Introduction
One of the most challenging problems in modern cosmology is the tension between different estimates of the present expansion rate H 0 H_0 H 0 . This Hubble tension describes the discrepancy between the expansion rate of the Universe inferred through local distance measurements (e.g. H 0 S H 0 E S = 73.17 ± 0.86 H_0^{\mathrm{SH0ES}} = 73.17 \pm0.86 H 0 SH0ES = 73.17 ± 0.86 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 from SH0ES; Breval et al. 2024) and the predicted expansion rate based on calibrating the parameters of the Lambda cold dark matter (Λ \Lambda Λ CDM) standard cosmological paradigm (Efstathiou, Sutherland & Maddox 1990; Ostriker & Steinhardt 1995) using the Cosmic Microwave Background (CMB), which gives a far lower value (e.g. SPT-3G + ACT DR6 + Planck obtained that H 0 S P A = 67.19 ± 0.38 H_0^{\mathrm{SPA}} = 67.19 \pm0.38 H 0 SPA = 67.19 ± 0.38 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 ; SPT-3G Collaboration [107 ] ). These results are in a 6.4 σ 6.4\sigma 6.4 σ tension with each other. An even higher tension has recently been claimed by the Local Distance Network based on combining several popular local determinations and taking into account their covariance (H0DN Collaboration 2026). However, this does not properly model selection effects or distance priors (Desmond et al. 2026; Stiskalek et al. 2026). The three possible explanations for the tension are unaccounted-for systematics in the distance ladder, statistical inadequacy of the inferences (e.g. Becker et al. 2015; Feeney et al. 2018; Desmond et al. 2026; Percival 2026; Stiskalek et al. 2026; Stiskalek & Desmond 2026), and new physics beyond the standard Λ \Lambda Λ CDM model (e.g. Valentino et al. 2025, and references therein). New physics could primarily affect the Universe prior to recombination or have little impact until more recent epochs, leading to a distinction between early-time and late-time solutions.
A useful expansion probe is provided by supernovae (SNe) of Type Ia (SNIa). These thermonuclear explosions occur when a carbon–oxygen white dwarf in a binary system reaches a mass close to 1.2 M ⊙ 1.2\,M_{\odot} 1.2 M ⊙ , as predicted by models like the dynamically driven double-degenerate double-detonation (D6 ^{6} 6 ) scenario (Shen et al. 2018a,b; Bauer et al. 2021). Since SNIa occur at nearly the same mass, they are good standardisable candles, i.e., the peak absolute magnitude M B M_B M B in the B B B band can be estimated independently of the redshift z z z (Phillips 1993; Tripp 1998; Phillips et al. 1999). Due to the known H 0 H_0 H 0 –M B M_B M B degeneracy (Cai & Wang 2026; Wang & Wang 2026), SNIa require an absolute calibration to measure H 0 H_0 H 0 . Since SNIa are very luminous (M B ∼ − 19.3 M_B \sim-19.3 M B ∼ − 19.3 ), they can be detected out to z ≈ 2 z \approx2 z ≈ 2 (Scolnic et al. 2022). However, it is difficult to access all of this redshift range using just one survey. Modern SNIa catalogues like Pantheon+ (Scolnic et al. 2022), Union3 (Rubin et al. 2025), and Unite (Camilleri et al. 2026) typically combine different surveys. This may create problems due to the use of different filters, selection functions, and relative calibration offsets (Betoule et al. 2014; Scolnic et al. 2014, 2015; Efstathiou 2025). A homogeneous catalogue with only one survey and one selection function would be far preferable for cosmological inference.
This is provided by the Dark Energy Survey (DES; DES Collaboration 2016) through its Dovekie catalogue (Popovic et al. 2026). This single-instrument dataset with a single filter system is typically complemented by legacy SNIa from older surveys to improve the leverage on the shape of the expansion history at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 , where there are very few SNIa in DES (though it has SNIa down to z = 0.073 z = 0.073 z = 0.073 ; DES Collaboration 2024). However, including the legacy SNIa risks the aforementioned problems with combining surveys, and assessing this legacy sample relative to the DES sample is a main aim of this paper. Our baseline analysis involves only SNIa in the DES data, which we label simply DES. We also consider an extension including the legacy SNIa (DES + legacy), which corresponds to what other authors typically mean by “DES” (e.g. DESI Collaboration 2025). This distinguishes our work from similar studies (most notably Popovic & Sullivan 2026), which generally do not distinguish between legacy and DES SNIa.
Our main motivation is the presently open discussion on whether the legacy SNIa have been correctly recalibrated onto the DES photometric filter system, something which cannot be directly verified through further observations given the transient nature of SNIa (Efstathiou 2025; Huang et al. 2025; Ormondroyd et al. 2026). Since the legacy SNIa lie in z ∈ [ 0.025 , 0.087 ] z \in[0.025, 0.087] z ∈ [ 0.025 , 0.087 ] , they provide a critical lever on the value of H 0 H_0 H 0 , making it crucial that they be unbiased. This motivates further analyses where we calculate the possible offset between the legacy SNIa and the main DES sample. This is also a crucial issue for the evidence for evolving dark energy, which is largely driven by the legacy SNIa; allowing an offset provides a better fit to the data in a Bayesian sense than allowing a time-variation in the equation of state of dark energy (Ormondroyd et al. 2026). In § 4.5 , we explore the extent to which the legacy SNIa affect the evidence for evolving dark energy.
There are two main routes to determine the absolute SNIa magnitude M S N I a M_{\mathrm{SNIa}} M SNIa . The first route is through the cosmic distance ladder, usually with three rungs. The first rung involves geometry to calibrate distance indicators like Cepheid variable stars (Leavitt & Pickering 1912; Riess et al. 2022; Breuval et al. 2024), the tip of the red giant branch (TRGB; Lee et al. 1993; Madore & Freedman 1995; Freedman et al. 2019; Hoyt et al. 2024), or the J-region Asymptotic Giant Branch (JAGB; Madore & Freedman 2020; Lee et al. 2025). The second rung is where the M B M_B M B calibration happens using SNIa host galaxies near enough to contain a first rung indicator. The third rung contains SNIa in the Hubble flow at z ≳ 0.023 z \gtrsim0.023 z ≳ 0.023 , where peculiar velocities are not important in Λ \Lambda Λ CDM, allowing a determination of H 0 H_0 H 0 from the slope of the Hubble diagram. Using this approach, the SH0ES collaboration derived a value of M B = − 19.253 ± 0.027 M_B = -19.253 \pm0.027 M B = − 19.253 ± 0.027 (Riess et al. 2022; Scolnic et al. 2022).
The second route to M S N I a M_{\mathrm{SNIa}} M SNIa is through the inverse distance ladder (IDL). This involves anchoring M S N I a M_{\mathrm{SNIa}} M SNIa using our understanding of the early Universe to set the baryon acoustic oscillation (BAO) comoving sound horizon r d r_d r d at the end of the baryon drag epoch (z d ≈ 1060 z_d \approx1060 z d ≈ 1060 ; Hu & Sugiyama 1996; Eisenstein & Hu 1998) and then propagating this to intermediate redshifts using the BAO evident from galaxy clustering, which nowadays is typically obtained from the Dark Energy Spectroscopic Instrument (DESI) Data Release 2 (DR2; [33 ] ). One common assumption which we follow is to adopt the Planck constraint of r d = 147.05 ± 0.30 r_d = 147.05 \pm0.30 r d = 147.05 ± 0.30 comoving Mpc (cMpc) from Planck Collaboration VI [100 ] (hereafter Planck [100 ] ). This explicitly adopts Λ \Lambda Λ CDM at early times, so the IDL approach implicitly assumes that the Hubble tension, if due to new physics, must be solved at late times. Under this assumption, M S N I a ≃ − 19.30 M_{\mathrm{SNIa}} \simeq-19.30 M SNIa ≃ − 19.30 to − 19.40 -19.40 − 19.40 (Camilleri et al. 2025; Popovic & Sullivan 2026), which is significantly lower than suggested by the local distance ladder. This is because a lower H 0 H_0 H 0 value requires SNIa to be more distant at a given z z z , which for given apparent magnitude requires the SNIa to be brighter. The discrepancy with the locally determined M B M_B M B leads to the usual conclusion that a successful solution to the Hubble tension requires a smaller r d r_d r d , which is possible only through new pre-recombination physics (Cai & Wang 2026; Sabogal et al. 2026; Tiwari et al. 2026).
Although the expansion history is nowadays mainly probed through the distance–redshift relation, it is possible to bypass distances altogether by instead focusing on the lookback time. The elapsed time between any two redshifts can be found using the cosmic chronometer (CC) technique, which involves comparing massive passive elliptical galaxies at those redshifts to determine the relative age (Jimenez & Loeb 2002; Moresco et al. 2012; Moresco 2015; Moresco et al. 2016, 2022; Moresco 2024). Since we can relate distance and lookback time using the speed of light, we can use CC datasets to calibrate M B M_B M B without assuming any value for r d r_d r d , thereby obtaining an estimate for H 0 H_0 H 0 (Cogato et al. 2024; Guo et al. 2025).
Early-time solutions to the Hubble tension seek to resolve it by decreasing r d r_d r d , which requires modifying the redshift dependence of the sound horizon and/or the redshift of recombination (Mirpoorian et al. 2025). Since BAO measurements are most sensitive to the product H 0 r d H_0r_d H 0 r d , reducing r d r_d r d allows for an increased H 0 H_0 H 0 . However, these solutions face at least seven problems (Vagnozzi 2023). One of these is that they would imply a significantly reduced age of the Universe (Vagnozzi et al. 2022). This prediction is in significant tension with the observed ages of the oldest Galactic stars and globular clusters (Montalbán et al. 2021; Valcin et al. 2020, 2021; Cimatti & Moresco 2023; Souza et al. 2024; Valcin et al. 2025, 2026; Banik et al. 2026b). These features cannot be explained through a simple change to r d r_d r d . Moreover, currently known early-time solutions struggle with Big Bang nucleosynthesis (BBN) due to their need for a generically higher baryon density to match the CMB anisotropies, which affects primordial light element abundances (Giovanetti 2026; Launders et al. 2026; Poulin et al. 2026). More generally, the success of Λ \Lambda Λ CDM in fitting the CMB anisotropies by now limits the scope for substantial modifications prior to recombination ([107 ] (Planck 2020; Calabrese et al. 2025)), including in models where electron density fluctuations cause inhomogeneous recombination (Lynch et al. 2026).
Late-time solutions to the Hubble tension avoid these problems by assuming standard pre-recombination physics, but such models have their own problems at z ≲ 1 z \lesssim1 z ≲ 1 (Krishnan et al. 2021a). In particular, Bansal & Huterer 2026 find that a model modifying only the late-time expansion history cannot simultaneously fit H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES , the DESI BAO data, and relative SNIa distances. A fit is possible only if M S N I a M_{\mathrm{SNIa}} M SNIa takes a sharp step at very low redshift (z ≲ 0.01 z \lesssim0.01 z ≲ 0.01 ) or the distance-duality relation is broken. Both escapes are trivial since they decouple the calibrated SNIa from the remaining data. Importantly, their analysis rests on the Pantheon+ catalogue. A step in M B M_B M B at low redshift is degenerate with precisely the relative calibration offsets between surveys described above, given the difficulty of covering a very wide redshift range using the same survey. Their most successful late-time magnitude step model has a transition in the redshift range where a heterogeneous compilation is least able to separate new physics from filter-related changes to the zero point. This motivates the test we present here. We note that astrophysically it is difficult to understand why SNIa might abruptly change in intrinsic brightness, with terrestrial data placing strong constraints on models with a sharp change in G G G in the geologically recent past (Banik, Desmond & Samaras 2025). Such models also face difficulties with the mass–radius relation of red dwarf stars due to their long thermal timescales (Sakstein & Desmond 2026).
Our main goal is to assess if the IDL can accommodate a late-time solution to the Hubble tension, i.e., if the assumption of standard pre-recombination physics is compatible with the high H 0 H_0 H 0 required to match various local constraints (Di Valentino et al. 2025; H0DN Collaboration 2026). The observational data we use for this purpose are described in § 2 . We then explain the IDL, our Bayesian model comparison of late-time solutions to the Hubble tension, and consistency tests (§ 3 ). We go on to present our results (§ 4 ) and discuss their implications (§ 5 ). We conclude in § 6 . In Appendix A, we present the Alcock–Paczynski (AP) consistency test between DESI DR2 BAO and DES-Dovekie SNIa. Appendix B reports how the inferred cosmological parameters depend on the polynomial order. Appendix C contains all the posteriors from our analyses. In Appendix D, we use ΛCDM mocks to better understand how our analysis is affected by the limited DES coverage at low redshift.
Observational data
DES-Dovekie SNIa
The DES-Dovekie compilation of SNe provides 1820 spectroscopically and photometrically classified events with their full statistical plus systematic covariance (Popovic et al. 2026). The intrinsic scatter contribution to this covariance is calibrated so that ΛCDM yields an acceptable fit to the Hubble diagram. This is because we currently lack sufficiently large samples of SNIa with redshift-independent distances from which the intrinsic scatter can be inferred, as is usually done with other indicators like Cepheids. The total SNIa uncertainties are therefore scaled to a reference cosmology, which means that an acceptable χ 2 \chi^2 χ 2 under ΛCDM in SNIa-dominated analyses like ours does not by itself demonstrate that ΛCDM is an adequate description of the data. We consider two SN samples, which we label DES and DES + legacy throughout this paper. Our baseline is the DES sample, which uses the 1623 DES SNe alone (IDSURVEY = 10), leading to a redshift range of z ∈ [ 0.073 , 1.14 ] z \in[0.073, 1.14] z ∈ [ 0.073 , 1.14 ] – though in practice there is limited constraining power at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 (see figures 3 and 4 of DES Collaboration 2024). We select the SNe with magnitude uncertainty σ m < 1 \sigma_m < 1 σ m < 1 , removing 82 SNe as these are likely contaminants (e.g. other types of SN) given they are down-weighted by the Bayesian Estimation Applied to Multiple Species (BEAMS) analysis (Kunz et al. 2007). They carry ≈ 10 − 5 \approx10^{-5} ≈ 1 0 − 5 of the total statistical weight, so removing them does not significantly change the results.
The DES + legacy sample adds an external low-redshift anchor from legacy surveys, which contribute a further 197 spectroscopically confirmed SNIa with z ∈ [ 0.025 , 0.087 ] z \in[0.025, 0.087] z ∈ [ 0.025 , 0.087 ] , lowering the redshift floor from z ≈ 0.073 → 0.025 z \approx0.073 \to0.025 z ≈ 0.073 → 0.025 . It is important to stress that both DES and legacy SNIa are uncalibrated as there is no independent determination of their absolute magnitude in the DES filters. We denote this absolute magnitude M D E S M_{\rm DES} M DES rather than M B M_B M B to differentiate it from the rest-frame Johnson B B B -band in which the SH0ES M B M_B M B is defined. The reported distance modulus μ ( z ) = m ( z ) − M D E S \mu(z) = m(z)-M_{\rm DES} μ ( z ) = m ( z ) − M DES assumes a fiducial M D E S = − 19.327 M_{\rm DES} = -19.327 M DES = − 19.327 , but this can be arbitrarily varied. We infer M D E S M_{\rm DES} M DES jointly with the other parameters, which implicitly imposes a calibration via the IDL. The predicted μ = 25 + 5 log 10 D L / M p c \mu= 25 + 5\log_{10}D_L/{\rm Mpc} μ = 25 + 5 log 10 D L / Mpc , with the luminosity distance D L D_L D L found as follows:
D L = ( 1 + z h e l ) D M ( z H D ) , (1) D_L = (1+z_{\rm hel})D_M(z_{\rm HD}), \tag*{(1)} D L = ( 1 + z hel ) D M ( z HD ) , (1)
D M ( z ) = c ∫ 0 z d z ~ H ( z ~ ) , (2) D_M(z) = c\int_0^z \frac{\mathrm{d}\tilde{z}}{H(\tilde{z})}, \tag*{(2)} D M ( z ) = c ∫ 0 z H ( z ~ ) d z ~ , (2)
where D M D_M D M is the comoving distance, z h e l z_{\rm hel} z hel is the heliocentric redshift of each SNIa, and z H D z_{\rm HD} z HD is its “Hubble diagram” redshift obtained after corrections for peculiar velocities.
To study possible differences between the main DES sample and the legacy SNIa, we also consider adding a systematic offset of Δ M l e g a c y \Delta M_{\rm legacy} Δ M legacy to the absolute magnitude of the legacy SNIa. This measures possible systematics when correcting the legacy SNIa for survey completeness effects or converting the observed legacy magnitudes to what would have been measured in the DES filters – despite the lack of direct measurements in those filters. We label this as DES + legacy (Δ M l e g a c y \Delta M_{\rm legacy} Δ M legacy ) and explain it further in § 3.6 .
DESI DR2 BAO
The BAO is a statistical standard ruler arising from matter fluctuations in the primordial Universe (Eisenstein & Hu 1998). We consider the latest DESI data release (DESI DR2; DESI Collaboration 2025). They report six measurements of D M / r d D_M/r_d D M / r d , six of D H / r d D_H/r_d D H / r d at the same redshifts, and one of D V / r d D_V/r_d D V / r d at a lower redshift:
D H ( z ) = c H ( z ) , (3) D_H(z) = \frac{c}{H(z)}, \tag*{(3)} D H ( z ) = H ( z ) c , (3)
D V ( z ) = [ z D M 2 ( z ) D H ( z ) ] 1 / 3 , (4) D_V(z) = \left[zD_M^2(z)D_H(z)\right]^{1/3}, \tag*{(4)} D V ( z ) = [ z D M 2 ( z ) D H ( z ) ] 1/3 , (4)
r d = ∫ z d ∞ c s ( z ~ ) H ( z ~ ) d z ~ , (5) r_d = \int_{z_d}^{\infty}\frac{c_s(\tilde{z})}{H(\tilde{z})}\,\mathrm{d}\tilde{z}, \tag*{(5)} r d = ∫ z d ∞ H ( z ~ ) c s ( z ~ ) d z ~ , (5)
where c s c_s c s is the sound speed in the baryon–photon plasma. We use the full DESI covariance matrix, which has a block-diagonal form because measurements at different redshifts use different galaxies, making them independent.
CC
We use 15 CC measurements to constrain the expansion history without assumptions about M D E S M_{\rm DES} M DES or r d r_d r d (Moresco et al. 2012; Moresco 2015; Moresco et al. 2016). CC measurements constrain the Hubble–Lemaître parameter via the following relation
H ( z ) = − z ˙ 1 + z , (6) H(z) = -\frac{\dot{z}}{1+z}, \tag*{(6)} H ( z ) = − 1 + z z ˙ , (6)
where an overdot denotes a time derivative. z ˙ \dot{z} z ˙ is measured through the differential age evolution of early-type galaxies (Jimenez & Loeb 2002). We use the full covariance reported in Moresco et al. (2020). The errors quoted in CC compilations like Moresco et al. (2022) omit the stellar-population modelling systematic. We find that this is crucial as it can significantly change the cosmological parameter posteriors. In particular, the error on r d r_d r d roughly doubles when considering the full covariance rather than only the diagonal terms. Since the covariance with other CC measurements is not known, we do not use the other 18 of the 33 available CC measurements (Moresco et al. 2020; Wang et al. 2026).
We use the CC data for an extended analysis in which a direct comparison between differential age H ( z ) H(z) H ( z ) measurements and D H / r d D_H/r_d D H / r d BAO measurements can independently calibrate r d r_d r d (§ 4.3 ). This affords an IDL independent of the CMB.
Methodology
Model setup
The cosmic expansion history a ( t ) a(t) a ( t ) can be approximated as a Taylor series around the present epoch, where a a a is the cosmic scale factor normalised to unity today and t t t is the time since the Big Bang. The use of a Taylor series minimises assumptions about the expansion history, making the results model-independent. The series expansion is commonly based on redshift, but since we consider data out to z = 2.33 z = 2.33 z = 2.33 , we instead expand in the lookback scale factor a ~ = 1 − a \tilde{a} = 1 - a a ~ = 1 − a , limiting the range to 0–1. We therefore use the following series expansion:
E ( a ~ ) = 1 + ∑ k = 1 k max C k a ~ k , a ~ = z 1 + z = 1 − a , (7) E(\tilde{a}) = 1 + \sum_{k=1}^{k_{\max}} C_k \tilde{a}^{k}, \qquad\tilde{a} = \frac{z}{1+z} = 1-a, \tag*{(7)} E ( a ~ ) = 1 + k = 1 ∑ k m a x C k a ~ k , a ~ = 1 + z z = 1 − a , (7)
where E ( a ~ ) ≡ H ( a ~ ) / H 0 E(\tilde{a}) \equiv H(\tilde{a})/H_0 E ( a ~ ) ≡ H ( a ~ ) / H 0 is the normalised expansion rate and k max k_{\max} k m a x is the order of the polynomial with coefficients C k C_k C k . Using the lookback scale factor a ~ \tilde{a} a ~ instead of z z z leads to a narrower range because a ~ < 1 ∀ z \tilde{a} < 1\ \forall z a ~ < 1 ∀ z . This is important here because DESI reaches out to z = 2.33 z = 2.33 z = 2.33 , which is too large for a convergent Taylor series since it converges only for ∣ z ∣ < 1 |z| < 1 ∣ z ∣ < 1 (Cattoën & Visser 2007). This means that expanding in powers of z z z and fitting the model with data at z > 1 z > 1 z > 1 is not a cosmographic expansion but rather a polynomial fit, which is clearly model-dependent (e.g. with fixed asymptotic behaviour). In this case, the parameters H 0 H_0 H 0 , q 0 q_0 q 0 , j 0 j_0 j 0 , . . . do not represent the standard cosmographic parameters but rather are parameters of the polynomial model. This problem does not arise in our setup because a ~ ≤ 0.70 \tilde{a} \le0.70 a ~ ≤ 0.70 . Since the z ≪ 1 z \ll1 z ≪ 1 correspondence is a ~ = z − z 2 + O ( z 3 ) \tilde{a} = z-z^2+O(z^3) a ~ = z − z 2 + O ( z 3 ) , the deceleration parameter can be found from q 0 = C 1 − 1 q_0 = C_1-1 q 0 = C 1 − 1 .
A crucial aspect for a cosmographic analysis is the prior on the polynomial coefficients C k C_k C k , but it is difficult to estimate an appropriate prior range. Instead of directly bounding each C k C_k C k separately, we set a prior on the shape of the expansion history by requiring the following:
0 ≤ E ( a ~ ) ≤ 2 + 7 a ~ , 0 ≤ a ~ ≤ a ~ max , (8) 0 \le E(\tilde{a}) \le2 + 7\tilde{a}, \qquad0 \le\tilde{a} \le\tilde{a}_{\max}, \tag*{(8)} 0 ≤ E ( a ~ ) ≤ 2 + 7 a ~ , 0 ≤ a ~ ≤ a ~ m a x , (8)
where a ~ max = 0.70 \tilde{a}_{\max} = 0.70 a ~ m a x = 0.70 is around the largest lookback scale factor considered by the data (the DESI DR2 BAO point at z = 2.33 z = 2.33 z = 2.33 ). The lower limit imposes an expanding universe. The upper limit is twice our Λ \Lambda Λ CDM-inspired linear model of 1 + 3.5 a ~ 1 + 3.5\tilde{a} 1 + 3.5 a ~ , which approximately joins the two ends of the flat Λ \Lambda Λ CDM E ( a ~ ) E(\tilde{a}) E ( a ~ ) relation over this range with the SPA measurements (Figure 1 ). This means the upper limit lies between ≈ 2 × \approx2\times ≈ 2 × and 3.3 × 3.3\times 3.3 × the Λ \Lambda Λ CDM expansion rate at every a ~ \tilde{a} a ~ in this range. The broad allowed range of E E E permits any plausible late-time departure from Λ \Lambda Λ CDM. We sample our E E E prior (Eq. (8) ) using a uniform prior on E E E over the allowed range ( 0 (0 ( 0 to 2 + 7 a ~ ) 2 + 7\tilde{a}) 2 + 7 a ~ ) at each of k max k_{\max} k m a x control points, which we place at the Chebyshev–Lobatto nodes a ~ j = a ~ max [ 1 − cos ( j π / k max ) ] / 2 \tilde{a}_j = \tilde{a}_{\max}[1-\cos{(j\pi/k_{\max})}]/2 a ~ j = a ~ m a x [ 1 − cos ( j π / k m a x ) ] /2 , where j = 1 … k max j = 1\ldots k_{\max} j = 1 … k m a x . Given also that E ≡ 1 E \equiv1 E ≡ 1 today, this defines a unique polynomial of order k max k_{\max} k m a x through a fixed linear transformation dependent only on k max k_{\max} k m a x . We then reject the polynomial if it goes outside our prior limits anywhere in the interval 0 ≤ a ~ ≤ a ~ max 0 \le\tilde{a} \le\tilde{a}_{\max} 0 ≤ a ~ ≤ a ~ m a x . This a joint prior that is uniform in the C k C_k C k over the range allowed by Eq. (8) , although the marginal prior on each C k C_k C k is not uniform (§ 4.9 ). We show in § 4.9 that this condition does not truncate our posteriors, which give relatively narrow ranges across which the priors are essentially flat.
