Astrophysics

The second term in the bending of light

Starlight grazing the Sun is bent by 1.75 arcseconds, and the formula that gives it, four times GM over c² times the impact parameter, is the first term of a series. The second term is eleven millionths of an arcsecond at the Sun's edge. It is too small to have been seen, and it is not small enough to ignore: it is half the size of the change any surviving rival to general relativity could make, so a measurement good enough to test the theory further must compute it first. It also has the odd property that its size depends on which distance it is written in.

Assumes: The bend Newton got half right · The bend a moving planet gives

In 1919 the deflection of starlight by the Sun was measured as somewhere between 1.6 and 2 arcseconds, and that was enough to decide between Newton and Einstein. A century later it is measured to a part in ten thousand, by radio interferometers watching quasars slide past the Sun and by spacecraft timing signals that skim it. The formula every one of those measurements was compared with is the one the bend Newton got half right derived:

α≈4GMc2b,\alpha \approx \frac{4GM}{c^2 b},

with bb the impact parameter, the distance by which the ray would miss the centre of the mass if it went straight. For the Sun’s limb that is 1.75 arcseconds.

The “approximately” is easy to read past. The formula is exact in one limit only, when GM/c2bGM/c^2b is vanishingly small, and for any real ray it is the first term of a series. The series has a second term, a third, and so on, each smaller than the last by roughly another factor of GM/c2bGM/c^2b. For a ray past the Sun that factor is two millionths, so the second term is a few hundred thousand times smaller than the first. That sounds like a reason never to think about it. This essay is about why it is not, and about a property of the second term that the first does not have: its size depends on which distance it is written in.

The bend, computed without approximating it

A ray of light passing a spherical mass follows a null geodesic of Schwarzschild’s geometry, and its orbit equation can be integrated exactly. Measure lengths in units of m=GM/c2m = GM/c^2, which is 1,477 metres for the Sun and 4.4 millimetres for the Earth, and write u=1/ru = 1/r. A ray coming in from infinity with impact parameter bb reaches a closest approach r0r_0, related to bb by b2=r03/(r0−2m)b^2 = r_0^3/(r_0 - 2m), and the total angle it turns through is

α(b)=2∫0u0du1/b2−u2+2mu3−π.\alpha(b) = 2\int_0^{u_0} \frac{du}{\sqrt{1/b^2 - u^2 + 2m u^3}} - \pi .

Without the 2mu32mu^3 term the integral is exactly π/2\pi/2 each way and the ray goes straight. With it the ray turns, by an amount that is an ordinary definite integral with a square-root singularity at the turning point, which a change of variable removes. Nothing about it needs a weak field. It holds for a ray that barely grazes the Sun and for one that loops round a black hole.

Light bent by a mass, exactly and to two orders. The deflection of a ray passing a mass, against its impact parameter in units of GM/c², integrated exactly from the Schwarzschild orbit equation and compared with the first-order 4GM/c²b and with the series carried to second order. Far out the three agree. At b = 10 the exact bend is 0.590 radians and the first-order term misses it by 32 per cent; adding the second-order term brings it within 12 per cent. Near b = 3√3 GM/c², the photon sphere, the exact bend grows without limit while every term of the series stays finite.
Fig. 1 The deflection of a ray against its impact parameter in units of GM/c2GM/c^2, integrated exactly (solid) and approximated by the first-order 4GM/c2b4GM/c^2b and by the series taken to second order. Far from the mass the three are one line. At b=10 GM/c2b = 10\,GM/c^2 the exact bend is 0.59 radians, a third larger than the first-order value, and the second-order term closes most of the gap. At b=33 GM/c2b = 3\sqrt{3}\,GM/c^2, the photon sphere, the exact bend grows without limit.

On a logarithmic scale the first-order formula is a straight line of slope minus one, and the exact deflection lies on it across most of the picture. Far out, where every star and quasar that has ever been watched past the Sun lies, the first term is the whole story to the precision of the drawing. Nearer in the exact curve rises above the line, gently at a hundred mass units and steeply at ten, and at 33 m3\sqrt{3}\,m, about 5.2 mass units, it turns vertical. That is the photon sphere, where the circle light cannot leave found a ray that circles the mass indefinitely: a ray aimed exactly at it winds round forever, and one aimed a hair outside winds round several times before escaping. The first-order formula knows nothing of this. At the photon sphere it predicts a bend of about 0.77 radians, a gentle forty-four degrees, for a ray whose real deflection is unbounded.

