\)), so nothing is lost; what is gained is that the numbers may be compared directly with the modern ones without a second act of translation. The uncertainty conventions are those of Measurement, SI Units, and the Theory of Errors and [JCGM:2008]: a parenthesised figure is a standard uncertainty on the last digits quoted.
Historical context and the prediction under test
The experiment tests a law that had been believed for over a century and never once checked in a laboratory. This section states what the law asserts, shows why the astronomical evidence for it is structurally incapable of yielding the constant that appears in it, and records the one earlier attempt to weigh the Earth by gravitation alone.
Universal gravitation without a constant
Newton's law of universal gravitation, as this treatise states it, is
with \(\hat{\vect{r}}_{12}\) the unit vector from body 1 to body 2, so that the force on body 1 points towards body 2. The constant \(G\) has the SI dimension \(\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), which is already an unusual thing for a coupling constant: it is not dimensionless, and no combination of the other constants of nature produces it.
That form is not Newton's. The Principia argues throughout in proportions—the force is “as the product of the masses and reciprocally as the square of the distance”—and no symbol for a constant of proportionality appears anywhere in it [Newton:1687]. Nor could one have been assigned a value, because the eighteenth century possessed no measurement from which to extract it. That is not an accident of technique but a structural feature of the astronomical evidence, and it is worth stating exactly.
Let a system of bodies of masses \(m_{a}\) move under Equation (33.1) alone. Then every trajectory of the system depends on \(G\) and on the masses only through the \(n\) products \(Gm_{a}\). In particular, no observation of orbital motion—however precise, and of however many bodies—can separate \(G\) from a mass. For a small body in orbit about a much larger one this specialises to Kepler's third law in the form
whose right-hand side contains \(G\) nowhere except in that product. Rests on Equation (33.1).
Derives Proposition 33.2. The equation of motion of body \(b\) is, by Equation (33.1) and the second law of Newtonian Dynamics,
The inertial mass \(m_{b}\) has cancelled—itself an experimental fact and not a theorem, as Experiment: Free Fall and Projectile Motion insists—and what survives on the right is the set of products \(Gm_{a}\) and nothing else. The initial-value problem for Equation (33.3) therefore has the same solution for any \((G,m_{a})\) with the products fixed, so no trajectory distinguishes them. For the two-body case, reduction to the relative coordinate replaces \(Gm_{a}\) by \(G(M+m)\) and the closed elliptical orbit gives Equation (33.2) (Central Forces and Statics).
The consequence for practice is sharp. Astronomy measures the products \(GM\) for the Sun, the Earth and the planets far more precisely than any laboratory measures \(G\) — and cannot convert any of them into a mass, because the conversion needs exactly the factor the observations do not contain. To weigh the Earth, one must measure a gravitational force between two bodies whose masses are known by weighing, which means bodies small enough to put on a balance. That is the whole rationale of the experiment.
∎The eighteenth-century way of putting the same point was to ask for the mean density of the Earth relative to water, a pure number then called the “density of the Earth”. Newton himself had guessed it in Book III of the Principia, reasoning from the density of surface rocks and the compression expected at depth, at five to six times that of water [Newton:1687]—a guess that turned out to be remarkably good, and that framed the problem for the next century as one of weighing the Earth rather than of measuring a constant.
Two ordinary bodies on a bench attract one another. The attraction is far too weak to notice — a lead sphere of some \(160\,\mathrm{kg}\) pulls a companion of under a kilogram with a force of order \(1.5\times 10^{-7}\,\mathrm{N}\), some tens of millions of times smaller than the weight of the smaller body — but it is measurable, it is proportional to the product of the two masses and to the inverse square of the distance between their centres, and it is directed along the line joining them. That the law Newton inferred from the Moon and the planets [Newton:1687] also holds between metre-scale masses in a closed room is not a deduction from the celestial evidence; it is a separate observation, and it is the observation this chapter reports. Rests on Equation (33.1) and Proposition 33.2.
Derivation. Derives Phenomenon 33.3. The magnitude asserted follows from the law together with the measured constant. Taking \(G=6.67430\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) [Tiesinga:2021], a source sphere of \(M=157.9\,\mathrm{kg}\) and a suspended ball of \(m=0.730\,\mathrm{kg}\) whose centres stand \(r=0.2248\,\mathrm{m}\) apart attract with
against a weight \(mg=7.16\,\mathrm{N}\) for the very same ball: a ratio of \(2.1\times 10^{-8}\). No balance weighs a difference that small directly. Each sphere is treated here as a point mass at its centre, which for a homogeneous sphere is exact and not an approximation (Central Forces and Statics); the correction for the actual inhomogeneity of a cast lead sphere is one of the systematics of Section 33.2.3.
The smallness of the ratio is the design constraint of the whole experiment. It is why the measurement is made as a torque against a torsion fibre rather than as a force against gravity (Section 33.2.1); why the apparatus must be sealed against air currents, whose forces on a \(50\,\mathrm{mm}\) sphere are orders of magnitude larger (Section 33.2.3); and — ultimately — why \(G\) remains the least precisely known of the constants (Section 33.7.4).
∎The Schiehallion experiment
One serious attempt preceded Cavendish. In 1774 Nevil Maskelyne, the Astronomer Royal, took a zenith sector to the isolated Perthshire mountain Schiehallion and measured the apparent latitude of two stations, one on the north flank and one on the south, by star transits. The mountain's own attraction pulls the plumb-line sideways at each station, in opposite senses, so the astronomical difference in latitude exceeds the geodetic difference measured by survey. The excess he found was \(11.6\) arcseconds, that is \(5.62\times 10^{-5}\,\mathrm{rad}\) of total deflection shared between the two sides [Maskelyne:1775].
Converting a deflection into a density requires the mountain's volume and shape, and it was Charles Hutton who supplied them, from an elaborate survey of the massif; to carry out the integration he invented the device of joining points of equal height by a line, which is the contour of every topographic map since. Hutton's reduction gave a mean density for the Earth of about \(4.5\) times that of water, that is about \(4.5\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) [Hutton:1778].