Figure 1. Illustration of our prior on the shape of the expansion history E ( a ~ ) ≡ H ( a ~ ) / H 0 E(\tilde{a}) \equiv H(\tilde{a})/H_0 E ( a ~ ) ≡ H ( a ~ ) / H 0 (Eq. (7) ). The thick orange line shows flat Λ \Lambda Λ CDM with Ω m = 0.3175 \Omega_{\mathrm{m}} = 0.3175 Ω m = 0.3175 , which has negligible errors for the purpose of this plot (SPT-3G Collaboration [107 ] ). The dashed blue line shows the best quartic fit to the Λ \Lambda Λ CDM prediction. The dotted black line at E = 1 + 3.5 a ~ E = 1 + 3.5\tilde{a} E = 1 + 3.5 a ~ approximately connects the fixed point at a ~ = 0 \tilde{a} = 0 a ~ = 0 to the Λ \Lambda Λ CDM prediction at a ~ = 0.7 \tilde{a} = 0.7 a ~ = 0.7 . The solid black line at twice this level ( 2 + 7 a ~ ) (2 + 7\tilde{a}) ( 2 + 7 a ~ ) shows our prior upper limit on E E E , above which the grey shading indicates the excluded region. We also require E ≥ 0 E \ge0 E ≥ 0 to enforce an expanding universe. Our limits apply only to the range 0 ≤ a ~ ≤ 0.70 0 \le\tilde{a} \le0.70 0 ≤ a ~ ≤ 0.70 as we do not use data outside this range. For the AP test in Appendix A, we further restrict to 0 ≤ a ~ ≤ 0.53 0 \le\tilde{a} \le0.53 0 ≤ a ~ ≤ 0.53 for the same reason. We adopt a uniform prior on E E E at k max k_{\max} k m a x control points in a ~ \tilde{a} a ~ and then check that the unique polynomial through these points with E = 1 E = 1 E = 1 today remains within the above bounds.
Table 1 shows the prior on every parameter considered in this paper. The best-fitting parameters can be found using the least-squares approach, where χ 2 \chi^2 χ 2 is assumed to vary quadratically with the model parameters near its minimum, corresponding to a locally Gaussian approximation to the likelihood. This approach does not require us to explicitly set prior ranges. However, we do not use least squares for our baseline results because it is preferable to relax the Gaussian assumption. We therefore sample the posterior with the No-U-Turn Sampler (NUTS) as implemented in numpyro (Hoffman & Gelman 2014; Bingham et al. 2019; Phan et al. 2019). We ensure convergence with the Gelman-Rubin diagnostic R − 1 < 0.02 R - 1 < 0.02 R − 1 < 0.02 (Gelman & Rubin 1992; Brooks & Gelman 1998). Results obtained in this way are similar to those using least squares, further demonstrating that our priors are broad enough (§ 4.9 ). Since we consider a large number of analysis variants, we use least squares for many of our variants to reduce the computational cost compared to NUTS, given the numerically similar results.
Table 1. Priors on the parameters considered in this paper. The shape of the expansion history is required to lie within the bounds specified in Eq. (8) , with E E E assigned a uniform prior over its allowed range at each of the k max k_{\max} k m a x control points in a ~ \tilde{a} a ~ (§ 3.1 ). This is followed by a check that the resulting polynomial (Eq. (7) ) remains within bounds for all 0 ≤ a ~ ≤ 0.70 0 \le\tilde{a} \le0.70 0 ≤ a ~ ≤ 0.70 (see the text). In analyses with a more rigorous CMB comparison, ω b = Ω b h 2 \omega_b = \Omega_b h^2 ω b = Ω b h 2 is the physical baryon density and Ω b c \Omega_{\mathrm{bc}} Ω bc the present density parameter of baryons + CDM, from which we derive r d r_d r d (§ 3.3 ).
We also fit the flat Λ \Lambda Λ CDM model for comparison. This involves replacing the cosmographic expansion of Eq. (7) with the following:
H ( z ) = H 0 Ω m ( 1 + z ) 3 + ( 1 − Ω m ) (9) H(z) = H_0\sqrt{\Omega_{\mathrm{m}}(1+z)^3+(1-\Omega_{\mathrm{m}})} \tag*{(9)} H ( z ) = H 0 Ω m ( 1 + z ) 3 + ( 1 − Ω m ) (9)
where Ω m \Omega_{\mathrm{m}} Ω m is the matter density parameter. All other aspects of the analysis remain unchanged, but the free parameters are now Ω m \Omega_{\mathrm{m}} Ω m , H 0 H_0 H 0 , M D E S M_{\mathrm{DES}} M DES , and r d r_{\mathrm{d}} r d . The deceleration parameter is the derived quantity q 0 = 3 2 Ω m − 1 q_0 = \frac{3}{2}\Omega_{\mathrm{m}} - 1 q 0 = 2 3 Ω m − 1 . Eq. (9) has much less flexibility in the shape of H ( z ) H(z) H ( z ) , allowing us to check how conclusions like the predicted H 0 H_0 H 0 and any calibration offset in the legacy SNIa (§ 3.6 ) depend on the extra flexibility afforded by our model-agnostic cosmographic fit (Eq. (7) ). If there is a departure from flat Λ \Lambda Λ CDM behaviour at low redshift, Eq. (9) is too restrictive, leading to a divergence between results assuming it and results with a more flexible parametrisation.
The cosmographic expansion in Eq. (7) or the flat Λ \Lambda Λ CDM version (Eq. (9) ) is the only description of the background in our analysis. It enters every likelihood: the CC directly through H ( z ) H(z) H ( z ) , the SNIa through the luminosity distance D L D_{\mathrm{L}} D L (Eq. (1) ), and the BAO through D M D_{\mathrm{M}} D M , D H D_{\mathrm{H}} D H , and D V D_{\mathrm{V}} D V , which are the integrals and ratios of the same H ( z ) H(z) H ( z ) . The C k C_k C k coefficients (or Ω m \Omega_{\mathrm{m}} Ω m for Λ \Lambda Λ CDM), M D E S M_{\mathrm{DES}} M DES , and r d r_{\mathrm{d}} r d are sampled jointly, so the three types of observation are coupled through a common expansion history and cannot drift apart. Since we consider both luminosity distances to SNIa and angular diameter distances for the BAO, our results assume the Etherington distance-duality relation D L = ( 1 + z ) D M D_{\mathrm{L}} = (1+z)D_{\mathrm{M}} D L = ( 1 + z ) D M ([39 ] ).
Throughout the paper, our baseline truncates Eq. (7) at k max = 4 k_{\max}=4 k m a x = 4 , the order with the highest Bayesian evidence for DES. We test how the results depend on this choice by repeating the analysis for k max = 2 k_{\max}=2 k m a x = 2 –6 6 6 and computing the Bayesian evidence Z ( k max ) Z(k_{\max}) Z ( k m a x ) of each order. We use the nested sampling algorithm dynesty (Skilling 2004, 2006; Speagle 2020; Koposov et al. 2024) with random slice sampling, 1000 live points and a stopping criterion of Δ ln Z = 0.1 \Delta\ln Z=0.1 Δ ln Z = 0.1 , under the same prior on the expansion history as the rest of this paper. Assigning every order the same prior probability, the weighted posterior probability of order k max k_{\max} k m a x is as follows:
w ( k max ) = Z ( k max ) ∑ k = 2 6 Z ( k ) (10) w(k_{\max})=\frac{Z(k_{\max})}{\sum_{k=2}^{6}Z(k)} \tag*{(10)} w ( k m a x ) = ∑ k = 2 6 Z ( k ) Z ( k m a x ) (10)
We present the results using different k max k_{\max} k m a x as a robustness test of our baseline in § 4.9 , where we then average the results for different k max k_{\max} k m a x using the above weights to get the Bayesian model average (BMA; see Appendix B).
The basic idea behind the IDL is to calibrate datasets assuming the validity of Λ \Lambda Λ CDM prior to recombination. In our setup, the total χ 2 \chi^2 χ 2 is therefore
χ t o t a l 2 = χ S N I a 2 + χ B A O 2 + { χ C M B 2 (configuration A) , χ C C 2 (configuration B) . (11) \chi^2_{\mathrm{total}}=\chi^2_{\mathrm{SNIa}}+\chi^2_{\mathrm{BAO}}+
\begin{cases}
\chi^2_{\mathrm{CMB}} & \text{(configuration A)},\\
\chi^2_{\mathrm{CC}} & \text{(configuration B)}.
\end{cases} \tag*{(11)} χ total 2 = χ SNIa 2 + χ BAO 2 + { χ CMB 2 χ CC 2 (configuration A) , (configuration B) . (11)
where χ S N I a 2 \chi^2_{\mathrm{SNIa}} χ SNIa 2 is the contribution of SNIa from DES (possibly supplemented by the legacy SNIa), χ B A O 2 \chi^2_{\mathrm{BAO}} χ BAO 2 is the contribution of DESI DR2 BAO measurements, χ C M B 2 \chi^2_{\mathrm{CMB}} χ CMB 2 is the CMB constraint imposed via a Gaussian prior of r d = 147.05 ± 0.30 r_{\mathrm{d}}=147.05\pm0.30 r d = 147.05 ± 0.30 Mpc ([100 ] ), and χ C C 2 \chi^2_{\mathrm{CC}} χ CC 2 is the CC contribution. We use either the CMB constraint on r d r_{\mathrm{d}} r d assuming Λ \Lambda Λ CDM prior to recombination, or we use CC and impose only a broad uninformative prior on r d r_{\mathrm{d}} r d . We break the well-known H 0 H_0 H 0 –M D E S M_{\mathrm{DES}} M DES degeneracy in the SNIa by requiring consistency with BAO, but the BAO data itself suffers from a D M H V − r d D_{\mathrm{M}}H_{\mathrm{V}}^{-r_{\mathrm{d}}} D M H V − r d degeneracy. This is because the observables that the BAO angular scale and redshift depth, which measure distances relative to r d r_{\mathrm{d}} r d .
We obtain an independent calibration of r d r_{\mathrm{d}} r d using two different approaches:
Baseline (configuration A): Assuming standard pre-recombination physics, which requires r d = 147.05 ± 0.30 r_{\mathrm{d}}=147.05\pm0.30 r d = 147.05 ± 0.30 cMpc (TT, TE, EE + low-ℓ \ell ℓ modes in the CMB; [100 ] ); or
Extension (configuration B): Comparing CC measurements of H ( z ) H(z) H ( z ) to BAO measurements of D H / r d D_{\mathrm{H}}/r_{\mathrm{d}} D H / r d , with a wide uniform r d r_{\mathrm{d}} r d prior covering 120–175 cMpc. This approach does not assume Λ \Lambda Λ CDM prior to recombination, instead relying on stellar physics.
Our baseline assumes Λ \Lambda Λ CDM behaviour at high redshift. The extension relaxes this assumption, limiting potential systematic error while also reducing the statistical constraining power of the analysis. The extended case can still be used as a robustness check of the baseline assumption. Importantly, the former’s ability to constrain r d r_{\mathrm{d}} r d independently of the CMB allows it to test whether early-time solutions to the Hubble tension can be accommodated without considering their impact on the CMB. Since these models have no new physics after recombination, BAO datasets tightly constrain H 0 r d H_0r_{\mathrm{d}} H 0 r d , leading to a low predicted r d r_{\mathrm{d}} r d . We can compare this to the r d r_{\mathrm{d}} r d inferred by our extension analysis. This calibration also sets absolute BAO distances, which can be used to calibrate M D E S M_{\mathrm{DES}} M DES via the Etherington relation.
Redshift-floor sweep
To check if our results are driven by SNIa at especially low redshifts, we conduct a sweep of the redshift floor z f l o o r z_{\mathrm{floor}} z floor below which SNIa are neglected, recalibrating the IDL each time. Some studies report a descending trend in H 0 H_0 H 0 with redshift of the data used to measure it (Krishnan et al. 2020, 2021b; Lopez-Hernandez & De-Santiago 2025; Jia et al. 2025, 2026). Unlike those studies which infer H 0 H_0 H 0 from data in a narrow redshift range centred on z z z and thereby obtain a curve of H 0 ( z ) H_0(z) H 0 ( z ) , our approach uses a redshift floor, so we denote the results as H 0 ( z f l o o r ) H_0(z_{\mathrm{floor}}) H 0 ( z floor ) . A clear trend in H 0 ( z f l o o r ) H_0(z_{\mathrm{floor}}) H 0 ( z floor ) which becomes asymptotically flat would indicate that the lowest redshift SNIa carry the cosmological parameters away from the more robust result from higher redshift SNIa. On the other hand, a flat trend in H 0 ( z f l o o r ) H_0(z_{\mathrm{floor}}) H 0 ( z floor ) at the low-redshift end would indicate that different SNIa subsets prefer much the same present expansion rate, though error bars should widen at high z f l o o r z_{\mathrm{floor}} z floor as more data is excluded.
Bayesian evidence for a late-time solution to the Hubble tension
Alternative models
To assess if the data prefers a late-time solution to the Hubble tension over Λ \Lambda Λ CDM, we consider the following phenomenological expansion history (Nájera, Banik, Desmond & Kalaitzidis 2026):
H ( z ) = H 0 , Λ C D M Ω m ( 1 + z ) 3 + ( 1 − Ω m ) + Δ H ( z ; A , z 1 ) , H(z)=H_{0,\Lambda\mathrm{CDM}}\sqrt{\Omega_{\mathrm{m}}(1+z)^3+(1-\Omega_{\mathrm{m}})}+\Delta H(z;A,z_1), H ( z ) = H 0 , Λ CDM Ω m ( 1 + z ) 3 + ( 1 − Ω m ) + Δ H ( z ; A , z 1 ) ,
Δ H ( z ; A , z 1 ) = A exp ( − z / z 1 ) . (12) \Delta H(z;A,z_1)=A\exp(-z/z_1). \tag*{(12)} Δ H ( z ; A , z 1 ) = A exp ( − z / z 1 ) . (12)
where A A A is the amplitude and z 1 z_1 z 1 is the redshift scale of Δ H \Delta H Δ H , an extra contribution to H ( z ) H(z) H ( z ) . A solution to the Hubble tension should have A ≈ 6 A \approx6 A ≈ 6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , while we expect z 1 ≲ 0.2 z_1 \lesssim0.2 z 1 ≲ 0.2 (Nájera et al. 2026). This exp model represents a flexible late-time solution to the Hubble tension in which H 0 = H 0 , Λ C D M + A H_0 = H_{0,\Lambda\mathrm{CDM}} + A H 0 = H 0 , Λ CDM + A . We adopt A ∼ U ( − 10 , 10 ) A \sim\mathcal{U}(-10, 10) A ∼ U ( − 10 , 10 ) km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (Table 1 ). This range comfortably covers the A ≈ 6 A \approx6 A ≈ 6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 needed to solve the Hubble tension, and it is symmetric about zero, so it does not favour a positive A A A a priori.
We also consider the Chevallier-Polarski-Linder (CPL; Chevallier & Polarski 2001; Linder 2003, 2008) model in which the dark energy equation of state is linear in a a a (w = w 0 + w a e a w = w_0 + w_a e^a w = w 0 + w a e a ), so the expansion history is as follows:
H ( z ) = H 0 Ω m ( 1 + z ) 3 + ( 1 − Ω m ) f ( z ) , (13) H(z) = H_0 \sqrt{\Omega_\mathrm{m}(1+z)^3 + (1-\Omega_\mathrm{m})f(z)}, \tag*{(13)} H ( z ) = H 0 Ω m ( 1 + z ) 3 + ( 1 − Ω m ) f ( z ) , (13)
f ( z ) = ( 1 + z ) 3 ( 1 + w 0 + w a ) exp ( − 3 w a e a ) . (14) f(z) = (1+z)^{3(1+w_0+w_a)}\exp(-3w_a e^a). \tag*{(14)} f ( z ) = ( 1 + z ) 3 ( 1 + w 0 + w a ) exp ( − 3 w a e a ) . (14)
We can recover Λ \Lambda Λ CDM behaviour by setting w 0 = − 1 w_0=-1 w 0 = − 1 and w a = 0 w_a=0 w a = 0 , in which case w = − 1 w=-1 w = − 1 and f ( z ) = 1 ∀ z f(z)=1\ \forall z f ( z ) = 1 ∀ z , indicating a constant dark energy contribution.
The Bayesian evidence of model M i M_i M i is defined as follows:
Z ( M i ) ≡ ∫ d θ π ( θ ∣ M i ) L ( D ∣ θ , M i ) , (15) Z(M_i) \equiv\int\mathrm{d}\theta\ \pi(\theta|M_i)\mathcal{L}(D|\theta,M_i), \tag*{(15)} Z ( M i ) ≡ ∫ d θ π ( θ ∣ M i ) L ( D ∣ θ , M i ) , (15)
where π \pi π is the prior and L \mathcal{L} L is the likelihood of the data D D D if the vector of model parameters is θ \theta θ . We compare the exp and CPL models to Λ \Lambda Λ CDM through the Bayes factor
B 10 ≡ Z ( M 1 ) Z ( M 0 ) , (16) B_{10} \equiv\frac{Z(M_1)}{Z(M_0)}, \tag*{(16)} B 10 ≡ Z ( M 0 ) Z ( M 1 ) , (16)
where M 0 M_0 M 0 is Λ \Lambda Λ CDM and M 1 M_1 M 1 is an alternative, which in our case is either the exp model (Eq. (12) ) or the CPL model (Eq. (14) ). Both models are phenomenological, with the exp model aimed at the Hubble tension (Nájera et al. 2026) and the CPL model aimed at the preference of DESI DR2 BAO for evolving dark energy (DESI Collaboration 2025). The Bayesian evidence automatically applies an Occam penalty for the two extra parameters that these models have compared to Λ \Lambda Λ CDM. Since the exp and CPL models have the same number of free parameters as each other, Bayesian evidence differences between them are not directly related to the number of free parameters. Even so, the size of the Occam penalty depends on how much of the prior range the data exclude, and therefore on the adopted prior ranges (given in Table 1 ).
To compare these models on an equal footing, we must use the same dataset in all cases. We use the SNIa, BAO, and CC data described in § 2 , possibly supplemented by local anchors on H 0 H_0 H 0 . As described below, we supplement these with additional constraints from the early Universe.
χ t o t a l 2 = χ S N I a 2 + χ B A O 2 + χ C M B 2 + χ B B N 2 + χ C C 2 ( + χ a n c h o r 2 ) , (17) \chi^2_{\mathrm{total}} = \chi^2_{\mathrm{SNIa}} + \chi^2_{\mathrm{BAO}} + \chi^2_{\mathrm{CMB}} + \chi^2_{\mathrm{BBN}} + \chi^2_{\mathrm{CC}} \left(+\chi^2_{\mathrm{anchor}}\right), \tag*{(17)} χ total 2 = χ SNIa 2 + χ BAO 2 + χ CMB 2 + χ BBN 2 + χ CC 2 ( + χ anchor 2 ) , (17)
with the anchor term entering only where stated (§ 3.4.1 ). Since we are interested in a model comparison, we need the Bayesian evidence, so for this analysis we cannot use NUTS. We therefore use the nested sampling algorithm dynesty (Skilling 2004, 2006; Feroz et al. 2009; Speagle 2020; Koposov et al. 2022) with dynamic nested sampling, 800 live points and a stopping criterion of Δ ln Z ^ = 0.01 \Delta\ln\hat{Z}=0.01 Δ ln Z ^ = 0.01 , where Z ^ \hat{Z} Z ^ denotes the approximate remaining Bayesian evidence.
Compressed CMB likelihood
In the rest of this paper, the CMB is included only via the Planck prior on r d r_d r d because the cosmographic expansion cannot be extended to high redshifts. Since this is possible for both the exp and CPL models, we can implement the CMB constraint in a more rigorous manner. We calculate the CMB term in Eq. (17) following the compressed CMB likelihood approach of Banik et al. (2026a), whose protocol we summarise here. We no longer impose a prior on r d r_d r d because it must now be found self-consistently using the physics of sound waves in the baryon–photon plasma prior to decoupling. This requires both the baryon and CDM densities, so we sample the present density parameter of baryons + C D M +\mathrm{CDM} + CDM (Ω b c \Omega_\mathrm{bc} Ω bc ) and the physical baryon density ω b ≡ Ω b h 2 \omega_\mathrm{b}\equiv\Omega_\mathrm{b}h^2 ω b ≡ Ω b h 2 from their prior ranges, where h ≡ H 0 / ( 100 k m s − 1 M p c − 1 ) h\equiv H_0/(100\ \mathrm{km\ s}^{-1}\mathrm{Mpc}^{-1}) h ≡ H 0 / ( 100 km s − 1 Mpc − 1 ) and ω x ≡ Ω x h 2 \omega_x\equiv\Omega_xh^2 ω x ≡ Ω x h 2 for any component x x x (Table 1 ). In the exp model, these densities refer to the underlying Λ \Lambda Λ CDM background, so h h h is set by H 0 , Λ C D M H_{0,\Lambda\mathrm{CDM}} H 0 , Λ CDM rather than H 0 H_0 H 0 . This is because Δ H \Delta H Δ H (Eq. (12) ) is completely negligible long before recombination for any z 1 z_1 z 1 in our prior range, so it has no effect on the expansion history around the time of recombination. However, the Δ H \Delta H Δ H term is important when calculating the comoving distance to recombination, which affects the angular scale of its anisotropies.
Within this framework, the CMB constraints are expressed through the parameter vector x ≡ ( 100 θ ⋆ , ω b , ω b c ) T \boldsymbol{x}\equiv(100\theta_\star,\omega_\mathrm{b},\omega_\mathrm{bc})^\mathrm{T} x ≡ ( 100 θ ⋆ , ω b , ω bc ) T , where θ ⋆ ≡ r ⋆ / D c ( z ⋆ ) \theta_\star\equiv r_\star/D_\mathrm{c}(z_\star) θ ⋆ ≡ r ⋆ / D c ( z ⋆ ) is the angular size of the comoving sound horizon r ⋆ r_\star r ⋆ at recombination, D c ( z ⋆ ) D_\mathrm{c}(z_\star) D c ( z ⋆ ) is the comoving distance to recombination, and ⋆ \star ⋆ subscripts denote quantities evaluated at that epoch (Lemos & Lewis 2023). CMB observations indicate that x \boldsymbol{x} x has a mean value of μ C M B \boldsymbol{\mu}_\mathrm{CMB} μ CMB and covariance matrix C C M B \mathbf{C}_\mathrm{CMB} C CMB , both of which were obtained by Banik et al. (2026a) from the chains of SPT-3G Collaboration [107 ] . The methodology involves computing the autocorrelation time τ i \tau_i τ i of each parameter i i i in x \boldsymbol{x} x , finding max ( τ i ) \max(\tau_i) max ( τ i ) across these parameters, and removing 3 max ( τ i ) 3\max(\tau_i) 3 max ( τ i ) steps from the start of the chains to account for burn-in (Nájera et al. 2026). The analysis of SPT-3G Collaboration [107 ] combines Planck 2020, SPT-3G Collaboration [107 ] , and ACT DR6 (Louis et al. 2025). We find that this SPA combination can be summarised as follows:
μ C M B = ( 1.04161 , 0.02240 , 0.14265 ) T , \boldsymbol{\mu}_{\mathrm{CMB}} = (1.04161,\,0.02240,\,0.14265)^\mathrm{T}, μ CMB = ( 1.04161 , 0.02240 , 0.14265 ) T ,
C C M B = 10 − 9 [ 53.8482 1.10723 − 28.7005 1.10723 9.20447 − 17.6016 − 28.7005 − 17.6016 843.132 ] . (18) \mathbf{C}_{\mathrm{CMB}} = 10^{-9}
\begin{bmatrix}
53.8482 & 1.10723 & -28.7005 \\
1.10723 & 9.20447 & -17.6016 \\
-28.7005 & -17.6016 & 843.132
\end{bmatrix}. \tag*{(18)} C CMB = 1 0 − 9 53.8482 1.10723 − 28.7005 1.10723 9.20447 − 17.6016 − 28.7005 − 17.6016 843.132 . (18)
The CMB term is then χ C M B 2 = ( x − μ C M B ) T C C M B − 1 ( x − μ C M B ) \chi^2_{\mathrm{CMB}}=(\boldsymbol{x}-\boldsymbol{\mu}_{\mathrm{CMB}})^\mathrm{T}\mathbf{C}_{\mathrm{CMB}}^{-1}(\boldsymbol{x}-\boldsymbol{\mu}_{\mathrm{CMB}}) χ CMB 2 = ( x − μ CMB ) T C CMB − 1 ( x − μ CMB ) , with x \boldsymbol{x} x predicted by each model. We obtain D c ( z ⋆ ) D_\mathrm{c}(z_\star) D c ( z ⋆ ) by integrating c / H ( z ) c/H(z) c / H ( z ) for each model all the way to recombination. This means that a late-time departure from Λ \Lambda Λ CDM changes the predicted θ ⋆ \theta_\star θ ⋆ even though it leaves the early universe untouched, which is precisely how the CMB constrains late-time solutions to the Hubble tension. In particular, any extra expansion at low redshift has to be compensated for in the expansion history in order to preserve the comoving distance to recombination.
Because the integral extends to z ⋆ ≈ 1090 z_\star\approx1090 z ⋆ ≈ 1090 , the expansion history now includes photons and neutrinos. We follow Banik et al. (2026a) in assuming three neutrino species with masses of ( 50 , 10 , 0 ) (50,10,0) ( 50 , 10 , 0 ) meV/c 2 c^2 c 2 and N e f f = 3.044 N_\mathrm{eff}=3.044 N eff = 3.044 , with the two massive species changing from relativistic to non-relativistic behaviour with a sharpness parameter k ν = 2 k_\nu=2 k ν = 2 (see their appendix B). The photon density follows from blackbody physics at the CMB temperature of T C M B = 2.72548 T_\mathrm{CMB}=2.72548 T CMB = 2.72548 K (Fixsen 2009). The matter density relevant at late times is thus Ω m = Ω b c + Ω ν \Omega_\mathrm{m}=\Omega_\mathrm{bc}+\Omega_\nu Ω m = Ω bc + Ω ν , which includes the contribution Ω ν \Omega_\nu Ω ν of massive neutrinos. We note that radiation changes H ( z ) H(z) H ( z ) by less than 0.04% over the redshift range of the SNIa, BAO, and CC, so it matters only for the CMB, where its impact is substantial.
