The last parsec, and the stars that are not there

Nearly every massive galaxy has a black hole at its centre — sometimes revealed only by a flare that puts a ceiling on its mass —, and nearly every massive galaxy has been through at least one major merger. Those two facts together demand that pairs of massive black holes are common, and that they are either merging or sitting around. Which of the two matters a great deal, because a merging pair is the loudest gravitational wave source in the universe and a stalled pair is silent.

The pair has to lose almost all of its orbital angular momentum before it can merge, and there is a range of separations where nothing obvious can take any.

Three regimes and a gap #

The descent happens in stages, and only the middle one is problematic.

Kiloparsecs to a few parsecs: friction. A massive object moving through a sea of stars raises a wake behind it and is dragged by its own wake. The drag is efficient while the black hole is heavy compared with the stars it passes and while there are many of them, and it brings the two holes together over a few hundred million years — the same mechanism that decides whether two galaxies merge at all. Below the hard-binary separation: three-body ejection. The pair becomes a bound binary when its separation drops below

$$a_h \simeq \frac{G m_2}{4\sigma^2},$$

which for a hundred-million-solar-mass hole in a nucleus with a two-hundred-kilometre-per-second dispersion is a few parsecs. Below that, a star wandering into the binary is not merely deflected: it is slingshotted out with more energy than it arrived with, and the binary tightens by the difference. That is the only mechanism available.

Below about a thousandth of a parsec: radiation. Once the pair is tight enough, gravitational waves take over and the merger follows quickly. Quickly is relative — the radiation timescale falls as the fourth power of the separation, so a factor of ten inwards is a factor of ten thousand faster, and the last decade of the descent takes a negligible fraction of the total.

The gap is between the second and third. The three-body mechanism has to carry the binary across three decades in separation, and its rate depends on how many stars are available to be thrown.

The loss cone #

To interact with the binary, a star has to come within roughly the binary’s own separation of it. On an orbit through the nucleus, that means the star’s angular momentum about the centre has to be very small — it must be on a nearly radial orbit.

The set of such orbits is a narrow cone in velocity space, and it is called the loss cone because stars entering it are lost: they are ejected, and they do not come back.

As the binary tightens, the required angular momentum falls, and the cone narrows. Meanwhile the binary is emptying it. In a smooth, spherical nucleus the only thing that refills the cone is two-body relaxation — the slow diffusion of stellar orbits caused by the accumulation of small deflections — and that is very slow in a massive galaxy.

The relaxation time in the nucleus of a massive elliptical galaxy is far longer than the age of the universe. So the loss cone empties, stays empty, and the hardening stops. That is the final-parsec problem: the pair is stuck at about a parsec, with a merger time longer than a Hubble time.

The consequence that is observable #

A binary that has spent a long time ejecting stars leaves a mark, and the mark is a hole in the middle of the galaxy.

The most massive elliptical galaxies have cores — regions where the stellar surface brightness profile flattens instead of continuing to rise inwards, as it does in every less massive galaxy. The stellar mass missing from those cores is comparable to a few times the mass of the central black hole, which is what a binary would eject while hardening from the hard-binary separation to where radiation takes over. So the process runs. Whether it runs to completion is the open question, and the observed cores establish only that the binaries got as far as the hardening phase.

The cores carry one further piece of information that is often overlooked. Scouring does not merely remove stars; it removes the ones on low-angular-momentum orbits preferentially, because those are the ones that reach the binary. A scoured core should therefore be tangentially biased — its remaining stars should be on orbits that avoid the centre — and that bias is measurable from the shape of the line-of-sight velocity distribution rather than merely its width. Where it has been measured, the bias is present, and its radial extent matches the size of the photometric core.

That is the most direct evidence available that the mechanism described here has operated, and it is evidence about the past rather than about the present: it says that a binary was there and was throwing stars, not that it is still there.

Why the crossing is the whole answer #

The opening figure is worth reading carefully, because it makes a point about optimisation that the prose above states less sharply.

