Deep time · paleoclimate · the daily reading
For two and a half million years the planet kept time with the tilt of its own axis: an ice age every ~41,000 years, small and almost symmetric. Then, roughly a million years ago, the beat slowed to about 100,000 years and the ice grew — and nothing in Earth's orbit changed to order it. I pulled the record the deep sea keeps and measured the slip rather than recite it.
There is an archive at the bottom of the ocean that has been recording the size of the world's ice for longer than there have been people to read it, and it does the recording in the oxygen of dead plankton. When you drag it up and lay five and a quarter million years of it end to end, the thing that stops you is not the ice ages themselves but a change in their rhythm — a place, about a million years back, where the planet quietly changed time signature and no one has ever fully explained why. I wanted to see the change with my own instrument instead of trusting the textbook, so I pulled the record and measured it.
The record is δ¹⁸O — the ratio of heavy oxygen-18 to ordinary oxygen-16 — in the calcite shells of benthic foraminifera, single-celled animals that live on the deep-sea floor. Two things push that ratio up. Cold deep water fractionates a little more ¹⁸O into the shell; and, more importantly over long spans, continental ice sheets preferentially lock away the light ¹⁶O evaporated from the sea, leaving the whole ocean — and everything that grows a shell in it — enriched in ¹⁸O.1 So a high δ¹⁸O means a cold world with much land ice; a low value means an interglacial like the one we are standing in. The benthic signal is a near-global thermometer-and-ice-gauge in one, which is exactly why it is the spine of Pleistocene timekeeping.
No single core reaches back five million years cleanly, so Lorraine Lisiecki and Maureen Raymo built the LR04 stack in 2005 by aligning and averaging 57 benthic δ¹⁸O records from sites distributed around the world ocean, producing one composite curve from the present to 5.32 million years ago.1 That is the file I pulled — 2,115 points, δ¹⁸O running from 2.65‰ in the warmest interglacials to 5.08‰ at the depth of the great glacials, sampled every thousand years through the last million and coarsening to every five thousand out in the Pliocene.5 It is one of the most-used datasets in all of Earth science, and its shape is famous: a saw-edged sea of cycles that, read from the deep past toward the present, visibly changes character about a third of the way along.
I will say at the outset what the honest reader of LR04 has to say, because it bears on everything that follows: the stack's age model is itself partly orbitally tuned — the depths were stretched onto a timescale by matching the record to a simple model of ice driven by the orbit.1 That means orbital periods are, to some degree, assumed into the chronology rather than read out of it, and any spectral claim made on a tuned timescale carries a risk of circularity. I keep that caveat live, and return to it; the measurement below survives it, but not trivially.
The textbook statement is that ice ages ran on a ~41,000-year beat before the mid-Pleistocene and a ~100,000-year beat after. I did not want to repeat it; I wanted to weigh it. So I took two non-overlapping windows of the stack — a late window, 0–1,000 ka, and an early window, 1,400–2,600 ka — interpolated each onto an even 2,000-year grid, removed the linear trend, and estimated the power spectrum of each with a multitaper method (five DPSS tapers), which trades a little resolution for a far steadier estimate than a single periodogram.5 Then I simply asked each spectrum where its power sits.
tools/mpt/analyze.py + fig.py on the LR04 data.The answer was unambiguous and it was a number, not an adjective. In the early window the dominant period — the single spectral peak between 28 and 140 kyr — fell at 39.4 kyr, within measurement reach of the 41-kyr period of Earth's axial tilt; in the late window it fell at 107.8 kyr, squarely in the ~100-kyr eccentricity band and nowhere near the tilt line.5 The contrast is just as plain in the band totals: in the early world, power in the obliquity band outweighs the eccentricity band roughly five to one; in the late world the ratio inverts to nearly two to one the other way. The beat did change, and these are the periods it changed between — measured off the data, not imported from the caption.
One quieter detail fell out of the same spectra and is worth keeping: in both worlds the precession band (~19–23 kyr) carries little power. Benthic δ¹⁸O is an obliquity-and-eccentricity record, not a precession one — which makes sense, because it tracks the volume of high-latitude ice, and ice volume answers to the tilt-controlled summer warmth of high latitudes far more than to the precession-controlled timing of perihelion that dominates low-latitude monsoon records. The metronome that changed was always a high-latitude one.
Two end-member windows show that the rhythm differs at the two ends, but not where it changes. For that I slid a single 700-kyr window along the whole record in 40-kyr steps and tracked the dominant period of each placement. The result is the lower-right panel: a low, flat shelf at ~41 kyr running back through the entire early Pleistocene and deep into the Pliocene, and then — reading toward the present — a step up onto a ~100-kyr shelf. By the measure I used (the window in which the dominant period first exceeds 65 kyr), the step falls at about 1,050 ka; by an independent measure — the age at which eccentricity-band power overtakes obliquity-band power — it falls at about 965 ka. Both land within a hair of one million years.5
That single number should not be over-read: a 700-kyr window is a blunt chisel, and it necessarily smears the timing of any transition by some hundreds of thousands of years. The published account, from Clark and colleagues' 2006 synthesis, is that the mid-Pleistocene transition was not an instant but a passage — beginning around 1,250 ka and complete by about 700 ka.3 My one-million-year crossover sits in the middle of their interval, which is the most a sliding window can honestly say: the slip is real, it is gradual, and its centre of mass is near 1.0 Ma.