The result is low by some \(18\,\mathrm{\%}\) against the modern figure, and the reason is instructive: a mountain is an attractor whose mass distribution and internal density are not known, and Hutton had to assume a rock density for the whole massif. The error enters the answer in full. A lead sphere, by contrast, can be weighed on a balance and its dimensions measured with callipers, so that its mass and the position of its centre are known to the precision of ordinary metrology. That, and not any improvement in sensitivity, is what makes the closed-laboratory experiment the decisive one: it replaces an attractor that must be modelled by one that can be weighed.
Michell's apparatus
The instrument was not Cavendish's. It was conceived and built by the Reverend John Michell, who died before he could use it; it passed to Francis John Hyde Wollaston and thence to Cavendish, who rebuilt it substantially before beginning work. Cavendish says so in the opening pages of his paper and credits Michell with the design throughout [Cavendish:1798]. The same instrument had been arrived at independently, and for a different purpose, by Coulomb, whose memoir on the torsion of metal wires established the linear restoring law and applied it to electrostatics [Coulomb:1784] [Coulomb:1785].
Why torsion? Because it is the only elastic degree of freedom soft enough. A body suspended from a fine fibre is held against gravity by the fibre's tension—an enormously stiff constraint—while being almost free to rotate about the fibre's axis, and the two stiffnesses can differ by ten orders of magnitude. Numerically, the balance described in Section 33.3 has a torsional stiffness of order \(6\times 10^{-5}\,\mathrm{N}\,\mathrm{m}/\mathrm{rad}\), which turns the \(1.5\times 10^{-7}\,\mathrm{N}\) of Equation (33.4) into a displacement of a few millimetres at the end of a metre-long beam: a signal that a telescope and a scale can read. The same force applied against the suspension's vertical stiffness would move the ball by a distance far below atomic dimensions.
Torsion pendulum theory
Everything the instrument reports is an angle. This section sets out the theory that converts angles into forces, and—because that theory is what the modern discrepancy turns on—states carefully which of its ingredients are theorems and which are experimental facts about a piece of wire.
The torsion pendulum
A fine wire or fibre, twisted through a small angle about its own axis, resists with a torque proportional to the angle, and the constant of proportionality is a property of the fibre alone — not of the body hung from it, and not of the amplitude [Coulomb:1784]. In consequence a body suspended from such a fibre oscillates torsionally with a period independent of how far it was displaced, and that period, together with the body's moment of inertia, suffices to determine the constant. This is what makes the torsion balance an absolute instrument: its force scale is fixed by a length, a mass distribution and a time, with no separate force standard anywhere in the chain. Rests on Equation (9.106).
Derivation. Derives Phenomenon 33.4. Let the restoring torque be \(\tau=-\kappa\theta\), with \(\kappa\) in \(\mathrm{N}\,\mathrm{m}/\mathrm{rad}\), and let the suspended body have moment of inertia \(I\) about the fibre axis. The rotational form of the second law (Rigid Bodies and Rotating Frames) gives
which is Equation (9.106) with \(\lambda^{2}=\kappa/I\) and time in place of the spatial variable. Its solutions are therefore those of Equation (9.108), \(\theta(t)=A\cos\omega t+B\sin\omega t\) with \(\omega=\sqrt{\kappa/I}\), and every one of them has the same, amplitude-independent period
From Equations (9.108) and (33.5) (reading off \(\omega=\sqrt{\kappa/I}\) from the harmonic solution and inverting \(T=2\pi/\omega\)). Here \(I\) is computed from the measured masses and lengths of the beam and its balls, and \(T\) is timed; \(\kappa\) follows without any force having been applied to the instrument at all. An equilibrium deflection \(\theta\) produced by the source masses then reports the gravitational torque as \(\kappa\theta\), and hence the force.
Two cautions belong with the result, and both return later. The linearity is an experimental fact about the fibre, not a theorem, and it holds only for small twists; and its slow departures — creep, and the frequency-dependent anelasticity that makes \(\kappa\) differ between a static deflection and an oscillation — are precisely the systematic of Remark 33.10.
∎For a cylindrical wire of radius \(a\) and length \(\ell\) made of a material of shear modulus \(\mu\), the elastic theory of Continuum Mechanics and Elasticity gives
The fourth power is the design lever: halving the radius divides the stiffness by sixteen and multiplies the deflection by the same factor, at the cost of a suspension that must still carry the weight of the beam in tension. Cavendish's wire was of copper–silver; Boys later replaced it with drawn quartz [Boys:1895], which is both finer and far more nearly perfectly elastic — the single largest instrumental advance between 1798 and the twentieth century. The symbol \(\mu\) is used here for the shear modulus to keep \(G\) free for the gravitational constant. Rests on Equation (33.6).
Real fibres dissipate, and the balance is in practice a damped oscillator: with a retarding torque \(-b\dot{\theta}\) the equation of motion becomes
which is Equation (17.88) with the rotational variables in place of the linear ones. Its free solutions decay as \(\ee^{-\Gamma t}\) with \(\Gamma=b/(2I)\), and the quality factor \(Q=I\omega_{0}/b=\omega_{0}/(2\Gamma)\) counts, up to a factor \(2\pi\), the number of radians of oscillation in one decay time (Example 17.75). For the fibres used in this experiment \(Q\) runs from a few hundred to a few thousand: the free swing persists for hours, which is convenient for timing the period and inconvenient for waiting out a deflection — a tension resolved in Section 33.4.