The dashed line between them is the series carried one step further, and it is already much better. At b=10 mb = 10\,m the first term misses by 32 per cent and the first two by 12. Each further term pushes the agreement further in. None of them, as the last figure below shows, ever reaches the photon sphere.

What remains when the first term is taken away

The second term can be read off the integral rather than taken on trust. Subtract the first-order bend from the exact one and divide by the square of m/bm/b. If the deflection were exactly 4m/b4m/b, the result would be zero; if there is a second term proportional to (m/b)2(m/b)^2, the result will settle on its coefficient as bb grows.

What is left when the first-order bend is taken away. The exact deflection minus 4GM/c²b, divided by the square of GM/c²b, against the impact parameter. If the first-order formula were the whole answer this would be zero. It settles instead on 15π/4 = 11.78, the second-order coefficient, approached from above along a line whose slope is the third-order coefficient, 128/3. At b = 1000 GM/c² the residual is 11.82. The coefficients are read off the integration, not assumed.
Fig. 2 The exact deflection minus 4GM/c2b4GM/c^2b, divided by (GM/c2b)2(GM/c^2b)^2, against the impact parameter. It does not go to zero; it settles on 15π/4=11.7815\pi/4 = 11.78, approaching from above along the line 15π/4+(128/3) GM/c2b15\pi/4 + (128/3)\,GM/c^2b whose slope is the third coefficient. At b=1000 GM/c2b = 1000\,GM/c^2 the residual reads 11.82.

It settles on 15π/415\pi/4, which is 11.78. The approach is from above, because the next term is positive too, and the dashed line through the curve is the second coefficient plus the third times m/bm/b, with the third coefficient equal to 128/3128/3. Written out, with ε=GM/c2b\varepsilon = GM/c^2b, the series is

α=4ε+15π4ε2+1283ε3+3465π64ε4+⋯\alpha = 4\varepsilon + \frac{15\pi}{4}\varepsilon^2 + \frac{128}{3}\varepsilon^3 + \frac{3465\pi}{64}\varepsilon^4 + \cdots

The odd-order coefficients are rational and the even-order ones carry a π\pi, and each is larger than the one before. Light is the limiting case of an orbit with infinite speed at infinity, and the same integral, with a finite energy in place of the infinite one, gives the advance of a perihelion that the orbit that does not come back to itself computed for massive bodies. The deflection is that calculation pushed to its fastest possible traveller.

The residual curve also shows how slowly the second term emerges in practice. At a thousand mass units it is still four tenths of a per cent away from its limit. For the Sun, at nearly half a million mass units, the limit is reached to six decimal places, and the integral itself can no longer be evaluated directly in sixteen-digit arithmetic: the deflection is a few millionths of a radian sitting on top of π\pi, and the second term is a millionth of that. Near the Sun only the series is usable, which is part of why its coefficients matter.

One bend, three second terms

Here is the property that sets the second term apart. The first-order formula, 4GM/c24GM/c^2 over a distance, gives the same answer whatever the distance is called, because at first order every reasonable definition of “how close the ray came” agrees. The ray’s closest approach r0r_0 and its impact parameter bb differ only by about one mass unit, a kilometre and a half for the Sun, against a solar radius of seven hundred thousand kilometres. At second order that kilometre and a half matters.

One bend, three second-order coefficients. The same exact deflection expanded in three different distances, each with its first-order term 4GM/c² over that distance removed and the remainder divided by the square. In the impact parameter b the coefficient settles on 15π/4 = 11.78; in the closest-approach radius of Schwarzschild's coordinates on 15π/4 − 4 = 7.78; in the closest-approach radius of isotropic coordinates on 15π/4 − 8 = 3.78. The first-order term is the same in all three. The second-order term depends on what the distance is, which is why it is quoted against b, the one distance that is defined far from the mass.
Fig. 3 The exact deflection expanded in three distances, with the first-order term 4GM/c24GM/c^2 over that distance removed and the remainder divided by its square. In the impact parameter the coefficient tends to 15π/4=11.7815\pi/4 = 11.78; in Schwarzschild’s closest-approach radius r0r_0 to 15π/4−4=7.7815\pi/4 - 4 = 7.78; in the closest-approach radius of isotropic coordinates to 15π/4−8=3.7815\pi/4 - 8 = 3.78.

