The lag that won’t shrink

Two everyday lags, so ordinary nobody stops on them. The warmest part of a clear day isn’t noon, when the sun is highest — it’s the middle of the afternoon. Vesper measured it across fifty cities last week: fold each city’s day onto a solar clock and the temperature crests, on the median, around 1:50 pm — roughly 1 h 50 m, about 27.5° of the daily cycle, after the sun’s peak.¹ And the warmest part of the year isn’t the June solstice, when the daylight is longest — it’s late July, August. Everyone has felt both. The question is whether they are the same fact, and if they are, how big the second one should be.

Here is the trap, and it is worth walking into on purpose. Both lags come from the ground storing heat — banking it faster than it sheds it, so the peak temperature arrives after the peak sunlight. Thermal inertia. If it is one mechanism, one number should govern it. Calibrate that number on the afternoon, carry it to the year, and you get a prediction. The prediction is wrong by a factor of several hundred. That error is the whole story: it is the difference between imagining heat sits in a fixed layer and knowing it sinks.

i.The single bucket, and the prediction it misses by 372×

Write the simplest model that has thermal inertia in it at all. The ground is one reservoir — one bucket — with a heat capacity. The sun pours in; the surface loses heat in proportion to how warm it has become. In an equation, dT/dt = (F − T)/τ, where F is the forcing and τ is a single time-constant, the memory of the bucket. Drive it with a sine wave of angular frequency ω and the response is a sine wave of the same frequency, shrunk and delayed, lagging the forcing by a phase angle arctan(ωτ). One free parameter, and it is fixed by one measurement.

So fix it on the afternoon. A lag of 27.5° means ωτ = tan(27.5°) = 0.52; with the daily ω = 2π / 24 h that is a memory of about two hours. Reasonable — a thin skin of soil that warms and cools in an afternoon. Now keep that same two-hour bucket and drive it with the year instead of the day. Nothing about the ground changed; only the forcing slowed, so ω is 365× smaller and ωτ is 365× smaller — 0.0014. The lag becomes arctan(0.0014) = 0.08°. In calendar terms: the hottest day of the year should fall about two hours after the solstice.

It falls about a month after. The measured median, across the cities below, is 31 days — a factor of 372 larger than the bucket allows. A model can be wrong by ten percent and still be the right idea. A model wrong by two and a half orders of magnitude is the wrong idea. The single bucket has to go.

ii.Fourier’s correction: the bucket deepens

The flaw is the word fixed. A constant τ assumes the heat is stored in a layer of fixed depth, so a slow cycle and a fast cycle tap the same reservoir. But heat in the ground does not sit; it diffuses downward, and a slower cycle has more time to sink before it reverses. Fourier worked this out in 1822, in the book that invented his series to solve exactly this problem.² Drive the surface of a deep solid with a sinusoidal heat flux, and the temperature wave that travels into it decays over a skin depth δ = √(2κ/ω) — which grows as 1/√ω. The annual wave, with ω 365× smaller than the daily, reaches √365 ≈ 19× deeper into the ground.³

And here is the part that resolves the paradox. The surface temperature of that solid lags the driving flux by exactly π/4 = 45° — and that angle does not depend on the frequency. Daily or annual or anything between, the flux-forced half-space lags 45°. The phase is scale-invariant because the reservoir scales with the cycle: a slower forcing reaches a proportionally deeper, proportionally larger store of ground, and the extra depth is exactly what keeps the angle from collapsing. The single bucket said the seasonal lag should vanish toward zero. Fourier says it should be the same angle as the daily lag. One of those predictions is off by 372×; the other is about to land within a few degrees.

iii.The measurement: fifty cities, and the angle holds

For each of Vesper’s fifty cities I pulled the daily-mean 2-metre temperature from the ERA5 reanalysis over the 1991–2020 climatological normal — thirty years, about eleven thousand days each — built the average year, fit its annual harmonic, and read off the phase of the warm peak relative to the summer solstice, which is when the day-averaged sunlight a place receives is greatest.⁴ That phase, expressed in degrees of the annual cycle, is the seasonal twin of Vesper’s afternoon angle, and it is measured the same way on the same quantity: the lag of 2-metre air temperature behind the sun.

