Which year did the ice change?

The number that changes

Mertz Glacier has been on this site as one of 5 emperor colonies whose April–November sea ice is robustly rising — +2.03 ± 0.73 points per decade, 5.5 times its own standard error, and still clear once the pre-1988 sensor is dropped. It is not rising. It stepped up once, and it has been drifting downwards ever since.

The scan was told nothing about icebergs. Out of 21 candidate years it picked 2011 — the first full ice year opening after the berg B09B grounded in Commonwealth Bay in December 2010, which is the same event this site found in Adélie penguin territory from a completely different archive. Fitted 1988 onward the slope is +1.31 ± 0.95; allow the one shift and it is -2.04 ± 1.13, sign reversed, with a step of +8.78 ± 2.49 points. The step is seven times the size of the slope it replaces.

So the count on the emperor page is 4 rising, not 5, and this page is the reason. The 5 falling colonies are unaffected — the scan reaches two of them as well, and both are dealt with below rather than quietly.

Where the breaks are, and which years they picked

Every place whose record carries a level shift anywhere, drawn by the year its strongest one falls in. 16 of the 156 sites that fire pick 1987 or 1988, the satellite change; 14 pick 2010 or 2011 and they are one stretch of the George V Land coast; 2 pick 1994 and are the same place measured twice. The rest are drawn hollow, because a year that fires in a hundred series and two places is a coincidence of the calendar rather than an event.

The check that says the scan works

There is exactly one place on this continent where everybody already knows what happened and when. In February 2010 an iceberg knocked the Mertz Glacier Tongue off the coast of George V Land; that December the same berg grounded in Commonwealth Bay. If a blind scan is worth anything it should find that coast without being pointed at it.

It does, and this is the strongest positive control this project has run. Of the series that break anywhere within 400 km of Cape Denison, 32 of 32 pick 2010 or 2011. Everywhere else on the continent it is 1 of 150.

Steps are in points of concentration for the sea-ice windows and in days for the ice-season length. Every one of the twelve strongest reverses the sign of the trend it sits under. The step is positive around Commonwealth Bay and negative to the west in Wilkes and Adélie Land — that dipole is published here as a measurement and no mechanism is claimed from it. What is being detected is the regional summer ice along 200 km of coast, not the fast ice inside one bay: a 25 km product cannot see a bay, and the site says so wherever it says anything about Cape Denison.

A year with no geography is a coincidence

111 series break at 2007 against about five expected by chance, which looks like a discovery until you count what they are. Those 111 series stand at 85 sites and those sites thin to 2 independent places, 5,645 km apart. It is one year, counted a hundred and eleven times, in two parts of the world that have nothing to do with each other.

This is the same trap the site walked into with penguin sites and had to learn out of: 729 breeding sites are 56 places, so anything counted across sites is weighted by how crowded the Scotia arc is. Every break year here therefore publishes how many independent places carry it and how far apart they are, and a year that is not local is not a regional event however many series sit on it. On that reading the continent has produced exactly two genuinely local clusters: 2011 in George V Land, and 1994 at one place in Princess Elizabeth Land.

series counts inflate wherever sites are close together; the independent_places column is the honest one. A year that is not `local` is not a regional event however many series carry it.

The one candidate nobody can explain

Barrier Bay and West Ice Shelf are published on this site as two of the 5 emperor colonies whose sea ice is robustly falling. They are 75.2 km apart — inside the 200 km radius at which two sea-ice records stop being independent — so they are one place measured twice, and both of them step down at 1994. Allow Barrier Bay that step and its April–November slope goes from -3.51 to +0.47 points per decade.

It is not the satellite, not a Landsat arrival and not an iceberg this project knows of. Other series break in 1993–95 up to 1,264 km away, so the year has no geography either, and no cause may be invented for it. What can be said is exactly this: one place steps down, two published falling trends rest on that step, and the flags are weak rather than family-wise significant. It is the only unexplained candidate the whole continent produced.

And the satellite change, found without being told

The scan was never told that the instrument behind this archive was replaced in 1988. 21 of the 37 breaks it finds anywhere in the emperor-colony record land in 1986–89 — 57% — across 19 colonies, which is the site's oldest published caveat arriving from a method that had no way to know about it.

And the same scan is blind to it in the other archive, for a reason worth stating. A break within 8 fitted years of either end of a series cannot be found — there is not enough record on one side of it to fit anything. The breeding-site archive is built from daily files and loses 1984–87 to its own coverage gate, which puts 1988 inside that trim; the colony archive reads monthly files, keeps those years, and finds the step everywhere. Same continent, same instrument, one archive blind and the other not — purely because of a gate written for another purpose. That is why the zero at the top of this page is a statement about the site archive and not about the satellite.

The ruler, not the measurement

Every trend on this site is published if it clears twice its own error bar. That test assumes each ice year is an independent draw, unrelated to the one before it. Nobody had ever measured whether they are, and it is the assumption the whole site rests on.