From deflection to attraction
Let a beam carry two small balls of mass \(m\) at distance \(d\) from the suspension axis, and let two source spheres of mass \(M\) be placed one beside each ball, on opposite sides of the beam, with centre-to-centre separation \(r\) and the line of centres perpendicular to the beam. Let \(T\) be the free torsional period and \(s=d\theta\) the linear displacement of a ball from its undisturbed position. Then, in the idealisation that the moment of inertia is that of the two balls alone and that each sphere attracts only its neighbour,
Neither the mass \(m\) of the suspended balls nor the half-length \(d\) of the beam appears: the instrument needs no knowledge of what is hanging from it. Rests on Equations (33.1) and (33.6).
Derives Proposition 33.6. Each source sphere pulls its neighbouring ball with the force Equation (33.1), directed perpendicular to the beam and therefore wholly tangential; the two forces act at opposite ends and in opposite senses, so their torques add:
At equilibrium this is balanced by the fibre, \(\tau_{g}=\kappa\theta\), and \(\kappa\) is eliminated in favour of the measured period by Equation (33.6) with \(I=2md^{2}\):
From Equations (33.6) and (33.10) (substituting \(\kappa=4\pi^{2}I/T^{2}\) with \(I=2md^{2}\), so that \(m\) cancels between torque and inertia). Both \(m\) and one power of \(d\) have cancelled. Writing \(s=d\theta\) for the displacement actually read on the scale removes the second, and solving for \(G\) gives Equation (33.9).
The cancellation of \(m\) is not a coincidence and is worth naming: the gravitational torque is proportional to \(m\) and so is the inertia that sets the period, so the suspended mass drops out of the ratio. This is the same cancellation of gravitational against inertial mass that makes free fall universal (Experiment: Free Fall and Projectile Motion), appearing here as a property of the instrument rather than of the trajectory.
∎Equation (33.9) is a two-line formula and it is not the experiment. Four corrections are needed before it becomes one, and each was computed by Cavendish himself [Cavendish:1798] and systematised by Boys [Boys:1895]:
-
The beam's own inertia. \(I=2md^{2}+I_{\text{beam}}\), and \(G\) is proportional to \(I\), so a beam of a few hundred grams changes the answer by several percent. This is the largest of the four and it is measured, not estimated: the beam is weighed and its dimensions taken.
-
Cross-attraction. Each source sphere also pulls the far ball, along a direction nearly parallel to the beam, which contributes little torque but not none, and in the opposite sense.
-
The rest of the apparatus. The case, the mounting and the suspension wire all attract the balls; only the difference between the two source-mass positions survives the reversal of Remark 33.8, but the source spheres' attraction on the beam does not cancel and must be computed.
-
Finite size. A homogeneous sphere attracts an external point exactly as a point mass at its centre (Central Forces and Statics), so sphericity costs nothing; departures from homogeneity in a cast lead sphere do.
Applied to Cavendish's own headline numbers — \(r=0.2248\,\mathrm{m}\), \(s=4.1\times 10^{-3}\,\mathrm{m}\), \(M=157.9\,\mathrm{kg}\), \(T=9.0\times 10^{2}\,\mathrm{s}\) — Equation (33.9) alone returns \(G\approx6.33\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), five percent below the modern value. Correction 1 has the right sign and the right size to account for essentially all of that shortfall. It is the difference between a five-percent estimate and Cavendish's one-percent result. Rests on Equation (33.9).
The beam has an unknown rest position, set by the fibre's own untwisted orientation, and that position drifts. Cavendish therefore never measures a deflection from a zero: he swings the source spheres from one side of the beam to the other, reversing the sign of \(\tau_{g}\), and measures the difference of the two equilibrium positions, which is \(2s\). Any static offset, and any drift slow compared with a reversal cycle, cancels identically in the difference, while the signal doubles. Every torsion measurement of \(G\) since has been differential in the same way; the modern versions modulate the source masses continuously and detect at the modulation frequency, which is the same idea with the Fourier analysis of Fourier Analysis and Integral Transforms placed after the detector. Rests on Equation (33.9).
Let the source masses be placed instead in two configurations, “near” and “far”, chosen so that the gravitational torque on the beam varies with the deflection angle, with gradients \(\pp\tau_{g}/\pp\theta=G C_{1}\) and \(G C_{2}\) respectively, where \(C_{1}\ne C_{2}\) are geometrical factors carrying the masses and distances. Let \(\omega_{1},\omega_{2}\) be the measured angular frequencies of free torsional oscillation in the two configurations. Then
in which the fibre constant \(\kappa\) has cancelled entirely. Rests on Equations (33.5) and (33.10).
Derives Proposition 33.9. Expanding the total torque about the equilibrium angle, the effective stiffness is the fibre's stiffness reduced by the gravitational gradient,
Subtracting the second instance from the first removes \(\kappa\) identically: \(I(\omega_{2}^{2}-\omega_{1}^{2})=G(C_{1}-C_{2})\), which is Equation (33.12). Nothing about the fibre survives the subtraction — provided there is one \(\kappa\) to subtract.
∎The attraction of the method is that it replaces a displacement, read against a scale to perhaps a part in \(10^{3}\), by a frequency, which can be timed to a part in \(10^{6}\) or better; it is the method of Heyl [Heyl:1930] and of Luther and Towler [Luther:1982]. Its weakness is exactly the step that looks like its strength. The cancellation of \(\kappa\) in Equation (33.13) assumes one number \(\kappa\); if the fibre's stiffness depends on the frequency at which it is interrogated, the two configurations do not share it, and the difference survives as a bias. That is Remark 33.10.
Systematics of the method
Cavendish's paper is, among other things, the first systematic error budget in experimental physics, and the discipline of it is why the result stands. He identified and pursued three nuisances [Cavendish:1798].
Magnetism. The lead spheres might carry iron, and the balls might respond to it. He tested by rotating the source spheres about their own axes between runs, which reverses any magnetic moment they carry relative to the beam while leaving the gravitational configuration untouched, and found no change.
Air currents. The room was closed, kept dark, and the observer stayed outside it: the source spheres were swung by a pulley and cord operated from outside, and the beam was read through telescopes sighted at the case through small apertures, the scales lit by lamps. Nothing that could warm or stir the air was allowed inside with the instrument.
