Why do two good measurements of the universe disagree?

The Planck satellite mapped the light released when the universe was about 380,000 years old. The pattern of hot and cold spots in that light encodes the contents and geometry of the early universe, and running the standard cosmological model forward from it predicts today’s expansion rate.

The 2018 result: a Hubble constant of 67.4 plus or minus 0.5 kilometres per second per megaparsec. The paper adds that it finds "no compelling evidence for extensions to the base-ΛCDM model".

The other method does not assume a model of the early universe at all. It builds a ladder: measure distances to nearby Cepheid variable stars by geometry, use Cepheids to calibrate type Ia supernovae in galaxies that contain both, then use those supernovae to reach far enough that the expansion dominates.

The 2022 measurement, using Hubble Space Telescope observations of Cepheids in the host galaxies of 42 supernovae, gives 73.04 plus or minus 1.04. The uncertainty is quoted including systematics.

Five or six units apart, with quoted uncertainties of half a unit and one. In 2009 the two routes gave 74.2 and 71.9, each uncertain by more than the gap between them, and it was reasonable to expect the two to converge as the measurements improved.

They improved. They did not converge. The SH0ES team puts the difference at five standard deviations, which is the threshold particle physicists use to call something a discovery rather than a fluctuation.

The obvious candidate is an unknown systematic in the distance ladder: crowding in the fields where Cepheids are measured, the metallicity dependence of the period-luminosity relation, or supernova corrections that differ between star-forming and quieter host galaxies.

Each of these has been tested hard: the 2022 SH0ES paper alone runs nearly seventy variants of its analysis. Alternative rungs, including tip-of-the-red-giant-branch distances and gravitational lensing time delays, have been used to route around the suspect steps, and this is where the two camps part. The SH0ES team finds red giants agree with its Cepheids, giving 72.53 in combination. The Chicago-Carnegie team, measuring red giants with the Hubble and James Webb telescopes, gets 70.39 and calls it consistent with the standard model; SH0ES puts the difference down to which supernovae each sample happens to include. Six lensed quasars, analysed without any distance ladder at all, gave 73.3 in 2019. That precision leaned on assumptions about the lensing galaxies’ mass, and a 2025 analysis of eight lenses that loosened them gave 71.6, uncertain by three or four units either way.

The early-universe number is not a direct measurement of today’s expansion. It is a prediction, produced by evolving the standard model of cosmology forward for 13.8 billion years. If that model is missing something, the prediction can be precise and wrong.

The review that classifies the proposals sorts them into early dark energy, late dark energy, extra relativistic degrees of freedom, extra interactions, modified gravity and six other families, and cites more than a thousand papers doing it. None has made a strong case. Models that ease the tension on Planck’s data alone often disagree with other probes, and many ease it by widening the error bar rather than by moving the number.

Nobody knows. That sentence is the reason this question is on this site rather than in a textbook: it is not a gap in public understanding, it is a gap in the field’s.

What would settle it is more independent routes to the same number. Gravitational-wave sirens, which measure distance directly from the physics of a merger without any ladder at all, are one such route: the first, in 2017, gave a value consistent with both camps and an uncertainty far too wide to choose between them. There are not yet enough of them.

If it is a systematic, the fix will be dull and specific: a recalibration, a better treatment of one population of stars, and the two numbers will slide together over a few years. It may take more than one: the 2021 review of the tension knew of no single proposed error that the data had not already ruled out.

If it is physics, it will show up in other places at once, because a change to the model moves more than one number. That is the useful property of this kind of tension. A real crack is never local.