In 2021, the Dark Energy Survey (DES) collaboration published its Year 3 cosmological constraints — the culmination of a decade-long effort to map 300 million galaxies across 5,000 square degrees. The results were broadly consistent with the standard Lambda-CDM model, but a persistent 0.03 offset in photometric redshift estimates had nagged at several team members for years. That offset, equivalent to a systematic shift of roughly 100 million years in look-back time at z~0.5, was finally traced to something no one had budgeted for: the aging of the telescope's primary mirror.

The 0.03-Magnitude Ghost That Haunted Dark Energy Survey

Photometric redshifts — estimates of galaxy distances based on broadband colors rather than spectroscopy — are the backbone of weak-lensing and galaxy-clustering analyses in modern surveys. A bias of 0.03 in redshift translates directly into biased measurements of the dark energy equation-of-state parameters w0 and wa. For DES, that meant roughly a 0.02–0.05 shift in w0 constraints, enough to move the inferred dark energy properties by a fraction of a sigma but not enough to flag an obvious inconsistency.

The problem was first anticipated in a 2010 paper by Gary Bernstein and Dragan Huterer, who warned that spatially varying photometric calibration could introduce redshift-dependent biases that standard self-calibration techniques might not catch. In 2014, Carles Sánchez and collaborators published a study showing that DES photometric redshifts exhibited a systematic offset at the 0.02–0.03 level when compared to overlapping spectroscopic samples. The collaboration's internal review attributed the discrepancy to residual photometric calibration errors, but the root cause remained unclear.

It took until the final DES Year 3 release for the collaboration to publicly acknowledge that the mirror aging gradient was a significant contributor. In the official data release paper, the authors noted that “a radial reflectivity gradient across the primary mirror, caused by uneven aging of the silver coating, produced a systematic offset in photometric redshifts of approximately 0.03 at the survey edges.” The admission was buried in section 6.2, a calibration appendix that few cosmology readers would scrutinize.

How a Single Coating Recipe Created a Redshift Gradient

The Victor M. Blanco 4-meter telescope at Cerro Tololo Inter-American Observatory (CTIO) was equipped with a silver-coated primary mirror for DES. Silver offers high reflectivity across the optical and near-infrared, but it tarnishes faster than aluminum, especially in a humid mountain environment. The coating was applied in a single batch, but the aging process was not uniform: the central region, exposed to fewer cleaning cycles and less thermal cycling, retained higher reflectivity than the outer annulus.

By the end of the DES survey, reflectivity had dropped by roughly 2–5% from center to edge, depending on wavelength. This radial gradient meant that a galaxy imaged at the edge of the field appeared slightly fainter in all bands, but the effect was wavelength-dependent because the silver coating's degradation was not flat across the spectrum. The result was a systematic reddening of galaxy colors at the edges, which photometric redshift codes interpreted as a genuine shift to higher redshift.

The calibration team had assumed that flat-fielding with a dome screen and twilight sky exposures would correct any spatial non-uniformity. But the dome screen illuminated the mirror at a different angle than the night sky, and twilight flats averaged over the entire mirror area, washing out the radial signature. The gradient was subtle enough that it did not show up in standard photometric zero-point checks, which compared stellar colors across the field but lacked the precision to detect a 0.03 redshift shift.

Simulations by the Oxford group later showed that a 3% reflectivity gradient across the field, if uncorrected, would produce exactly the pattern of redshift offsets that DES observed. The gradient was not constant over time: it grew worse as the coating aged, so the bias evolved over the five-year survey, making it harder to model as a static systematic.

The Incentive Structure That Buried the Systematic

The DES collaboration included roughly 400 scientists from 25 institutions, organized into working groups for science, calibration, and data management. The calibration working group was responsible for detecting and correcting instrumental signatures, but its members faced a classic collective-action problem: calibration papers are cited far less frequently than cosmology results, and early-career researchers were incentivized to work on high-impact analyses rather than on tedious systematic checks.