Thermal convection. This proved to be the real one, and the manner of its discovery is the model. Cavendish found an attraction that varied between runs in a way the geometry could not explain, and supposed—while thinking it improbable—that a temperature difference between the massive lead spheres and the case was driving a slow current of air past the balls. He then did the decisive thing: rather than argue about it, he made the suspected effect large. He placed lamps beneath the source spheres to heat them deliberately, put thermometers into the spheres themselves and read them through the telescope by light reflected from a convex mirror, and watched the beam move. The effect was there, and it was large. He then established the cause of the ordinary, unheated case: overnight the massive spheres cooled more slowly than the thin case, so a residual temperature difference—and a convection current—was present in every run. Diagnosing a systematic by amplifying it deliberately and then measuring its coefficient is the technique the whole subject still uses.
The modern list is longer and the terms are smaller, but the shape is the same: seismic and acoustic background, the local gravitational background of moving people and vehicles, electrostatic patch potentials between the beam and the case, residual magnetic susceptibility, the metrology of the source masses and their internal density inhomogeneity, and the elastic behaviour of the suspension [Rothleitner:2017]. The last of these is the one that has resisted.
A real fibre is not perfectly elastic: it dissipates, and dissipation and dispersion are not independent. The torsional response \(\theta(\omega)=\hat\chi(\omega)\tau(\omega)\) is causal, so Theorem 17.78 ties the frequency dependence of the real part of the stiffness to the loss, and a fibre with a frequency-independent loss angle \(1/Q\) necessarily has a stiffness that creeps logarithmically with frequency. Kuroda's observation [Kuroda:1995] is that this converts the apparently exact cancellation of \(\kappa\) in Equation (33.12) into an upward fractional bias of the time-of-swing result,
with \(Q\) the quality factor of the torsion mode. For a tungsten fibre with \(Q\approx2\times 10^{3}\) that is about \(160\) parts per million — the same order as the spread among the modern determinations of Section 33.7.4, and an order of magnitude larger than the uncertainty several of them quote. The static-deflection method of Proposition 33.6 does not share the bias, because there \(\kappa\) is not required to be the same at two frequencies; and the angular-acceleration method of Section 33.7.2 avoids the fibre altogether by keeping it untwisted. That three methods with different exposures to one effect still disagree is why the anelastic explanation, though the leading candidate, is not a resolution. Rests on Theorem 17.78 and Equation (33.12).
Apparatus
The instrument is a horizontal beam suspended at its midpoint by a fine wire, carrying a small lead ball at each end; two large lead spheres on a swivelling frame can be brought up to the balls from either side without the operator entering the room. The beam, the balls and the scales are enclosed in a narrow mahogany case, and the case stands in a closed shed, read from outside through telescopes. Every dimension is collected in Table 33.1, converted from the inches and pounds of the original per Remark 33.1.
| part | quantity | value |
|---|---|---|
| beam | length | \(1.83\,\mathrm{m}\) |
| half-length $d$ | \(0.914\,\mathrm{m}\) | |
| suspended balls | diameter | \(5.08\times 10^{-2}\,\mathrm{m}\) |
| mass $m$ (each) | \(0.730\,\mathrm{kg}\) | |
| source spheres | diameter | \(0.305\,\mathrm{m}\) |
| mass $M$ (each) | \(157.9\,\mathrm{kg}\) | |
| centre separation $r$ | \(0.2248\,\mathrm{m}\) | |
| suspension (i) | torsional period $T$ | \(9.0\times 10^{2}\,\mathrm{s}\) |
| stiffness $\kappa$ | \(6\times 10^{-5}\,\mathrm{N}\,\mathrm{m}/\mathrm{rad}\) | |
| suspension (ii) | torsional period $T$ | \(4.5\times 10^{2}\,\mathrm{s}\) |
| stiffness $\kappa$ | \(2.4\times 10^{-4}\,\mathrm{N}\,\mathrm{m}/\mathrm{rad}\) | |
| scale | division | \(1.27\times 10^{-3}\,\mathrm{m}\) |
| reading resolution | \(2.5\times 10^{-4}\,\mathrm{m}\) |
Three features of the design carry the experiment and are worth naming separately. The case is narrow, so that the air inside it has little room to circulate. The source spheres never enter the case: they act through its wall, so that the massive, slowly cooling bodies are thermally outside the volume the balls live in — and even so the residual coupling was Cavendish's dominant systematic. And the readout is remote: telescopes with verniers, sighted from outside the closed room on ivory scales fixed to the ends of the beam and illuminated by lamps, so that the observer's own body heat never reaches the instrument.
Procedure
A run proceeds as follows. The source spheres are set on one side; the beam, disturbed by the change, begins a slow torsional oscillation about its new equilibrium. Rather than wait for it to come to rest — which at \(Q\) of order \(10^{3}\) and a period of a quarter of an hour would take days — the successive turning points are recorded through the telescope and the equilibrium is inferred from them. The spheres are then swung to the other side and the procedure repeated, and the difference of the two inferred equilibria is \(2s\) (Remark 33.8). Separately, with the spheres removed, the free oscillation is timed over many periods to fix \(T\) and hence \(\kappa\).
Let a lightly damped oscillation about an unknown equilibrium \(\theta_{\text{eq}}\) have successive turning points \(\theta_{1},\theta_{2},\theta_{3}\), let \(A\) be the displacement of the first of them from equilibrium, and let \(q=\ee^{-\Gamma T/2}\in(0,1)\) be the amplitude ratio between successive extrema. Then
The three-point estimator is therefore in error only at second order in the decay per half period, the two-point estimator at first order. Rests on Equation (33.8).
