Did a burst of reionization in the dark ages leave a detectable imprint on cosmology?

A mild preference rather than a detection

The paper asks a deliberately narrow question: do current cosmological data prefer a short episode of extra ionization localized at some redshift after recombination, on top of the standard history in which early stars reionize the universe at z ∼ 10–20? Fitting a burst model to a broad data suite, the authors find Δχ² between about 6.9 and 8.3 relative to a matched ΛCDM control, which they translate into roughly a 2–2.5σ preference. That falls well short of evidence, and the authors say so directly.

The result is stable in amplitude but not in timing. Across all three Planck reprocessings tested, the extra optical depth sits near Δτ_burst ≈ 0.012, with a 95% interval of 0.001 < Δτ_burst < 0.022, and the total optical depth lands near τ ≈ 0.073. The preferred burst redshift, by contrast, ranges from z ∼ 45 to z ∼ 165 depending on which high-ℓ likelihood is used.

Why a burst adds optical depth but little polarization

The model treats the burst as a source term added to the standard recombination calculation, parameterized by an amplitude A_x, a central redshift z_x, and a fixed width σ_x = 0.1. The source switches off at z_x, after which ordinary recombination resumes. Because the injection rate peaks slightly before the cutoff, the free-electron fraction x_e(z) peaks 4–8% above z_x at the best-fit amplitudes.

Ionization does not fall back to the unperturbed value. Once the recombination timescale exceeds the Hubble time, the remaining electrons freeze out at a floor x_e^freeze = H/(α_rec n_H). Each burst therefore relaxes only down to that curve, leaving a residual free-electron fraction visible at low redshift. This is the mechanism by which an early, temporary episode still contributes to the integrated optical depth τ that the CMB measures.

The same mechanism explains why the burst is harder to see in polarization than a late source of equal τ. The authors state that a redshift-localized contribution adds more weakly to low-ℓ polarization signals than later sources do, because its signature is spread differently across multipoles. The optical depth is accumulated, but the low-ℓ EE bump that usually anchors τ is not proportionally enhanced.

Where the preference comes from depends on the data reduction

The authors test the low-ℓ EE likelihood alone first, since it covers exactly the multipole range (ℓ ≲ 30) most sensitive to when reionization happened and carries no high-ℓ foreground nuisance parameters. Using LoLLiPoP, the burst improves the fit by Δχ² ≈ 4.4 with Δτ_burst ≈ 0.010 and z_x essentially unconstrained. Using SRoll2, the improvement is only Δχ² ≈ 0.4.

Adding the full suite — Planck low-ℓ TT, ACT DR6, SPT-3G, DESI BAO, Pantheon+, and a high-ℓ Planck likelihood — raises the preference in every combination: Δχ² = 6.90 for SRoll2+plik, 8.33 for LoLLiPoP+CamSpec, and 7.29 for LoLLiPoP+HiLLiPoP. The authors caution that partitioning this total across datasets is misleading, because the parameters at the minimum already reflect a joint compromise.

Their drop-one analysis shows the structure of the preference. For PR4, removing LoLLiPoP collapses the preference to Δχ² = 0.30 (CamSpec) or 0.06 (HiLLiPoP), but the authors call this misleading: the ΛCDM fit responds by sliding z_reio up to ∼11.5 to reach τ ≈ 0.095, a value in tension with independent observations. Capping z_reio at a conservative value such as z_reio < 10 restores the burst, because the fit first saturates the low-redshift optical depth τ_low-z ≡ τ_eff − Δτ_burst and then adds a burst to bring the total back up. The model is therefore highly sensitive to any prior on τ_low-z and z_reio.

For PR3 the same caveat applies in reverse: removing SRoll2 drops the preference from 6.90 to 0.38, but there too the ΛCDM fit answers by raising τ to 0.098. The authors conclude that the preference is really a tension between the low-ℓ EE likelihood and the high-ℓ data, not a contribution attributable to any single dataset.

The burst epoch splits by high-ℓ likelihood

The three reprocessings agree on Δτ_burst and on total τ but disagree on when the burst happened. HiLLiPoP favors z ∼ 45, CamSpec z ∼ 100, and plik z ∼ 165. Marginalizing over the full posterior, the median z_x is 102 for CamSpec against 62 for HiLLiPoP.