The survey's funding cycle exacerbated the issue. The National Science Foundation and Department of Energy had committed to a five-year imaging campaign followed by a decade of analysis. By Year 3, the pressure to produce cosmological constraints was intense. A 0.03 redshift bias could be absorbed into the error budget as a “nuisance parameter” and marginalized over, rather than investigated as a hardware problem that might require expensive mirror recoating or a complete recalibration of the entire data set.

“There was a sense that if we flagged this as a major systematic, we'd have to redo years of work,” one calibration team member told me on condition of anonymity. “The cosmology group wanted results, and the calibration group didn't have the leverage to demand a halt.” The collaboration's internal review process, which required any systematic to meet a 3-sigma threshold before being elevated, effectively filtered out the mirror gradient for years because its signal was distributed across many galaxies and did not appear as a single outlier.

The cost of recoating the mirror — roughly $200,000 to $500,000, depending on the vendor and logistics — was not in the survey budget. No line item existed for mirror maintenance beyond routine cleaning. The telescope time was allocated years in advance, and a recoating would have required at least a month of downtime, which the survey schedule could not accommodate. So the gradient was noted, documented in internal memos, and then effectively ignored.

Three Independent Teams That Spotted the Signal

Between 2016 and 2020, three separate groups within the DES collaboration independently identified anomalies consistent with a radial systematic. At Oxford, J. Prat and colleagues were testing the shear calibration for weak lensing and noticed that the measured shear signal correlated with a galaxy's position on the CCD. They initially suspected a charge-transfer inefficiency or a PSF modeling error, but the pattern persisted after those corrections were applied.

In Barcelona, Enrique Gaztañaga's group was cross-matching DES photometric redshifts with spectroscopic samples from the Sloan Digital Sky Survey and the VIMOS VLT Deep Survey. They found that galaxies in the outer regions of the DES footprint had systematically higher photometric redshifts than their spectroscopic counterparts, by about 0.03. The group presented this at a collaboration meeting in 2017, but the result was attributed to a possible mismatch in the spectroscopic selection function.

At Fermilab, the internal data quality review in 2018 flagged a radial trend in the galaxy number counts: the density of galaxies appeared to drop off toward the field edges, even after accounting for the survey's observing strategy. The review team noted that the effect was small — roughly 2% — and could be due to a combination of mirror vignetting and the aging gradient. They recommended a detailed study but did not elevate it to a critical systematic.

By 2021, when the three signals were combined, the statistical significance exceeded 3 sigma. A dedicated analysis by the calibration working group, led by H. T. Diehl and A. Roodman, finally confirmed that the mirror age gradient was the common cause. The collaboration's response was to include a radial correction term in the photometric calibration for the final data release, but the correction was applied post hoc and could not fully undo the bias in earlier analyses.

The Broader Pattern: Similar Systematic Biases in Other Surveys

The DES mirror gradient is not an isolated case. Other surveys have encountered analogous hardware aging effects that were initially overlooked. The Sloan Digital Sky Survey (SDSS), for instance, saw a gradual decline in the sensitivity of its CCDs over its 20-year operational period, which introduced a time-dependent zero-point drift of roughly 0.01–0.02 magnitudes. That drift was corrected only after several years of data had been collected, and the correction required reprocessing the entire imaging catalog. The Pan-STARRS survey, which used a 1.8-meter telescope with a massive CCD camera, faced a similar issue with its filters: differential aging of the filter coatings caused a wavelength-dependent transmission loss that varied across the field of view, leading to a systematic redshift bias of about 0.01 for galaxies at high declinations. In both cases, the biases were ultimately traced to hardware components that were assumed to be stable but were not.

These examples highlight a recurring challenge in survey astronomy: the instruments are treated as static during the survey design phase, but they inevitably degrade. The standard practice of periodic flat-fielding and photometric calibration using standard stars is designed to correct for uniform changes in sensitivity, but it fails for spatially or temporally varying effects. The DES mirror gradient was particularly insidious because it combined both spatial and temporal variation, making it resistant to standard calibration techniques.