We compute r ⋆ r_\star r ⋆ using the analytic result of Hu & Sugiyama (1996):
r ⋆ = 2 c 3 a e q Ω b c H 0 2 R e q ln ( 1 + R ⋆ + R ⋆ + R e q 1 + R e q ) . (19) r_\star= \frac{2c}{3}\sqrt{\frac{a_\mathrm{eq}}{\Omega_\mathrm{bc}H_0^2R_\mathrm{eq}}}\ln\left(\frac{\sqrt{1+R_\star}+\sqrt{R_\star+R_\mathrm{eq}}}{1+\sqrt{R_\mathrm{eq}}}\right). \tag*{(19)} r ⋆ = 3 2 c Ω bc H 0 2 R eq a eq ln ( 1 + R eq 1 + R ⋆ + R ⋆ + R eq ) . (19)
where the subscript eq denotes matter–radiation equality (treating neutrinos as radiation) and R ≡ 3 ρ b / ( 4 ρ γ ) \mathcal{R} \equiv3\rho_b/(4\rho_\gamma) R ≡ 3 ρ b / ( 4 ρ γ ) is the ratio between the baryon and photon densities, which sets the sound speed of the photon–baryon fluid. The BAO sound horizon r d r_d r d follows from the same expression evaluated at the end of the drag epoch, i.e. using R d \mathcal{R}_d R d instead of R ⋆ \mathcal{R}_\star R ⋆ . This requires the scale factors at recombination and at the end of the drag epoch, which we obtain from a ⋆ = T C M B / T ⋆ a_\star= T_{\rm CMB}/T_\star a ⋆ = T CMB / T ⋆ and a d = T C M B / T d a_d = T_{\rm CMB}/T_d a d = T CMB / T d , respectively, where T T T denotes the photon temperature at the epoch indicated by the subscript. We find T ⋆ T_\star T ⋆ and T d T_d T d using quadratic fits to a grid of CLASS runs (Blas et al. 2011), as described in section 2.6 and appendix A of Banik et al. (2026a):
T ⋆ = T ⋆ r e f + d T g ⋆ + 1 2 d T H ⋆ d , T_\star= T_\star^{\rm ref} + \mathbf{d}^{T}\mathbf{g}_\star+ \frac{1}{2}\mathbf{d}^{T}\mathbf{H}_\star\mathbf{d}, T ⋆ = T ⋆ ref + d T g ⋆ + 2 1 d T H ⋆ d ,
d ≡ ( ω b , ω b c ) T − ( 0.0225 , 0.1450 ) T , (20) \mathbf{d} \equiv(\omega_b,\omega_{bc})^{T} - (0.0225,0.1450)^{T}, \tag*{(20)} d ≡ ( ω b , ω b c ) T − ( 0.0225 , 0.1450 ) T , (20)
where T ⋆ r e f = 2970.29 T_\star^{\rm ref} = 2970.29 T ⋆ ref = 2970.29 K. The temperature T d T_d T d at the end of the drag epoch is found in the same way, with T d r e f = 2892.74 T_d^{\rm ref} = 2892.74 T d ref = 2892.74 K. The gradients and Hessians are as follows, with temperatures in K:
g ⋆ = ( − 3188.92 194.09 ) , H ⋆ = ( 265960.27 − 7860.93 − 7860.93 − 488.36 ) , \mathbf{g}_\star=
\begin{pmatrix}
-3188.92\\
194.09
\end{pmatrix},
\qquad
\mathbf{H}_\star=
\begin{pmatrix}
265960.27 & -7860.93\\
-7860.93 & -488.36
\end{pmatrix}, g ⋆ = ( − 3188.92 194.09 ) , H ⋆ = ( 265960.27 − 7860.93 − 7860.93 − 488.36 ) ,
g d = ( 5925.64 203.34 ) , H d = ( − 189076.80 − 4767.87 − 4767.87 − 626.49 ) . (21) \mathbf{g}_d =
\begin{pmatrix}
5925.64\\
203.34
\end{pmatrix},
\qquad
\mathbf{H}_d =
\begin{pmatrix}
-189076.80 & -4767.87\\
-4767.87 & -626.49
\end{pmatrix}. \tag*{(21)} g d = ( 5925.64 203.34 ) , H d = ( − 189076.80 − 4767.87 − 4767.87 − 626.49 ) . (21)
These fits reproduce the CLASS temperatures to within 0.8 mK across the grid, a negligible fraction of T ⋆ ≈ 2970 T_\star\approx2970 T ⋆ ≈ 2970 K.
Following Banik et al. (2026a), we also add the BBN constraint to Eq. (17) through a Gaussian prior on ω b = 0.021915 ± 0.000215 \omega_b = 0.021915 \pm0.000215 ω b = 0.021915 ± 0.000215 (Giovanetti et al. 2025). In flat Λ \Lambda Λ CDM, this compressed likelihood recovers the H 0 H_0 H 0 and its uncertainty inferred by SPT-3G Collaboration [107 ] from the full CMB power spectra to within 0.01 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , demonstrating that it retains the information relevant to the background expansion (see appendix C of Banik et al. 2026a). We stress that Eq. (19) neglects dark energy before recombination. This is an excellent approximation for Λ \Lambda Λ CDM and the exp model, but not for CPL models with w 0 + w a > 0 w_0 + w_a > 0 w 0 + w a > 0 as the dark energy density in such models grows faster than the matter density towards high redshift. However, the posterior assigns these models a probability below 10 − 34 10^{-34} 1 0 − 34 in every case we consider, so this approximation has no impact on our results.
Local H 0 H_0 H 0 anchors
To assess the feasibility of a late-time solution to the Hubble tension, we consider the case with and without anchors on H 0 H_0 H 0 from the standard distance ladder. We use either the SH0ES constraint of H 0 = 73.17 ± 0.86 H_0 = 73.17 \pm0.86 H 0 = 73.17 ± 0.86 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 from the traditional Cepheid–SNIa route (Leavitt & Pickering 1912; Breuval et al. 2024), or we use three anchors which are independent of Cepheid and SNIa absolute magnitudes, thereby making them independent of some of the main contributors to the Hubble tension. We assume all three are independent of each other because they come from non-overlapping datasets, as follows:
The TRGB (Lee, Freedman & Madore 1993; Madore & Freedman 1995) is used to calibrate distances to galaxies with measured surface brightness fluctuations (SBF; Tonry & Schneider 1988). Our adopted TRGB + SBF value is H 0 = 73.8 ± 0.7 H_0 = 73.8 \pm0.7 H 0 = 73.8 ± 0.7 (stat) ± 2.3 \pm2.3 ± 2.3 (sys) km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (Jensen et al. 2025). We combine the statistical and systematic errors in quadrature to obtain a single uncertainty of 2.4 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 .
The GW170817 standard siren [136 ] (Abbott et al. 2017) is used to obtain a luminosity distance.
This is combined with a peculiar velocity correction and the very long baseline interferometry (VLBI) constraint on the inclination to obtain an estimated H 0 = 68. 3 − 4.5 + 4.6 H_0 = 68.3^{+4.6}_{-4.5} H 0 = 68. 3 − 4.5 + 4.6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (Mukherjee et al. 2021), as listed in table 2 of Alfradique et al. (2026). We adopt the larger uncertainty of 4.6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 to be conservative. We do not use the headline result of Alfradique et al. (2026) as this also combines the single bright siren with 17 dark sirens, which would significantly complicate our analysis while only slightly improving the constraining power.
(iii) Type II SNe (SNII) with the expanding photosphere method (EPM; Vogl et al. 2025). Their study gives an estimated H 0 = 74.9 ± 2.7 H_0 = 74.9 \pm2.7 H 0 = 74.9 ± 2.7 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 if we assume the statistical and systematic errors are comparable, which the authors indicate is a reasonable conservative assumption.
We stress that adding SH0ES and legacy SNIa to the same analysis risks overlapping data. This is the reason why we lead with the non-SH0ES anchors, which we can use even if we also include the legacy SNIa. However, if the latter are excluded, there is negligible overlap between the SNIa used by DES and by SH0ES (Riess et al. 2022).
There is an important caveat that we must take into account for the exp model. Measurements of H 0 H_0 H 0 from the above anchors compute it by fitting the luminosity distances with their second-order kinematic expansion in z z z , assuming that q 0 = − 0.55 q_0 = -0.55 q 0 = − 0.55 and the jerk parameter j 0 = 1 j_0 = 1 j 0 = 1 , allowing us to fit even a single data point. This approach is similar to that used in the literature, since probes at very low redshift usually lack enough data to constrain higher order cosmographic parameters. H 0 H_0 H 0 can then be obtained only by extrapolating from the redshift of the data down to z → 0 z \to0 z → 0 . The consequence is that if z 1 ≲ z max z_1 \lesssim z_{\max} z 1 ≲ z m a x , the maximum redshift of the data considered, then the exp model H 0 H_0 H 0 does not correspond to that measured from the Taylor series. For these cases, we therefore do not compare the model H 0 H_0 H 0 directly with the published H 0 H_0 H 0 anchor. We instead calculate the exp model distance moduli at the redshifts of the real data and fit them with the same kinematic expansion, giving each object equal weight, from which we can read off H 0 H_0 H 0 (similar to section 4.1.1 of Nájera et al. 2026). This calculates the value of H 0 H_0 H 0 that would have been reported by observers from distance measurements to anchors with the same redshift distribution as the real data, if the universe followed the phenomenological exp cosmology. The same quantity decides whether a model solves the Hubble tension. Since SH0ES measure H 0 H_0 H 0 from SNIa at z = 0.023 z = 0.023 z = 0.023 –0.15 0.15 0.15 , a model whose excess expansion rate is redshifted below z ≈ 0.05 z \approx0.05 z ≈ 0.05 predicts a lower SH0ES measurement than its H ( 0 ) H(0) H ( 0 ) , because the observers fit a smooth Hubble law across a window in which most of the excess has already decayed. We therefore quote, alongside H ( 0 ) H(0) H ( 0 ) , the value H 0 w i n H_0^{\rm win} H 0 win that the SH0ES analysis would infer from each exp posterior (Table 4). Throughout, H ( 0 ) H(0) H ( 0 ) remains the expansion rate today, while H 0 w i n H_0^{\rm win} H 0 win decides whether a model reproduces the SH0ES measurement.
Alcock–Paczynski consistency
We subject the IDL to the AP consistency test (Alcock & Paczynski 1979). This is based on the AP ratio F A P ≡ D M / D H F_{\rm AP} \equiv D_M/D_H F AP ≡ D M / D H (Appendix A). Using SNIa alone and assuming smoothness of the distance–redshift relation combined with the Etherington relation, we predict F A P F_{\rm AP} F AP at three redshifts where particularly precise BAO measurements of F A P F_{\rm AP} F AP are available from DESI DR2. Taking the ratio causes r d r_d r d to cancel out, so the AP test is independent of the absolute calibration of the BAO sound horizon. It is therefore also independent of the SNIa absolute magnitude – and thus H 0 H_0 H 0 . The AP test instead probes the shape of the reconstructed distance–redshift relation, or equivalently of our cosmographic E ( 1 − a ) E(1-a) E ( 1 − a ) expansion.
Systematics with the legacy SNIa
The DES-Dovekie compilation (Popovic et al. 2026) combines the homogeneous, single-instrument DES catalogue of SNe (z ∈ [ 0.073 , 1.14 ] z \in[0.073, 1.14] z ∈ [ 0.073 , 1.14 ] ) with an external low-redshift anchor assembled from several legacy surveys (z ∈ [ 0.025 , 0.087 ] z \in[0.025, 0.087] z ∈ [ 0.025 , 0.087 ] ). The legacy SNIa come almost entirely from the Foundation sample (117 SNIa) or from the Center for Astrophysics CfA3 and CfA4 samples (72), with a further 8 from the Carnegie Supernova Project (CSP-I). We denote the CfA and CSP samples together as CfA/CSP. As noted in § 2 , the homogeneous DES sample has practically no SNIa at z ≲ 0.1 z \lesssim0.1 z ≲ 0.1 , but all 197 legacy SNIa lie here. This makes the legacy SNIa an important lever on the z → 0 z \to0 z → 0 region, where BAO data is scant (though see table 1 of Banik & Kalaitzidis 2025, and references therein). We therefore test whether the legacy SNIa follow the same distance–redshift relation as DES. We conduct another analysis to test what this means for H 0 H_0 H 0 .
We stress that the redshift distributions of the legacy SNIa and the main DES sample have very little overlap (see figure 3 of DES Collaboration 2024). This means our test relies on a downward extrapolation of the DES sample, not a comparison of SNIa apparent magnitudes at the same redshift between the two subsamples. The legacy SNIa span z ∈ [ 0.025 , 0.087 ] z \in[0.025, 0.087] z ∈ [ 0.025 , 0.087 ] while DES begins at z = 0.073 z = 0.073 z = 0.073 , so only 2 DES SNe and 10 legacy SNIa lie in the overlap region. Given the large number of legacy SNIa with modest error bars, the uncertainty in Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy is dominated by extrapolation uncertainty, i.e., how much freedom the DES reference has below its own floor. The main uncertainty in Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy is thus not from the legacy SNIa themselves – at least if we take their uncertainties at face value.
Our test for systematics in the legacy SNIa is designed to require no absolute calibration. We proceed in three steps:
We fit the shape of the DES Hubble diagram (using only DES SNIa) with a rigid fourth-order polynomial as the baseline model (Eq. (7) ), the same expansion additionally constrained by the DESI result for F A P ≡ D M / D H F_{\mathrm{AP}} \equiv D_M/D_H F AP ≡ D M / D H (which is independent of r d r_d r d ; see § 3.5 ), rigid Λ \Lambda Λ CDM with free Ω m \Omega_m Ω m (Eq. (9) ), and rigid Λ \Lambda Λ CDM likewise constrained by F A P F_{\mathrm{AP}} F AP .
We extrapolate the fitted relation below the DES floor (z = 0.073 z = 0.073 z = 0.073 ) down to the redshifts of the legacy SNIa, allowing us to calculate the residuals r i ≡ m l e g a c y , i − m p r e d ( z i ) r_i \equiv m_{\mathrm{legacy},i} - m_{\mathrm{pred}}(z_i) r i ≡ m legacy , i − m pred ( z i ) between observed legacy SNIa and our extrapolated predictions.
We take Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy to be the weighted mean of the r i r_i r i under the full statistical + systematic covariance of the compilation, including the legacy–DES cross-covariance.
It is important to highlight why this approach is calibration-free, since this property is what makes Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy comparable across the four parametrisations. Rescaling H ↦ λ H H \mapsto\lambda H H ↦ λ H maps D M ↦ D M / λ D_M \mapsto D_M/\lambda D M ↦ D M / λ and hence m ↦ m − 5 log 10 λ m \mapsto m - 5\log_{10}\lambda m ↦ m − 5 log 10 λ , so that ( H 0 → λ H 0 , M → M + 5 log 10 λ ) (H_0 \to\lambda H_0, M \to M + 5\log_{10}\lambda) ( H 0 → λ H 0 , M → M + 5 log 10 λ ) is an exact flat direction of the fit. We therefore fix H 0 = 70 k m s − 1 M p c − 1 H_0 = 70\ \mathrm{km\ s}^{-1}\mathrm{Mpc}^{-1} H 0 = 70 km s − 1 Mpc − 1 and let M D E S M_{\mathrm{DES}} M DES float freely. Every predicted apparent magnitude is invariant along this direction, making the fitted M D E S M_{\mathrm{DES}} M DES meaningless by construction. This means that no absolute scale enters Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy , making it independent of the sound horizon and the local distance ladder. In particular, Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy inherits no part of the calibration chain that later produces H 0 H_0 H 0 .
To constrain H 0 H_0 H 0 , we cannot simply add the Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy obtained above to the legacy SNIa because H 0 H_0 H 0 and Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy are correlated. We therefore conduct a separate analysis where we sample Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy alongside
the other parameters. The constraints we apply are SNIa + BAO + Planck 2020 r d r_d r d . We consider the cosmographic expansion as well as Λ \Lambda Λ CDM. For comparison, we also report results for the two cases that do not have Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy as a model parameter: excluding the legacy SNIa altogether, or including them at face value (Δ M l e g a c y = 0 \Delta M_{\mathrm{legacy}} = 0 Δ M legacy = 0 ). These results highlight the role of the legacy SNIa in H 0 H_0 H 0 determinations via the IDL.
Results
The observed distance–redshift relation
To gain a first impression of whether the Hubble tension might plausibly be solved at late times, we can consider the ratio between observed distances from SNIa and BAO and those predicted in the flat Λ \Lambda Λ CDM cosmology that best fits the CMB (the SPA cosmology; SPT-3G Collaboration [107 ] ). The results are shown in Figure 2 , which plots DESI DR2 BAO measurements of D V / r d D_V/r_d D V / r d and binned D L D_L D L from SNIa, using uniform bins in ln z \ln z ln z . The SNIa are calibrated according to the best-fit value of M D E S M_{\mathrm{DES}} M DES , which has very little uncertainty due to the need for consistency with the BAO. This is evident in the excellent agreement between SNIa and BAO measurements in the regime where they overlap (z = 0.3 z = 0.3 z = 0.3 –1 1 1 ). In particular, the DES SNIa have a weighted mean ratio of D L D_L D L relative to the SPA prediction of 0.981 ± 0.003 0.981 \pm0.003 0.981 ± 0.003 , while the corresponding result for DESI DR2 BAO measurements of D V / r d D_V/r_d D V / r d is 0.987 ± 0.003 0.987 \pm0.003 0.987 ± 0.003 . This agreement is inevitable in the IDL approach since it is how we calibrate M D E S M_{\mathrm{DES}} M DES . However, it does not guarantee consistency over the wide redshift range where there is an overlap between SNIa and BAO. We revisit this issue in § 3.5 using the AP test, which does not rely on an assumed r d r_d r d . The solid lines show our fourth-order cosmographic expansion (Eq. (7) ). At each redshift, the shaded band shows the 1 σ 1\sigma 1 σ uncertainty using Gaussian equivalent percentiles (i.e., the 15.87–84.13% range).
Figure 2. Observed distances divided by the flat Λ \Lambda Λ CDM prediction with the SPA parameters (H 0 = 67.19 k m s − 1 M p c − 1 H_0 = 67.19\,\mathrm{km\,s^{-1}\,Mpc^{-1}} H 0 = 67.19 km s − 1 Mp c − 1 and Ω m = 0.3175 \Omega_{\mathrm{m}} = 0.3175 Ω m = 0.3175 ; SPT-3G Collaboration [107 ] ) and r d = 147.05 c M p c r_{\mathrm{d}} = 147.05\,\mathrm{cMpc} r d = 147.05 cMpc (Planck [100 ] ). We neglect the small uncertainties in these parameters and in M D E S M_{\mathrm{DES}} M DES , which is fixed at the best fit of each analysis. Faint grey (magenta) points show individual DES (legacy) SNIa at their Hubble diagram redshift. Circles show inverse variance-weighted mean SNIa luminosity distances relative to SPA using the full DES-Dovekie covariance for DES (black), legacy at face value (red), and legacy recalibrated by the jointly inferred Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (blue; § 3.6 ), which lowers them by ≈ 3 % \approx3\% ≈ 3% . The SNIa are binned uniformly ln z \ln z ln z bins. Magenta squares show DESI DR2 D V / r d D_{\mathrm{V}}/r_{\mathrm{d}} D V / r d , combining D M D_{\mathrm{M}} D M and D H D_{\mathrm{H}} D H using their published covariance. Lines in the same colour as the data show the best-fitting k max = 4 k_{\max} = 4 k m a x = 4 cosmography (Eq. (7) ) up to z = 2.33 z = 2.33 z = 2.33 , with 1 σ 1\sigma 1 σ Gaussian-equivalent bands. Results are extrapolated below the lowest SNIa of each sample, as marked by the lines becoming dotted and a short gap in the band. The green band is H 0 S P A / H 0 S H 0 E S H_0^{\mathrm{SPA}}/H_0^{\mathrm{SH0ES}} H 0 SPA / H 0 SH0ES , the level that a late-time solution to the Hubble tension must reach as z → 0 z \to0 z → 0 .
As we move towards lower redshift, Figure 2 shows that there is a drastic difference depending on whether we use only SNIa from the homogeneous DES-Dovekie catalogue, also include the legacy SNIa normally included in the catalogue at face value, or include them after recalibrating them by the best-fitting value of Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy in our analysis where this is inferred jointly with the other parameters (§ 3.6 ). The legacy SNIa cover the crucial range of z = 0.025 z = 0.025 z = 0.025 –0.087 0.087 0.087 , but they sit systematically above where we would expect them to sit based on the DES SNIa. Recalibrating the legacy SNIa naturally reduces this discrepancy using an offset of Δ M l e g a c y = + 0.065 \Delta M_{\mathrm{legacy}} = +0.065 Δ M legacy = + 0.065 mag. Since H 0 H_0 H 0 is inversely related to the distance scale in the z → 0 z \to0 z → 0 limit, these differences can be expected to significantly affect our conclusions regarding the IDL prediction for H 0 H_0 H 0 . To highlight this, we add a green horizontal band at H 0 S P A / H 0 S H 0 E S = 0.918 ± 0.012 H_0^{\mathrm{SPA}}/H_0^{\mathrm{SH0ES}} = 0.918 \pm0.012 H 0 SPA / H 0 SH0ES = 0.918 ± 0.012 , which is the level that must be reached as z → 0 z \to0 z → 0 in order to solve the Hubble tension. Our results suggest that this is quite likely if we only consider the homogeneous DES catalogue, marginally possible if we include the legacy SNIa after recalibrating them to line up with the extrapolated relation constrained by BAO + DES, and not possible if we take the legacy SNIa at face value. The rest of this study elaborates this basic picture without much altering it.
Reconstructed expansion history
We now turn to the expansion history a ( z ) = H ( z ) / ( 1 + z ) a(z) = H(z)/(1+z) a ( z ) = H ( z ) / ( 1 + z ) inferred by fitting our cosmographic expansion (Eq. (7) ) to the datasets mentioned in § 2 . The median result for each analysis is shown in Figure 3 , with the shaded bands showing the 1 σ 1\sigma 1 σ range using Gaussian equivalent percentiles. We use a solid line to illustrate the range covered by the underlying SNIa dataset, while the dotted line shows the z → 0 z \to0 z → 0 extrapolation. The orange line without uncertainty shows the Λ \Lambda Λ CDM prediction based on the H 0 H_0 H 0 and Ω m \Omega_{\mathrm{m}} Ω m required to fit the latest CMB observations (SPT-3G Collaboration [107 ] ). The reconstructed a ˙ \dot a a ˙ is higher than this expectation at z ≈ 0.4 z \approx0.4 z ≈ 0.4 . This discrepancy narrows towards lower redshift – possibly even reversing sign. At the very lowest redshifts probed by the SNIa, a ˙ \dot a a ˙ rises sharply towards lower z z z . This rise is most pronounced with DES alone and least pronounced with DES + legacy. Allowing recalibration of the legacy SNIa returns an intermediate result (§ 3.6 ).
Figure 3. The median reconstructed a ˙ ( z ) \dot a(z) a ˙ ( z ) in our cosmographic expansion with k max = 4 k_{\max} = 4 k m a x = 4 (Eq. (7) ) constraining SNIa from DES only (solid black line), DES + legacy (solid red line), or DES + legacy allowing the legacy SNIa to be recalibrated by Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (dashed blue line at an intermediate level). The shaded bands show the 1 σ 1\sigma 1 σ uncertainty using Gaussian equivalent percentiles. The dotted lines in the same colour show the extrapolation below the lowest redshift SNIa in each case, which we mark with a small gap in the error band. The solid orange line without uncertainty shows the flat Λ \Lambda Λ CDM prediction using parameters that best fit the CMB (SPT-3G Collaboration [107 ] ). The inset shows the z < 0.2 z < 0.2 z < 0.2 region in more detail.
The inset to Figure 3 shows the z < 0.2 z < 0.2 z < 0.2 regime in greater detail. It is clear that if the legacy SNIa are taken at face value and assumed to be free of systematics, then the Hubble tension will be very difficult to solve at late times. But if the legacy SNIa are excluded or we assume that systematics with them can be rectified using a single recalibration parameter, then a late-time solution becomes much more plausible and may even be preferred. We note that since we do not apply a full CMB comparison and only consider r d r_{\mathrm{d}} r d , the a ˙ \dot a a ˙ reconstructions shown here likely need to be reduced by a few percent to preserve the comoving distance to recombination, which is related to the precisely measured angular scale of the CMB anisotropies. Roughly speaking, one can think of this as constraining c ∫ d z / H c \int dz/H c ∫ d z / H up to very high redshift, so a higher a ˙ \dot a a ˙ at some redshift must be compensated by a lower a ˙ \dot a a ˙ elsewhere. Given the integral is barely affected by the z ≲ 0.1 z \lesssim0.1 z ≲ 0.1 regime crucial to the Hubble tension, a compensating reduction in the overall normalisation designed to preserve the comoving distance to recombination would not qualitatively change our conclusions.
A more detailed comparison with the CMB is not possible in our cosmographic approach because it cannot be extrapolated to very high redshift where E → ∞ E \to\infty E → ∞ , but this is possible in the exp and CPL models considered in § 3.3.1 .
Cosmological parameters
Our results in Figure 3 show moderate deviations from the SPA cosmology at z ≲ 0.8 z \lesssim0.8 z ≲ 0.8 , even though we adopt a Planck prior of r d = 147.05 ± 0.30 c M p c r_{\mathrm{d}} = 147.05 \pm0.30\,\mathrm{cMpc} r d = 147.05 ± 0.30 cMpc and BAO data show at most modest deviations from Λ \Lambda Λ CDM predictions calibrated using the CMB (Banik et al. 2026a). These predictions are evidently accurate at high redshift, so any deviations at low redshift should be apparent in both the Hubble and deceleration parameters. We therefore use Figure 4 to show the joint posterior of H 0 H_0 H 0 and q 0 = C 1 − 1 q_0 = C_1 - 1 q 0 = C 1 − 1 . The black contours show that without the legacy SNIa, the parameters are indeed significantly different to the Λ \Lambda Λ CDM expectation, which we show using an orange star (neglecting the very small uncertainties; SPT-3G Collaboration [107 ] ). Remarkably, the predicted H 0 H_0 H 0 with our DES-only cosmographic IDL is very much in line with the H 0 H_0 H 0 anchors mentioned in § 3.3 , even slightly exceeding the central value of H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES (Brout et al. 2024). The high H 0 H_0 H 0 is related to a very negative inferred q 0 q_0 q 0 , which implies the expansion rate accelerates more than expected at late times.
