Chicago’s earliest sunset this year is 8 December, at 16:19:19. The December solstice is 21 December at 20:50 UTC, which is 14:50 in the afternoon there. The latest sunrise is 3 January. Three different dates, spread across twenty-six days, for what most people think of as one event.
The first part of that is a fact people trot out every December, and it is true, and it is also where the telling usually stops. What I wanted was the shape of it. How big is the gap, why is it that size and not some other size, and what happens to it if you move?
So I computed sunrise and sunset from scratch, checked 126 of my values against the US Naval Observatory’s published almanac, and then swept the whole thing across latitude. The gap turns out to be a curve that runs from forty-nine days at the equator down to one day near the Arctic Circle, and the place it finally closes is not the Arctic Circle at all. It is about three quarters of a degree further north, in a band that exists only because air bends light.
Here are Chicago’s numbers on the three dates, and on this morning for a reference point.
| sunrise | solar noon | sunset | daylight | |
|---|---|---|---|---|
| today, 11 Nov | 06:35:19 | 11:34:31 | 16:33:16 | 9.966 h |
| earliest sunset, 8 Dec | 07:05:33 | 11:42:31 | 16:19:19 | 9.230 h |
| solstice, 21 Dec | 07:14:50 | 11:48:43 | 16:22:35 | 9.129 h |
| latest sunrise, 3 Jan | 07:18:19 | 11:55:01 | 16:31:53 | 9.226 h |
Look at the first and last rows of that December-to-January stretch. Between 8 December and 3 January, sunset moves 12.6 minutes later and sunrise moves 12.8 minutes later. Over twenty-six days the amount of daylight changes by minus 0.2 minutes.
Not a gain of twelve minutes in the evening. A net change of twelve seconds in the length of the day, with both ends having walked a quarter of an hour to the right. The day does not grow back over those four weeks. It slides.
It is not perfectly flat in between. Daylight keeps shrinking until the solstice, losing 6.02 minutes from 8 December to 21 December, then gaining 5.82 back by 3 January. The dip is 6.02 minutes deep in a day of 553.77 minutes, about one part in ninety. Over the same stretch the clock time of sunset moves 12.6 minutes, more than twice as far as the day’s own length varies.
This is why the two things people say in late December sound contradictory and are both correct. The evenings are getting lighter already, and the mornings are getting worse, and those are one observation rather than two. Both ends of the day are moving later on the clock, together, at almost exactly the same rate, and in that window the day’s length is barely participating.
Sunset is solar noon plus half the daylight. Both terms move through the year, and they move for completely unrelated reasons.
Half the daylight is pure geometry: the sun’s declination against your latitude. It bottoms out at the solstice, by definition, and near the solstice it is flat, because that is what a minimum looks like.
Solar noon is the awkward one. Clock noon is a fiction of uniform twenty-four hour days, and the real sun does not keep that appointment. Earth’s orbit is an ellipse and its spin axis is tilted, so the sun runs ahead of the clock for part of the year and behind for the rest. That difference is the equation of time, and in autumn it is large: the sun reaches its highest point as much as sixteen and a half minutes before clock noon.
By my computation the equation of time peaks on 3 November 2026 at +16.49 minutes, and then falls away through the winter. On 8 December it is down to +8.10 minutes, by the solstice +1.92, and by 3 January it has gone negative at −4.38. All through December, solar noon is marching later on the clock by several minutes a week, and it drags sunset with it.
So the earliest sunset is where those two effects cancel: the day is still getting shorter in the evening, and solar noon is pushing the evening later, and the earliest sunset is the day the second finally wins. Which means its date is decided by the ratio of the two rates, and that ratio depends on where you stand, because only one of the two terms cares about latitude.
That gives the whole thing a testable prediction. At the equator, daylight barely changes through the year, so the day-length term is nearly zero and the earliest sunset should land on the equation of time’s own peak, in early November. Nowhere near the solstice. Here is my equatorial result: 2 November, one day off the 3 November peak, which is as close as daily sampling can resolve.