The total time to merge from the hard-binary separation is the time spent hardening plus the time spent radiating. Hardening is fast when the orbit is wide and slow when it is tight; radiation is the reverse. So the total, as a function of where the handover happens, has a minimum — and the handover happens where the two rates are equal, which is exactly that minimum. Nothing chooses it; it is where the system ends up.

That makes the total time a function of one thing: the density of stars available to be thrown, which sets the hardening rate. And the dependence is weaker than it might look. The crossing separation moves as the density to the minus one fifth, and the total time moves as the density to the minus four fifths — so starving the supply by a factor of a hundred lengthens the merger by a factor of forty, not by a factor of a hundred.

That fifth-root softening is the reason the problem is a marginal one rather than a hopeless one. A nucleus with a fully replenished loss cone merges comfortably inside a Hubble time; one that is starved by two orders of magnitude does not; and the boundary between the two falls right in the middle of the plausible range, which is why the question has stayed open for thirty years.

What refills the cone #

Every proposed resolution supplies stars, and the differences are about how.

Triaxiality. A galaxy nucleus is not spherical. In a triaxial potential, angular momentum about the centre is not conserved for an individual star, and a large population of orbits — box orbits, which pass arbitrarily close to the centre — exists that has no counterpart in a spherical system. Such orbits refill the loss cone geometrically rather than by diffusion, and they do it on an orbital timescale rather than a relaxation timescale. Merger remnants are triaxial, so this is not an exotic assumption.

Gas. A merger drives gas to the centre, and a gas disc exerts a torque on an embedded binary in the same way a protoplanetary disc does on a planet. The mechanism works, and its difficulty is that the amount of gas required is substantial and the geometry matters — a disc can also stall a binary at a separation where the torques balance.

A third hole. Galaxies merge repeatedly, so a stalled binary can be joined by a third black hole from the next merger. Three-body dynamics then either merges two of them or ejects the lightest, and either way the stall is resolved.

The three are not equally satisfying, and the reason is that they have different failure modes rather than different probabilities. Triaxiality is the strongest because it costs nothing: every merger remnant has it, no free parameter is being set, and the refilling rate follows from the shape of the potential rather than from an assumption about what else is present. Gas is the most powerful and the least general — it works decisively in a gas-rich merger and does nothing in a dry one, and the most massive ellipticals, which are where the problem is worst, are exactly the systems with the least gas. A third hole resolves the stall but does not shorten it: the binary waits for the next merger, which is a wait of billions of years, so this mechanism predicts a population that merges eventually rather than one that merges promptly.

The distinction matters for what a measurement would show. Triaxiality and gas move the merger to early times and produce a background dominated by binaries that crossed quickly; a third hole spreads the mergers out and suppresses the signal at the frequencies where a pulsar array is most sensitive. So the shape of the spectrum, not merely its amplitude, carries information about which of the three is doing the work.

What would settle it #

The question is whether massive black hole binaries merge, and the direct test is to listen for them.

A pair of hundred-million-solar-mass holes at a separation of a thousandth of a parsec radiates gravitational waves at a period of years. No interferometer can hear that; the wavelength is measured in light years. What can hear it is an array of pulsars.

The signal expected is not one binary but the superposition of all of them across cosmic time, and its amplitude depends directly on whether the population merges or stalls. A universe in which most pairs stall at a parsec produces a much weaker background than one in which they all merge — so the measured amplitude is a measurement of how the final parsec is crossed, integrated over every galaxy. It is an unusual kind of measurement: nothing about any individual system is recovered, and what is constrained is a rate averaged over a population that cannot be enumerated. That is a weakness when a specific galaxy is the question and a considerable strength when the question is whether the mechanism works at all, because a single stalled pair proves nothing and a suppressed background across the whole sky would. Two lines of evidence bear on whether the stall actually happens, and they operate at opposite ends of the population: one looks for individual pairs and finds none where it matters, and one measures the whole population at once.

The two regimes the problem sits between are worth drawing at a second mass and a second dispersion, since the gap between them is what the whole essay is about.

The pairs that have been looked for #

A sub-parsec binary in a distant galaxy cannot be resolved, so the searches look for periodicity instead, and it is worth setting out what they have and have not found.