Figure 4. The joint inference on H 0 H_0 H 0 and q 0 q_0 q 0 from our k max = 4 k_{\max}=4 k m a x = 4 cosmographic expansion using SNIa from DES (black), DES + legacy (red), and DES + legacy with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (blue). In all cases, the darker shading shows the 1 σ 1\sigma 1 σ region and the lighter shading shows the 2 σ 2\sigma 2 σ region. The orange star shows the Λ \Lambda Λ CDM expectation based on the latest CMB observations [107 ] . The vertical green band shows H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES (Breuval et al. 2024).
Adding the legacy SNIa tightens the constraints considerably, but it also shifts both H 0 H_0 H 0 and q 0 q_0 q 0 very close to the CMB-calibrated Λ \Lambda Λ CDM prediction (Figure 4 ). The 2 σ 2\sigma 2 σ constraints with and without the legacy SNIa barely overlap in the ( H 0 , q 0 ) (H_0, q_0) ( H 0 , q 0 ) plane, which is unusual given the latter is strictly a subset of the former. This suggests that the estimated DES filter system apparent magnitudes of the legacy SNIa might have unrecognised systematics. If we allow for this possibility using our analysis with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (§ 4.8 ), we get intermediate results.
Our main results are presented in Table 2 , which we split into different sections to indicate if we are using configuration A or B (Eq. (11) ), whether the legacy SNIa are included, and whether we parametrise the expansion history using our fourth-order cosmographic expansion (Eq. (7) ) or flat Λ \Lambda Λ CDM (Eq. (9) ). Our baseline (cosmographic configuration A) result is that the predicted H 0 = 76.25 ± 2.87 k m s − 1 M p c − 1 H_0 = 76.25 \pm2.87\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 76.25 ± 2.87 km s − 1 Mp c − 1 for DES. This is only 1.03 σ 1.03\sigma 1.03 σ above SH0ES (Breuval et al. 2024), but it lies 3.13 σ 3.13\sigma 3.13 σ above the latest CMB-constrained flat Λ \Lambda Λ CDM prediction [107 ] . This falls to 2.22 σ 2.22\sigma 2.22 σ once we calibrate the significance on mock data, which reveal an upward bias in our estimator (§ 4.9 ). Since we do not use constraints from the standard distance ladder, our IDL result is not a reformulation of the classic Hubble tension, although it can be considered a rediscovery of it using the IDL.
Table 2. Summary of the most important cosmological parameters in different setups. Configuration A (SNIa + BAO + Planck r d r_d r d ) pins the sound horizon to the Planck 2020 prior (r d = 147.05 ± 0.30 r_d = 147.05 \pm0.30 r d = 147.05 ± 0.30 cMpc). Configuration B (SNIa + BAO + CC) replaces the r d r_d r d prior with CC datasets, so r d r_d r d is derived without CMB input. In the first block for each configuration, we obtain the deceleration parameter q 0 = C 1 − 1 q_0 = C_1 - 1 q 0 = C 1 − 1 in our cosmographic IDL analysis with k max = 4 k_{\max}=4 k m a x = 4 (Eq. (7) ). The second block of each configuration repeats the IDL assuming flat Λ \Lambda Λ CDM (Eq. (9) ), for which we obtain q 0 = 3 2 Ω m − 1 q_0 = \frac{3}{2}\Omega_m - 1 q 0 = 2 3 Ω m − 1 . The final column shows the reduced χ 2 \chi^2 χ 2 using the posterior mean value of each parameter. We quote the posterior mean and standard deviation, except for markedly non-Gaussian posteriors (mode and mean differing by more than 0.3 σ 0.3\sigma 0.3 σ , or the two sides of the 68% highest posterior density interval differing by more than 0.3 σ 0.3\sigma 0.3 σ ), for which we quote the mode and that interval.
Importantly, the tension with SPA depends on our choice of k max = 4 k_{\max}=4 k m a x = 4 , since a Bayesian model average (BMA) over k max = 2 k_{\max}=2 k m a x = 2 –6 6 6 weighted by the Bayesian evidence yields a larger uncertainty on H 0 H_0 H 0 , which means that DES alone does not determine H 0 H_0 H 0 precisely enough to rediscover the Hubble tension (§ 4.9 ). We stress that the larger uncertainty in the BMA compared to a fixed choice of k max = 4 k_{\max}=4 k m a x = 4 is partly due to SNIa from the homogeneous DES catalogue having a redshift floor of 0.073 and lacking effective coverage at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 (with barely any SNIa at z < 0.1 z < 0.1 z < 0.1 ), requiring us to extrapolate the behaviour here — which is then dependent on the assumed k max k_{\max} k m a x . We can reduce this extrapolation error if we extend the low-redshift coverage by adding the legacy SNIa, in which case the predicted H 0 H_0 H 0 using our cosmographic IDL drops to 67.98 ± 0.69 k m s − 1 M p c − 1 67.98 \pm0.69\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 67.98 ± 0.69 km s − 1 Mp c − 1 . This is only 1.00 σ 1.00\sigma 1.00 σ above H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA [107 ] , but lies 4.69 σ 4.69\sigma 4.69 σ below SH0ES (Breuval et al. 2024). Adding the legacy SNIa therefore completely changes the result.
To reduce sensitivity to the assumption of Λ \Lambda Λ CDM prior to recombination, we can drop the Planck prior on r d r_d r d and set the absolute scale of the expansion history using CC datasets. This is our configuration B, which yields H 0 = 75.7 ± 5.3 k m s − 1 M p c − 1 H_0 = 75.7 \pm5.3\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 75.7 ± 5.3 km s − 1 Mp c − 1 for DES and 67.6 ± 3.6 k m s − 1 M p c − 1 67.6 \pm3.6\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 67.6 ± 3.6 km s − 1 Mp c − 1 for DES + legacy. The inferred r d r_d r d lies within 0.10 σ 0.10\sigma 0.10 σ (0.16 σ 0.16\sigma 0.16 σ ) of the standard Planck value for DES (DES + legacy). A clearer picture will emerge with further improvements to CC measurements, especially at low redshift where leverage from DES remains limited (Wang et al. 2026).
Our constraints rely on an absolute calibration of DES SNIa, for which we use DESI DR2 BAO calibrated assuming a standard r d r_d r d . In all cases, the results in Table 2 are consistent with an absolute SNIa calibration of M D E S ≈ − 19.38 M_{\mathrm{DES}} \approx-19.38 M DES ≈ − 19.38 , with a narrow range of − 19.380 ± 0.012 -19.380 \pm0.012 − 19.380 ± 0.012 for DES alone to − 19.378 ± 0.012 -19.378 \pm0.012 − 19.378 ± 0.012 for DES + legacy. These results are in 4.30 σ 4.30\sigma 4.30 σ and 4.23 σ 4.23\sigma 4.23 σ tension, respectively, with the SH0ES Cepheid-based result of M B = − 19.253 ± 0.027 M_B = -19.253 \pm0.027 M B = − 19.253 ± 0.027 (Riess et al. 2022). The result would differ slightly if we had instead used the slightly higher and more precise H 0 H_0 H 0 measurement from Breuval et al. (2024), but those authors do not directly publish a value for M B M_B M B . We note that when using the traditional Cepheid–SNIa route to H 0 H_0 H 0 as an anchor on some of our models, we adopt the Breuval et al. (2024) constraint because it improves upon Riess et al. (2022) by adding the Small Magellanic Cloud as a fourth geometric anchor to reduce uncertainties, but otherwise the two studies use the same formalism. Figure 2 shows that our result for M D E S M_{\mathrm{DES}} M DES is largely set by BAO data at z = 0.51 z = 0.51 z = 0.51 –0.93 0.93 0.93 combined with the Planck prior on r d r_d r d , so M D E S M_{\mathrm{DES}} M DES is not dependent on how we extrapolate below the DES floor. This explains why M D E S M_{\mathrm{DES}} M DES and its uncertainty are barely affected by inclusion of the legacy SNIa, even though these significantly extend the coverage towards lower redshift.
The large discrepancy between our clear result for M D E S M_{\mathrm{DES}} M DES and the SH0ES result for M B M_B M B could arise due to the use of different photometric filter systems. Unlike the SH0ES result based on an absolute calibration of SNIa close enough to have a Cepheid distance, SNIa in DES are too distant for this approach. It is therefore difficult to directly test whether M D E S M_{\mathrm{DES}} M DES might plausibly be − 19.4 -19.4 − 19.4 , as required by our analysis if standard physics applies prior to recombination (configuration A). In configuration B where this assumption is relaxed and we replace the Planck prior on r d r_d r d with cosmology-independent CC constraints, the preferred value of M D E S M_{\mathrm{DES}} M DES remains very close to − 19.4 -19.4 − 19.4 , but the uncertainty is considerably larger, reducing the discrepancy with the SH0ES result to ≈ 1.1 σ \approx1.1\sigma ≈ 1.1 σ –1.2 σ 1.2\sigma 1.2 σ .
As a further check that the M D E S M_{\mathrm{DES}} M DES constraint does not depend on our particular cosmographic expansion (Eq. (7) ), we repeat our IDL analysis using instead the flat Λ \Lambda Λ CDM form of the expansion history (Eq. (9) ). This is shown in Table 2 as the second block of results for each configuration. The M D E S M_{\mathrm{DES}} M DES calibration remains unchanged, sitting at − 19.377 ± 0.012 -19.377 \pm0.012 − 19.377 ± 0.012 (− 19.379 ± 0.012 -19.379 \pm0.012 − 19.379 ± 0.012 ) for DES (DES + legacy) in configuration A. The equivalent result for configuration B is − 19.40 ± 0.12 -19.40 \pm0.12 − 19.40 ± 0.12 (− 19.40 ± 0.13 -19.40 \pm0.13 − 19.40 ± 0.13 for DES (DES + legacy). This matches the calibration of our cosmographic IDL to within 0.02 mag. The offset from the SH0ES M B M_B M B therefore remains much the same: it is 4.20 σ 4.20\sigma 4.20 σ (4.26 σ 4.26\sigma 4.26 σ ) for configuration A with DES (DES + legacy). The equivalent result for configuration B is 1.20 σ 1.20\sigma 1.20 σ (1.11 σ 1.11\sigma 1.11 σ ) for DES (DES + legacy). The lower tension is due to wider error bars, since CC datasets have a constraining power that is presently far weaker than the CMB under the assumption of flat Λ \Lambda Λ CDM prior to recombination, an assumption that is not required for CC. The quality of fit to the data remains the same regardless of whether we use flat Λ \Lambda Λ CDM or our cosmographic expansion, with reduced χ 2 = 0.92 \chi^2 = 0.92 χ 2 = 0.92 –0.94 0.94 0.94 . We note that since we have N d a t a = 1555 N_{\mathrm{data}} = 1555 N data = 1555 (1752) data points without (with) the legacy SNIa and the number of model parameters is negligible in comparison, we expect the reduced χ 2 \chi^2 χ 2 to approximately follow a Gaussian distribution centred on 1 with dispersion 0.036 (0.034) in these cases. The reduced χ 2 \chi^2 χ 2 drops from 0.94 to 0.92 upon excluding the legacy SNIa in all analysis variants, indicating a slightly better fit quality without the legacy SNIa.
In our analyses with the tight Planck prior on r d r_d r d (configuration A), the posterior remains the same as the prior because there is no information to shift the absolute scale of H ( z ) H(z) H ( z ) from uncalibrated SNIa or from BAO data that inherit their absolute calibration from r d r_d r d . To relax the assumption of standard physics prior to recombination, we can relax the Planck prior on r d r_d r d and constrain it with CC. This alternative route (configuration B) yields r d = 148.48 ± 8.39 r_d = 148.48 \pm8.39 r d = 148.48 ± 8.39 (148.51 ± 8.68 \pm8.68 ± 8.68 ) cMpc in flat Λ \Lambda Λ CDM with SNIa from DES (DES + legacy). These results are within 0.2 σ 0.2\sigma 0.2 σ of the model-dependent Planck value of r d r_d r d . We note that our results without this prior are not as precise as in Guo et al. (2025), who used a somewhat different approach and a larger number of CC measurements, combined as here with uncalibrated BAO and SNIa. Interestingly, their predicted H 0 = 68. 4 − 0.8 + 1.0 H_0 = 68.4^{+1.0}_{-0.8} H 0 = 68. 4 − 0.8 + 1.0 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 lies 3.6 σ 3.6\sigma 3.6 σ below the central SH0ES value.
Redshift-floor sweep
To check if our results are particularly sensitive to SNIa at the low-redshift end, we truncate the SNIa to only those with z > z f l o o r z > z_{\mathrm{floor}} z > z floor . The top panel of Figure 5 shows how the predicted H 0 H_0 H 0 depends on z f l o o r z_{\mathrm{floor}} z floor . The black line shows no obvious trend with only DES SNIa – all results are ≈ 2.5 σ \approx2.5\sigma ≈ 2.5 σ –3.5 σ 3.5\sigma 3.5 σ above H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA until we get to high z f l o o r z_{\mathrm{floor}} z floor , when the error bars become large due to the small number of SNIa. The predicted H 0 H_0 H 0 is always above the SH0ES value (Breuvall et al. 2024), but the level of tension is below 2 σ 2\sigma 2 σ except at z f l o o r = 0.15 z_{\mathrm{floor}} = 0.15 z floor = 0.15 , where it is 2.1 σ 2.1\sigma 2.1 σ . Importantly, the tension with SH0ES is only 1.03 σ 1.03\sigma 1.03 σ when we consider DES without any redshift floor (or equivalently z f l o o r < 0.073 z_{\mathrm{floor}} < 0.073 z floor < 0.073 ). Thus, the high predicted H 0 H_0 H 0 in our baseline analysis (without any redshift floor) agrees reasonably well with local H 0 H_0 H 0 anchors, which are generally in good agreement with SH0ES (Di Valentino et al. 2025; H0DN Collaboration 2026).
Figure 5. The solid lines show the recovered H 0 H_0 H 0 (top ) and q 0 q_0 q 0 (bottom ) in our k max = 4 k_{\max}=4 k m a x = 4 cosmographic IDL constrained by DES-Dovekie SNIa + DESI DR2 BAO + Planck r d r_d r d , shown as a function of the redshift floor z f l o o r z_{\mathrm{floor}} z floor below which SNIa are excluded and the sample re-fitted. In both panels, we present results with DES (black), DES + legacy (red), and DES + legacy with free recalibration of the legacy SNIa by Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (blue; Eq. (22) ). The dashed vertical lines show the lower limit of each dataset without truncation. The upper limits of z = 0.087 z=0.087 z = 0.087 for the legacy SNIa and z = 1.14 z=1.14 z = 1.14 for DES are not shown here. At redshift floors which reject all legacy SNIa, we show results only for DES. In the top panel, the horizontal dotted lines with 1 σ 1\sigma 1 σ error bands show H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES (Breuval et al. 2024) and H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA (SPT-3G Collaboration [107 ] ), as indicated. In the bottom panel, the horizontal dotted orange line shows q 0 S P A q_0^{\mathrm{SPA}} q 0 SPA .
The bottom panel of Figure 5 shows that the high H 0 H_0 H 0 is associated with a very low deceleration parameter of q 0 ≈ − 3 q_0 \approx-3 q 0 ≈ − 3 . This is expected since the expansion history needs to be similar to that in the Planck cosmology at high redshift, so we can get elevated H 0 H_0 H 0 only with extra acceleration at late times. Similarly to the results for H 0 H_0 H 0 , the black line shows no obvious trend in q 0 q_0 q 0 with z f l o o r z_{\mathrm{floor}} z floor when considering only the homogeneous DES catalogue.
These results change substantially upon including the legacy SNIa at low redshift, as shown by the red curve in each panel of Figure 5 . As we start to include legacy SNIa by decreasing z f l o o r z_{\mathrm{floor}} z floor , the predicted H 0 H_0 H 0 drops sharply until it becomes consistent with H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA . This behaviour is mirrored for q 0 q_0 q 0 , which is far below the SPA value of 3 2 Ω m − 1 = − 0.52 \frac{3}{2}\Omega_m - 1 = -0.52 2 3 Ω m − 1 = − 0.52 at high z f l o o r z_{\mathrm{floor}} z floor , but rises sharply towards and ultimately becomes consistent with it as z f l o o r → 0 z_{\mathrm{floor}} \to0 z floor → 0 . Adding the legacy SNIa considerably reduces the formal uncertainties, so the predicted H 0 H_0 H 0 at the lowest z f l o o r z_{\mathrm{floor}} z floor values lies well below SH0ES and other local H 0 H_0 H 0 anchors. This tension is clearly driven by the legacy SNIa. We note that when we set z f l o o r = 0.073 z_{\mathrm{floor}} = 0.073 z floor = 0.073 in our DES + legacy analysis to include all DES SNIa while minimising the number of legacy SNIa, we end up including 10 legacy SNIa. Adding this handful of legacy SNIa should have a negligible impact on the results, but in fact it decreases H 0 H_0 H 0 by 4.45 4.45 4.45 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which is 1.6 × 1.6\times 1.6 × the uncertainty prior to adding the legacy SNIa. This suggests a direct tension between DES and the legacy SNIa in the narrow redshift range where they overlap.
While we do of course expect the addition of more data – especially at low redshift – to tighten the constraints, such large parameter shifts are not expected on the basis of the stable DES-only results. This is apparent from repeating our redshift-floor sweep test on the 10 4 10^4 1 0 4 mock datasets presented in Appendix D, which we generate assuming correctness of the published BAO and SNIa covariance matrices and analyse similarly to the real data. Despite the scatter between the H 0 ( z f l o o r ) H_0(z_{\mathrm{floor}}) H 0 ( z floor ) curves and a gradual upwards drift due to our cosmographic expansion not being exact, it is difficult to find a mock realisation of Λ \Lambda Λ CDM where the behaviour resembles that of the real data (Figure D2 ). The sharp feature in the DES + legacy result for H 0 ( z f l o o r ) H_0(z_{\mathrm{floor}}) H 0 ( z floor ) at precisely the redshift where the catalogue transitions from legacy SNIa to DES suggests a systematic affecting the legacy SNIa rather than a genuine transition at low redshift. The DES-only result for H 0 ( z f l o o r ) H_0(z_{\mathrm{floor}}) H 0 ( z floor ) is fairly smooth out to z f l o o r = 0.3 z_{\mathrm{floor}}=0.3 z floor = 0.3 , so the sharp feature at z ≈ 0.08 z\approx0.08 z ≈ 0.08 in the DES + legacy case is not expected either in Λ \Lambda Λ CDM mocks or from analysis of the homogeneous DES catalogue.
Figure D2. Top: The results of our redshift-floor sweep for H 0 H_0 H 0 (left panels ) and q 0 q_0 q 0 (right panels ) in our Λ \Lambda Λ CDM mocks where the SNIa redshift distribution matches DES (black) and DES + legacy (red). The true values are shown using horizontal dotted orange lines in the top panels. The thin lines show 500 of the mocks in each sample’s own colour, using a fixed stride through the ensemble. The median of all 10 4 10^4 1 0 4 mocks is shown as a thick solid line without points. The thin solid lines around it in the same colour show the 1 σ 1\sigma 1 σ range using Gaussian equivalent percentiles, while the dotted lines show the 2 σ 2\sigma 2 σ range. The result with the real data, obtained by least squares as for the mocks, is shown using a thick solid line with points. The inset focuses on the region covered by the legacy SNIa. Their redshift floor is shown with a vertical dotted red line, while the DES redshift floor is shown with a vertical dotted black line. We do not show the DES + legacy results beyond the redshift range of the legacy SNIa, as only DES SNIa are available there. Bottom: The uncertainty is shown in two ways, both plotted as ± σ \pm\sigma ± σ error bars around a central value of zero. The vertical scale is the same as the top panels. For clarity, the redshift is dithered symmetrically relative to the actually adopted z f l o o r z_{\mathrm{floor}} z floor . The dotted bars show half the 16th–84th percentile range of the best-fitting value between mocks. The solid bars show the median value across the mocks of the uncertainty returned by our analysis from each mock. Since not all mocks yield the same uncertainty, we use shaded rectangles in the colour of each sample to show the 1 σ 1\sigma 1 σ Gaussian equivalent range in the uncertainty returned by each mock. These are only visible at high z f l o o r z_{\mathrm{floor}} z floor because different mocks give results with a very similar precision.
Bayesian model comparison of late-time solutions
The Bayesian evidence of each considered model (exp and CPL) is shown in Table 3 relative to the Λ \Lambda Λ CDM control, whose uncertainty contributes a systematic component of 0.13 to all the stated ln B 10 \ln B_{10} ln B 10 uncertainties. Without relying on any local H 0 H_0 H 0 anchor, ∣ ln B 10 ∣ ≤ 1.22 \lvert\ln B_{10}\rvert\leq1.22 ∣ ln B 10 ∣ ≤ 1.22 for the exp model, implying at most weak evidence for or against it – though it is generally disfavoured in this case. The results for the CPL model are critically dependent on whether we include the legacy SNIa. Without them, CPL is disfavoured by 1.79 natural units (nats). Including them at face value leads to a 1.70 nat preference for CPL, but including them and allowing an arbitrary recalibration reverts this to a 1.39 nat preference against CPL. These results clarify that without local H 0 H_0 H 0 anchors, the Bayesian evidence for CPL largely comes from adding the legacy SNIa at face value, which we have argued elsewhere is most likely not appropriate (cf. Ormondroyd et al. 2026).
Table 3. Table 3. Parameter constraints for the exp (Eq. (12) ) and CPL (Eq. (14) ) models and results of Bayesian model comparison against Λ \Lambda Λ CDM. A positive ln B 10 \ln B_{10} ln B 10 value for a model means it is preferred over Λ \Lambda Λ CDM. All relative ln B \ln B ln B uncertainties include a systematic contribution of 0.13 from the Λ \Lambda Λ CDM analysis. We use the DES-Dovekie + DESI DR2 BAO + CC + CMB + BBN constraint in all cases (§ 3.3 ). We also consider the SH0ES anchor (Breuval et al. 2024) or the three H 0 H_0 H 0 anchors that are independent of the SNIa absolute magnitude, imposed jointly and labelled “Excl. SNIa”. We use three different cases for the DES-Dovekie sample: DES alone, DES + legacy, and DES + legacy with a magnitude offset of Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy on the legacy SNIa. The w 0 / A w_0/A w 0 / A column shows w 0 w_0 w 0 for CPL and A A A for the exp model, while the w a / z 1 w_a/z_1 w a / z 1 column shows w a w_a w a for CPL and z 1 z_1 z 1 for the exp model. Missing model parameters are marked with a dash, while “Unc.” indicates parameters that are largely unconstrained by the data, with a posterior spanning essentially the whole prior range (Table 1 ). Uncertainties follow the convention of Table 2 , and one-sided limits enclose 68% of the posterior. The triangle plots corresponding to these results are shown in Appendix C.
The predicted H 0 H_0 H 0 in the CPL model (without H 0 H_0 H 0 anchors) is always similar to H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA , but this is true in the exp model only if we include the legacy SNIa at face value. If we neglect them altogether or add them with arbitrary recalibration, the predicted H 0 H_0 H 0 is almost 69 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which given the uncertainties is marginally compatible with H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES (e.g. DES alone gives H 0 = 68.83 ± 2.02 H_0=68.83\pm2.02 H 0 = 68.83 ± 2.02 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which is within 2 σ 2\sigma 2 σ of SH0ES). It is therefore possible to combine local H 0 H_0 H 0 anchors with the other data only for the exp model and even then only if the legacy SNIa are excluded or recalibrated. We cannot add local H 0 H_0 H 0 anchors if we use the CPL model, regardless of whether we add the legacy SNIa. If we take these at face value, we cannot add local H 0 H_0 H 0 anchors for either model – or indeed for any model with standard r d r_d r d .
The CPL model was not proposed as a solution to the Hubble tension, but rather as a way to better fit available BAO and uncalibrated SNIa data consistently with the CMB. We find that when constraining the CPL model with DES + legacy, ( w 0 , w a ) (w_0,w_a) ( w 0 , w a ) is in 3.36 σ 3.36\sigma 3.36 σ tension with the Λ \Lambda Λ CDM limit, which we obtain from the improvement in the fit when w 0 w_0 w 0 and w a w_a w a are freed (Δ χ 2 = 14.32 \Delta\chi^2=14.32 Δ χ 2 = 14.32 ), converted to a Gaussian significance for two additional parameters. However, this w 0 w_0 w 0 –w a w_a w a tension gets reduced to 2.13 σ 2.13\sigma 2.13 σ upon removing the legacy SNIa, while allowing an arbitrary recalibration of them relative to DES leads to 2.21 σ 2.21\sigma 2.21 σ tension. Both cases are plausibly consistent with Λ \Lambda Λ CDM. The departure from Λ \Lambda Λ CDM expectations claimed by DESI Collaboration (2025) is therefore reliant on the legacy SNIa (Ormondroyd et al. 2026). Indeed, the actual BAO observables (angular scale and redshift depth) can be fit jointly with the CMB at an acceptable total χ 2 \chi^2 χ 2 if SNIa are neglected (Banik et al. 2026a). On the other hand, their work found that Λ \Lambda Λ CDM faces a 3.07 σ 3.07\sigma 3.07 σ CMB–BAO tension in the ( H 0 r d , Ω m ) (H_0r_d,\Omega_m) ( H 0 r d , Ω m ) parameter space without considering SNIa. Combined with our result in Figure 2 , this is suggestive of a growing anomaly between CMB-calibrated Λ \Lambda Λ CDM predictions and observations as we move towards lower redshift.
