Dynamic or Systematic? Bayesian model selection between dark energy and supernova biases

A. N. Ormondroyd, Will Handley, M. P. Hobson, A. Lasenby, Yallup, D.arXiv 2025 · DOI · arXiv

TL;DR

Is dark energy actually evolving, or are our "standard candles" just a bit flickering? This paper demonstrates that the recent, much-hyped evidence for dynamical dark energy might actually be a simple systematic error—a small magnitude offset—in the supernova data. By applying rigorous Bayesian model selection, the authors show that a standard $\Lambda$CDM model with a calibration correction is a much better fit for the universe than complex new physics.

What Problem Does This Paper Address?

For decades, the "Cosmological Constant" ($\Lambda$CDM) has been the gold standard, positing that dark energy is constant ($w = -1$). However, recent data from the Dark Energy Survey (DES-5Y) and the Dark Energy Spectroscopic Instrument (DESI) seemed to hint at something more exotic: "dynamical" dark energy that changes over cosmic time.

The catch? Some researchers, notably George Efstathiou, suggested this "discovery" was actually caused by a systematic bias in the DES pipeline. Specifically, the low-redshift supernovae (the "anchors" of the distance ladder) came from older legacy surveys and might have a 0.04 magnitude offset compared to the high-redshift supernovae found by DES itself. This paper seeks to settle the debate using Bayesian evidence: Does the data actually prefer a complex dark energy model, or does it prefer a simple model with a calibration offset?

Methods

The authors employ several sophisticated statistical tools to compare models:

  • Flexknot Reconstruction: A "model-independent" approach that uses a linear spline with a variable number of "knots" to reconstruct the dark energy equation of state, $w(z)$. This allows the data to "speak for itself" rather than forcing it into a specific formula.
  • Bayesian Model Selection: They calculate Bayes factors (the ratio of evidences) to determine if adding a new parameter (like a magnitude offset) is justified by the improvement in fit, or if it's just "overfitting."
  • Nested Bridge Sampling (NBS): A novel, efficient method for calculating these Bayes factors by comparing the similarity between the posteriors of nested models.
  • JAX Implementation: To speed up the heavy math, they ported their likelihoods to JAX, allowing them to run complex nested sampling on GPUs in seconds rather than tens of minutes on traditional CPUs.

Datasets

The study utilizes two primary cosmological pillars:

  1. DES-5Y Supernovae: Over 1,600 Type Ia supernovae, including 194 low-redshift "anchors" from legacy surveys (CfA, Carnegie Supernova Project, and the Foundation Sample).
  2. DESI DR2 BAO: Baryon Acoustic Oscillation measurements, which provide a "standard ruler" for the expansion of the universe, acting as a crucial cross-check for the supernova "standard candles."

Key Insights and Ideas

  • Occam’s Razor Wins: The "flexknot" model (which can represent almost any dark energy behavior) is statistically "beaten" by a simple $\Lambda$CDM model once a single magnitude offset parameter ($\Delta m_B$) is introduced.
  • The "Decisive" Evidence: While the evidence for an offset is suggestive using supernova data alone, it becomes "decisive" (a log-Bayes factor of ~4.1) when combined with DESI BAO data. This suggests that the "tension" between different datasets was largely driven by this calibration issue.
  • Systematic vs. Physics: The paper highlights a recurring theme in precision cosmology: what looks like "New Physics" is often "Old Systematics." If you don't account for the 0.04 magnitude shift in low-z supernovae, the statistics "force" the dark energy model to evolve to compensate for the discrepancy.

Key Results

  • The Offset is Real: The calculated offset $\Delta m_B$ is consistently around -0.04, exactly as predicted by critics of the original DES-5Y analysis.
  • Tension Reduction: Without the offset, $\Lambda$CDM is the most "discrepant" model when compared to DESI data. With the offset, $\Lambda$CDM's tension is significantly reduced, making it once again a top-tier fit for the observed universe.
  • Performance Leap: The JAX-based pipeline allowed the team to perform model selection almost instantaneously. This suggests a future where complex Bayesian analysis can be done in real-time during data collection.

Limitations and Future Work

The authors acknowledge that simply adding a "magnitude offset" parameter is a diagnostic tool, not a complete fix. A proper resolution requires the supernova collaborations to dive back into their pipelines to find the physical root of the selection effects or calibration errors.

Furthermore, while the JAX/NBS approach worked perfectly for simple models like $\Lambda$CDM and CPL, the authors noted that more work is needed to ensure these high-speed tools can handle the "multi-modal" (complex, multi-peaked) landscapes of free-form flexknot reconstructions. Future work will likely focus on making these robust Bayesian tools standard for the next generation of "Big Data" cosmology surveys.

Summary updated

Abstract

DES-5Y supernovae, combined with DESI BAO, appear to favour Chevallier-Polarski-Linder $(w_0, w_a)$ dynamical dark energy over $Λ$CDM. arXiv:2408.07175 suggested that this is driven by a systematic in the DES pipeline, which particularly affects the low-redshift supernovae brought in from legacy surveys. It is difficult to investigate these data in isolation, however, as the complicated supernovae pipelines must properly account for selection effects. In this work, we discover that the Bayesian evidence previously found for flexknot dark energy (arXiv:2503.17342) is beaten by a magnitude offset between the low- and high-redshift supernovae. In addition, we find that the possible tension between DES-5Y and DESI is significantly reduced by such an offset. We also take the opportunity to trial Nested Bridge Sampling with Sequential Monte Carlo as an alternative method for calculating Bayes factors.
BibTeX
@misc{ormondroyd2025dynamic,
  author = {A. N. Ormondroyd and Will Handley and M. P. Hobson and A. Lasenby and Yallup, D.},
  title = {{Dynamic or Systematic? Bayesian model selection between dark energy and supernova biases}},
  year = {2025},
  eprint = {2509.13220},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.13220},
  howpublished = {arXiv},
}
Figure