Derives Proposition 33.11. The free solution of Equation (33.8) is \(\theta(t)=\theta_{\text{eq}} +A\ee^{-\Gamma\left(t-t_{1}\right)}\cos\omega_{d}(t-t_{1})\), whose turning points are separated by half a period, so that with \(q=\ee^{-\Gamma T/2}\)
Substituting the three values into the two estimators,
and \(1-2q+q^{2}=(1-q)^{2}\), which is Equation (33.15). To size the gain, note that \(\Gamma T/2=\pi\Gamma/\omega_{0}=\pi/(2Q)\) to leading order, so \(1-q\approx\pi/(2Q)\) and the ratio of the two errors is
For \(Q\approx10^{3}\) that is a suppression by three orders of magnitude, which is why three readings taken over one and a half slow periods do the work of waiting days for the beam to come to rest.
∎This is what Cavendish does, in words rather than symbols: he takes the mean of the extremes with the middle one counted twice. It is the reason the experiment was feasible at all with an instrument whose oscillation outlasts the observer's patience by two orders of magnitude.
The published series ran from August 1797 to May 1798 and comprises seventeen experiments yielding twenty-nine determinations of the mean density [Cavendish:1798]. The first three experiments, six determinations made in August 1797, used the original suspension wire; Cavendish judged it insufficiently stiff — the beam was too easily disturbed — and replaced it with a wire of about four times the torsional stiffness, on which the remaining fourteen experiments and twenty-three determinations were made. The change is visible in the data as a halving of the period and a roughly fourfold reduction of the deflection — exactly the covariation Equation (33.6) and Equation (33.9) demand of a stiffness raised by a factor of four, and the reason two periods and two deflections appear in Table 33.1. Cavendish tabulates the two groups separately and does not combine them into a weighted mean; the grand average quoted in Section 33.5 is an unweighted one over all twenty-nine determinations.
Observations and data
| quantity | value | |
|---|---|---|
| signal | deflection $s$, wire (i) | \(4.1\times 10^{-3}\,\mathrm{m}\) |
| deflection $s$, wire (ii) | \(7.6\times 10^{-4}\,\mathrm{m}\) | |
| gravitational force per ball | \(1.5\times 10^{-7}\,\mathrm{N}\) | |
| gravitational torque on the beam | \(2.8\times 10^{-7}\,\mathrm{N}\,\mathrm{m}\) | |
| determinations (experiments) | \(29\) (\(17\)) | |
| result | mean density, as published | \(5.480(38)\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) |
| mean density, corrected | \(5.448(33)\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) | |
| relative uncertainty | \(6.1\times 10^{-3}\) | |
| implied $G$ | \(6.745(41)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) | |
| modern | mean density of the Earth | \(5.5134\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) |
| $G$ (CODATA 2022) | \(6.67430(15)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) | |
| relative uncertainty | \(2.2\times 10^{-5}\) | |
| Cavendish's $G$ is high by | \(1.1\,\mathrm{\%}\) |
Two things in Table 33.2 deserve comment before the physics is drawn out of them. The first is the arithmetic slip. The number \(5.480\) printed by Cavendish is not what his own twenty-nine determinations average to; the correct mean is \(5.448\), as Baily found in 1821 [Poynting:1894]. The book quotes both, because a reader who checks the primary source must be able to see why it disagrees with the modern secondary literature, and because the correction moves the answer away from the modern value — his error happened to be in the helpful direction, which is worth knowing when reading claims of eighteenth-century accuracy.
The second is the quoted uncertainty. The \(\pm0.033\) is the standard error of the mean of twenty-nine determinations scattering by a few percent, that is a statistical uncertainty only (Theorem 2.7). It contains no allowance for the systematic effects of Section 33.2.3, and the actual departure from the modern value, \(1.1\,\mathrm{\%}\), is close to twice the quoted error — which is the honest reading: Cavendish's statistical uncertainty was smaller than his systematic one, as it is for very nearly every measurement of \(G\) since.
The mean density of the Earth is about five and a half times that of water, roughly \(5.5\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\), whereas the rocks accessible at the surface run between about \(2.5\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) and \(3.0\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) [Cavendish:1798]. The interior must therefore consist of something substantially denser than anything the crust exposes. This was the first quantitative evidence that the planet is differentiated, and the beginning of the line of inference that ends in the seismological picture of an iron core; it was also the result the torsion-balance experiment was performed to obtain, the gravitational constant being a later restatement of the same measurement. Rests on Equation (33.1) and Proposition 33.6.
Derivation. Derives Phenomenon 33.12. Treat the Earth as a sphere of radius \(R\) and mass \(M\). A body of mass \(m\) at the surface has weight \(GMm/R^{2}=mg\), so that \(M=gR^{2}/G\) and the mean density is
From Equation (33.1) (equating the weight \(mg\) to the attraction of a sphere of mass \(M\) at radius \(R\) and dividing by the volume). Every quantity on the right is measured without any assumption about the interior: \(g\) by dropping bodies (Experiment: Free Fall and Projectile Motion), \(R\) by geodesy, and \(G\) — the only difficult one — by this experiment. With \(g=9.80665\,\mathrm{m}/\mathrm{s}^{2}\) [BIPM:2019], \(R=6.371\times 10^{6}\,\mathrm{m}\) and \(G=6.67430\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) [Tiesinga:2021],
Both are low by \(0.14\,\mathrm{\%}\) against the accepted \(5.5134\times 10^{3}\,\mathrm{kg}/\mathrm{m}^{3}\) and \(5.9722\times 10^{24}\,\mathrm{kg}\), and the discrepancy is not an error: \(g_{0}\) is a conventional surface value that already includes the centrifugal reduction of the Earth's rotation, and the planet is an oblate spheroid rather than the sphere assumed here. At the \(1\,\mathrm{\%}\) level at which this chapter works, the spherical idealisation is entirely adequate; at the level of the modern numbers it is not, which is why the modern mean density is computed from \(GM_{\oplus}\) and \(R\) rather than from \(g\).
Read in the other direction this is exactly the eighteenth-century programme of “weighing the Earth”: a measurement carried out entirely between laboratory masses delivers the mass of the planet, because \(G\) is the one factor that no astronomical observation can supply (Proposition 33.2).