When adding the SNIa-independent H 0 H_0 H 0 anchors, the exp model is disfavoured compared to Λ \Lambda Λ CDM if we add the legacy SNIa at face value, but this dataset combination is not valid because the legacy SNIa cannot be reconciled with local H 0 H_0 H 0 measurements in any plausible model with standard r d r_d r d . In the other cases, the non-SHOES anchors on H 0 H_0 H 0 lead to a preference for the exp model, especially for DES alone (+1.54 nats). In this case, the inferred H 0 = 70.0 0 − 4.98 + 4.78 H_0=70.00^{+4.78}_{-4.98} H 0 = 70.0 0 − 4.98 + 4.78 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , consistent with a full solution to the Hubble tension when judged by H ( 0 ) H(0) H ( 0 ) . However, the excess expansion rate in this model decays on a scale z 1 < 0.086 z_1<0.086 z 1 < 0.086 , so much of it lies below the Hubble-flow window of SH0ES. The H 0 H_0 H 0 that SH0ES would measure in this model is therefore only H 0 w i n = 70.23 ± 0.99 H_0^{\mathrm{win}}=70.23\pm0.99 H 0 win = 70.23 ± 0.99 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which is 2.25 σ 2.25\sigma 2.25 σ below the published value (Table 4 ). This model thus reduces the Hubble tension substantially, from 6.4 σ 6.4\sigma 6.4 σ for SPA, but it does not remove it. With DES + legacy allowing recalibration, the free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy raises H 0 H_0 H 0 to 69.7 5 − 0.98 + 1.30 69.75^{+1.30}_{-0.98} 69.7 5 − 0.98 + 1.30 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which is still short of a full solution, since SH0ES would measure only H 0 w i n = 69.61 ± 0.76 H_0^{\mathrm{win}}=69.61\pm0.76 H 0 win = 69.61 ± 0.76 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 in this model, 3.10 σ 3.10\sigma 3.10 σ below the published value. The DES-only case gives a low inferred z 1 < 0.086 z_1<0.086 z 1 < 0.086 at the 1 σ 1\sigma 1 σ confidence level (C.L.), which suggests that the solution is localised at quite low redshift (as also found by Nájera et al. 2026). We discuss this further in § 4.6 .
Table 4. Impact of the z 1 z_1 z 1 prior on the exp model results. We repeat each analysis in Table 3 using a uniform prior on z 1 z_1 z 1 or on ln z 1 \ln z_1 ln z 1 , the latter being more representative of a local structure due to its unknown size (see the text). All results shown here are relative to the Λ \Lambda Λ CDM control. For clarity, we omit the numerical uncertainty on ln B 10 \ln B_{10} ln B 10 of ± 0.19 \pm0.19 ± 0.19 . To make the results with the two z 1 z_1 z 1 priors more comparable, we show the mean and standard deviation of A A A and H 0 = H 0 , Λ C D M + A H_0 = H_{0,\Lambda\mathrm{CDM}} + A H 0 = H 0 , Λ CDM + A . H 0 w i n H_0^{\mathrm{win}} H 0 win is the H 0 H_0 H 0 that SH0ES would infer if the Universe followed each posterior, i.e. the kinematic fit of Section 3.3.3 applied over the redshifts of the SH0ES Hubble-flow SNIa (z = 0.023 z = 0.023 z = 0.023 –0.149 0.149 0.149 ).
Only when we add the SH0ES anchor on H 0 H_0 H 0 do we get a strong preference for the exp model, with ln B 10 = + 10.61 \ln B_{10} = +10.61 ln B 10 = + 10.61 for DES and + 7.87 +7.87 + 7.87 for DES + legacy with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy . In these cases, the model comes close to reproducing the SH0ES measurement itself, with H 0 w i n = 71.93 ± 0.67 k m s − 1 M p c − 1 H_0^{\mathrm{win}} = 71.93 \pm0.67\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 win = 71.93 ± 0.67 km s − 1 Mp c − 1 for DES and 71.05 ± 0.75 k m s − 1 M p c − 1 71.05 \pm0.75\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 71.05 ± 0.75 km s − 1 Mp c − 1 for DES + legacy with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy , i.e. 1.14 σ 1.14\sigma 1.14 σ and 1.86 σ 1.86\sigma 1.86 σ below SH0ES. Their H 0 H_0 H 0 is higher (75.49 ± 1.57 75.49 \pm1.57 75.49 ± 1.57 and 72.30 ± 1.58 k m s − 1 M p c − 1 72.30 \pm1.58\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 72.30 ± 1.58 km s − 1 Mp c − 1 ) because the excess expansion rate peaks at z = 0 z = 0 z = 0 and has largely decayed across the SH0ES window. We stress that the anchor is now part of the fit, so these values are not predictions. But even with the strong SH0ES anchor, the DES + legacy case (without Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy ) still disfavours the exp model relative to Λ \Lambda Λ CDM by 1.78 nats and predicts a low H 0 = 68.84 ± 0.70 k m s − 1 M p c − 1 H_0 = 68.84 \pm0.70\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 68.84 ± 0.70 km s − 1 Mp c − 1 .
The CPL model is at most weakly favoured by the Bayesian evidence if we add the non-SH0ES local anchors on H 0 H_0 H 0 , with the best case being for DES + legacy without recalibration (0.87 nat preference for CPL). Adding the SH0ES anchor leads to a clear preference for the CPL model, which is 4.06 nats for DES alone. This is far weaker than the 10.61 nat preference for the exp model over Λ \Lambda Λ CDM with the same dataset combination. Moreover, the inferred H 0 H_0 H 0 in the CPL + SH0ES case barely exceeds 70 k m s − 1 M p c − 1 \mathrm{km\ s^{-1}\ Mpc^{-1}} km s − 1 Mp c − 1 and has a small uncertainty, which still leads to considerable tension with SH0ES if
we discard the legacy SNIa or try to rectify potential problems with them by allowing an arbitrary recalibration. The low H 0 H_0 H 0 problem is worsened if we add the legacy SNIa at face value, since then the inferred H 0 = 68.87 ± 0.48 k m s − 1 M p c − 1 H_0 = 68.87 \pm0.48\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 68.87 ± 0.48 km s − 1 Mp c − 1 , far below H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES . We attribute this to any deviation from Λ \Lambda Λ CDM in the CPL model declining too gradually with redshift – the linear dependence of w w w upon a a a prevents a rapid convergence to − 1 -1 − 1 . The CPL model also has a theoretical problem in that matching the local H 0 H_0 H 0 anchor requires w < − 1 w < -1 w < − 1 today and/or in the recent past. This implies a phantom fluid, violating the weak energy condition (Sen & Scherrer 2008; Lewis & Chamberlain 2025).
Log-uniform prior on the exp model z 1 z_1 z 1
Our analysis implicitly assumes that the solution to the Hubble tension lies at the background level. In this case, since it is not known a priori when deviations from Λ \Lambda Λ CDM would start to set in, a uniform prior on the exp model z 1 z_1 z 1 is appropriate. But if the Hubble tension is actually caused by a local underdensity or void (Haslbauer et al. 2020; Banik et al. 2026a,c), then z 1 z_1 z 1 provides an indirect measure of the void size, since the gravitational impact of a void would persist even after the density has returned to the cosmic mean (Banik & Kalaitzidis 2025). Even so, a smaller void can generally be expected to show up in our exp model analysis as a lower z 1 z_1 z 1 . Because we do not know a priori how large a void we might be in, a scale-free prior is arguably a better choice. We therefore revisit the above exp model results using a uniform prior on ln z 1 \ln z_1 ln z 1 , while keeping the limits unchanged. We note that a model with z 1 < 0.001 z_1 < 0.001 z 1 < 0.001 cannot realistically be expected to solve the Hubble tension because this would imply essentially no change to the Λ \Lambda Λ CDM distance–redshift relation in the SH0ES range (z = 0.023 z = 0.023 z = 0.023 –0.15 0.15 0.15 ; Riess et al. 2022). The upper limit is less important because we will see that the results imply a low z 1 z_1 z 1 .
Table 4 shows the impact of changing the z 1 z_1 z 1 prior from U ( z 1 ) → U ( ln z 1 ) \mathcal{U}(z_1) \rightarrow\mathcal{U}(\ln z_1) U ( z 1 ) → U ( ln z 1 ) . We begin with the case without a local H 0 H_0 H 0 anchor, so the inferred H 0 H_0 H 0 can be considered a genuine prediction of the IDL approach constrained by SNIa + BAO + CC + CMB + BBN. With SNIa from DES alone, A A A remains very consistent with zero for both priors, though the uncertainty rises from 2 to 5 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 when going from the uniform prior to the log-uniform prior. These changes directly affect the predicted H 0 H_0 H 0 , which is 68.83 ± 2.02 68.83 \pm2.02 68.83 ± 2.02 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with the uniform prior and 68.93 ± 5.02 68.93 \pm5.02 68.93 ± 5.02 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with the log-uniform prior. While these results are more in line with A = 0 A = 0 A = 0 and H 0 = H 0 S P A H_0 = H_0^{\mathrm{SPA}} H 0 = H 0 SPA , they cannot exclude the possibility that A ≈ 6 A \approx6 A ≈ 6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 and H 0 = H 0 S H 0 E S H_0 = H_0^{\mathrm{SH0ES}} H 0 = H 0 SH0ES —especially with the log-uniform prior. The picture remains similar upon including the legacy SNIa with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy , which again shows no preference for A > 0 A > 0 A > 0 . The Bayesian evidences are also not appreciably different from that of Λ \Lambda Λ CDM.
The results are quite different with DES + legacy regardless of the prior choice. We infer A < 0 A < 0 A < 0 with either z 1 z_1 z 1 prior. In particular, A = − 2.17 ± 1.65 A = -2.17 \pm1.65 A = − 2.17 ± 1.65 (− 4.20 ± 2.89 -4.20 \pm2.89 − 4.20 ± 2.89 ) km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with the uniform prior on z 1 z_1 z 1 (ln z 1 \ln z_1 ln z 1 ). Although the Bayesian evidence no longer disfavours the exp model and actually prefers it with the log-uniform prior, this reflects a preference for A < 0 A < 0 A < 0 , which is strongly incompatible with the A ≈ + 6 A \approx+6 A ≈ + 6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 required to solve the Hubble tension. This is evident in that the inferred H 0 H_0 H 0 is 66.21 ± 1.70 66.21 \pm1.70 66.21 ± 1.70 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with the uniform z 1 z_1 z 1 prior, which is lower than the 67.94 ± 0.24 67.94 \pm0.24 67.94 ± 0.24 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 we obtain for flat Λ \Lambda Λ CDM from the same data. With the log-uniform prior on z 1 z_1 z 1 , the inferred H 0 H_0 H 0 drops to just 63.92 ± 2.92 63.92 \pm2.92 63.92 ± 2.92 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . In both cases, the results are considerably below the SH0ES measurement. We note that a sharp reduction in predicted H 0 H_0 H 0 upon including the legacy SNIa is also evident from our sweep of the redshift floor using a model-independent cosmographic expansion (Figure 5 ). We are able to recover this feature in the context of our exp model due to it allowing for a steep feature in the expansion history at z ≲ z 1 z \lesssim z_1 z ≲ z 1 . Despite the preference for A < 0 A < 0 A < 0 , we are still not able to meaningfully constrain z 1 z_1 z 1 with DES + legacy if we adopt a uniform prior. With a log-uniform prior, we obtain z 1 < 0.019 z_1 < 0.019 z 1 < 0.019 at the 1 σ 1\sigma 1 σ C.L., which is considerably tighter than the z 1 < 0.070 z_1 < 0.070 z 1 < 0.070 given by the prior alone (Table 4 ). This shows that while z 1 z_1 z 1 remains essentially unconstrained with the uniform prior, the log prior suggests a rather low z 1 z_1 z 1 . Since A A A is still plausibly consistent with zero, we suggest that there is no clear preference for the exp model when we include the legacy SNIa at face value.
Once we impose a local H 0 H_0 H 0 anchor, the inferred H 0 H_0 H 0 is no longer a prediction, so the important quantity is now the Bayesian evidence. We find that using a log-uniform prior on z 1 z_1 z 1 increases ln B 10 \ln B_{10} ln B 10 in nearly all cases, typically by about 0.7. The increases exceed the numerical uncertainty of ± 0.19 \pm0.19 ± 0.19 , often by a considerable amount. The only decrease occurs if we combine the SH0ES anchor with DES + legacy allowing a free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy , in which case ln B 10 \ln B_{10} ln B 10 decreases by just 0.06. The systematic increase in the Bayesian evidence in the remaining cases is caused by a log-uniform prior increasing the prior weight on low values of z 1 z_1 z 1 , which are preferred even for a uniform prior. As a result, the DES-only case with SNIa-independent anchors shifts from ln B 10 = + 1.54 \ln B_{10} = +1.54 ln B 10 = + 1.54 to + 2.53 +2.53 + 2.53 , which is now a clear preference for the exp model. This is related to a corresponding rise in H 0 H_0 H 0 from 72.53 72.53 72.53 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 to 74.23 74.23 74.23 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . The Bayesian evidence for the DES-only case with the strong SH0ES anchor strengthens from + 10.61 +10.61 + 10.61 to + 11.26 +11.26 + 11.26 , while A A A rises from 8.00 ± 1.46 8.00 \pm1.46 8.00 ± 1.46 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 to 8.25 ± 1.32 8.25 \pm1.32 8.25 ± 1.32 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . In the DES + legacy case with the SH0ES anchor, changing from a uniform prior to a log-uniform prior reduces H 0 H_0 H 0 from 68.84 68.84 68.84 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 to 68.28 68.28 68.28 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , demonstrating that even the tight SH0ES anchor cannot pull the legacy-dominated fit up to H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES regardless of the z 1 z_1 z 1 prior.
Our overall conclusions regarding the exp model do not change much depending on whether we use a uniform prior on z 1 z_1 z 1 or on ln z 1 \ln z_1 ln z 1 . A late-time solution with A ≈ 6 A \approx6 A ≈ 6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 is preferred only when we impose a local H 0 H_0 H 0 anchor and do not consider the legacy SNIa, though there is a weak preference if we include them allowing recalibration. Without any local H 0 H_0 H 0 anchor, the data are consistent with a late-time solution to the Hubble tension, but do not prefer the high required A A A . With the strong SH0ES anchor and SNIa only from DES, we infer that A ≈ 8 A \approx8 A ≈ 8 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 and z 1 ≈ 0.03 z_1 \approx0.03 z 1 ≈ 0.03 with either z 1 z_1 z 1 prior. We stress, however, that our prior on A A A bounds both values. Since most of the SH0ES Hubble-flow SNIa lie at z > z 1 z > z_1 z > z 1 , the distances they measure respond mainly to the integrated excess ∫ Δ H d z = A z 1 \int\Delta H\,\mathrm{d}z = A z_1 ∫ Δ H d z = A z 1 rather than to A A A and z 1 z_1 z 1 separately. The likelihood therefore has a ridge along which A A A rises as z 1 z_1 z 1 falls, with A z 1 = 0.2 9 − 0.08 + 0.12 A z_1 = 0.29^{+0.12}_{-0.08} A z 1 = 0.2 9 − 0.08 + 0.12 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . The upper edge of our prior at A = 10 A = 10 A = 10 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 truncates this ridge, and 31% of the posterior mass lies at A > 9 A > 9 A > 9 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (36% with the log-uniform prior on z 1 z_1 z 1 ). A wider prior would thus raise A A A and lower z 1 z_1 z 1 . It would also change ln B 10 \ln B_{10} ln B 10 , both through the larger prior volume, which lowers it by ln 2 \ln2 ln 2 for each doubling of the range of A A A , and through the additional likelihood along the ridge, which raises it.
AP consistency test
Most of our results are heavily influenced by the tight Planck prior on r d r_d r d . As discussed in § 3.5 , the AP test avoids this dependency by using the AP ratio F A P = D M / D H F_{\mathrm{AP}} = D_{\mathrm{M}}/D_{\mathrm{H}} F AP = D M / D H . In Appendix A, we apply the AP test to check whether the shape of the expansion history is consistent between SNIa and BAO at three redshifts where particularly tight BAO results are available and there is good SNIa coverage. We find consistency for both DES and DES + legacy, with the p p p -value (likelihood of a higher χ 2 \chi^2 χ 2 ) being 9.7% and 10.5%, respectively. There is a mild ≈ 2 σ \approx2\sigma ≈ 2 σ deficit in both cases at z = 0.510 z = 0.510 z = 0.510 , while at z = 0.934 z = 0.934 z = 0.934 , there is a mild excess of 2.0 σ 2.0\sigma 2.0 σ for DES and 1.8 σ 1.8\sigma 1.8 σ for DES + legacy. However, there is excellent agreement at z = 0.706 z = 0.706 z = 0.706 . The acceptable p p p -values indicate that fitting our cosmographic expansion to SNIa predicts F A P F_{\mathrm{AP}} F AP values that are in line with BAO measurements regardless of r d r_d r d , justifying our approach of combining SNIa and BAO datasets assuming the Etherington relation. However, there is some tension between the DESI DR2 BAO measurements at z = 0.706 z = 0.706 z = 0.706 and previous BAO results at a similar redshift (section 3.6 of Banik et al. 2026a). It will therefore be important to revisit these results with future BAO measurements.
Systematics with the legacy SNIa
Adding the legacy SNIa to the homogeneous DES sample shifts the inferred cosmological parameters in a manner that would not be expected from the latter (Figure 4 ). This might be due to systematics with the legacy SNIa. We test this possibility by adding an offset parameter Δ M l e g a c y ∼ U ( − 1 , 1 ) \Delta M_{\mathrm{legacy}} \sim\mathcal{U}(-1,1) Δ M legacy ∼ U ( − 1 , 1 ) mag to the absolute magnitude of the legacy SNIa, which therefore becomes as follows:
M l e g a c y = M D E S + Δ M l e g a c y . (22) M_{\mathrm{legacy}} = M_{\mathrm{DES}} + \Delta M_{\mathrm{legacy}}. \tag*{(22)} M legacy = M DES + Δ M legacy . (22)
A higher Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy corresponds to intrinsically fainter legacy SNIa, which for fixed apparent magnitude implies that they must be closer to us. Since the redshifts are spectroscopically confirmed, this would lead to a higher H 0 H_0 H 0 .
Our results for Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy are shown in Figure 6 for the four considered cases discussed in § 3.6 , namely cosmography, cosmography + AP, Λ \Lambda Λ CDM, and Λ \Lambda Λ CDM + AP. We consistently infer an offset of Δ M l e g a c y = + 0.05 \Delta M_{\mathrm{legacy}} = +0.05 Δ M legacy = + 0.05 to + 0.07 +0.07 + 0.07 mag, with the detection significance ranging from 1.2 σ 1.2\sigma 1.2 σ for the more flexible cosmographic expansion, to 2.9 σ 2.9\sigma 2.9 σ for the less flexible Λ \Lambda Λ CDM parametrisation with the AP consistency requirement. All four analyses agree on the magnitude of the offset, but differences in the uncertainty create differences in how confident each analysis can be that Δ M l e g a c y > 0 \Delta M_{\mathrm{legacy}} > 0 Δ M legacy > 0 . In our baseline analysis where we use a cosmographic IDL and infer Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy alongside the other parameters, we find that Δ M l e g a c y = + 0.0646 ± 0.0391 \Delta M_{\mathrm{legacy}} = +0.0646 \pm0.0391 Δ M legacy = + 0.0646 ± 0.0391 mag, a 1.65 σ 1.65\sigma 1.65 σ offset (cf. Figure 5 ).
Figure 6. The relative offset Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (Eq. (22) ) of the legacy SNIa from the DES-Dovekie relation extrapolated to z → 0 z \to0 z → 0 , for the four analyses described in § 3.6 . The vertical grey band shows the ≈ 0.04 \approx0.04 ≈ 0.04 mag offset independently reported by Efstathiou (2025).
To gain further insight into how a calibration offset might have arisen, we split the legacy SNIa into two subsamples and infer a separate Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy for each. Our aim is to find if any particular subset of the legacy SNIa has a calibration offset Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy that differs from the offset Δ M l e g a c y r e s t \Delta M_{\mathrm{legacy}}^{\mathrm{rest}} Δ M legacy rest that we obtain for the remaining legacy SNIa. Figure 7 shows Δ M l e g a c y − Δ M l e g a c y r e s t \Delta M_{\mathrm{legacy}} - \Delta M_{\mathrm{legacy}}^{\mathrm{rest}} Δ M legacy − Δ M legacy rest for a few possible splits. The legacy SNIa largely come from Foundation or from CfA. The former have a positive Δ M l e g a c y − Δ M l e g a c y r e s t \Delta M_{\mathrm{legacy}} - \Delta M_{\mathrm{legacy}}^{\mathrm{rest}} Δ M legacy − Δ M legacy rest . Since ‘rest’ in this context largely means CfA, this behaviour is mirrored in that case, explaining why the CfA SNIa have a negative Δ M l e g a c y − Δ M l e g a c y r e s t \Delta M_{\mathrm{legacy}} - \Delta M_{\mathrm{legacy}}^{\mathrm{rest}} Δ M legacy − Δ M legacy rest by almost the same amount and with a similar uncertainty. If the offset is survey-specific, the underlying issue may be the completeness correction for each survey or the conversion between different photometric filter
Figure 7. Test of whether the legacy offset depends on which subset of the legacy SNIa we consider, based on splitting the legacy SNIa into two subsamples according to the distinguishing feature indicated at the left (e.g., the first row splits off the Foundation SNIa from the remaining legacy SNIa). The values show Δ M l e g a c y − Δ M l e g a c y r e s t \Delta M_{\mathrm{legacy}}-\Delta M_{\mathrm{legacy}}^{\mathrm{rest}} Δ M legacy − Δ M legacy rest , the inferred offset for the subsample of interest and the rest of the legacy SNIa, respectively. We jointly infer both offset parameters. Note this does not test whether the legacy SNIa are offset from the main DES sample (which is shown in Figure 6 ), but rather whether the offset depends on subsample within the legacy set.
systems, rather than any problem with the legacy SNIa observations in their original passband.
We also check for any redshift-dependent trend in Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy by splitting the legacy SNIa in redshift. A trend here might arise due to the greater importance of peculiar velocity at low redshift, while survey incompleteness may be an issue towards the high-redshift end of the z = 0.025 z = 0.025 z = 0.025 –0.087 0.087 0.087 range covered by the legacy SNIa. We therefore split the legacy SNIa into subsamples below and above a redshift threshold of 0.05 or 0.073, the lowest redshift SNIa in the main DES sample. Figure 7 shows that in both cases, Δ M l e g a c y − Δ M l e g a c y r e s t \Delta M_{\mathrm{legacy}} - \Delta M_{\mathrm{legacy}}^{\mathrm{rest}} Δ M legacy − Δ M legacy rest is very consistent with 0. We therefore conclude that the difference between the main DES sample and the legacy SNIa is not driven by any particular redshift bin.
The legacy SNIa calibration offset is directly related to H 0 H_0 H 0 because intrinsically fainter SNIa must be closer to have the same apparent magnitude, leading to a higher H 0 H_0 H 0 . We explore what impact this might have on the IDL using the approach described in § 3.6 . The results are presented in Figure 8 and summarised in Table 5 . For our cosmographic expansion, allowing Δ M l e g a c y ≠ 0 \Delta M_{\mathrm{legacy}}\ne0 Δ M legacy = 0 increases H 0 H_0 H 0 from 67.98 ± 0.69 k m s − 1 M p c − 1 67.98\pm0.69\ {\rm km\,s^{-1}\,Mpc^{-1}} 67.98 ± 0.69 km s − 1 Mp c − 1 to 71.03 ± 1.99 k m s − 1 M p c − 1 71.03\pm1.99\ {\rm km\,s^{-1}\,Mpc^{-1}} 71.03 ± 1.99 km s − 1 Mp c − 1 , which is now only 0.99 σ 0.99\sigma 0.99 σ below SH0ES. Allowing the Foundation and CfA/CSP legacy SNIa to have different offsets leaves this result almost unchanged, with H 0 = 70.91 ± 2.00 k m s − 1 M p c − 1 H_0=70.91\pm2.00\ {\rm km\,s^{-1}\,Mpc^{-1}} H 0 = 70.91 ± 2.00 km s − 1 Mp c − 1 . This shows that while the shape suggests a 1 σ 1\sigma 1 σ calibration difference between these legacy datasets, a flexible cosmographic expansion along with a single legacy offset parameter absorbs nearly all the difference with H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA . The full IDL thus gains little from allowing separate offset parameters for Foundation and CfA/CSP.
Figure 8. Top: H 0 H_0 H 0 inferred from the IDL. Circles show results from our cosmographic analysis with k max = 4 k_{\max}=4 k m a x = 4 (Eq. (7) ). Squares show results using flat Λ \Lambda Λ CDM (Eq. (9) ). The vertical shaded bands show H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA (SPT-3G Collaboration [107 ] ) and H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES (Breuval et al. 2024), as indicated in the legend. The top row excludes legacy SNIa, while the middle row includes them without correction (Δ M l e g a c y = 0 \Delta M_{\mathrm{legacy}}=0 Δ M legacy = 0 ). The bottom row includes the legacy SNIa allowing free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy . Bottom: Our inferred Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy when we sample it jointly with r d r_d r d and M D E S M_{\mathrm{DES}} M DES , assuming our cosmographic IDL. The vertical grey band shows the range reported by Efstathiou (2025). Both panels adopt the Planck 2020 prior on r d r_d r d .
Table 5. Inferred H 0 H_0 H 0 in our baseline analysis (SNIa + DESI DR2 BAO + Planck 2020 r d r_d r d ) with the legacy SNIa excluded, included without correction, or included after correction by Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy to align them with the DES relation extrapolated to z → 0 z \to0 z → 0 . The final row shows Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy .
The preference for H 0 > H 0 S P A H_0>H_0^{\mathrm{SPA}} H 0 > H 0 SPA is still discernible in flat Λ \Lambda Λ CDM, but it is far weaker due to the less flexible expansion history (Eq. (9) ). In this case, H 0 H_0 H 0 rises from 68.60 ± 0.46 k m s − 1 M p c − 1 68.60\pm0.46\ {\rm km\,s^{-1}\,Mpc^{-1}} 68.60 ± 0.46 km s − 1 Mp c − 1 without Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy to 69.02 ± 0.48 k m s − 1 M p c − 1 69.02\pm0.48\ {\rm km\,s^{-1}\,Mpc^{-1}} 69.02 ± 0.48 km s − 1 Mp c − 1 with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy . The very small difference arises because fixing both the shape of the expansion history and its normalisation assuming Λ \Lambda Λ CDM necessarily implies a Planck-like H 0 H_0 H 0 regardless of whether the legacy SNIa can be recalibrated. However, assuming Λ \Lambda Λ CDM at late times automatically excludes the possibility of solving the Hubble tension at late times. The fact that H 0 > H 0 S P A H_0>H_0^{\mathrm{SPA}} H 0 > H 0 SPA even in this case may be hinting at a late-time solution.