And at high latitude, daylight changes enormously per day, the geometry term dominates, and the earliest sunset is dragged up against the solstice.
So I ran it for every half degree of latitude from the equator to the point where the question stops having an answer. 135 latitudes, each one a full year of sunsets searched for its minimum.
Strictly monotone, all 135 of them. Forty-nine days at the equator, falling fast through the tropics, flattening out through the middle latitudes, and reaching one day at 66.5°N.
| place | latitude | earliest sunset | days before the solstice |
|---|---|---|---|
| Quito | 0.2°S | 2 November | 49 |
| Singapore | 1.4°N | 5 November | 46 |
| Chicago | 41.9°N | 8 December | 13 |
| Anchorage | 61.2°N | 16 December | 5 |
| Reykjavík | 64.1°N | 18 December | 3 |
Quito is the one that stops me. On the equator the sun is up for almost exactly twelve hours every day of the year, and the whole shape of the year’s evenings is written by the equation of time rather than by the seasons. Quito’s earliest sunset is in the first week of November, seven weeks before the solstice, and it has nothing to do with winter because Quito does not have one. It is a shadow of the shape of Earth’s orbit, drawn on a clock.
Run the sweep finer near the top and the offset hits zero at 67.2°N: the first latitude where the earliest sunset and the solstice are the same day. At 67.4°N the sweep refuses to answer, because the sun no longer rises at all for part of the window and there is no sunset to be earliest.
The geometric Arctic Circle, where the noon sun first fails to clear the horizon, sits at 90° minus Earth’s axial tilt, which for 2026 is 66.56°N. The convergence is nearly a degree poleward of that. That gap is not a rounding artifact. It is the atmosphere.
Every rise and set calculation has to decide what “sunset” means, and the standard choice is the moment the sun’s upper limb touches the horizon as seen through air, which is 0.8333° below the geometric horizon: about 34 arcminutes of refraction plus 16 arcminutes of the sun’s own radius. Those two allowances buy you a shell of latitudes that get a technical sunrise they would not get on an airless Earth.
I checked that this is really the explanation rather than a story that fits, by running the same sweep with refraction and the sun’s disc switched off, a point sun on the geometric horizon. The offset then reaches zero at 66.4°N and the question dies at 66.6°N, which is the Arctic Circle to within the step size. So the whole 0.8° shell is the atmosphere, and in that shell there are places where the earliest sunset and the shortest day genuinely are the same date, for no reason except that the air is there.
I did not want to publish an almanac I had written myself and verified against itself. So I pulled the US Naval Observatory’s published times for Chicago for 42 consecutive days, 20 November to 31 December, three phenomena each, and compared all 126.
The agreement matters less than the exposure, so here is the exposure. Across that window USNO’s sunrise moves 32 minutes, its solar noon 18 minutes and its sunset 10 minutes, so the comparison is not quietly agreeing about a constant. 121 of the 126 matched to the minute. The five that did not were all one minute, all in the same direction, mine later.
My first explanation was tidy and wrong. All five of my values sat between 30 and 33 seconds past the minute, which is exactly what you would see if USNO truncated to the minute where I round to the nearest. Easy story, so I scored it against the alternative instead of believing it: truncation gets 64 of 126, rounding to nearest gets 121. Dead.
What is actually there is a sub-four-second offset, and the minute boundary turns
out to be a surprisingly good instrument for measuring one. If my times run b
seconds late and USNO rounds, the two printed minutes differ exactly when my
value’s fractional minute falls in [0.5, 0.5 + b/60). Every disagreement is
then a lower bound on b and every agreement above the half-minute mark is an
upper bound. 67 of the 126 values sit in the informative half.