The three curves describe one deflection of one ray. Written in the impact parameter, the second coefficient is 15π/415\pi/4. Written in the radius of closest approach as Schwarzschild’s coordinates measure it, it is 15π/4−415\pi/4 - 4. Written in the closest approach in isotropic coordinates, the form of the same geometry in which space looks the same in every direction at each point, it is 15π/4−815\pi/4 - 8. The arithmetic of the difference is short: bb exceeds r0r_0 by about mm, so 4m/b4m/b is 4m/r04m/r_0 less 4m2/r024m^2/r_0^2, and that −4-4 is moved from one term into the next. The isotropic radius is smaller again by about mm, which moves another 4.

So the question “how big is the second-order bend?” has no answer until a distance is chosen, and the answer is not a matter of taste. The radius of closest approach is a coordinate. It labels a point deep in the curved region, and two coordinate systems that both describe the Sun correctly assign that point different numbers. The impact parameter is not a coordinate in the same sense. It is the ratio of the photon’s angular momentum to its momentum, both conserved, both measured far away where space is flat, and it is the quantity an astronomer infers from where a star is on the sky relative to the Sun’s centre. A prediction meant for comparison with a measurement has to be written in something a measurement can reach. At first order the distinction was invisible, and that is exactly why it is easy to forget: the same carelessness that costs nothing at one order costs the whole answer at the next.

There is a general lesson here about series in a strong field. Two clocks that disagree about the fall found two clocks giving different answers to “how long does it take?”, each correct, because each answered a different question. Coefficients beyond the first in any relativistic expansion carry the same caution. They mean something only alongside a statement of what is being expanded in, and a table of them that omits that statement is a table of numbers about nothing.

Eleven microarcseconds, and why they are not negligible

For the Sun the second term is

α2=15π4(GM⊙c2b)2,\alpha_2 = \frac{15\pi}{4}\left(\frac{GM_\odot}{c^2 b}\right)^2 ,

which at the limb, with GM⊙/c2b=2.12×10−6GM_\odot/c^2b = 2.12\times10^{-6}, is 5.3×10−115.3\times10^{-11} radians, or 10.9 microarcseconds. A microarcsecond is the angle a coin on the Moon subtends from the Earth. The best measurements of the Sun’s deflection, by radio interferometry of quasars, reach tens to hundreds of microarcseconds near the Sun; the Gaia satellite measures stellar positions to tens of microarcseconds at best, and does not look within forty-five degrees of the Sun at all. So the second term has never been seen, and it is not about to be.

But the reason anyone measures light bending is not to confirm a number already known. It is to test whether gravity is the theory that predicts that number, and here the second term earns its place. Alternatives to general relativity are catalogued by a small set of parameters that describe, among other things, how much curvature of space a unit of mass produces. The most important is called γ\gamma, and it enters the first-order bend as

α1=1+γ2⋅4GMc2b.\alpha_1 = \frac{1 + \gamma}{2}\cdot\frac{4GM}{c^2b}.

General relativity has γ=1\gamma = 1. The half of the bend a slow body never feels traced the factor of two in the bend to exactly this term: half the deflection is the time-warping a Newtonian picture also contains, and the other half is the curvature of space that γ\gamma measures. Radio tracking of the Cassini spacecraft in 2002, using the delay of signals that passed near the Sun, found γ−1=(2.1±2.3)×10−5\gamma - 1 = (2.1 \pm 2.3)\times10^{-5}. Missions proposed to do better aim at 10−610^{-6} or even 10−810^{-8}.

The Sun's second-order bend against a change in γ. The second-order term in the bending of starlight by the Sun, in microarcseconds, against the ray's closest approach in solar radii: 10.9 µas at the limb, falling as the inverse square. Beside it, the change in the first-order bend if the parameter γ, which measures how much space curvature a unit of mass makes, differed from general relativity's value by 2.3 × 10⁻⁵ (about the present bound) or by 10⁻⁶. At the limb the second-order term is 0.54 of the first and 13 times the second, so a measurement of γ at either level has to model it.
Fig. 4 The Sun’s second-order bend in microarcseconds against the ray’s closest approach in solar radii, 10.9 µas at the limb and falling as the inverse square, beside the change in the first-order bend that a departure of γ\gamma from 1 by 2.3×10−52.3\times10^{-5} or by 10−610^{-6} would make, falling as the inverse first power.

The figure sets the second term beside the signal such a measurement looks for. A change in γ\gamma of 2.3×10−52.3\times10^{-5}, the size of Cassini’s uncertainty, changes the limb deflection by 20 microarcseconds. The second-order term is 11 of them, more than half. A change in γ\gamma of 10−610^{-6} shifts the limb deflection by under a microarcsecond, and the second-order term is thirteen times that. Any experiment that sets out to measure γ\gamma to a part in a million by bending light near the Sun, and models the bend as 4GM/c2b4GM/c^2b, will find a departure from general relativity of more than ten parts in a million that is nothing but the second term it left out.