Our results suggest that a calibration offset with the legacy SNIa contributes ≃ 3 k m s − 1 M p c − 1 \simeq3\ {\rm km\,s^{-1}\,Mpc^{-1}} ≃ 3 km s − 1 Mp c − 1 to the H 0 H_0 H 0 gap between local anchors and the IDL including legacy SNIa. Allowing these to be recalibrated relative to the homogeneous DES sample relaxes the tension with SH0ES from 4.69 σ 4.69\sigma 4.69 σ to just 0.99 σ 0.99\sigma 0.99 σ given the uncertainty in Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy . Averaging over the truncation order rather than selecting k max = 4 k_{\max}=4 k m a x = 4 raises this to 1.43 σ 1.43\sigma 1.43 σ (Appendix B2). Remarkably, the H 0 H_0 H 0 value recovered by the IDL using DES without legacy SNIa is very similar to the standard distance ladder result obtained by anchoring the DESI Fundamental Plane using the distance to the Coma Cluster (Scolnic et al. 2025) or with a broader range of anchors (Carr et al. 2026). This agreement survives and in the latter case even improves upon averaging over the cosmographic order k max k_{\max} k m a x , which gives H 0 = 74.77 ± 4.83 k m s − 1 M p c − 1 H_0=74.77\pm4.83\ {\rm km\,s^{-1}\,Mpc^{-1}} H 0 = 74.77 ± 4.83 km s − 1 Mp c − 1 (Table B1 ). In this case the Bayesian evidence assigns 20 % 20\% 20% of the probability to k max = 5 k_{\max}=5 k m a x = 5 , which gives a lower and more uncertain result for H 0 H_0 H 0 , as discussed next.
Table B1. Sweep over the truncation order k max k_{\max} k m a x of our cosmographic expansion (Eq. (7) ) in configuration A. The column w k w_k w k is the weighted posterior probability of order k k k (Eq. (10) ). The χ min 2 \chi^2_{\min} χ m i n 2 column is the minimum χ 2 \chi^2 χ 2 at fixed k max k_{\max} k m a x without any prior on the polynomial coefficients. Δ χ 2 \Delta\chi^2 Δ χ 2 is measured from the row with highest evidence (shown in bold), so the difference between − Δ χ 2 / 2 -\Delta\chi^2/2 − Δ χ 2 /2 and Δ ln Z \Delta\ln Z Δ ln Z is the change in the Occam penalty factor between the two orders. The third block repeats the DES + legacy analysis with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (§ 3.6 ). The Λ \Lambda Λ CDM row of each block is our flat Λ \Lambda Λ CDM fit (Eq. (9) ), shown for reference only. Every row quotes the posterior mean and standard deviation, so a few entries differ slightly from the mode-based values in Table 2 . The row labelled “BMA” is the Bayesian model average over k max = 2 k_{\max}=2 k m a x = 2 –6 (Eq. (27) ). Its uncertainty contains both the width of each posterior and the spread of the central values between orders. At k max = 4 k_{\max}=4 k m a x = 4 , the nested sampling reproduces the NUTS results of Table 2 to within 0.13 standard deviations in H 0 H_0 H 0 . In the DES + legacy blocks, k max = 4 k_{\max}=4 k m a x = 4 and 5 carry almost all of the weight and place q 0 q_0 q 0 on opposite sides of zero, so the BMA q 0 q_0 q 0 is bimodal. We therefore show its 1 σ 1\sigma 1 σ highest posterior density region as the union of two intervals.
Robustness tests
The most striking feature of the results in Table 2 is the substantial difference between the H 0 H_0 H 0 predictions of our cosmographic IDL constrained by DES and by DES + legacy, which we denote H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy , respectively. The poorer low-redshift coverage with DES alone forces our cosmographic expansion to extrapolate across the z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 regime, though it has SNIa down to z = 0.073 z = 0.073 z = 0.073 . Since our cosmographic expansion does not exactly match the true H ( 1 − a ) H(1-a) H ( 1 − a ) for any finite k max k_{\max} k m a x , it inevitably introduces some systematic error into the predicted H 0 H_0 H 0 , even in the absence of measurement errors.
This can be investigated using mock data. The size of the effect depends on the unknown true expansion history; for illustration we quantify it in the flat ΛCDM cosmology that best fits the CMB (SPT-3G Collaboration [107 ] ). We set up 10 4 10^4 1 0 4 mock datasets assuming this SPA cosmology, with the mock SNIa having the same redshift distribution and magnitude uncertainties as the real SNIa in DES or DES + legacy (further details are provided in Appendix D). The difference Δ H 0 ≡ H 0 D E S − H 0 D E S + l e g a c y \Delta H_0 \equiv H_0^{\mathrm{DES}}-H_0^{\mathrm{DES+legacy}} Δ H 0 ≡ H 0 DES − H 0 DES + legacy exceeds the observed 8.27 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 in only 1.24% of the mocks (Figure D1 ). This corresponds to a Gaussian-equivalent significance of 2.24 σ 2.24\sigma 2.24 σ for a one-tailed test, which is appropriate because the sign of the Hubble tension is well known, so the ‘alternative hypothesis’ to ΛCDM would be that Δ H 0 ≈ 6 \Delta H_0 \approx6 Δ H 0 ≈ 6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 and the legacy SNIa are subject to systematics. The observed Δ H 0 \Delta H_0 Δ H 0 is therefore unlikely to arise from random fluctuations in a ΛCDM cosmology coupled with extrapolation error due to the limited DES coverage at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 .
Figure D1. Top: Joint distribution of H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy , inferred from applying our fourth-order cosmographic IDL analysis to 10 4 10^4 1 0 4 mock datasets assuming Λ \Lambda Λ CDM with the SPA parameters ([107 ] ). The true H 0 H_0 H 0 is indicated using solid lines. The contours show the 1 σ 1\sigma 1 σ and 2 σ 2\sigma 2 σ regions, outside which individual points are shown with higher opacity. The dashed grey line is the line of equality. The dotted red line is offset upwards by the observed Δ H 0 = 8.27 k m s − 1 M p c − 1 \Delta H_0=8.27\ \mathrm{km\,s^{-1}\,Mpc^{-1}} Δ H 0 = 8.27 km s − 1 Mp c − 1 so it intersects the observed point, shown here as the red star. This is obtained using NUTS, but the open blue diamond shows that the corresponding least-squares result (labelled ‘LS’) is almost identical, so we use least-squares for the mocks. The green plus shows the result of our noiseless mock. Bottom: The distribution of Δ H 0 = H 0 D E S − H 0 D E S + l e g a c y \Delta H_0=H_0^{\mathrm{DES}}-H_0^{\mathrm{DES+legacy}} Δ H 0 = H 0 DES − H 0 DES + legacy from the same mocks. The solid black line shows the mean result, while the dotted black line shows the result in the noiseless mock. The solid red line shows the observed Δ H 0 \Delta H_0 Δ H 0 using NUTS, while the dotted red line shows the almost identical result using least-squares. The red shading at the right shows the region where Δ H 0 \Delta H_0 Δ H 0 exceeds the observed value.
However, our mocks suggest that these issues can provide a partial explanation given their tendency to find Δ H 0 > 0 \Delta H_0 > 0 Δ H 0 > 0 . This bias is related to a generic overestimation of H 0 H_0 H 0 by ≈ 1 σ \approx1\sigma ≈ 1 σ , even if the DES coverage at low redshift is further limited by a redshift floor (Figure D2 ). As shown by our noise-free mock, the bias is a truncation error arising from differences between the assumed ΛCDM expansion history and our fourth-order cosmographic expansion. We do not know how accurately this can represent the real cosmological model, especially as a late-time solution to the Hubble tension must by definition differ from ΛCDM. Even so, our mocks quantify the effect. The scatter of H 0 D E S H_0^{\mathrm{DES}} H 0 DES across them (2.86 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 ) matches our stated uncertainty (2.87 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 ), so the error bar captures the statistical uncertainty. However, their mean exceeds the input H 0 H_0 H 0 by 2.76 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , so the ± 1 σ \pm1\sigma ± 1 σ (± 2 σ \pm2\sigma ± 2 σ ) interval contains the true value in only 49 (85)% of the mocks. Calibrated on the mocks, our H 0 D E S = 76.25 H_0^{\mathrm{DES}} = 76.25 H 0 DES = 76.25 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 is reached in 133 of the 10 4 10^4 1 0 4 realisations, i.e. 2.22 σ 2.22\sigma 2.22 σ for a one-tailed test, rather than the 3.13 σ 3.13\sigma 3.13 σ implied by the formal uncertainty. If we subtract the bias, H 0 D E S H_0^{\mathrm{DES}} H 0 DES becomes 73.49± \pm ± 2.87 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . This is 2.18 σ 2.18\sigma 2.18 σ above SPA and in excellent agreement with SH0ES (0.11 σ 0.11\sigma 0.11 σ ). We note that this correction assumes the SPA cosmology, so it illustrates the size of the bias rather than correcting for it in general.
The truncation error can be reduced with a higher-order expansion, but this makes the model more flexible, which inflates the uncertainties and can lower the Bayesian evidence. We therefore test how our results depend on the order by repeating the analysis for k max = 2 k_{\max}=2 k m a x = 2 –6. The results are shown in Figure 9 , with further details in Appendix B. Although H 0 H_0 H 0 changes significantly between orders, k max = 4 k_{\max}=4 k m a x = 4 is the lowest order that provides an adequate fit to the data. In fact, k max = 2 k_{\max}=2 k m a x = 2 and 3 raise χ min 2 \chi^2_{\min} χ m i n 2 by 210 and 30 for DES because they cannot describe the shape of the Hubble diagram across the DES redshift range. With k max = 5 k_{\max}=5 k m a x = 5 , we instead obtain H 0 = 68.52 ± 4.99 H_0=68.52\pm4.99 H 0 = 68.52 ± 4.99 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 for DES and H 0 = 66.14 ± 0.90 H_0=66.14\pm0.90 H 0 = 66.14 ± 0.90 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 for DES + legacy. This means that DES with k max = 5 k_{\max}=5 k m a x = 5 or 6 gives results that are consistent with both SH0ES and SPA because of their larger uncertainties, whereas k max = 4 k_{\max}=4 k m a x = 4 is consistent only with SH0ES. Even so, the higher orders remain consistent with the k max = 4 k_{\max}=4 k m a x = 4 result within the 1.58 σ 1.58\sigma 1.58 σ C.L. for DES and the 1.61 σ 1.61\sigma 1.61 σ C.L. for DES + legacy.
Figure 9. Polynomial order sweep for our cosmographic IDL in configuration A, using DES (black), DES + legacy (red), and DES + legacy with the legacy zero point Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy free (blue). Top: Bayesian evidence relative to the most probable order of each sample, Δ ln Z ≡ ln Z − ln Z max \Delta\ln Z \equiv\ln Z-\ln Z_{\max} Δ ln Z ≡ ln Z − ln Z m a x . The grey band marks ∣ Δ ln Z ∣ < 1 |\Delta\ln Z|<1 ∣Δ ln Z ∣ < 1 , inside which two orders are not distinguished (Trotta 2008). Stars and the vertical dotted line mark our baseline choice of k max = 4 k_{\max}=4 k m a x = 4 . Bottom: The posterior on H 0 H_0 H 0 at each order, with the Bayesian model average over k max = 2 k_{\max}=2 k m a x = 2 –6 shown as diamonds in the column labelled “BMA” (Appendix B2). The horizontal green and orange bands show H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES (Breuval et al. 2024) and H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA (SPT-3G Collaboration [107 ] ), respectively.
The fifth-order results emphasise that without the legacy SNIa, the IDL is plausibly consistent with a late-time solution to the Hubble
tension. They also highlight that the lack of effective DES coverage at low redshift makes it difficult to be sure what H 0 H_0 H 0 to expect from the IDL, so the high H 0 D E S H_0^{\mathrm{DES}} H 0 DES with k max = 4 k_{\max}=4 k m a x = 4 does not clearly confirm the Hubble tension.
The Bayesian evidence of each order (Eq. (10) ) quantifies how strongly the data support our baseline. The first result to highlight is that k max = 4 k_{\max}=4 k m a x = 4 is the most probable order for DES, carrying 80% of the posterior probability in configuration A and 83% in configuration B. The evidence at this order exceeds that at k max = 2 k_{\max}=2 k m a x = 2 and 3 by 97 and 10 nats, respectively. Beyond k max = 4 k_{\max}=4 k m a x = 4 , the Bayesian evidence falls again, by 1.4 at k max = 5 k_{\max}=5 k m a x = 5 and by 4.8 at k max = 6 k_{\max}=6 k m a x = 6 . This is because the fifth coefficient improves χ min 2 \chi^2_{\min} χ m i n 2 by only 2.4, which does not compensate for the extra volume prior. This is different for DES + legacy, where k max = 5 k_{\max}=5 k m a x = 5 improves χ min 2 \chi^2_{\min} χ m i n 2 by 9.3 and is preferred over k max = 4 k_{\max}=4 k m a x = 4 by Δ ln Z = 1.1 \Delta\ln Z=1.1 Δ ln Z = 1.1 , but even this preference is only weak evidence on the Jeffreys scale (Trotta 2008). Once Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy is free, the two orders are indistinguishable, with Δ ln Z = 0.1 \Delta\ln Z=0.1 Δ ln Z = 0.1 . Thus, the data support our baseline for DES, and they do not significantly disfavour it for DES + legacy regardless of whether we allow recalibration of the legacy SNIa, allowing us to keep the two samples on an equal footing. However, the evidence assigns 20% of the probability to the higher orders for DES, while for DES + legacy it slightly prefers k max = 5 k_{\max}=5 k m a x = 5 , so we cannot treat the fourth order as certain in either case.
One way to deal with the uncertain polynomial order k max k_{\max} k m a x is to combine the results at different orders by averaging over the considered k max = 2 k_{\max}=2 k m a x = 2 –6 range with weights given by their posterior probabilities (Eq. (10) ). We present the BMA in Appendix B2. For DES alone, it gives H 0 = 74.77 ± 4.83 k m s − 1 M p c − 1 H_0 = 74.77 \pm4.83\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 74.77 ± 4.83 km s − 1 Mp c − 1 , compared with 76.43 ± 3.01 k m s − 1 M p c − 1 76.43 \pm3.01\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 76.43 ± 3.01 km s − 1 Mp c − 1 at k max = 4 k_{\max}=4 k m a x = 4 from the same nested sampling. The central value moves little because k max = 4 k_{\max}=4 k m a x = 4 carries 80% of the probability. However, the uncertainty grows 1.6× because the average includes the spread between orders, which in this case includes the 7.9 k m s − 1 M p c − 1 \mathrm{km\ s^{-1}\ Mpc^{-1}} km s − 1 Mp c − 1 lower H 0 H_0 H 0 at k max = 5 k_{\max}=5 k m a x = 5 . As a result, the tension with SPA falls from 3.13 σ 3.13\sigma 3.13 σ to 1.56 σ 1.56\sigma 1.56 σ . Once we include the legacy SNIa, k max = 5 k_{\max}=5 k m a x = 5 carries 73% of the probability and the BMA gives H 0 = 66.61 ± 1.18 k m s − 1 M p c − 1 H_0 = 66.61 \pm1.18\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 66.61 ± 1.18 km s − 1 Mp c − 1 . This is 1.37 k m s − 1 M p c − 1 \mathrm{km\ s^{-1}\ Mpc^{-1}} km s − 1 Mp c − 1 below our k max = 4 k_{\max}=4 k m a x = 4 result, but it remains 4.50 σ 4.50\sigma 4.50 σ below SH0ES, so our conclusions on the legacy SNIa remain unchanged. After recalibrating the legacy SNIa, the BMA gives H 0 = 69.50 ± 2.43 k m s − 1 M p c − 1 H_0 = 69.50 \pm2.43\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 69.50 ± 2.43 km s − 1 Mp c − 1 , which is only 1.43 σ 1.43\sigma 1.43 σ below SH0ES. Our fiducial results focus on the same polynomial order (4th) for all SNIa configurations to help isolate the effect of including the legacy sample without simultaneously changing the flexibility of the reconstruction, as would occur if we had used the BMA.
It is also important to consider how our results are affected by the prior, which could matter for the DES-alone case because DES does not constrain the expansion history below z = 0.073 z=0.073 z = 0.073 . Our prior bounds only the expansion history itself (Eq. (8) ), with no posterior approaching the prior limits (§ 3.1 ). In particular, the best-fitting expansion history for DES stays at least 0.92 in E E E from both limits, so prior truncation effects should be negligible. We can also check the effect of the prior by using the least-squares method instead of NUTS, approximating that χ 2 \chi^2 χ 2 near its minimum varies quadratically with the model parameters. Applying the least-squares method to DES in terms of the C k C_k C k rather than E E E at the control points gives H 0 = 76.70 ± 2.97 k m s − 1 M p c − 1 H_0 = 76.70 \pm2.97\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 76.70 ± 2.97 km s − 1 Mp c − 1 , while DES + legacy gives instead 68.00 ± 0.71 k m s − 1 M p c − 1 68.00 \pm0.71\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 68.00 ± 0.71 km s − 1 Mp c − 1 . Both results are fully consistent with our NUTS inference, which we treat as our baseline because it relaxes the locally quadratic assumption. The similar results when using least squares with the C k C_k C k instead of NUTS with uniform priors on E E E at the control points indicate that the parametrisation used has little impact on the results.
Although our prior is uniform in the C k C_k C k over the allowed region, the envelope of Eq. (8) couples the different orders, making the marginal prior on each C k C_k C k bounded and unimodal rather than uniform. However, it is very broad, with σ C 1 = σ q 0 = 17.3 \sigma_{C_1}=\sigma_{q_0}=17.3 σ C 1 = σ q 0 = 17.3 , σ C 2 = 148 \sigma_{C_2}=148 σ C 2 = 148 , σ C 3 = 378 \sigma_{C_3}=378 σ C 3 = 378 , and σ C 4 = 293 \sigma_{C_4}=293 σ C 4 = 293 , where σ x \sigma_x σ x is the standard deviation of the prior on x x x . This is 24–73× the width of the corresponding posterior in all four of our cosmographic analyses (DES and DES + legacy in configurations A and B). As a result, the marginal prior density changes by less than 4% across the ± 2 σ \pm2\sigma ± 2 σ range of any posterior. This is because every posterior sample stays at least 0.8 in E E E from both limits of Eq. (8) . Thus, our prior neither truncates the posteriors nor appreciably shapes them. In fact, we obtain virtually the same results with other priors. In configuration A, independent uniform priors on the dimensional coefficients H 0 C k H_0C_k H 0 C k , over ranges much wider than the posteriors, give H 0 = 76.83 ± 2.88 k m s − 1 M p c − 1 H_0=76.83\pm2.88\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 76.83 ± 2.88 km s − 1 Mp c − 1 for DES and 68.02 ± 0.70 k m s − 1 M p c − 1 68.02\pm0.70\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 68.02 ± 0.70 km s − 1 Mp c − 1 for DES + legacy, against 76.25 ± 2.87 76.25\pm2.87 76.25 ± 2.87 and 67.98 ± 0.69 k m s − 1 M p c − 1 67.98\pm0.69\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 67.98 ± 0.69 km s − 1 Mp c − 1 , respectively, with our baseline prior. Moreover, a least-squares fit in the standard cosmographic parameters q 0 q_0 q 0 , j 0 j_0 j 0 , s 0 s_0 s 0 , and l 0 l_0 l 0 (which places no limits on them) returns H 0 = 76.70 ± 2.97 k m s − 1 M p c − 1 H_0=76.70\pm2.97\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 76.70 ± 2.97 km s − 1 Mp c − 1 and 68.00 ± 0.71 k m s − 1 M p c − 1 68.00\pm0.71\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 68.00 ± 0.71 km s − 1 Mp c − 1 for DES and DES + legacy, respectively, which is essentially identical to the above least-squares fit in the C k C_k C k . Even so, we adopt the envelope because it is simpler to implement and its edges are known in advance, since they are set by the expansion history itself rather than by a separate range on each C k C_k C k that would have to be tuned for every k max k_{\max} k m a x and then checked against the posteriors. This is important for the Bayesian evidence, since the Occam penalty at each order depends on the prior volume. With separate ranges on the C k C_k C k , this volume would depend on how we tune the ranges at each k max k_{\max} k m a x , whereas the envelope applies the same physical bounds at every order, so the comparison between orders (Appendix B2) does not involve any order-by-order tuning.
Discussion
Cosmological implications
Our IDL result for the absolute SNIa magnitude (M D E S = − 19.380 ± 0.012 M_{\mathrm{DES}}=-19.380\pm0.012 M DES = − 19.380 ± 0.012 ) is in tension with the SH0ES result using Cepheid distances to SNIa host galaxies (M B = − 19.253 ± 0.027 M_B=-19.253\pm0.027 M B = − 19.253 ± 0.027 ; (Riess et al. 2022)). This could be viewed as merely a manifestation of the Hubble tension, given the opposite starting points of the two approaches. We argue instead for an alternative interpretation, recognising that the DES and SH0ES catalogues operate in different passbands. SH0ES standardises SNIa into a rest-frame Johnson B B B -band ((Riess et al. 2022; Scolnic et al. 2022)), often calibrating their results using ground-based optical filters and HST WFC3 filters like F160W, F555W, and F814W (the numbers indicate central wavelengths in nm). By contrast, DES-Dovekie operates natively in Dark Energy Camera (DECam) passbands (g , r , i , z g,r,i,z g , r , i , z ; (Flaugher et al. 2015; Burke et al. 2018)). Converting DECam light curves into a standardised B B B -band magnitude requires spectral energy distribution (SED) template fitting (e.g., via SALT2 or SALT3) and colour conversions ((Guy et al. 2007; Kenworthy et al. 2021)). Unrecognised systematic shifts in filter transmission curves or colour-term corrections routinely introduce O ( 0.01 ) O(0.01) O ( 0.01 ) mag shifts ((Stubbs & Tonry 2006; Brout et al. 2022)). Furthermore, zero-point standards (e.g., AB vs. Vega definitions, CALSPEC white dwarf flux standards) across different telescopes (DECam at Cerro Tololo vs. ground-based legacy instruments or HST) carry non-negligible cross-calibration uncertainties ((Oke & Gunn 1983; Bohlin et al. 2014)). Another complication is that white dwarfs are quite different from SNIa in terms of colour and especially in terms of brightness, with nearby white dwarfs and other standard stars perhaps causing saturation effects that would not arise with a more distant SNIa near the limit of a legacy survey. Given that Pantheon+ comprises a wide array of legacy surveys, it is difficult to be sure of the correct relative calibration between them ((Scolnic et al. 2015; Brout et al. 2022)). DES-Dovekie has the major advantage of being a uniform dataset based on a single survey, provided we exclude the legacy SNIa ((Popovic et al. 2026)). This is the main reason we treat the DES-only case as our baseline and DES + legacy as an extension.
Building our IDL using only SNIa from the homogeneous DES catalogue predicts H 0 D E S = 76.25 ± 2.87 k m s − 1 M p c − 1 H_0^{\mathrm{DES}}=76.25\pm2.87\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 DES = 76.25 ± 2.87 km s − 1 Mp c − 1 , which is 1.03 σ 1.03\sigma 1.03 σ above the local SH0ES value (H 0 S H O E S = 73.17 ± 0.86 k m s − 1 M p c − 1 H_0^{\mathrm{SHOES}}=73.17\pm0.86\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 SHOES = 73.17 ± 0.86 km s − 1 Mp c − 1 ; (Breuval et al. 2024)). This result can be interpreted as a rediscovery of the Hubble tension using the IDL to derive H 0 H_0 H 0 instead of the usual approach based on the standard distance ladder, though without the latter, we obtain a smaller significance of 3.04 σ 3.04\sigma 3.04 σ relative to Planck 2020 (3.13σ \sigma σ relative to H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA , or 2.22σ \sigma σ once calibrated on our Λ \Lambda Λ CDM mocks). This holds for k max = 4 k_{\max}=4 k m a x = 4 , which is the order with the highest Bayesian evidence for DES (80% of the posterior probability). Using instead the BMA over k max = 2 k_{\max}=2 k m a x = 2 –6 reduces the tension with SPA to 1.56 σ 1.56\sigma 1.56 σ (§ 4.9 ). Importantly, this is not the usual Hubble tension between low- and high-redshift measurements because H 0 D E S H_0^{\mathrm{DES}} H 0 DES inherits its absolute scale from the CMB through Planck [100 ] prior on r d r_d r d . The tension between H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA compares two CMB-calibrated predictions that differ only in whether ΛCDM holds after recombination. This makes it a measure of departures from the ΛCDM expansion history at late times, not merely of a discrepancy between the local and early Universe. However, we stress that H 0 D E S H_0^{\mathrm{DES}} H 0 DES requires extrapolation below the DES redshift floor of 0.073. Reducing the redshift floor to 0.025 by adding the legacy SNIa decreases the predicted H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy to 67.98 ± 0.69 k m s − 1 M p c − 1 67.98 \pm0.69\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 67.98 ± 0.69 km s − 1 Mp c − 1 , which is now 4.69 σ 4.69\sigma 4.69 σ below H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES . The 197 legacy SNIa are therefore central to the question of whether the IDL is consistent with SH0ES.