∎Interpretation
Cavendish never wrote a value for \(G\), and it would have been an odd thing for him to do: the constant had no name, no symbol and no recognised status as a quantity to be measured. What he reported was a density ratio. Inverting Equation (33.17) with the modern \(g\) and \(R\) turns his corrected \(5.448\) into
which stands \(1.1\,\mathrm{\%}\) above the CODATA 2022 value [Mohr:2025]. The reframing—from a density of the Earth to a constant of nature—was made a century later, in Boys's paper, which is titled “On the Newtonian constant of gravitation” and treats the Earth's density as a by-product rather than the object [Boys:1895], and in Poynting's synthesis of the whole line of work [Poynting:1894].
The evidential point for this part of the book is the one Central Forces and Statics needs and cannot supply: that the law inferred from the heavens holds unchanged between bodies on a bench. It is worth being precise about what is and is not established. The experiment fixes the constant, for lead, at a separation of about \(0.2\,\mathrm{m}\). It does not, on its own, test the inverse square at laboratory distances to any precision — for that one varies \(r\), which Cavendish did not systematically do, and which the modern short-range experiments of The Equivalence Principle and Classical Tests pursue; nor does it test composition dependence, which had to wait for Heyl's gold, platinum and glass [Heyl:1930] and, far more sharply, for the Eötvös-type balances of the same chapter. What the combination of the bench measurement with the celestial one does test, and sharply, is that a single constant serves both.
The constant measured between lead spheres in a closed room is the same constant that governs the orbits of the Moon and the planets [Cavendish:1798]. No scaling factor, no dependence on the material of the attracting bodies, and no departure from the inverse square is needed to pass from a separation of \(0.2\,\mathrm{m}\) to one of \(3.8\times 10^{8}\,\mathrm{m}\) — nine orders of magnitude in distance, and more than twenty in mass. Universality of this kind is not implied by the celestial evidence alone, which fixes only the products \(GM\) and never \(G\) and \(M\) separately; it is the laboratory measurement that separates the two and thereby makes the statement testable at all. Rests on Proposition 33.2 and Phenomenon 33.12.
Derivation. Derives Phenomenon 33.13. Take \(G\) from the torsion balance and the Earth's mass \(M=5.964\times 10^{24}\,\mathrm{kg}\) from Equation (33.17), both obtained without any reference to the Moon. Newton's law together with the circular-orbit condition of Central Forces and Statics then predicts the lunar month with nothing further adjusted: Equation (33.2) for a circular orbit of radius \(r\) about a centre of total mass \(M+m\) reads
With \(r=3.844\times 10^{8}\,\mathrm{m}\), \(M=5.964\times 10^{24}\,\mathrm{kg}\) and the Moon's \(m=7.346\times 10^{22}\,\mathrm{kg}\) this gives \(T=2.359\times 10^{6}\,\mathrm{s}\), that is \(27.30\) days, against an observed sidereal month of \(27.3217\) days. The agreement to \(0.08\,\mathrm{\%}\) — the residual being the eccentricity of the orbit and the solar perturbation, not the constant — is the check that the laboratory \(G\) and the celestial one are a single number. Note what the check consumes: the laboratory \(G\) enters twice, once in converting \(g\) and \(R\) into \(M\) and once in Equation (33.19) itself, and the two occurrences do not cancel, since the first carries \(G^{-1}\) and the second \(G^{-1/2}\). The test is therefore genuine and not circular.
That same constant is carried unchanged into the field equations of The Einstein Field Equations, where it appears in the coupling \(8\pi G/c^{4}\); the far sharper tests of the universality of gravitational free fall are the subject of The Equivalence Principle and Classical Tests.
∎Modern determinations and their discrepancies
Classic redeterminations
Boys reworked the experiment in 1895 with two changes that between them define the modern instrument [Boys:1895]. The second was the suspension: a drawn quartz fibre in place of drawn metal, which is finer, stronger in tension, and far closer to perfectly elastic — the material answer to Remark 33.5 and, in part, to Remark 33.10. The first was to make the whole instrument small: his beam was a few centimetres long against Cavendish's \(1.83\,\mathrm{m}\). He obtained \(G\approx6.658\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), and, as the title of his paper announces, he reported it as a constant of nature.
The reason a small instrument does as well is a scaling argument, and it is the one that made Boys's design possible. Take two geometrically similar balances, the second smaller by a factor \(\lambda\), and load each suspension to the same fraction of the breaking stress of its material. The suspended weight scales as \(\lambda^{3}\), so the fibre's cross-sectional area does too and its radius as \(\lambda^{3/2}\); by Equation (33.7) the torsional stiffness then scales as \((\lambda^{3/2})^{4}/\lambda=\lambda^{5}\). The moment of inertia \(2md^{2}\) scales as \(\lambda^{3}\lambda^{2}=\lambda^{5}\) as well, so by Equation (33.6) the period is unchanged; and the gravitational torque Equation (33.10), with \(M,m\propto\lambda^{3}\) and \(d,r\propto\lambda\), scales as \(\lambda^{5}\) likewise. Hence
The angular deflection — which is what a mirror and a distant scale read, and which does not care how long the beam is — is independent of the size of the apparatus. Shrinking it therefore costs nothing in signal and buys a great deal in disturbance: less air inside the case to convect, a smaller temperature difference across it, and a smaller and more easily controlled room around it. The strength of quartz is what makes the premise of the argument, a fibre loaded to a fixed fraction of breaking stress, achievable at small scale. Rests on Equations (33.7) and (33.10).
Poynting had meanwhile attacked the same quantity with a common balance rather than a torsion fibre, weighing a mass against the attraction of a large lead sphere placed beneath one pan and then the other; his Adams Prize essay is also where the correction to Cavendish's arithmetic is recorded [Poynting:1894].