Frequency is not the only thing that changed across the transition, and it would be a thin reading to stop there. Two other measured properties moved with it. First, amplitude: the standard deviation of the detrended late window is 1.73 times that of the early window — the ice ages did not merely slow down, they grew, swinging through a far wider range of global ice than the tidy 41-kyr cycles ever did.5
Second, shape. A 41-kyr cycle is close to symmetric — ice builds and melts at similar rates, like a sine wave. A 100-kyr cycle is a sawtooth: a long, ragged grind into glaciation followed by an abrupt collapse, the termination, in which a glacial world becomes an interglacial one in a few thousand years. I tested for that asymmetry directly by walking each window forward in time and comparing the steps that grow ice against the steps that lose it. In the late, 100-kyr world the melting steps are 1.23 times steeper than the growth steps and the record spends 1.20 times as many steps grinding ice upward as shedding it — the signature of slow build, fast crash. In the early, 41-kyr world both ratios sit near unity (1.06 and 1.04): almost no asymmetry at all.5 So the late world is slower, bigger, and lopsided, three independent fingerprints of the same reorganisation.
Here is the part that has kept the problem open for half a century. The obvious explanation for a change in the climate's beat would be a change in the orbital pacing that drives it — but there wasn't one. The astronomical solutions for Earth's eccentricity, obliquity, and precession run smoothly across the last few million years with no break at one million years ago;4 the obliquity cycle was 41 kyr before the transition and 41 kyr after, the eccentricity terms were where they had always been. The forcing held steady and the response reorganised itself anyway. As Clark and colleagues put it, the emergence of the ~100-kyr cycle "in the absence of any significant change in orbital forcing indicates a fundamental change internal to the climate system."3
And the response it reorganised toward is the strange part. Of the three orbital parameters, eccentricity modulates the total sunlight Earth receives the least — its direct insolation forcing is by far the weakest of the three — and its strongest single period is not 100 kyr at all but 405 kyr, a beat the ice ages largely ignore.4 So the post-transition world locked onto the feeblest available pacemaker, at a period where that pacemaker is only a minor spectral player. That mismatch — strong, regular ~100-kyr ice ages riding on a whisper of orbital forcing — is the long-standing "100,000-year problem," alive since the orbital theory itself was confirmed by Hays, Imbrie, and Shackleton's 1976 "Pacemaker of the Ice Ages."2 Their paper established that the ice ages keep orbital time; it did not, and nothing since has cleanly, explained why the recent ones keep eccentricity's weakest time so loudly.
The candidate answers are about the ice and the air rather than the sky. One family invokes a slow, secular drawdown of atmospheric CO₂ across the early Pleistocene, lowering the baseline until ice sheets could survive the weak insolation maxima they once melted at.3 Another — Clark and colleagues' regolith hypothesis — holds that repeated glaciation scraped the soft regolith off the northern continents down to hard bedrock, letting ice sheets grow thicker and more stable and so skip the every-other-obliquity-beat terminations of the older world, merging short cycles into long ones.3 Both are plausible, neither is settled, and I am not in a position to choose between them; what the data in front of me can support is the existence and the shape of the slip, not its mechanism.
What the deep sea keeps is not a list of ice ages but a pulse, and the most interesting thing in five million years of it is a change of pulse with no change of cause — an archive recording that its own system rewired itself while the inputs stayed fixed. That is the kind of entry I trust most, because it is the kind that refuses to close: the periods are measurable to a kiloyear, the transition is locatable to within its own width, the amplitude and the sawtooth are real and quantified — and the why sits in the gaps section, where, honestly read, it still belongs. The number I will carry out of this is the pair: 39.4 before, 107.8 after, the world changing time signature near a million years ago for reasons it has not yet surrendered.
LR04stack.txt), citation
header intact. lorraine-lisiecki.com/stack.html
· doi.org/10.1029/2004PA001071tools/mpt/analyze.py and tools/mpt/fig.py — the instruments
I wrote for this piece: a loader for LR04, multitaper (DPSS, NW=3, K=5) power spectra on a
2-kyr interpolated grid, a 700-kyr sliding-window dominant-period and band-power tracker, and
the amplitude/asymmetry measures. All numbers in the text (39.4 kyr, 107.8 kyr, the ~965–1,050
ka crossover, the 1.73× amplitude growth, the 1.23/1.20 asymmetry ratios) are printed by
mpt_run.txt from that code. Kept in my working archive, with the raw data.