The two curves fall differently with distance, and that is a second reason the term cannot be ignored but can be handled. The second-order term falls as 1/b21/b^2 and the γ\gamma signal as 1/b1/b, so a measurement at several distances from the Sun can tell them apart: the ratio of the second-order term to the first-order signal shrinks in proportion to bb. At five solar radii the second term has fallen to 0.4 microarcseconds. An experiment can choose to work far enough out that it no longer matters, at the cost of a smaller signal, or close in, at the cost of having to compute it. What it cannot do is pretend it is zero.

What the second term tests

In general relativity the second coefficient is fixed by the same geometry as the first, and so it tests nothing new on its own. In the wider family of theories it does. The parameterised post-Newtonian framework that defines γ\gamma also defines β\beta, which measures how nonlinear gravity is, how much the gravity of gravitational energy itself contributes. Extending the framework to second order in the bend adds a parameter often written ε\varepsilon, describing the next term in the curvature of space. The second-order deflection, written in the impact parameter, is then

α2=π[2(1+γ)−β+34ε](GMc2b)2,\alpha_2 = \pi\left[2(1+\gamma) - \beta + \tfrac34\varepsilon\right]\left(\frac{GM}{c^2b}\right)^2,

which with γ=β=ε=1\gamma = \beta = \varepsilon = 1 is 15π/415\pi/4. A measurement of the second term would therefore be the first light-bending test of β\beta, which is otherwise known from the advance of Mercury’s perihelion and from lunar laser ranging, and the only light-bending test of ε\varepsilon, which nothing else in the solar system constrains well. That is the honest scientific case for wanting it. It is also a case nobody yet knows how to meet: eleven microarcseconds at the solar limb, measured to a few per cent, is a precision of a fraction of a microarcsecond in the hardest place in the sky to observe.

The coefficient in this formula is the impact-parameter one. Written in the closest-approach radius, the bracket changes, because the conversion between the two distances itself involves γ\gamma; a parameter that is a coordinate artefact at first sight turns out to be absorbed into the definition of distance at second order. This is a point on which published expansions have disagreed with each other, not because anyone made an error in the geometry but because they expanded in different distances without saying so prominently. The cleanest way to avoid it is the one the previous section argued for: quote the bend against bb, which an observer far from the mass can define without reference to any coordinates near it.

Where the series runs out

Every term of the series is finite at every impact parameter. The exact bend is not. So the series must fail somewhere, and the place it fails says something about the kind of object a deflection is.

Where the series stops being the answer. The exact deflection against GM/c²b, from far away to the photon sphere at 1/(3√3) = 0.192, beside the series cut after one, two and four terms and the logarithm that describes the bend right at the photon sphere. Each added term carries the series further in, and every one of them stays finite at the photon sphere, where the exact bend diverges. Weak-field and strong-field descriptions are two ends of one integral, joined by nothing but the integral itself.
Fig. 5 The exact deflection against GM/c2bGM/c^2b, out to the photon sphere at 1/33=0.1921/3\sqrt{3} = 0.192, beside the series cut after one, two and four terms and the logarithm that describes the bend just outside the photon sphere. Each term carries the series further in; none follows the divergence.

Plotted against ε=m/b\varepsilon = m/b rather than bb, the exact curve starts along the first-order straight line and bends upward, gradually and then violently, as ε\varepsilon approaches 1/331/3\sqrt{3}. The series cut at two terms follows it to about ε=0.05\varepsilon = 0.05, a ray passing at twenty mass units. Four terms follow it to about 0.1. Beyond that the series lags, and no finite number of terms can catch a curve that goes to infinity at a finite point. The coefficients grow with order, roughly fourfold from each to the next among the four written above, so every added term buys less, and near the photon sphere a truncated series is a poor guide however many terms it carries.

Right at the photon sphere a different approximation takes over. A ray whose impact parameter exceeds the critical value bc=33 mb_c = 3\sqrt{3}\,m by a small fraction circles the mass a number of times that grows as the logarithm of how close it came, and its deflection is

α≈−ln⁡ ⁣(bbc−1)+ln⁡ ⁣[216 (7−43)]−π,\alpha \approx -\ln\!\left(\frac{b}{b_c} - 1\right) + \ln\!\left[216\,(7 - 4\sqrt{3})\right] - \pi ,

the curve that rises off the top of the figure. Weak-field and strong-field approximations are two ends of one exact integral. Between them, around ten to twenty mass units, neither is good, and only the integral serves. That is the region where light leaving the surface of a neutron star travels, which is why the star whose far side can be seen had to integrate the bend rather than expand it; a neutron star’s surface sits about six mass units from its centre, where a grazing ray’s bend is half as large again as 4GM/c2b4GM/c^2b says.