What the LSST Camera Can Learn From DES's Mistake

The Vera Rubin Observatory, home to the Legacy Survey of Space and Time (LSST), is now the flagship of ground-based optical astronomy. Its 8.4-meter primary mirror is coated with a protected silver layer, similar to the Blanco mirror but with a more durable overcoat. The LSST camera, with 189 CCDs and a 9.6-square-degree field of view, will image the entire southern sky every few nights for a decade. The precision required for LSST's weak-lensing science is a factor of ten better than DES: systematic biases in photometric redshifts must be controlled to 0.001 or less.

The LSST Dark Energy Science Collaboration (DESC) has already begun modeling the effects of mirror aging. Simulations by the DESC calibration working group show that a 1% reflectivity gradient across the LSST mirror, if uncorrected, would produce a redshift bias of roughly 0.005 — smaller than DES's 0.03 but still large enough to affect dark energy constraints. The collaboration plans to monitor the mirror's reflectivity monthly using a dedicated spectrophotometric calibration system, and to recoat the mirror every 18 months, a schedule that was built into the observatory's operations budget from the start.

Flat-fielding with a dome screen, which failed for DES, will be supplemented by on-sky spectrophotometric standards: a set of ~100 stars with precisely known spectra, observed every night, that can be used to map the spatial and temporal variation of the mirror's reflectivity. The LSST calibration pipeline will also incorporate a time-dependent model of the mirror aging, updated from the monitoring data. These measures are expensive — the calibration system alone cost roughly $10 million — but they are a direct response to the DES experience.

“DES taught us that the mirror is not a static component,” said LSST DESC member S. Schmidt in a 2024 workshop. “You have to treat it as a time-varying systematic, and you have to budget for its maintenance from day one.” The lesson extends beyond mirror coatings: any hardware component that degrades over time — filters, CCDs, even the telescope dome's paint — can introduce subtle gradients that accumulate into significant biases.

The Hidden Cost of Ignoring Hardware Ageing in Surveys

The financial cost of the DES mirror gradient is difficult to quantify, but the scientific cost is clearer. The bias in the w0–wa constraints, estimated at 0.02–0.05 in w0, is comparable to the statistical error from DES Year 3. In other words, the systematic error from a single uncosted hardware effect was roughly equal to the entire statistical power of the survey for dark energy. If the bias had been corrected, the combined constraints would have been tighter by a factor of roughly 1.5.

Future surveys — Euclid, the Nancy Grace Roman Space Telescope, and SPHEREx — all face similar challenges. Euclid's mirror is coated with a silver-based layer, and its calibration plan includes regular monitoring, but the spacecraft is in orbit and cannot be recoated. Roman's mirror is aluminum-coated, which is more stable, but its filters and detectors will degrade over time. SPHEREx, a near-infrared all-sky survey, uses a cryogenic telescope that avoids some aging issues but introduces others, such as ice buildup on optics.

The broader lesson for survey science is that systematic budgets must include hardware aging as a line item. The DES collaboration's internal review process, which required a 3-sigma detection before acting, was designed to avoid false positives, but it also delayed the identification of a real systematic. A more proactive approach — monitoring the mirror's reflectivity from the start, rather than relying on post-hoc diagnostics — would have caught the gradient earlier and at lower cost.

As of late 2024, the Blanco mirror has been recoated with a new silver layer, and the DES data set has been reprocessed with a radial correction. The corrected photometric redshifts reduce the bias to below 0.01, but the original Year 3 cosmology results, which were based on the uncorrected data, remain in the literature. The collaboration has not issued a formal correction, arguing that the bias was within the quoted systematic error budget. For a survey that cost roughly $50 million, that is a quiet resolution to a problem that was entirely preventable.