The observed 8.27 k m s − 1 M p c − 1 8.27\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 8.27 km s − 1 Mp c − 1 difference between H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy might arise due to the inaccuracy of extrapolating our cosmographic expansion to z → 0 z \to0 z → 0 to handle the lack of effective DES coverage at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 . We test this possibility by constructing mock datasets assuming flat ΛCDM with the SPA parameters SPT-3G Collaboration [107 ] . Our results in Appendix D show that such a large value for Δ H 0 ≡ H 0 D E S − H 0 D E S + l e g a c y \Delta H_0 \equiv H_0^{\mathrm{DES}} - H_0^{\mathrm{DES+legacy}} Δ H 0 ≡ H 0 DES − H 0 DES + legacy arises in only 1.24 % 1.24\% 1.24% of 10 4 10^4 1 0 4 mock datasets in which the SNIa have the same redshift distribution and magnitude uncertainties as DES or DES + legacy. The progressive loss of low-redshift SNIa does lead to a slight bias in the recovered H 0 H_0 H 0 due to our cosmographic expansion not being a perfect representation of ΛCDM, but our redshift-floor sweep with mock data indicates that the lack of low-redshift coverage only biases our results upwards by 1 σ 1\sigma 1 σ (Figure D2 ). The sharp rise in H 0 H_0 H 0 with z f l o o r z_{\mathrm{floor}} z floor evident in the real data (Figure 5 ) is thus unlikely to arise in ΛCDM due to random scatter and extrapolation error. We suggest instead that it is due to systematics with the legacy SNIa, as discussed in § 5.2 .
Our results are based on a convergent cosmographic expansion over the relevant redshift range (Eq. (7) ). This leaves the possibility of biases in the inferred H 0 H_0 H 0 coming from the choice of polynomial order (Appendix B). Using a higher order than our baseline of k max = 4 k_{\max} = 4 k m a x = 4 and restricting to SNIa from DES gives a lower H 0 H_0 H 0 compatible with H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA , albeit with a lower evidence and a larger uncertainty that also admits consistency with SH0ES. We stress that k max = 4 k_{\max} = 4 k m a x = 4 is the order preferred by the Bayesian evidence for DES (Eq. (10) ), whereas for DES + legacy the evidence cannot distinguish it from k max = 5 k_{\max} = 5 k m a x = 5 , so we use k max = 4 k_{\max} = 4 k m a x = 4 throughout for consistency. We can avoid the choice of order altogether by averaging over the orders with their posterior probabilities (see Appendix B2). This BMA approach gives H 0 = 74.77 ± 4.83 k m s − 1 M p c − 1 H_0 = 74.77 \pm4.83\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 74.77 ± 4.83 km s − 1 Mp c − 1 for DES, which is in 1.56 σ 1.56\sigma 1.56 σ tension with SPA, down from 3.13 σ 3.13\sigma 3.13 σ for k max = 4 k_{\max} = 4 k m a x = 4 . The high H 0 D E S H_0^{\mathrm{DES}} H 0 DES therefore depends on the order of the expansion. Once we propagate its uncertainty through to H 0 D E S H_0^{\mathrm{DES}} H 0 DES , we find that DES alone does not determine H 0 H_0 H 0 precisely enough to rediscover the Hubble tension, even though the central BMA value lies much closer to SH0ES than to SPA.
Our Bayesian model comparison shows that the data do not strongly favour or disfavour a phenomenological late-time solution to the Hubble tension (Eq. (12) ) if we do not consider any local H 0 H_0 H 0 anchor (see Table 3 ). The exp model becomes preferred only if we add the SH0ES anchor on H 0 H_0 H 0 and either exclude the legacy SNIa altogether or include them only with an arbitrary recalibration relative to the DES SNIa. For the DES + legacy case without recalibration, the legacy data taken at face value resist a high H 0 H_0 H 0 consistent with SH0ES. A late-time solution to the Hubble tension is highly unlikely only insofar as the legacy SNIa are accurately calibrated relative to DES SNIa.
The phenomenological exp model converges to ΛCDM rapidly due to its very low z 1 z_1 z 1 parameter (Eq. (12) ). This raises a fine-tuning issue concerning why the expansion history should closely follow ΛCDM for many Gyr and only recently start to appreciably deviate from it. One possibility is that the Hubble tension might instead be solved through outflow from a local underdensity or void ((Halsb et al. 2020; Banik & Kroupa 2020; Banik & Kalaitzidis 2025; Banik, Nájerа & Desmond 2026a)). This scenario requires enhanced structure growth on ≳ 100 M p c \gtrsim100\ \mathrm{Mpc} ≳ 100 Mpc scales to form a sufficiently large and deep void (for a review, see Banik et al. 2026c). In this context, faster convergence to ΛCDM would indicate a smaller void, which if anything poses a less serious fine-tuning argument because the required adjustment to ΛCDM structure growth would be more modest. We explore this possibility by using a log-uniform prior on z 1 z_1 z 1 , which in the context of a void model is related to the void size (§ 4.6 ). Since this is not known a priori, a scale-invariant prior may be more appropriate. The results do not change much, with the DES + SH0ES case preferring z 1 ≈ 0.03 z_1 \approx0.03 z 1 ≈ 0.03 and A ≈ 8 k m s − 1 M p c − 1 A \approx8\ \mathrm{km\ s^{-1}\ Mpc^{-1}} A ≈ 8 km s − 1 Mp c − 1 regardless of whether we use a uniform prior on z 1 z_1 z 1 or ln z 1 \ln z_1 ln z 1 . Even so, our prior on A A A bounds both values (§ 4.6 ), so z 1 ≈ 0.03 z_1 \approx0.03 z 1 ≈ 0.03 cannot be read as a measurement of the size of a void.
The different tests presented in this work paint a coherent picture: combining DES SNIa with the DESI DR2 BAO assuming the Planck r d r_d r d does not disfavour a late-time solution to the Hubble tension. This conclusion is not as robust as one would like given the limited coverage of DES at low redshift, requiring an extrapolation to reach the present epoch and thus predict H 0 H_0 H 0 . This can in principle be overcome by including legacy SNIa, which reach down to z = 0.025 z = 0.025 z = 0.025 , largely eliminating the extrapolation uncertainty. This completely changes the picture, now strongly disfavouring a late-time solution. If the legacy SNIa are accurate, the Hubble tension is unlikely to be solved at late times. However, we present several empirical arguments for systematics with the legacy SNIa – at least when recalibrated onto the DECam passbands, which they were not originally observed in.4
Systematics with the legacy SNIa
The results in § 4.8 suggest that the legacy SNIa are offset from the homogeneous DES sample by Δ M l e g a c y = 0.0646 ± 0.0391 m a g \Delta M_{\mathrm{legacy}} = 0.0646 \pm0.0391\ \mathrm{mag} Δ M legacy = 0.0646 ± 0.0391 mag , a 1.65 σ 1.65\sigma 1.65 σ offset that shows no dependence on redshift and at most a weak dependence on which legacy survey the SNIa come from. Importantly, the legacy SNIa already pull the predicted H 0 H_0 H 0 downwards when we include only the 10 that overlap in redshift with DES (z = 0.073 z = 0.073 z = 0.073 –0.087 0.087 0.087 ; see § 4.4 ). Since SNIa at the same redshift must trace the same expansion history, a late-time departure from ΛCDM would not be expected to make them disagree, whereas a difference in calibration would. This is why we consider the more plausible explanation to be a systematic in the legacy SNIa, in particular with how they were recalibrated onto the DECam passbands or in their selection functions. We now discuss how such an offset could arise.
Since SNIa are transients, legacy SNIa cannot be observed on a modern filter system, making it difficult to prove how bright the legacy SNIa would have appeared to a modern instrument that is sensitive enough to reach high enough redshift to allow calibration via the IDL. Although white dwarfs and other standard stars were observed on both filter systems, the relative calibration for SNIa may be different due to the very different SED. In particular, there can be an interplay between the filter transmission curve and SNIa spectral features, but this would not occur in the relatively featureless spectrum of a white dwarf. As a result, nm-scale shifts in the wavelength centring of each filter might conceivably matter more for SNIa than for white dwarfs. There can also be saturation effects when a modern instrument like DECam is used to observe standard stars that were easily observable on a typically less sensitive legacy instrument. This can be mitigated by reducing the DES exposure time, but that can create its own uncertainties with shutter timing. Moreover, there can be non-linearities in the response of a charge-coupled device (CCD) to incident light. While not as common a concern today, non-linear CCD response might be more of an issue with legacy surveys. Since the calibration standards are typically brighter than legacy SNIa at z > 0.025 z > 0.025 z > 0.025 , a calibration valid for the former may not be valid for the latter at < 0.1 < 0.1 < 0.1 mag accuracy. Another important concern is how the host galaxy contribution is subtracted from the SNIa light curve. For results to be comparable between two surveys, the host galaxy subtraction should be done similarly, but procedures can change over the decades and legacy instruments typically had a poorer angular resolution. Since SNIa often occur in relatively crowded parts of a galaxy, angular resolution can affect the estimated background light at the SNIa sky position. Differences in the background subtraction strategy can alter the SNIa colour and light curve shape, which would then affect its standardised magnitude via the Tripp formula Tripp (1998).
We note that systematics with recalibrating legacy SNIa onto the DECam filters would have little impact on the Hubble tension. While this relies on SNIa at z < 0.15 z < 0.15 z < 0.15 which do mostly come from legacy surveys, recalibration onto DECam filters is not required. Instead, the same legacy surveys were in many cases able to find SNIa in a sufficiently nearby host galaxy that its distance can be measured using Cepheids, TRGB, SBF, or other well-established techniques. This allows a measurement of the absolute SNIa magnitude on the relevant photometric system, from which one can derive H 0 H_0 H 0 Riess et al. (2022). We demonstrate this with a short reanalysis of SH0ES where we refit their public SH0ES data vector and covariance with a separate absolute magnitude M s M_s M s for each SNIa survey s s s instead of assuming a common M s ∀ s M_s \forall s M s ∀ s . Two surveys (SDSS and PS1MD) contain no SNIa in a Cepheid host and share no SNIa with any calibrated survey, so their M s M_s M s is unconstrained. We therefore remove their 57 Hubble-flow light curves. Since the SH0ES model is linear in its parameters, a generalised least-squares fit gives the exact posterior. Applying this to SNIa from the remaining surveys changes H 0 H_0 H 0 from the published 73.04 ± 1.04 k m s − 1 M p c − 1 73.04 \pm1.04\ {\rm km\,s^{-1}\,Mpc^{-1}} 73.04 ± 1.04 km s − 1 Mp c − 1 to 72.85 ± 1.04 k m s − 1 M p c − 1 72.85 \pm1.04\ {\rm km\,s^{-1}\,Mpc^{-1}} 72.85 ± 1.04 km s − 1 Mp c − 1 , which is still 5.1 σ 5.1\sigma 5.1 σ above H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA . Moreover, the data show no preference for survey-specific offsets. The 13 remaining surveys need at least one absolute magnitude, so a separate M s M_s M s per survey adds a further 12 parameters. Across the 297 remaining light curves (77 in Cepheid hosts and 220 in the Hubble flow), these extra M s M_s M s parameters improve χ 2 \chi^2 χ 2 by just 7.5, which is less than the improvement of about 12 expected from fitting noise alone. This confirms that even if we apply our baseline methodology of doubting cross-calibrations between different filters, the Hubble tension stays virtually unchanged.
Our results therefore lend credence to a late-time solution to the Hubble tension without providing any reason to doubt its significance. On the contrary, we avoid the local distance ladder altogether and still obtain a high H 0 H_0 H 0 that lies > 3 σ > 3\sigma > 3 σ above the CMB-derived Λ \Lambda Λ CDM prediction for our baseline k max = 4 k_{\max} = 4 k m a x = 4 (1.5σ \sigma σ for the BMA), yet is in good agreement with SH0ES and other local H 0 H_0 H 0 anchors (Section 3.3). While there could be deficiencies in the standard distance ladder and/or our implementation of the homogeneous DES IDL, the good agreement between the H 0 H_0 H 0 values from these quite distinct approaches lends support to the idea that a modification to the late-time expansion history (or redshift gradient) is responsible for both techniques showing an excess over H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA .
The above results are in our configuration A, which depends on the tight Planck 2020 prior on r d r_d r d . This relies on the validity of the Λ \Lambda Λ CDM prior to recombination. We relax this assumption in configuration B, which therefore does not assume a late-time solution to the Hubble tension and does not use the CMB or other high-redshift data. Instead, r d r_d r d is calibrated using cosmology-independent CC datasets, which could in principle lead to a low r d r_d r d characteristic of an early-time solution. Using DES SNIa and assuming Λ \Lambda Λ CDM at late times, this approach yields r d = 148.48 ± 8.39 r_d = 148.48 \pm8.39 r d = 148.48 ± 8.39 cMpc and H 0 = 69. 0 − 4.6 + 3.4 k m s − 1 M p c − 1 H_0 = 69.0^{+3.4}_{-4.6}\ {\rm km\,s^{-1}\,Mpc^{-1}} H 0 = 69. 0 − 4.6 + 3.4 km s − 1 Mp c − 1 (Table 2 ). The H 0 H_0 H 0 constraint is only 1.2 σ 1.2\sigma 1.2 σ below SH0ES Breuvall et al. (2024). We note that an early-time solution by definition preserves Λ \Lambda Λ CDM behaviour at late times, so it must reproduce the BAO constraint on H 0 r d = 10154 ± 73 k m s − 1 M p c − 1 H_0r_d = 10154 \pm73\ {\rm km\,s^{-1}\,Mpc^{-1}} H 0 r d = 10154 ± 73 km s − 1 Mp c − 1 (DESI Collaboration 2025). The SH0ES value of H 0 H_0 H 0 Breuvall et al. (2024) then requires r d = 138.8 ± 1.9 r_d = 138.8 \pm1.9 r d = 138.8 ± 1.9 cMpc. Our result is 1.1 σ 1.1\sigma 1.1 σ above this, mirroring the behaviour in H 0 H_0 H 0 . This is to be expected given the tight BAO constraint on H 0 r d H_0r_d H 0 r d . Including the legacy SNIa has little impact on this result. While our configuration B results for H 0 H_0 H 0 and r d r_d r d lie closer to the Planck 2020 values than to those required by an early-time solution, such a scenario is still in good agreement with SNIa + BAO + CC. This is mainly due to the weak constraining power of CC. However, early-time solutions face other difficulties, as discussed in § 1 .
Comparison with the literature
The IDL has a record of giving a low H 0 H_0 H 0 similar to Planck 2020. Anchoring the BAO scale to a standard early-Universe sound horizon and constraining H 0 H_0 H 0 from the expansion history shape mapped by SNIa to reach low redshift, Aubourg et al. (2015) first obtained a low H 0 H_0 H 0 from the Baryon Oscillation Spectroscopic Survey (BOSS) BAO, while Macaulay et al. (2019) predicted H 0 = 67.8 ± 1.3 k m s − 1 M p c − 1 H_0 = 67.8 \pm1.3\ {\rm km\,s^{-1}\,Mpc^{-1}} H 0 = 67.8 ± 1.3 km s − 1 Mp c − 1 with an IDL based on SNIa in DES-SN3YR. The more recent analyses of Camilleri et al. (2025) and Ling et al. (2025), which combine DES-SN5YR with DESI DR1 and DR2 BAO, respectively, return a comparably low H 0 H_0 H 0 . Across three generations of SNIa samples and two BAO surveys, the IDL gives H 0 ≈ 67 H_0 \approx67 H 0 ≈ 67 –68 k m s − 1 M p c − 1 68\ {\rm km\,s^{-1}\,Mpc^{-1}} 68 km s − 1 Mp c − 1 . This is near the Planck 2020 value, but well below the local value from SH0ES Breuvall et al. (2024). Our DES + legacy result is very much in line with these previous results. Sabogal et al. (2026) recently formalised this obstacle as a “shape wall”. If pre-recombination physics is standard, the sound horizon and the angular acoustic scale are fixed, so a late-time solution can raise H 0 H_0 H 0 only by changing the shape of the expansion history E ( z ) E(z) E ( z ) . Relative SNIa distances combined with BAO constrain this shape tightly enough to restrict any such increase in H 0 H_0 H 0 to only a few percent, well short of the ≈ 9 % \approx9\% ≈ 9% needed to reach H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES . This makes the IDL a challenge for late-time solutions to the Hubble tension.
The closest analysis to ours in both data and method is that of Popovic & Sullivan (2026), who run a cosmographic IDL on the same DES-Dovekie catalogue (Popovic et al. (2026)), calibrating the SNIa using DESI DR2 and a standard CMB sound horizon. Those authors conclude that the IDL leaves little room for a late-time resolution. There is a methodological difference with our approach in that their cosmographic expansion is a polynomial in z z z (following Visser 2004), whereas we expand in a ~ ≡ z / ( 1 + z ) \tilde a \equiv z/(1+z) a ~ ≡ z / ( 1 + z ) (Cattoën & Visser 2007), which can be thought of as the lookback scale factor 1 − a 1-a 1 − a (Eq. (7) ). This matters at high redshift because a series expansion in z z z has a radius of convergence of ∣ z ∣ < 1 |z| < 1 ∣ z ∣ < 1 . This is insufficient to handle DES given it has SNIa out to z ≈ 1.14 z \approx1.14 z ≈ 1.14 , while DESI DR2 BAO measurements go out to z = 2.33 z = 2.33 z = 2.33 . An expansion in z z z is therefore not a model-independent cosmographic expansion, but rather a polynomial model for H ( z ) H(z) H ( z ) , with asymptotic behaviour set by k max k_{\max} k m a x . Eq. (5) of Popovic & Sullivan (2026) shows that those authors use k max = 4 k_{\max}=4 k m a x = 4 , implying an asymptotic H ∝ z 4 H \propto z^4 H ∝ z 4 behaviour – far steeper than the z 3 / 2 z^{3/2} z 3/2 behaviour predicted by ΛCDM (Eq. (9) ). Their recovered H 0 H_0 H 0 is thus tied to this specific polynomial ansatz for H ( z ) H(z) H ( z ) rather than providing a model-independent estimate of H 0 H_0 H 0 . Our parametrisation has a ~ < 1 ∀ z \tilde{a}<1\ \forall z a ~ < 1 ∀ z , keeping the expansion convergent throughout. Another problem with the Popovic & Sullivan (2026) analysis is their use of different polynomials for H ( z ) H(z) H ( z ) and D L ( z ) D_L(z) D L ( z ) , making the analysis internally inconsistent since c / H ( z ) c/H(z) c / H ( z ) must be integrated to obtain D M ( z ) D_M(z) D M ( z ) and thus D L ( z ) D_L(z) D L ( z ) . While using two separate polynomials is a good approximation at low redshift, it is not accurate given the DES and DESI range of redshifts (up to z = 2.33 z=2.33 z = 2.33 ).
Leaving aside these problems with a cosmographic expansion in z z z , once we adopt the same methodology as Popovic & Sullivan (2026) and use the same data, the results agree: by using the same likelihood, we recover their results within ≈ 1 σ \approx1\sigma ≈ 1 σ . Those authors report a nominal H 0 = 66.82 ± 0.74 k m s − 1 M p c − 1 H_0=66.82 \pm0.74\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 66.82 ± 0.74 km s − 1 Mp c − 1 from a sample that includes the low-redshift legacy SNIa. This corresponds to our DES + legacy configuration, for which we obtain H 0 = 67.98 ± 0.69 k m s − 1 M p c − 1 H_0=67.98 \pm0.69\ \mathrm{km\ s^{-1}\ Mpc^{-1}} H 0 = 67.98 ± 0.69 km s − 1 Mp c − 1 . These results are consistent at the 1.15 σ 1.15\sigma 1.15 σ C.L. and disfavour a late-time solution. Popovic & Sullivan (2026) argue that moving the inferred H 0 H_0 H 0 from 67 to 73 k m s − 1 M p c − 1 \mathrm{km\ s^{-1}\ Mpc^{-1}} km s − 1 Mp c − 1 demands a magnitude gradient d M D E S / d z ≈ 0.2 \mathrm{d}M_{\mathrm{DES}}/\mathrm{d}z \approx0.2 d M DES / d z ≈ 0.2 mag, with a flat ΛCDM fit after such a shift returning Ω m ≈ 0.23 \Omega_{\mathrm{m}} \approx0.23 Ω m ≈ 0.23 , a value that is 10 σ 10\sigma 10 σ discrepant with other probes. However, this argument presupposes the validity of ΛCDM, so any departure from its predicted shape for the expansion history must be absorbed by Ω m \Omega_{\mathrm{m}} Ω m . Since a late-time solution requires a different shape, assuming the validity of ΛCDM at late times produces a circular argument. A late-time solution is by definition a departure from the ΛCDM background, so inferring a non-standard effective Ω m \Omega_{\mathrm{m}} Ω m under an imposed ΛCDM fit is the expected signature of that departure rather than a falsification of it. Moreover, the probes against which a late-time solution was judged themselves assume a ΛCDM background, whose adequacy is precisely the question at issue. Most importantly, their nominal sample relies on the legacy SNIa, so their exclusion of late-time solutions holds only if the legacy SNIa are accurate.
The legacy “rung” has been contested in the literature, with several convergent lines of evidence pointing to a bias here. Within the DES-5Y SNIa compilation including the legacy SNIa, we measure a relative offset of Δ M l e g a c y = 0.0646 ± 0.0391 \Delta M_{\mathrm{legacy}} = 0.0646 \pm0.0391 Δ M legacy = 0.0646 ± 0.0391 mag between the legacy SNIa and DES (§ 4.8 ). This reproduces the offset first reported by Efstathiou (2025) from an independent comparison between Pantheon+ and DES. Watkins et al. (2026) reach the same conclusion from yet another angle, finding the high-redshift DES SNIa are well described by flat ΛCDM, but the low-redshift sample is systematically offset by ≈ 0.05 \approx0.05 ≈ 0.05 mag. They attribute this to unresolved systematics in the low-redshift sample rather than to new physics. The interpretation of the offset is contested: Vincenzi et al. (2025) reproduce the difference with Pantheon+ but attribute about 43% of it to improvements made to the DES-SN5YR analysis of intrinsic scatter and host galaxy properties. Since each DES release also reanalysed the legacy SNIa, improvements to the pipeline can have a non-negligible impact on the legacy SNIa, which after all can only be used alongside DES once they are recalibrated onto the DECam passbands and appropriate corrections are made for survey selection and other effects. The revision to host galaxy masses is larger for the legacy SNIa than for DES hosts, so these improvements mainly
affect the former. Vincenzi et al. (2025) attribute a further ≈ 38 % \approx38\% ≈ 38% of the difference with Pantheon+ to different selection functions for the DES subsets entering the two compilations, causing different bias corrections and thereby making a direct object-to-object comparison misleading (Efstathiou 2025). We find an offset in the same direction. The cause is not important for our purposes as the impact on H 0 H_0 H 0 is the same regardless.
Our analysis in § 4.8 suggests that around 3 k m s − 1 M p c − 1 3\ \mathrm{km\ s^{-1}\ Mpc^{-1}} 3 km s − 1 Mp c − 1 of the gap between H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy is caused by the legacy SNIa (Eq. (22) ). Further support for a systematic offset with the legacy SNIa is that such an offset would most plausibly come from conversion between photometric systems, in which case we might expect the offset to be survey-specific. Given the legacy SNIa largely come from two sources (Foundation and CfA/CSP), we can infer separate offsets for these. Our results in Figure 7 show only weak (1.0 σ 1.0\sigma 1.0 σ ) evidence for a difference between these two components of the legacy rung. Furthermore, when we sample all the cosmological parameters together with two legacy magnitude offsets, the posterior on H 0 H_0 H 0 remains virtually unchanged compared to having only one global offset for all the legacy SNIa. This shows that even if SNIa from Foundation and CfA/CSP require separate offsets, there is minimal impact on the predicted H 0 H_0 H 0 .
The legacy SNIa also explain why our conclusions differ from those of Bansal & Huterer (2026). Those authors ask whether a modification to only the late-time expansion history can simultaneously fit H 0 S H 0 E S H_0^{\mathrm{SH0ES}} H 0 SH0ES , the DESI BAO data, and relative SNIa distances, finding that this requires a sharp step in M S N I a M_{\mathrm{SNIa}} M SNIa at $z \lesssim 0.01 ({xref}Section 1). Their analysis rests on Pantheon+, which contains the same low-redshift legacy surveys as our DES + legacy configuration. In that configuration, we reach the conclusion that $H_0^{\mathrm{DES+legacy}}$ lies $4.69\sigma$ below $H_0^{\mathrm{SH0ES}}$. However, we reach a different verdict with the homogeneous DES sample, in which a single survey, filter set, and selection function remove the need to cross-calibrate between surveys. In this case, $H_0^{\mathrm{DES}}$ lies only $1.03\sigma$ above $H_0^{\mathrm{SH0ES}}$, while allowing the legacy SNIa to be recalibrated by $\Delta M_{\mathrm{legacy}}$ leaves $H_0$ only $0.99\sigma$ below $H_0^{\mathrm{SH0ES}}$. The answer to the question posed by {citet}Bansal & Huterer (2026) therefore hinges on the legacy SNIa, whose conversion onto the DECam passbands is precisely where a low-redshift step in $M_{\mathrm{SNIa}}$ could arise without any new physics. Even so, our DES-only result is not completely free of the legacy SNIa. We caution that although we have removed the legacy SNIa from our main analysis, they are still used to train the algorithm used to fit DES light curves and extract the most relevant parameters. To our knowledge, no analysis of DES has yet hidden the legacy magnitude–redshift relation from this training, so the size of the effect is currently unknown. This makes it important to reanalyse DES events using an algorithm trained without the use of legacy SNIa apparent magnitudes, perhaps restricting to only the light curve shape to help with event classification. Such a reanalysis would give the cleanest answer to the question posed by {citet}Bansal & Huterer (2026)` using only existing data.
The issue of possible systematics with the legacy SNIa can be addressed through reanalyses like that described above, but the best way is probably through homogeneous SNIa compilation that reaches a lower redshift than DES, which has little sensitivity at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 (though the closest SNIa it contains is at z = 0.073 z=0.073 z = 0.073 ). The crucial evidence will come from ongoing campaigns like the Dark Energy Bedrock All-Sky Supernova (DEBASS) program (Sherman et al. 2026), the Asteroid Terrestrial-impact Last Alert System (ATLAS; see the TITAN sample of Marlin et al. 2026), The Young Supernova Experiment (YSE; Jones et al. 2021; Aleo et al. 2023), and the Zwicky Transient Facility (ZTF; Rigault et al. 2025). Of particular importance is DEBASS given it uses DECam, the same instrument used for DES. This will simplify the treatment of any relative calibration offset, though selection effects for DEBASS differ from DES and must be handled with care.