Heyl's series at the Bureau of Standards in the 1920s introduced the time-of-swing method of Proposition 33.9 in its modern form and gave \(G=6.670(5)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), a relative uncertainty of \(7.5\times 10^{-4}\) [Heyl:1930]. He also did something nobody had: he ran the measurement with suspended masses of gold, of platinum and of optical glass, an explicit search for a dependence of gravitation on the material of the attracted body. Heyl and Chrzanowski's continuation of the work supplied the value on which the first CODATA recommendation rested [Rothleitner:2017].
The reference value for a generation was Luther and Towler's time-of-swing determination with tungsten source masses, \(G=6.6726(5)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), a relative uncertainty of \(7.5\times 10^{-5}\) [Luther:1982]. It dominated the recommended value until the 1990s, and it is a time-of-swing result, which is why Remark 33.10 landed as hard as it did when Kuroda published it in 1995 [Kuroda:1995].
Modern torsion-balance measurements
Gundlach and Merkowitz answered the anelasticity problem by removing the fibre from the measurement chain [Gundlach:2000]. Their balance sits on a turntable whose angular acceleration is servoed so as to keep the fibre's twist at zero: the feedback signal is then the angular acceleration required to follow the gravitational torque, and \(\kappa\) never enters, since the fibre is never twisted. Neither Equation (33.9) nor Equation (33.12) is used — a third method, whose systematics are dominated by the metrology of the source masses and by the turntable rather than by the suspension. They reported \(G=6.674215(92)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), \(14\) parts per million, then the most precise value ever obtained.
The BIPM group built a balance that could be operated in two modes on the same apparatus and the same masses: a servo mode, in which an electrostatic torque holds the beam and is itself the measurement, and a Cavendish mode, in which the static deflection is read as in Proposition 33.6. The two agreed with each other and gave \(G=6.67554(16)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), \(24\) parts per million, after the correction published as an erratum [Quinn:2013]. This value and Gundlach and Merkowitz's are the canonical irreconcilable pair: they differ by \(199\) parts per million with a combined uncertainty of \(28\), that is by more than seven standard deviations.
The HUST group then did what the field most needed, which is not another precise number but two of them from one laboratory: time-of-swing and angular-acceleration feedback, run on the same premises with the same source masses and by the same people, each reaching about \(12\) parts per million [Li:2018]. The two results are \(6.674184(78)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) and \(6.674484(78)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\): they differ from each other by \(45\) parts per million, about \(2.7\) standard deviations of the difference. Two methods, one laboratory, and still a disagreement — which is a stronger statement about the state of the subject than any single number in it.
Beyond torsion
Rosi and collaborators measured \(G\) with no fibre, no beam and no macroscopic proof mass at all: the test bodies are laser-cooled rubidium atoms in free fall, and the observable is the phase of a matter-wave interferometer, shifted by the gravitational field of tungsten source masses moved between two positions [Rosi:2014]. The result, \(G=6.67191(99)\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) at \(150\) parts per million, is an order of magnitude less precise than the best torsion values — and it is nevertheless one of the most useful numbers in the set, because its systematics have almost nothing in common with theirs. It lies \(544\) parts per million below the BIPM value, more than three standard deviations of the difference.
Other routes have been pursued: beam balances weighing a test mass against a movable source, free-fall and pendulum gradiometry, and Fabry–Pérot resonators, all surveyed with their uncertainty budgets in [Rothleitner:2017]. The lesson the survey draws is the one this section exists to make. The field does not need another measurement at \(10\) parts per million that shares a systematic with the measurements already there; it needs measurements whose dominant systematic is different, so that agreement means something. That is why a \(150\)-part-per-million atom-interferometric value carries weight disproportionate to its error bar.
The discrepancy, honestly stated
| year | laboratory and method | $G\times10^{11}$ | $u_{r}/10^{-6}$ | source |
|---|---|---|---|---|
| 1982 | NBS, ToS | $6.6726(5)$ | \(75\) | [Luther:1982] |
| 2000 | U.\ Washington, AAF | $6.674215(92)$ | \(14\) | [Gundlach:2000] |
| 2013 | BIPM, servo + Cav | $6.67554(16)$ | \(24\) | [Quinn:2013] |
| 2014 | LENS, AI | $6.67191(99)$ | \(150\) | [Rosi:2014] |
| 2018 | HUST, ToS | $6.674184(78)$ | \(11.6\) | [Li:2018] |
| 2018 | HUST, AAF | $6.674484(78)$ | \(11.6\) | [Li:2018] |
| 2022 | CODATA adjustment | $6.67430(15)$ | \(22\) | [Mohr:2025] |
Combine the six measurements of Table 33.3 as independent determinations of one quantity. The minimum-variance combination is the inverse-variance weighted mean Equation (2.4), which gives
a relative uncertainty of \(6.8\) parts per million. That number should not be believed, and the reason is visible in the same arithmetic. The residuals about it, in units of each measurement's own uncertainty and in the order of the table, are \(-3.56\), \(-1.82\), \(+7.24\), \(-2.50\), \(-2.54\) and \(+1.31\), and their sum of squares is
on \(6-1=5\) degrees of freedom. That the count is \(n-1\) and not \(n\) is Lemma 11.83 with the projector onto the complement of the weight vector; the expected value of such a \(\chi^{2}\) is \(5\) with a standard deviation of \(\sqrt{10}\) (Proposition 11.81), so the observed \(83\) corresponds to a \(p\)-value (Definition 11.75) of order \(10^{-16}\). The data are not a set of independent measurements of one quantity with the stated errors.
The Birge ratio \(\sqrt{\chi^{2}/(n-1)}=4.1\) is the factor by which every uncertainty would have to be multiplied to make the set self-consistent. CODATA, working over sixteen inputs rather than six, arrives at an expansion factor of \(3.9\) [Mohr:2025]. The agreement between the two numbers is a coincidence of arithmetic rather than a reproduction of the adjustment — different input sets, and CODATA chooses its factor to bring all residuals below two rather than to normalise \(\chi^{2}\) — but it shows where the recommended uncertainty comes from. It is not the uncertainty of the measurements. It is the width of their disagreement. Rests on Equation (2.4), Lemma 11.83 and Proposition 11.81.