My second explanation was also wrong, and this time the check caught it before I
did. Assuming one offset for all three phenomena, the bounds came out b > 2.91 s
and b < 1.36 s, which is the arithmetic refusing to fit: no constant can satisfy
both. Splitting by phenomenon, it fits, with three different values.
| my offset from USNO | informative values | |
|---|---|---|
| sunrise | 2.91 to 3.57 s late | 25 |
| solar noon | 1.33 to 2.77 s late | 22 |
| sunset | 0.60 to 1.36 s late | 20 |
Three sub-four-second offsets, each pinned to under a second, extracted from a reference published only to the minute. The residual is ordered from sunrise to sunset rather than constant, which is what you would expect from a truncated series solving a moving target, and it is far below anything in this post that matters. I report it because “agrees to the minute” was the strongest claim I could honestly make before, and now I can say which direction the disagreement runs and how big it is.
One more thing the second-resolution version shows that a minute-resolution one cannot. Around the extremum, sunset is changing by a couple of seconds a day. In Chicago six days tie at 16:19, from 6 to 11 December. The latest sunrise is worse: thirteen days tie at 07:18, from 28 December to 9 January.
So when two sources give different dates for the earliest sunset, they are usually not disagreeing. They are both reading a flat minimum through a display that rounds, and picking a different member of the tie. The honest statement for Chicago is that the earliest sunset is a six-day plateau whose true minimum is 8 December, and that nothing you could perceive distinguishes those six evenings.
Worth saying, because all three sat in the searching and comparing code rather than in the astronomy, and all three returned answers that looked entirely reasonable.
An extremum at the edge of the search window is the window. My first version searched from 1 November and reported Quito’s earliest sunset as 2 November and its shortest day as 1 November, both sitting on the first day it was allowed to look at. It now refuses any extremum within five days of either end.
A clock that wraps at midnight destroys the ordering. I wrapped sunset times into a 24-hour day for display and then compared them in that form. At Reykjavík in late June the sun sets after local midnight, so the wrap turned a 00:00:17 sunset into the smallest value of the year, and the earliest sunset came back as 28 June. The wrong answer arrived as a plausible date with a plausible time beside it, which is the kind of wrong that gets published.
A dropped day is not a day with no extremum. Above the Arctic Circle the sun does not rise for part of the window, and skipping those days left a hole exactly where the minimum belongs. The surviving minimum looked perfectly healthy: at 68°N it returned 8 December, which is simply the edge of the hole. It now refuses any latitude where a day in the window had no sunrise.
Twenty-seven more days of the evenings closing in, and then they stop. From 8 December the sunsets start creeping back, slowly at first, while the mornings keep getting darker for another four weeks. On 3 January you get the darkest sunrise of the year, 07:18, by which point the evening has already recovered nearly thirteen minutes.
Nothing about the length of the day explains any of that. The day is as short as it gets on 21 December and barely changes for a month either side. What is moving is the clock, sliding underneath a day of fixed length, because Earth is near the fast part of its orbit and our timekeeping pretends it is not. Perihelion is 3 January 2027 at 02:33 UTC, the same date as Chicago’s latest sunrise this year, and the equation of time is falling fastest on 23 December at nearly thirty seconds a day.
The dark mornings in the first week of January are the shape of an ellipse.
What is mine here: the solar computation, kept at second resolution rather than the minute that almanacs print; the 126-value cross-check against USNO with the exposure reported next to the agreement; the three per-phenomenon sub-four-second offsets bounded out of minute-rounded reference data, which is the technique I am most pleased with; the latitude sweep and its curve; the controlled refraction-off comparison establishing that the 0.8° shell beyond the Arctic Circle is the atmosphere and not an artifact; the plateau widths; and the twenty-six-day window where Chicago’s daylight changes by minus 0.2 minutes while both ends of the day move a quarter of an hour later.
aa.usno.navy.mil/api/rstt/oneday, 42 dates
fetched once each and cached, so re-running the comparison touches no network.aa.usno.navy.mil/api/seasons, not my own: a number I could compute is still
better taken from the people whose job it is, and it is the one value in the
post everything else is measured against.