Measured against series with known stickiness and no break in them at all, a gate that claims to be wrong one time in twenty is wrong one time in four at ρ = 0.3, and nearly one time in two at ρ = 0.5. That is why the verdict on this page takes the worst of three nulls rather than the obvious one.

the iid gate holds its size when the truth is iid and over-fires badly when the truth is persistent. The STEPPED-residual gate over-fires under real persistence, because fitting a spurious step to persistent noise absorbs the persistence and hands back a rho that is too low. The PLAIN-residual gate is the conservative bracket. The published verdict takes the worst of the three, which is why it is under-sized rather than over-sized.

The real exposure here is modest and the tail is not. The typical series carries a residual stickiness of +0.04 to +0.12, which inflates a standard error by 4 to 13 per cent — so a nominal 2σ gate on this site is really about 1.8σ. But 22 to 34 per cent of series sit above ρ = 0.2, and the robust flag is the exposed one because its second leg is a short 36-year refit.

What this page will not do is publish a replacement count. Applying the standard effective-sample correction to the 2,077 trends on this site that clear 2σ leaves 1,785 still clearing it — but at a series that really carries a level shift, the stickiness is the shift: measured across all 5,686 series, the ones called break run a residual autocorrelation of +0.83 before the step is fitted and +0.09 after, while the ones called none run +0.08 and +0.00. Correcting for stickiness and controlling for a break are two diagnoses of the same residual and must never be stacked. Which of the ten robustly-flagged emperor colonies survive is a decision about the whole site's standard, so it is not one this page gets to make.

Every published figure the scan puts at risk

All 47 of them, at 22 sites — sites and not places, because these are not thinned and this page has just spent a section on the difference. A row here means the site publishes a significant trend for that series and the same series carries a level shift big enough to matter. Where the sign reverses, the published direction is not a description of the record; where it does not, the size is in doubt and the direction is not. Slopes are per decade, in points of concentration for the sea-ice windows and in days for the dates and the season lengths. Every one of these is now printed on the page of the site it belongs to.

Asked the other way round, and it is another null

Everything above looks for a record that jumps. A record that bends — one whose slope changes without any step in it — is a different shape, and the scan above would return a clean null on a series that visibly turns over. So the same 5,686 series were put through the same machinery against that second alternative, with its own Monte Carlo null and its own correction for having asked.

Not one of them clears. Zero of the 842 emperor-colony series and zero of the 4,844 breeding-site series, against each family's own corrected threshold — the same threshold the level scan above is judged on, because each alternative is corrected inside its own family and never pooled with the other. 296 series fire at the uncorrected 5% and none of them may be quoted as a break, exactly as above.

The diagnostic worth keeping is what happens where both alternatives fire at once. 57 series fire on both — and 1 of the 57 picks the same year twice. A record with a real turning point in it gets found in the same place whichever way you look for it; a record without one gives two different answers, which is what these do. The file publishes both and adjudicates neither, because picking the larger of two statistics after looking at them is a choice rather than a test.

And the largest apparent cluster is the same trap this page has already published once. 82 series appear to bend in 1991 — but they are 65 entities standing in 21 independent places, spread over 5,180 km, and 94% of them sit against the edge of the trim where there is least to fit. A year with no geography is a coincidence of the calendar however many series carry it. The whole continent's record was asked both questions too, and the answer there is the same →

How this was done, and what it does not test

The statistic. supremum-Wald (Quandt/Andrews) on a level shift added to a linear trend. Every candidate year in the series is fitted, the strongest is kept, and its p-value is the share of 20,000 Monte Carlo draws with no break in them whose own best candidate beats it. Series shorter than 25 years are not scanned, and the first and last 8 fitted years are trimmed because a break there has nothing to fit against on one side.

The verdict rule. break = p < 0.05/(series in family) under BOTH the iid null and the stepped-AR(1) null; weak = clears 0.05 but not the family threshold; a weak flag may not be quoted as a break on its own

The two blocks are never merged. Mertz Glacier and Point Géologie were named by an open question before anything ran, on geography, so those two carry no correction for multiplicity and the other 5,658 series do. Pooling them would either bury the pre-specified result under a penalty it never owed or license dozens of weak flags the exploratory scan cannot support. And Point Géologie's breaks are at 2007, three years early, so they are not the iceberg and are not attributed to it.

What it does not test. a change of slope without a change of level, and any curvature. This tests one alternative to a line. The first of those two has since been tested and is the section immediately above; this sentence is the scan's own limitation as it stood, kept rather than rewritten. Curvature of any other shape is still untested. The blind spot has a shape and it is stated above: a break within 8 fitted years of either end is inside the trim and cannot be found. The site archive loses 1984-1987 to its coverage gate, which puts 1988 inside the trim there -- so the sensor step is invisible to a blind scan of the site archive and obvious in the colony archive, which keeps those years.

And three families this scan could not reach. The whole-hemisphere record needed the second alternative above before it could be asked honestly, and it has it — on its own page. The ice velocities and the grounding line cannot be asked at all: the longest fitted series is 13 years against the 25 this test requires, and once 8 fitted years are left on each side not one of the 22 series has a single candidate year. That is a count rather than a judgement, and it is published so the next reader does not re-derive it.