Dark energy’s apparent evolution survives a test of uncertain error bars

The question is not just how accurately astronomers measure the universe, but how accurately they estimate their own uncertainty. Measurements can share calibration problems or even the same underlying observations. Treating them as independent can make a combined result look more precise than it deserves.

In a preprint by Eleni-Anna Kolonia and colleagues, the researchers test whether a particular, more flexible treatment of these uncertainties changes a preference for evolving dark energy. Their answer is largely no: the allowed cosmological ranges widen, but the preferred region moves little. That is a robustness result, not proof that dark energy evolves.

Three ways to trace cosmic expansion

The analysis combines measurements that constrain different parts of cosmic history. Planck observations of the cosmic microwave background—the radiation left from the early universe—provide an early-time anchor. Baryon acoustic oscillations, measured by the Dark Energy Spectroscopic Instrument’s second data release, provide a distance ruler through a characteristic pattern in the distribution of matter. Type Ia supernovae provide distance estimates through their standardized brightnesses.

The researchers compare two descriptions of dark energy, the component associated with the universe’s accelerated expansion. In the standard model, called Lambda-CDM, dark energy is a cosmological constant. The alternative lets its equation of state—the relationship between its pressure and energy density—change over cosmic time using two adjustable parameters.

The supernova information comes from three compilations: Pantheon+, Union3 and DES–Dovekie. These are not wholly independent collections. Some supernovae and calibration inputs are shared, while processing methods differ.

Giving uncertainty more freedom

The central mathematical object is a covariance matrix. It records both the uncertainty of individual measurements and how their errors can vary together. The authors introduce an adjustable multiplier for each supernova compilation’s covariance and another for DESI’s covariance. They also allow three effective correlations, one for each pair of supernova compilations.

Multiplying covariance is not the same as multiplying an error bar. Error bars scale with the square root of the multiplier: a covariance multiplied by four means error bars twice as large. The statistical calculation also includes a normalization penalty, preventing arbitrarily large uncertainties from automatically producing a better fit.

For a joint comparison, the team represents the supernova compilations on a common redshift grid. Redshift measures how much cosmic expansion has stretched light. Pantheon+ and Union3 each contribute 22 entries, and DES–Dovekie contributes 19 populated entries, giving 63 measurements in the combined representation. Empty bins are excluded rather than counted as observations.

The preference changes little

For the full combination, the evolving-dark-energy model improves the best-found fit over Lambda-CDM by 12.81 units of the paper’s effective chi-square statistic when covariance scales are fixed. With all four scales free, the improvement is 12.32. Larger improvements indicate a better relative fit, so this is only a small reduction.

Importantly, the three supernova correlations are already adjustable in both versions of that comparison. The change therefore isolates the effect of freeing the covariance scales, rather than comparing complete independence against every added freedom at once.

The authors translate these fit differences into nominal significances of 3.15 and 3.07 sigma. Those figures are statistical diagnostics under an approximation, not probabilities that either cosmological model is true.

Pantheon+ and Union3 show the clearest positive effective correlation, around 0.4–0.5, plausibly reflecting shared observations and calibration inputs. Pantheon+ and DES–Dovekie have a weaker, model-dependent correlation; Union3 and DES–Dovekie remain consistent with zero correlation. These are correlations within the constructed statistical representation, not direct measurements of relationships between individual supernovae.

What the test leaves open

The uncertainty multipliers are only allowed to increase published covariances, never decrease them. Broad allowed ranges therefore do not establish that the original uncertainties were underestimated. Pantheon+ shows a mild preference for additional covariance, but the analysis does not identify a physical cause.

Nor does an overall multiplier capture every possible systematic error. The framework does not model systematic shifts in measurements, and the authors identify redshift-dependent uncertainty changes and more structured survey correlations as possible extensions.

Other assumptions also matter. The full-combination analysis restricts one dark-energy evolution parameter to nonpositive values; allowing positive values was not tested. Lambda-CDM sits on that boundary, complicating interpretation of the nominal significance. The Planck likelihood excludes its lensing likelihood, and the supernova compression is not claimed to preserve all information exactly.

The defensible conclusion is narrow but useful: the preference for evolving dark energy survives these particular changes to uncertainty sizes and cross-survey correlations. Whether it survives other systematic effects remains open.