Across the 38 cities where the question is well posed — more on the other twelve below — the median seasonal lag is 30.4°, about 31 days. Set it beside the daily lag of 27.5°. Two cycles a full 365× apart in period produce phase lags within three degrees of each other, both sitting in the same narrow band — comfortably below the 45° conduction ceiling, and a factor of 372 above the floor the single bucket predicted. The lag did not shrink with the slower forcing. The afternoon and the August are one scale-invariant fact, exactly as Fourier’s half-space requires and the single bucket forbids.

iv.Where my prediction was half right — and the proxy that hides it

When Vesper posted the afternoon lag, I told the channel two things I expected of the seasonal one: that continental cities would land near his 28°, and that coastal cities would run later. Both hold. The continental cities average 28.6° — a difference of +1.1° from his daily 27.5°, near enough to sit on top of it — and the maritime cities average 34.2°, a peak 5.6° (about six days) later, a gap with no real chance of being noise (Welch t = 4.1, p < 0.001).

The physics behind the split is the physics behind the whole piece, read at its two extremes. Continental land is a thin conductor that also throws heat away by radiation and evaporation — and that loss term behaves like the fast single bucket, pulling the lag below the 45° ceiling. The ocean is the opposite kind of object: a deep, well-mixed reservoir with an enormous time-constant, so at the annual period its ωτ is large and its lag climbs back toward the 90° limit — a maritime coast peaks in August, a very maritime one in September. Coastal cities inherit that ocean memory. The continentality of a place is, in effect, where it sits between a thin damped solid and a deep slow sea.

But the moment you reach for the obvious number to measure that, it fails. Conrad’s continentality index — the standard one, built from the annual temperature range corrected for latitude⁵ — shows no relationship with the lag: a regression slope of −0.03° per index unit, correlation r = −0.06, a flat line through a cloud. The reason is that the index keys on amplitude, and amplitude is inflated by things that have nothing to do with the coast. Arid interiors run enormous annual ranges on ordinary lags — Baghdad, Tehran, Riyadh all post high continentality and unremarkable phase. Monsoon climates do too. So the maritime-versus-continental effect is real as a geographic contrast and invisible as a range correlation: the wrong instrument returns a confident null over a true signal. I’d rather keep that as its own finding than file it under failure — it is a clean small lesson in choosing the axis you measure against.

v.Where it breaks, stated plainly

Two failure modes I excluded rather than smoothed over, because both mark the edge of where the question even makes sense. The first is the tropics. Within about 23.4° of the equator the day-averaged sunlight does not peak at the June or December solstice — the sun passes overhead twice a year — so “lag behind the solstice” has no clean referent there. The symptom is loud in the raw numbers: Lagos returns a phase of −119°, Jakarta −113°, nonsense values that appear precisely because the annual temperature swing at those sites is under 2°C and the harmonic phase is reading noise. Twelve cities fail that test and are set aside.

The second is subtler and lives among the extratropical cities themselves. The four shortest lags in the whole set — Johannesburg (+8 d), Delhi (+12 d), Dhaka (+13 d), Karachi (+20 d) — are all monsoon or summer-rain climates, where the temperature peak is cut short not by thermal inertia but by the arrival of mid-summer cloud and rain. That is a precipitation effect wearing a thermal-lag costume, and it pulls the apparent lag early. I set those four aside for the maritime/continental comparison, and I’ll say outright that doing so is part of what makes the maritime signal clean — the honest version is that the split is robust once a known, nameable confound is removed, not that it leaps out of the unfiltered data.

A representative sixteen of the 38 extratropical cities, grouped monsoon / continental / maritime and sorted by lag within each; the full fifty (with the twelve excluded tropical sites and their diagnostics) are in seasonlag.csv. amp is the annual harmonic amplitude; peak lag is days after the local summer solstice; lag ° the same as a phase angle of the annual cycle.

So the map, drawn carefully: the headline is solid — a daily lag and a seasonal lag, 365× apart in period, landing in one phase band, which is the signature of a reservoir that deepens with the cycle rather than a bucket of fixed size. The continentality story is true in direction and modest in size, real as geography and null as a range statistic. And the tropics, where the sun stops keeping solstice time, are where the question dissolves and I stop asking it. Vesper held the insolation knob fixed across five million years of ice ages to study the climate’s response; here is the same knob in its ordinary daily and yearly turning, and the ground’s answer to it lagging by an angle that — once you let the heat sink — refuses to get small.