The surprising connection is that the same coefficient, 15π/415\pi/4, which is invisible in the Sun’s light, is the first step of the description that ends with light wound round a black hole. Eleven microarcseconds at the solar limb and a ray circling a black hole a dozen times are not two phenomena with two theories. They are the two ends of one integral, the same 2mu32mu^3 in the same orbit equation, read at ε\varepsilon of two millionths and at ε\varepsilon of nineteen hundredths.

The limits of these pictures

Everything here assumes a single, spherical, non-rotating, isolated mass, which for the Sun is a good approximation and not an exact one. The Sun’s oblateness, measured by its quadrupole moment J2≈2×10−7J_2 \approx 2\times10^{-7}, adds a term to the limb deflection of order J2J_2 times the first-order bend, a fraction of a microarcsecond; its rotation adds a term of under a microarcsecond from the dragging of frames. The first falls faster with distance than the second-order term and the second as fast, so both matter mainly for rays grazing the limb, and both would have to be modelled in an experiment aiming at the second term. The planets add their own deflections, which the bend a moving planet gives found are measurable for Jupiter to degrees from its disc.

The deflection is computed in vacuum. Real rays near the Sun pass through the corona, whose plasma bends radio waves by an amount proportional to the square of the wavelength. Close to the Sun that refraction is far larger than eleven microarcseconds at any radio frequency, and the only remedy is to observe at several frequencies and fit away the part that depends on colour. Gravity bends every wavelength equally, which is what makes the separation possible. For optical light the plasma bending is negligible, but optical astrometry near the Sun has to contend with scattered sunlight, which is why no optical instrument has measured the bend at the limb except during eclipses.

The figures also show only the angle. A real measurement compares positions of a source seen near the Sun and far from it, so what is measured is a difference of deflections across the field of view, and the time delay of the delay that is not a bend has its own second-order term, which an experiment timing signals rather than measuring angles would need instead. Finally, the series is a statement about general relativity. What a second-order measurement would say about other theories depends on the parameterised framework above, which covers a broad class of metric theories but not every conceivable departure from Einstein; a theory outside the framework could hide at second order in ways the bracket does not describe.

Still open: whether anyone will see it

Two families of mission have been proposed that could reach the second term. One would put laser transmitters on spacecraft behind the Sun and measure the angles between their beams from a third craft with an interferometer, reaching a fraction of a microarcsecond; another would use a dedicated astrometric telescope with a coronagraph to measure star positions near the limb repeatedly. Neither has been built, and both face the same obstacles the corona and scattered light impose on every measurement close to the Sun. Whether the second-order term will be the first thing such a mission finds, or the first thing it has to subtract, depends on whether γ\gamma differs from one by more than a part in a million, which is unknown.

There is also a quieter open question, about what the lens equations used to model gravitational lenses should contain. Strong lenses in astronomy, the kind the lens with no focal length and the fifth image a galaxy hides described, are galaxies and clusters, for which GM/c2bGM/c^2b is of order 10−610^{-6} and the second term is a million times below the uncertainties in the mass models; the first term is used throughout. For lensing by the black holes at the centres of galaxies, imaged on scales of a few mass units, the full integral is necessary and nothing between is accurate. Which approximation is adequate for which purpose is a judgement made case by case.

The habit worth carrying away is to read a formula as the first term of something. The bending of light by a mass is 4GM/c2b4GM/c^2b plus (15π/4)(GM/c2b)2(15\pi/4)(GM/c^2b)^2 plus more, the second term is 10.9 microarcseconds at the Sun’s limb — more than half the size of the signal a present-day test of general relativity looks for — and its coefficient is 11.78 only when written in the impact parameter, becoming 7.78 or 3.78 in two radii that coordinates assign. A first-order formula hides which distance it was written in, because at first order the choice does not matter. At second order the choice is the answer.

Part 9 of 9

This essay is one argument about Light deflection. The others:

The objects named here

The third axis, after the field and the reading path: the things themselves, and every essay that touches each one.

AsymptoticsImpact parameterLight deflectionPerturbationPhoton spherePost-newtonian parameters