A profile-likelihood analysis, in which z_x is held fixed at each grid value and every other parameter is re-minimized, confirms this is not a posterior artifact: there is a genuine difference in preference between HiLLiPoP and either CamSpec or plik. The authors attribute the discrepancy to which 30 < ℓ < 100 data are included. At low z_x the three agree, because the burst does not directly change that multipole range; at higher z_x the burst signal moves into it, and the likelihoods diverge.

The authors also check the burst against existing reconstructions of the ionization history. A prior bound of τ(15,30) < 0.015 at 95% from one reconstruction does not conflict with their fits, because their best-fit histories give τ(15,30) = 0.0001–0.0009 over that window, one to two orders of magnitude below the bound, and essentially all of their Δτ_burst is deposited above z = 30 where that reconstruction has no support.

What the neutrino shift does and does not fix

The paper connects the burst to the reported preference for a sub-zero effective neutrino mass in some cosmological fits. Using the same joint dataset and allowing an additional burst without requiring an enhanced τ, the authors find that floating Σm_ν measurably shifts its posterior toward less negative values relative to ΛCDM, but does not resolve the tension outright.

The authors also note a numerical coincidence without claiming it means anything: an independent ΛCDM-only determination of τ from acoustic-peak damping gives τ = 0.080 ± 0.012, roughly 1.8σ above the standard low-ℓ polarization value, and lands near their own best-fit τ_eff of 0.0713–0.0760 depending on the high-ℓ likelihood. They stress that the comparison is the appropriate one, since τ_eff is the quantity appearing in the damping of the acoustic peaks, whereas a τ inferred from the low-ℓ EE reionization bump constrains τ_low-z instead.

Which future probe tests which redshift

The forecast depends sharply on where z_x actually sits, because an early burst leaves a redshift-dependent signature rather than a fixed-shape optical-depth increase. Its characteristic angular scale shifts as z_x changes, moving toward the acoustic-peak region at higher redshift.

If such a source is present at 40 ≲ z ≲ 150, LiteBIRD should be able to detect it, and Simons Observatory should also constrain it. The authors are more specific about the division of labor: Simons Observatory's Large Aperture Telescope is not expected to improve on Planck's own EE sensitivity over ℓ ∼ 30–100, but its strength lies at higher ℓ, where the best-fit models predict non-zero residuals extending into the acoustic-peak regime and SO's forecast noise is substantially below Planck's. The Small Aperture Telescopes aim to measure EE and BB down to ℓ ∼ 50 and may probe high-redshift ionization sources directly.

One caveat the authors raise themselves: at least part of any high-ℓ signature is expected simply from the shift in standard cosmological parameters, particularly the amplitude A_s, caused by accommodating the additional τ while remaining consistent with the observed CMB, rather than being a wholly independent signature of the burst. They have not attempted to separate the two.

A complementary route is 21 cm observation of neutral hydrogen, which traces the ionization state of the intergalactic medium at a given redshift rather than an integrated quantity like τ, so it is not limited by how large or small the CMB-preferred burst amplitude happens to be. The authors note that computing the 21 cm global signal requires a heating model alongside the ionization history, but a localized increase in x_e(z) would produce a redshift-localized modification of the spin temperature, potentially breaking the z_x degeneracy that CMB-only analyses leave open.

What to watch in the next data

The paper's own framing is that the data are far from definitive. The preference is model-agnostic: the authors leave the source unidentified and the burst redshift free, motivated by a not-quite-primordial black hole scenario but not tied to it. That breadth is a strength for constraining exotic high-redshift sources, and a limitation for anyone wanting a claim about a specific astrophysical population.

The practical priority follows from the split between robust and fragile quantities. Δτ_burst ≈ 0.012 and τ ≈ 0.073 survive every reprocessing tested; the epoch does not. A future measurement that tightens the low-ℓ EE region around ℓ ∼ 30–100, or a 21 cm detection of a redshift-localized ionization feature, would discriminate among the z_x values the current data allow. Until then, the honest summary is a 2–2.5σ hint whose amplitude is stable and whose timing is not.

Sources used in this answer