Periodic quasars. A binary embedded in an accretion flow modulates it: the two holes clear a cavity, streams cross it at the orbital period, and the accretion rate — and therefore the brightness — varies on that period. Long photometric surveys have produced a hundred or so candidates showing sinusoidal variability over several cycles.

The difficulty is that a quasar’s brightness varies anyway, stochastically, with a red spectrum — more power at longer timescales. A red-noise process observed for a few cycles produces apparent periodicity readily, and distinguishing a genuine periodicity from a chance one requires many more cycles than most candidates have. Several early candidates faded as the baselines lengthened.

Double-peaked emission lines. If both holes have their own broad-line regions, the line profile should show two components moving in antiphase at the orbital period. Systems with displaced or double-peaked broad lines are known, and the interpretation is contested: a single hole with a disc-like broad-line region produces double peaks too, and telling the two apart requires watching the peaks move over years.

A resolved pair. Radio interferometry can resolve milliarcsecond separations, and in one nearby galaxy two compact radio cores separated by a few parsecs have been imaged and are widely accepted as a genuine binary — at a separation well outside the problematic range, which is what makes it detectable and what makes it uninformative about the stall.

So the state of the direct searches is that pairs are found at separations where the physics is not in doubt, and no pair has been confirmed at the separation where it is.

A null result at sub-parsec separations is what both a stalling population and a rapidly merging one would produce, for opposite reasons, which is why the population-level measurement matters more than any individual system.

The individual searches are therefore inconclusive by construction, and the population measurement is not.

What the array has heard #

The measurement that constrains the population has now been made, and its result is the reason the subject has become active again.

Pulsar timing arrays have been accumulating data for two decades, and what they look for is the quadrupolar correlation between pairs of pulsars described above. In 2023 several independent collaborations reported evidence for exactly that correlation, at a significance of three to four sigma, in datasets spanning fifteen years and dozens of pulsars.

The amplitude is the interesting part. It sits at or slightly above the upper end of what standard models of the merging population predicted — which, taken at face value, says that massive black hole binaries do reach the gravitational-wave regime, in numbers, and that the final parsec is crossed rather than stalled.

Three caveats attach to reading it that way.

The amplitude depends on the population’s masses as well as on its merger rate, and the two are degenerate: a slightly heavier population merging less often produces the same background as a lighter one merging more.

The spectrum carries information the amplitude does not. A population that crosses the last parsec by ejecting stars loses energy to the stars as well as to radiation, which suppresses the background at the lowest frequencies relative to the pure radiation-driven case. Measuring that suppression would say directly which mechanism operates, and the current data do not resolve the spectral shape well enough.

And a background of this amplitude has other possible sources — cosmic strings, phase transitions in the early universe — whose spectra differ from the binaries’ and which the same measurement will eventually separate.

A thirty-year-old theoretical problem has acquired an observational constraint that appears to answer it, and turning that appearance into an answer requires another decade of the same measurement.

The measurement is also unusual in what it does not deliver. It says something about the whole population of massive black hole binaries across cosmic time and nothing whatever about any one of them, so a galaxy whose nucleus is stalled and a galaxy whose nucleus merged a billion years ago contribute to the same number and cannot be separated within it. That is the price of a background: it is a statement about an ensemble that no individual member could have supplied.

And the scattering experiment that Heggie’s law is a summary of, since the law is a statistical statement about individual events.

Where the ladder goes #

The next rung is what happens if the merger does complete. The remnant recoils: gravitational waves carry linear momentum away asymmetrically unless the two holes are identical and aligned, and the kick can reach a thousand kilometres a second — enough to eject a black hole from its galaxy entirely. Searching for displaced or offset active nuclei is the observational programme, and the candidates are few and contested.

The other direction runs to the smaller version of the same problem. Two stellar-mass black holes in a globular cluster face an identical accounting — friction, then ejection, then radiation — with the difference that the relaxation time in a cluster is short, so the loss cone refills and the binaries do merge. The eccentricity such a pair carries into a detector’s band is the fingerprint of having been assembled that way rather than having formed from two stars.