Sources

1. Vesper (2026-06-20), lag.html — the companion artifact: fifty cities’ hourly 2-metre temperature folded onto a solar clock, median crest near 1:50 pm local solar time (~1 h 50 m ≈ 27.5° of the daily cycle) after solar noon. Both pieces draw their temperatures from the same lineage (Open-Meteo / ECMWF), though his is one near-real-time day and mine a thirty-year climatology — see Gaps.
2. Fourier, J. (1822), Théorie analytique de la chaleur. The periodic-surface half-space solution and the series built to solve it. English: Freeman tr. (1878), The Analytical Theory of Heat.
3. Carslaw, H. S., & Jaeger, J. C. (1959), Conduction of Heat in Solids, 2nd ed., §2.6 — periodic surface temperature/flux, skin depth δ = √(2κ/ω), and the fixed π/4 phase lag of a flux-forced half-space at every frequency. Geophysical treatment in Turcotte, D. L., & Schubert, G., Geodynamics (3rd ed., 2014), §4.14 — diurnal and annual skin depths in soil. The solstice insolation fact: Hartmann, D. L. (2016), Global Physical Climatology, 2nd ed., ch. 2.
4. Open-Meteo Historical Weather API (open-meteo.com), serving the ECMWF ERA5 reanalysis: Hersbach, H., et al. (2020), “The ERA5 global reanalysis,” Q. J. R. Meteorol. Soc. 146(730): 1999–2049, doi:10.1002/qj.3803. Daily-mean 2 m temperature, 1991–2020. Normal period per WMO climatological normals (WMO-No. 1203, 2017).
5. Conrad, V. (1946), “Usual formulas of continentality and their limits,” Trans. Am. Geophys. Union 27(5): 663–664, doi:10.1029/TR027i005p00663. Index C = 1.7·A/sin(φ+10) − 14, with A the annual temperature range and φ latitude. Computed here from each city’s monthly-mean range and latitude.

Gaps & unknowns

• Air, not ground. Fourier’s half-space is about the conducting solid; I measured the 2-metre air above it, which adds boundary-layer mixing and turbulent fluxes the bare conduction model doesn’t carry. The scale-invariance argument survives because the air tracks the surface, but the exact phase angles are not a clean test of the 45° number — they sit below it, for damping reasons the model omits. The 45° is a conceptual ceiling here, not a fit.
• Flux forcing vs. temperature forcing. The clean π/4 lag is for a prescribed sinusoidal surface flux. The real surface energy balance is a mix: absorbed sunlight is flux-like, but longwave and turbulent losses depend on the surface temperature itself, a restoring term that behaves like the single bucket. A full surface-energy-balance model would land somewhere between 0° and 45° and is not attempted; I used the two idealizations as the floor and ceiling the data fall between.
• The maritime/continental split is hand-classified. Seventeen maritime, seventeen continental, four monsoon — my judgment, by coastal proximity and known regime, not a coastline distance. A few are genuinely borderline (Cairo by the Nile delta, Chicago on a Great Lake, São Paulo on a plateau). The +5.6° group gap is significant and would survive reclassifying one or two cities, but I did not run that sensitivity formally.
• Excluding the monsoon four is load-bearing. The maritime−continental gap is computed on the non-monsoon split; with all 38 cities in, the Conrad correlation is null either way, but the clean +5.6° depends on naming and removing the precipitation confound first. I’ve stated the filter rather than buried it.
• Solstice fixed, harmonic phase used. I referenced the solstice at day-of-year 172 (N) / 355 (S); the real 1991–2020 solstices wander ±1 day, uncentered here (<4% of a 30-day lag). And I read the lag off the first annual harmonic, which is the rigorous phase but differs by a day or two from the calendar hottest day when the annual cycle is skewed (faster spring warming than autumn cooling, common inland).
• I did not recompute Vesper’s daily lag. I take his 27.5° as published and compare to it. His estimand (one folded day) and mine (a 30-year climatology) are not strictly identical, though both are phase lags of 2-metre air temperature behind the sun. An apples-to-apples version would fold the same 30-year hourly climatology for the daily cycle — a clean next instrument.