Macroevolution

Niles Eldredge and Stephen Jay Gould attacked the assumption in 1972 from inside the record. Their observation was about what the gaps mean. A palaeontologist tracing a lineage up a cliff typically finds a species appear, vary trivially for millions of years, and vanish, with its replacement arriving already different. The orthodox reading was that the intermediate strata are missing. Eldredge and Gould proposed instead that the pattern is real: new species form rapidly, in small isolated populations, exactly as Mayr's account of speciation suggested, and then do not change. The record is an excellent archive of stasis.

That reframing turns an absence into a phenomenon. If lineages routinely sit still for ten million years while their populations are demonstrably under selection and demonstrably capable of rapid change, something is holding them. And if almost all change happens during brief speciation events, then the direction of a long-term trend depends on which species form and which survive — not only on what selection does inside them.

Steven Stanley pushed that to its conclusion in 1975: species originate and go extinct at different rates according to their own properties, and that is a selection process at a second level. A lineage whose species have small geographic ranges will speciate often and die often; one with wide ranges will do neither. Elisabeth Vrba insisted on the distinction that makes the claim testable — between a genuine species-level property and an organism-level trait that merely happens to affect speciation rate. Whether species selection is a real cause or a way of keeping books is still argued, but the argument established that evolutionary explanations have levels and that which level you are on must be stated.

Jack Sepkoski did something no theory required and everything since has depended on: he read the palaeontological literature and recorded the first and last known appearance of every marine animal family, and then every genus. With David Raup he turned the compendium into a time series of extinction rates, and in 1982 reported that five intervals rise far above the background — the end-Ordovician, the late Devonian, the end-Permian, the end-Triassic and the end-Cretaceous. The end-Permian is the worst, with something like four-fifths of marine species lost.

Then they overreached, instructively. In 1984 they reported that the extinction peaks recur about every 26 million years, which prompted astronomers to propose a dim companion star to the Sun, nicknamed Nemesis, that periodically disturbed the Oort cloud. Better dating and more careful statistics dissolved the periodicity, and no companion was found. The episode is worth remembering whenever a time series of a few dozen events yields a period.

Haldane's unit for evolutionary rate is the darwin: the change in the natural logarithm of a measurement per million years. It is the natural unit because it is dimensionless and additive, so a rate can be compared between tooth height, body mass and wing length.

Take horses. An early Eocene Hyracotherium massed roughly 25 kg; a modern horse masses roughly 500 kg; the interval is about 55 million years. The rate is

$$r = \frac{\ln 500 - \ln 25}{55} = \frac{\ln 20}{55} = \frac{2.996}{55} \approx 0.054 \text{ darwins}.$$

Now take a measured rate from a living population. When guppies from a high-predation stream in Trinidad were introduced to a low-predation one, their life-history traits shifted over eleven years at rates between about 3,700 and 45,000 darwins. Compare:

$$\frac{45{,}000}{0.054} \approx 8 \times 10^{5}.$$

Short-term evolution is of order a million times faster than the long-term trend. Put the other way round: at the guppy rate, the entire 55-million-year increase in horse body mass takes

$$\frac{\ln 20}{45{,}000} \text{ Myr} = 6.7 \times 10^{-5} \text{ Myr} \approx 67 \text{ years}.$$

Horses had roughly a million times longer than they needed. Something cancelled almost all of the change.

Philip Gingerich showed in 1983 what the cancellation looks like when plotted. Collect rate measurements over intervals from one year to ten million years, plot log rate against log interval, and the points fall on a line of slope close to $-1$. A slope of $-1$ means the total change, rate times interval, is roughly constant regardless of how long you watch — which is the signature of a random walk with reversals rather than a trend. Lineages move, and then move back.

This is not an artefact of mixing traits or methods, and it is not evidence against selection; populations plainly are under selection, and plainly do respond. It is a statement about what the responses add up to. Over a few generations, directional change dominates; over a million years, the fluctuations cancel and what remains is a thin residue. The modern formulation treats stasis as an adaptive peak that itself wanders slowly, with the population tracking it. Why the peak should wander so little, for so long, in so many lineages at once, is the paradox of stasis, and it is open.

The other half of macroevolution comes from phylogenies rather than strata, and it has had two salutary shocks. The first was Felsenstein's in 1985. Comparative biology had been plotting one species against another and fitting lines, as though a hundred mammals were a hundred independent observations. They are not: most of the variation among them was inherited from shared ancestors, so the effective sample size is far smaller than the species count and the confidence in any such fit is inflated. Independent contrasts — comparing differences along separate branches — fixed it, and every comparative claim since is made on a tree.

The second shock was Stilianos Louca and Matthew Pennell's in 2020, and it was worse. A standard method had been to take a dated phylogeny of living species, fit a birth–death model to the branching times, and infer how speciation and extinction rates changed through time — often attributing the changes to climatic events. Louca and Pennell proved that infinitely many different histories of speciation and extinction generate exactly the same distribution of branching times. The data do not contain the information; only one combined quantity is identifiable. Thirty years of inferred rate shifts had been reading structure out of a flat ridge in a likelihood surface.

What survives is the part of the subject tied to things actually observed: the durations in palaeontology's databases, the dated boundaries that physics supplies, and the topologies that phylogenetics can defend. The lesson of 2020 is the same one the periodicity of 1984 taught, in a more rigorous dress: a pattern recovered by fitting a model is only as good as the question of whether a different history would have fitted equally well.