The Newtonian constant of gravitation is known far less precisely than any other fundamental constant. The CODATA recommended value is \(G=6.67430\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) with a standard uncertainty of \(1.5\times 10^{-15}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), a relative uncertainty of \(2.2\times 10^{-5}\) [Tiesinga:2021] [Mohr:2025]; the same adjustment fixes the electron mass and the Rydberg constant to parts in \(10^{10}\) and better. The reason is not that the individual experiments are imprecise. Several claim uncertainties of tens of parts per million, and yet their central values scatter over several hundred parts per million — many times their own quoted errors — so that they cannot all be correct. The recommended uncertainty above is therefore not the uncertainty that any experiment reports: it is an inflated one, enlarged by the adjustment precisely because the inputs are mutually inconsistent [Mohr:2025]. Rests on Example 33.15 and Equation (33.9).
Derivation. Derives Phenomenon 33.16. Every clause of the statement is arithmetic on Table 33.3, and it is worth doing because the conclusion is often reported as an opinion about the field rather than as a measured property of its data.
Take the scatter first. The extreme values in the table are the BIPM determination at \(6.67554\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\) and the LENS atom-interferometric one at \(6.67191\times 10^{-11}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), whose difference is \(3.63\times 10^{-14}\,\mathrm{m}^{3}/\mathrm{kg}/\mathrm{s}^{2}\), or \(544\) parts per million of either. Discard the atom interferometer, the one entry whose own uncertainty is large, and the four torsion values still span \(203\) parts per million, from the same BIPM value down to the HUST time-of-swing one. Against that, the two most precise entries each claim \(11.6\) parts per million. The spread of the set exceeds the precision claimed within it by a factor between seventeen and forty-seven, which is the first clause of the statement: the experiments are not imprecise, they disagree.
That they disagree beyond their stated errors is Example 33.15, which is quoted here and not repeated: the \(\chi^{2}\) of Equation (33.22) is \(83\) on five degrees of freedom, whose \(p\)-value is of order \(10^{-16}\). No hypothesis under which the six are independent measurements of one quantity with the stated uncertainties survives that number.
The third clause — that the recommended uncertainty is an inflated one and not the uncertainty of any measurement — follows from a property of the weighted mean, without any knowledge of how CODATA performs its adjustment. By Equation (2.4) the inverse-variance combination of independent measurements has variance \(\hat{u}^{2}=\bigl(\sum_{i}u_{i}^{-2}\bigr)^{-1}\), so \(\hat{u}\le\min_{i}u_{i}\): combining consistent data can only sharpen the answer, never blunt it. Here \(\min_{i}u_{i}\) is \(11.6\) parts per million and the six inputs give \(\hat{u}=6.8\) parts per million, as Equation (33.21) records. The CODATA recommended value carries \(22\) parts per million. That is larger than the best single input, so the recommended uncertainty is provably not the uncertainty of a combination of the data as published; it is that combination enlarged, by a factor of \(3.2\) on the six values here and by the quoted \(3.9\) on the sixteen CODATA uses [Mohr:2025]. What the recommended figure measures is the width of the disagreement.
The remaining clause, that no other fundamental constant is known so poorly, is not derived here: it is read off the same adjustment, which fixes the electron mass and the Rydberg constant to parts in \(10^{10}\) and better [Tiesinga:2021] [Mohr:2025]. Nothing in this chapter establishes it independently.
What the derivation does not deliver is the cause. Every step above uses only the published values and their published uncertainties; none of it identifies which measurement is displaced, or by what. That question is open, and Remark 33.17 states how open.
∎No accepted account of the scatter exists, and none is constructed here. Resolving it would mean establishing which systematic effect displaces which measurements — frequency-dependent anelasticity of the suspension fibre and the metrology of the source masses are the leading candidates — and a reader is entitled to know how bad the situation is. The first thing to say is that the leading candidate does not fit. If fibre anelasticity (Remark 33.10) were the whole story, the time-of-swing values would sit systematically high and the angular-acceleration values low; instead the two HUST determinations, one of each kind, differ by \(45\) parts per million in the opposite sense to the naive prediction of Equation (33.14), and the BIPM value, which is not a time-of-swing measurement at all, is the highest in the set. The second thing is that no proposal of new physics is on the table here, and none should be: a discrepancy among six laboratory measurements of a laboratory quantity is overwhelmingly likely to be an unidentified systematic in one or more of them, and the honest statement is that nobody yet knows which. This remark is not a placeholder for a derivation the treatise owes. Phenomenon 33.16 asserts that the constant is poorly known, that the inputs are inconsistent and that the recommended uncertainty is inflated, and all three are derived above; the cause of the inconsistency is not asserted anywhere in this book, and so is not owed.
Primary references
The experiment itself: [Cavendish:1798], which credits Michell for the apparatus and contains the systematic investigations of Section 33.2.3. The precursor: [Maskelyne:1775] for the Schiehallion observations and [Hutton:1778] for their reduction. The torsion law and the independent invention of the instrument: [Coulomb:1784] [Coulomb:1785].
The reframing as a measurement of a constant, and the correction to Cavendish's arithmetic: [Boys:1895] [Poynting:1894]. Modern determinations: [Heyl:1930] [Luther:1982] [Gundlach:2000] [Quinn:2013] [Rosi:2014] [Li:2018], reviewed with their uncertainty budgets in [Rothleitner:2017]. The anelasticity systematic: [Kuroda:1995]. Recommended values and the expansion factor: [Tiesinga:2021] [Mohr:2025]; uncertainty conventions: [JCGM:2008] [Taylor:1997]; the conventional standard acceleration \(g_{0}\): [BIPM:2019].