Conclusions
We constructed the IDL using the homogeneous DES-Dovekie catalogue of uncalibrated SNIa (Popovic et al. 2026) and DESI DR2 BAO (DESI Collaboration 2025). We used a cosmographic expansion in the lookback scale factor a ~ ≡ z / ( 1 + z ) = 1 − a \tilde{a} \equiv z/(1+z)=1-a a ~ ≡ z / ( 1 + z ) = 1 − a . We adopted a polynomial order of k max = 4 k_{\max}=4 k m a x = 4 , the order preferred by the Bayesian evidence for DES, and checked our results against the BMA over k max = 2 k_{\max}=2 k m a x = 2 –6 6 6 . Our baseline analysis uses SNIa from DES alone (z ∈ [ 0.073 , 1.14 ] z \in[0.073,1.14] z ∈ [ 0.073 , 1.14 ] ). This excludes legacy SNIa that are normally included in the catalogue. Our main analyses calibrate the BAO assuming Λ \Lambda Λ CDM at early times through a Planck prior on r d r_d r d (Planck [100 ] ). The calibrated BAO results then allow a calibration of SNIa via the Etherington relation. We also conducted analyses where we relaxed the r d r_d r d prior and instead constrained it using CC (Moresco et al. 2022). In both configurations, we included the historical legacy SNIa as an optional extension (z ∈ [ 0.025 , 0.087 ] z \in[0.025,0.087] z ∈ [ 0.025 , 0.087 ] ). This allows us to assess their impact on the IDL and test them for possible systematics.
Our baseline result with the Planck prior on r d r_d r d is H 0 = 76.25 ± 2.87 H_0=76.25\pm2.87 H 0 = 76.25 ± 2.87 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which is fully consistent with SH0ES (Breuval et al. 2024) but in 3.04 σ 3.04\sigma 3.04 σ tension with Planck 2020 and 3.13 σ 3.13\sigma 3.13 σ tension with SPT-3G Collaboration [107 ] . This tension is specific to our baseline order. Averaging over k max = 2 k_{\max}=2 k m a x = 2 –6 6 6 with weights given by the Bayesian evidence gives H 0 = 74.77 ± 4.83 H_0=74.77\pm4.83 H 0 = 74.77 ± 4.83 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which reduces the tension with H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA to 1.56 σ 1.56\sigma 1.56 σ because the fifth order predicts a lower H 0 H_0 H 0 with a larger uncertainty. Excluding DES SNIa at low redshifts does not much affect the predicted H 0 H_0 H 0 , which for k max = 4 k_{\max}=4 k m a x = 4 remains stable at a value far above H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA out to the highest considered z f l o o r z_{\mathrm{floor}} z floor of 0.3 (Figure 5 ). Adding the legacy SNIa reduces H 0 H_0 H 0 to 67.98 ± 0.69 67.98\pm0.69 67.98 ± 0.69 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , only mildly above Planck and H 0 S P A H_0^{\mathrm{SPA}} H 0 SPA but 4.69 σ 4.69\sigma 4.69 σ below SH0ES (Breuval et al. 2024). Using the BMA instead changes H 0 H_0 H 0 to 66.61 ± 1.18 66.61\pm1.18 66.61 ± 1.18 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 and the tension with SH0ES to 4.50 σ 4.50\sigma 4.50 σ . The accuracy of the legacy SNIa is therefore central to assessing whether the IDL is compatible with a high H 0 H_0 H 0 , and thus whether the Hubble tension can be solved at late times. Replacing the Planck 2020 constraint on r d r_d r d with CC data increases the uncertainty due to the much weaker constraints on r d r_d r d , giving H 0 = 75.7 ± 5.3 H_0=75.7\pm5.3 H 0 = 75.7 ± 5.3 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with DES and 67.6 ± 3.6 67.6\pm3.6 67.6 ± 3.6 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with DES + legacy (74.1 ± 6.3 74.1\pm6.3 74.1 ± 6.3 and 66.2 ± 4.0 66.2\pm4.0 66.2 ± 4.0 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 for the BMA). The results are similar to the corresponding analyses with the Planck r d r_d r d prior, but the results are far less precise. We therefore focus only on results with the Planck r d r_d r d prior in what follows.
To see if the Δ H 0 = 8.27 \Delta H_0=8.27 Δ H 0 = 8.27 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 gap between our results with DES and DES + legacy can arise in Λ \Lambda Λ CDM, we set up mocks where we generated simulated Dovekie and DESI DR2 BAO catalogs sampled from Λ \Lambda Λ CDM with the SPA parameters that best fit the CMB (SPT-3G Collaboration [107 ] ). Only 1.24 % 1.24\% 1.24% of our 10 4 10^4 1 0 4 mocks exceed the observed value for Δ H 0 ≡ H 0 D E S − H 0 D E S + l e g a c y = 8.27 \Delta H_0 \equiv H_0^{\mathrm{DES}}-H_0^{\mathrm{DES+legacy}}=8.27 Δ H 0 ≡ H 0 DES − H 0 DES + legacy = 8.27 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (Appendix D). The BMA gives a similar gap of 8.16 8.16 8.16 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , although our mocks are analysed only at k max = 4 k_{\max}=4 k m a x = 4 . Assuming the expansion history has a shape similar to that in Λ \Lambda Λ CDM, the high observed Δ H 0 \Delta H_0 Δ H 0 is thus unlikely to arise from random errors and systematic uncertainties in extrapolating a cosmographic expansion from the DES range (z = 0.073 z=0.073 z = 0.073 –1.14 1.14 1.14 ) down to z → 0 z\to0 z → 0 . We suggest instead that the unusually high Δ H 0 \Delta H_0 Δ H 0 is driven by systematics with the legacy SNIa, as discussed below.
We find a consistent calibration of M D E S ≃ − 19.38 M_{\mathrm{DES}}\simeq-19.38 M DES ≃ − 19.38 for all polynomial orders with appreciable posterior weight (k max ≥ 4 k_{\max}\ge4 k m a x ≥ 4 ), regardless of the r d r_d r d prior. With the Planck prior on r d r_d r d , this is in 4.2 4.2 4.2 –4.3 σ 4.3\sigma 4.3 σ tension with the SH0ES Cepheid-based calibration of M B = − 19.253 ± 0.027 M_B=-19.253\pm0.027 M B = − 19.253 ± 0.027 (Riess et al. 2022). However, we must bear in mind that SH0ES standardises SNIa into the rest-frame Johnson B B B -band, whereas DES operates in DECam passbands. Furthermore, Pantheon+ (Scolnic et al. 2022) combines different surveys with their own zero-points, which might have systematic offsets between them. We therefore argue that the tension between M D E S M_{\mathrm{DES}} M DES from the IDL and M B M_B M B from Cepheid calibration does not necessarily represent the well-known Hubble tension. Instead, it could in principle be explained by differences in the passbands and relative calibration offsets.
We test the feasibility of a late-time solution to the Hubble tension using the phenomenological exp model (Eq. (12) ). This provides a flexible description of a late-time solution to the Hubble tension through an extra contribution to H ( z ) H(z) H ( z ) that enhances the late-time acceleration (Nájera et al. 2026). We found that the data are agnostic as to whether this model is preferred over Λ \Lambda Λ CDM based on the Bayesian evidence. Since our results prefer a high H 0 H_0 H 0 despite assuming the Planck prior on r d r_d r d , we can include local H 0 H_0 H 0 anchors into our analysis. If we add three H 0 H_0 H 0 anchors that are not based on SNIa (Mukherjee et al. 2021; Jensen et al. 2025; Vogl et al. 2025), we find that the Hubble tension cannot realistically be solved for the DES + legacy case as H 0 = 67.66 ± 1.01 H_0=67.66\pm1.01 H 0 = 67.66 ± 1.01 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . This changes when restricting to DES SNIa, since now H 0 = 70.0 0 − 0.98 + 4.78 H_0=70.00^{+4.78}_{-0.98} H 0 = 70.0 0 − 0.98 + 4.78 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , which would substantially reduce the Hubble tension at late times and is preferred over Λ \Lambda Λ CDM by the Bayesian evidence (Table 3 ). We note that SH0ES would measure only H 0 w i n = 70.23 ± 0.99 H_0^{\mathrm{win}}=70.23\pm0.99 H 0 win = 70.23 ± 0.99 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 in this model, 2.25 σ 2.25\sigma 2.25 σ below the published value, because much of the excess expansion rate lies below the SH0ES Hubble-flow window. The preference reaches around 11 nats if we use instead the strong SH0ES anchor on the local H 0 H_0 H 0 and exclude the legacy SNIa. It is therefore possible to solve, or at least significantly ameliorate, the Hubble tension at late times given DES, but such a scenario is highly disfavoured by DES + legacy, while including the legacy SNIa with arbitrary recalibration gives an intermediate result. Without the legacy SNIa, a successful solution has A ≃ 5 A\simeq5 A ≃ 5 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (H 0 ≃ 72.5 H_0\simeq72.5 H 0 ≃ 72.5 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 ) and z 1 < 0.086 z_1<0.086 z 1 < 0.086 at the 1 σ 1\sigma 1 σ C.L. When using the non-SNIa distance ladder anchors on H 0 H_0 H 0 . Upon using the SH0ES anchor instead, this changes to A ≃ 8 A\simeq8 A ≃ 8 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 (H 0 ≃ 75.5 H_0\simeq75.5 H 0 ≃ 75.5 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , or H 0 w i n ≃ 71.9 H_0^{\mathrm{win}}\simeq71.9 H 0 win ≃ 71.9 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 ) and z 1 ≃ 0.03 z_1\simeq0.03 z 1 ≃ 0.03 , though our prior on A A A bounds both values.
We also found that the Hubble tension cannot be solved with CPL, which always gives a low H 0 H_0 H 0 value that cannot be combined with the local anchors. Neglecting them, we found that CPL is preferred over Λ \Lambda Λ CDM only if the legacy SNIa are included at face value, but disfavoured by 1.79 nats if we exclude the legacy SNIa. These mixed results for the CPL model are a separate issue from whether there are deviations from Λ \Lambda Λ CDM, as the data seem to suggest even without considering SNIa at all (Banik et al. 2026a). A similar deviation is apparent in DES SNIa (Figure 2 ), though the underlying reason need not be evolving dark energy.
Our baseline is to use only SNIa from the homogeneous DES catalogue, avoiding the need to convert between filter systems. To test if the legacy SNIa might be subject to systematics in either the original observations or in how they have been recalibrated onto the DES photometric system, we predict the legacy SNIa apparent magnitudes based on DES SNIa. Our results suggest a systematic offset of Δ M l e g a c y = 0.0646 ± 0.0391 \Delta M_{\mathrm{legacy}}=0.0646\pm0.0391 Δ M legacy = 0.0646 ± 0.0391 mag (1.65 σ 1.65\sigma 1.65 σ ) in the legacy SNIa (0.053 ± 0.040 0.053 \pm0.040 0.053 ± 0.040 mag for the BMA), making them intrinsically fainter and thus closer to home. This is consistent with several previous studies (Efstathiou 2025; Huang et al. 2025; Ormondroyd et al. 2026; Watkins et al. 2026). Applying this correction to align the legacy SNIa with the main DES sample, the inferred H 0 H_0 H 0 rises to 71.03 ± 1.99 71.03 \pm1.99 71.03 ± 1.99 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 , only 0.99 σ 0.99\sigma 0.99 σ below SH0ES (69.50 ± 2.43 69.50 \pm2.43 69.50 ± 2.43 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 and 1.43 σ 1.43\sigma 1.43 σ for the BMA). Since the legacy SNIa largely come from Foundation and CfA/CSP, we applied separate offsets to legacy SNIa depending on their source. This revealed weak evidence for the offset being survey-specific, but a similar approach showed no evidence that the offset depends on redshift (Figure 7 ). Intra-data systematics might therefore undermine the conventional conclusion that SNIa strongly disfavour a late-time solution to the Hubble tension.
Ongoing homogeneous surveys like DEBASS (Sherman et al. 2026) will provide a new low-redshift calibration using the same instrument as DES, which should resolve this issue. Preliminary DEBASS DR0.5 distances do not show the ≈ 0.05 \approx0.05 ≈ 0.05 mag offset reported here and by other authors between the legacy low-redshift sample as given in the Dovekie catalogue and the extrapolated DES distance–redshift relation (Watkins et al. 2026). Their study tentatively suggests that the legacy offset arises from systematics in the legacy SNIa or how they were analysed, but this needs to be checked with a larger sample of SNIa and careful treatment of selection effects in the underlying discovery surveys. Until then, we cannot confidently assert that the IDL rules out a late-time solution to the Hubble tension.
ACKNOWLEDGEMENTS
All authors are supported by Royal Society University Research Fellowship grant 211046. Numerical computations were carried out on the SCIAMA High Performance Computing (HPC) cluster, which is supported by the Institute of Cosmology and Gravitation and the University of Portsmouth. The authors are grateful to Christopher Frohmaier for helpful discussions.
DATA AVAILABILITY
Our code is publicly available.5 The datasets used in this work are publicly available and come from the cited publications.
Alcock–Paczynski consistency
The AP test is an internal consistency test that is independent of both r d r_d r d and M D E S M_{\mathrm{DES}} M DES (§ 3.5 ). In this test, we compute the AP ratio F A P = D M / D H F_{\mathrm{AP}} = D_M/D_H F AP = D M / D H from the SNIa shape and compare it to the BAO results to check for mutual consistency. We implement this assuming only smoothness of the SNIa distance–redshift relation. We then compare the F A P F_{\mathrm{AP}} F AP SNIa predictions to DESI DR2 BAO observations at three redshifts where there is a good overlap between the datasets. In particular, DESI DR2 measurements at effective redshifts of z = 0.510 z = 0.510 z = 0.510 , 0.706 0.706 0.706 , and 0.934 0.934 0.934 lie within the range covered by the DES catalogue. We lack BAO measurements of F A P F_{\mathrm{AP}} F AP at lower redshift, while the DES catalogue of SNIa runs out at higher redshift. Reaching the next DESI DR2 BAO measurement at z = 1.32 z = 1.32 z = 1.32 would entail an extrapolation beyond the highest redshift DES SNIa at z = 1.14 z = 1.14 z = 1.14 . This leads to much larger uncertainties, so we do not report the result.
Methodology
Our predictions for F A P F_{\mathrm{AP}} F AP are based on the same cosmographic expansion as used throughout this work (Eq. (7) ). Both distances follow directly from it in a flat metric:
D M ( z ) = c ∫ 0 z d z ′ H ( z ′ ) , D H ( z ) = c H ( z ) . (23) D_M(z) = c \int_{0}^{z} \frac{\mathrm{d}z'}{H(z')}, \qquad D_H(z) = \frac{c}{H(z)}. \tag*{(23)} D M ( z ) = c ∫ 0 z H ( z ′ ) d z ′ , D H ( z ) = H ( z ) c . (23)
This implies D H = D M ′ D_H = D'_M D H = D M ′ , where a prime denotes differentiation with respect to z z z . In other words, the Hubble distance reported by DESI (based on the redshift depth of the BAO ruler) is exactly the derivative of the comoving distance (based on the angular scale of the BAO ruler). The SNIa reconstruction tells us the shape of D M ( z ) D_M(z) D M ( z ) , which we can differentiate to obtain D H D_H D H and thus F A P F_{\mathrm{AP}} F AP . Cancelling out the speed of light, we obtain that
F A P ( z ) ≡ D M ( z ) D H ( z ) = H ( z ) ∫ 0 z d z ~ H ( z ~ ) . (24) F_{\mathrm{AP}}(z) \equiv\frac{D_{\mathrm{M}}(z)}{D_{\mathrm{H}}(z)} = H(z) \int_0^z \frac{\mathrm{d}\tilde{z}}{H(\tilde{z})}. \tag*{(24)} F AP ( z ) ≡ D H ( z ) D M ( z ) = H ( z ) ∫ 0 z H ( z ~ ) d z ~ . (24)
To relate these results to our cosmographic expansion (Eq. (7) ), we recall that H ( z ) = H 0 E ( z ) H(z) = H_0 E(z) H ( z ) = H 0 E ( z ) , where
E ( z ) = 1 + ∑ k = 1 k max C k ( 1 − a ) k , E ( 0 ) = 1. (25) E(z) = 1 + \sum_{k=1}^{k_{\max}} C_k (1-a)^k,\qquad E(0) = 1. \tag*{(25)} E ( z ) = 1 + k = 1 ∑ k m a x C k ( 1 − a ) k , E ( 0 ) = 1. (25)
E ( z ) E(z) E ( z ) carries the shape of the expansion history, while H 0 H_0 H 0 carries its normalisation. Substituting into Eq. (24) , we obtain the following:
F A P ( z ) = H 0 E ( z ) ∫ 0 z d z ~ H 0 E ( z ~ ) = E ( z ) ∫ 0 z d z ~ E ( z ~ ) . (26) F_{\mathrm{AP}}(z) = H_0 E(z) \int_0^z \frac{\mathrm{d}\tilde{z}}{H_0 E(\tilde{z})} = E(z) \int_0^z \frac{\mathrm{d}\tilde{z}}{E(\tilde{z})}. \tag*{(26)} F AP ( z ) = H 0 E ( z ) ∫ 0 z H 0 E ( z ~ ) d z ~ = E ( z ) ∫ 0 z E ( z ~ ) d z ~ . (26)
This is the central equation for the AP test. Notice that since H 0 H_0 H 0 cancels out identically, F A P F_{\mathrm{AP}} F AP is a function only of the shape coefficients C k C_k C k .
The SNIa absolute magnitude cancels for the same reason. SNIa constrain the apparent magnitude m ( z ) = M D E S + 5 log 10 [ ( 1 + z ) D M ( z ) / M p c ] + 25 m(z) = M_{\mathrm{DES}} + 5\log_{10}[(1+z)D_{\mathrm{M}}(z)/\mathrm{Mpc}] + 25 m ( z ) = M DES + 5 log 10 [( 1 + z ) D M ( z ) / Mpc ] + 25 . Since D M ∝ 1 / H 0 D_{\mathrm{M}} \propto1/H_0 D M ∝ 1/ H 0 at fixed shape, SNIa alone measure only the shape coefficients together with the single degenerate combination M D E S − 5 log 10 H 0 M_{\mathrm{DES}} - 5\log_{10}H_0 M DES − 5 log 10 H 0 , never M D E S M_{\mathrm{DES}} M DES or H 0 H_0 H 0 separately. Holding M D E S M_{\mathrm{DES}} M DES fixed in the fit is therefore a gauge choice that leaves F A P F_{\mathrm{AP}} F AP untouched, so the calibration does not enter the prediction. A similar cancellation occurs on the BAO side, since a BAO survey measures D M / r d D_{\mathrm{M}}/r_d D M / r d and D H / r d D_{\mathrm{H}}/r_d D H / r d , so forming the ratio F A P = ( D M / r d ) / ( D H / r d ) F_{\mathrm{AP}} = (D_{\mathrm{M}}/r_d)/(D_{\mathrm{H}}/r_d) F AP = ( D M / r d ) / ( D H / r d ) causes r d r_d r d to divide out identically.
This cancellation is also what makes the AP test precise despite the SNIa sample being uncalibrated. The large M D E S M_{\mathrm{DES}} M DES uncertainty is entirely common-mode between the numerator and denominator in the AP ratio. This leaves only the error in the shape of the expansion history, which is constrained quite well by SNIa given their large number and wide redshift range.
To predict BAO measurements of F A P F_{\mathrm{AP}} F AP , we must by definition not consider the BAO data we are trying to predict. We therefore fit the cosmographic parameters using only SNIa, assuming some fixed M D E S M_{\mathrm{DES}} M DES . Since the SNIa only reach z = 1.144 z = 1.144 z = 1.144 (a ˉ = 0.534 \bar{a} = 0.534 a ˉ = 0.534 ), we impose the prior envelope of Eq. (8) up to a ˉ max = 0.534 \bar{a}_{\max} = 0.534 a ˉ m a x = 0.534 rather than a ˉ max = 0.70 \bar{a}_{\max} = 0.70 a ˉ m a x = 0.70 in this fit, so that the prior constrains only the range with available data. We then evaluate Eq. (26) by numerical quadrature on the actual posterior draws returned by the sampler. Every draw therefore carries the full covariance of the C k C_k C k through to F A P F_{\mathrm{AP}} F AP .
Results
Table A1 lists the SNIa-predicted and DESI-measured F A P F_{\mathrm{AP}} F AP at the three redshifts suitable for a comparison. We find that the SNIa constraint on the shape of the expansion history predicts AP ratios that are in good agreement with BAO measurements, as quantified by χ 2 = 6.33 \chi^2 = 6.33 χ 2 = 6.33 for DES and 6.13 for DES + legacy. The corresponding p p p -values of a higher χ 2 \chi^2 χ 2 are 9.7% and 10.5%, respectively. These mild levels of tension indicate that SNIa and BAO give consistent results for the shape of the expansion history. This agreement cannot be manufactured by a shared calibration, making it a genuine predictive success of the SNIa-derived shape.
Table A1. Result of the AP consistency test between DESI DR2 BAO and DES-Dovekie SNIa. Assuming only that the SNIa distance–redshift relation is smooth, we fit H ( z ) H(z) H ( z ) to the SNIa alone and predict F A P ≡ D M / D H F_{\mathrm{AP}} \equiv D_{\mathrm{M}}/D_{\mathrm{H}} F AP ≡ D M / D H at the three DESI BAO redshifts, all of which lie inside the SNIa range. The test carries no absolute calibration and no sound horizon assumption. Errors are evaluated on the actual posterior draws of the full SNIa catalogue.
Despite the overall agreement, the SNIa prediction at z = 0.510 z = 0.510 z = 0.510 falls below DESI, with a pull from the latter of − 1.91 σ -1.91\sigma − 1.91 σ for DES and − 1.98 σ -1.98\sigma − 1.98 σ for DES + legacy. This feature is robust as it carries the same sign and size in both samples. It is also stable against the polynomial order and whether we bin the SNIa in redshift. Although a 2 σ 2\sigma 2 σ tension is not significant on its own, there is also a similar tension of the opposite sign at z = 0.934 z = 0.934 z = 0.934 , where the pull is + 2.03 σ +2.03\sigma + 2.03 σ for DES and + 1.78 σ +1.78\sigma + 1.78 σ for DES + legacy. However, there is negligible pull at z = 0.706 z = 0.706 z = 0.706 . The tension is further mitigated by correlations between F A P F_{\mathrm{AP}} F AP predictions at different redshifts, which we account for when calculating the total χ 2 \chi^2 χ 2 .
Figure A1 shows the results relative to the fiducial SPA Λ \Lambda Λ CDM prediction. Given that BAO measurements are in reasonable agreement with the SPA cosmology (Banik et al. 2026a), the AP consistency between SNIa and BAO at z = 0.706 z = 0.706 z = 0.706 also indicates consistency here with the CMB-derived expectation in Λ \Lambda Λ CDM. However, the BAO measurements of F A P F_{\mathrm{AP}} F AP are not exactly in line with the CMB-calibrated Λ \Lambda Λ CDM expectation (cf. figure 3 of Banik & Kalaitzidis 2025, which uses the Planck cosmology). This causes the SNIa-derived predictions to be in less tension with the SPA cosmology than they are with DESI DR2 BAO – though even here the tension is modest. This justifies our approach of combining SNIa and BAO datasets using the Etherington relation.
Figure A1. The shaded bands show the SNIa-shape mean prediction and 1 σ 1\sigma 1 σ uncertainty for the AP ratio F A P ≡ D M / D H F_{\mathrm{AP}} \equiv D_{\mathrm{M}}/D_{\mathrm{H}} F AP ≡ D M / D H using either DES (black) or DES + legacy (red). The dark blue shapes with error bars show DESI DR2 measurements of F A P F_{\mathrm{AP}} F AP . The F A P F_{\mathrm{AP}} F AP values shown here are relative to the predicted value at that redshift in the SPA cosmology (Ω m = 0.3175 \Omega_{\mathrm{m}} = 0.3175 Ω m = 0.3175 ; [107 ] ).
Despite the good overall agreement, SNIa and BAO appear to follow opposite trends in the ratio F A P / F A P F C D M , S P A F_{\mathrm{AP}}/F_{\mathrm{AP}}^{\mathrm{FCDM,SPA}} F AP / F AP FCDM , SPA , which decreases with redshift when using BAO but increases with redshift when using SNIa. However, this is not too significant, especially given that the SNIa prediction becomes rather uncertain at z > 0.8 z > 0.8 z > 0.8 , reducing the discrepancy with Λ \Lambda Λ CDM. Since the tension is largely in the slope rather than the normalisation, we can try to quantify this by fitting F S N / F B A O = α + β ( z − z p ) F_{SN}/F_{BAO} = \alpha+ \beta(z-z_{\mathrm{p}}) F S N / F B A O = α + β ( z − z p ) over the three DESI redshifts, where z p = 0.614 z_{\mathrm{p}} = 0.614 z p = 0.614 is a suitable pivot redshift. This approach yields α − 1 = − 0.019 ± 0.021 \alpha- 1 = -0.019 \pm0.021 α − 1 = − 0.019 ± 0.021 and β = 0.41 ± 0.18 \beta= 0.41 \pm0.18 β = 0.41 ± 0.18 for DES, while α − 1 = − 0.023 ± 0.021 \alpha- 1 = -0.023 \pm0.021 α − 1 = − 0.023 ± 0.021 and β = 0.38 ± 0.17 \beta= 0.38 \pm0.17 β = 0.38 ± 0.17 for DES + legacy. The level is consistent with unity, but the shape tilt is around 2.2 σ 2.2\sigma 2.2 σ in both samples. This does not amount to a significant shape discrepancy because the three residuals give a total χ 2 = 6.33 \chi^2 = 6.33 χ 2 = 6.33 for DES and 6.13 6.13 6.13 for DES + legacy, which means the tension cannot exceed 2.52 σ 2.52\sigma 2.52 σ even if it could be collapsed onto one parameter.
Sweep over the cosmographic polynomial order
Results at a fixed order
The results of our polynomial order sweep are shown in Table B1 and Table B2 for configurations A and B, respectively. When using DES alone, the orders that give a high H 0 H_0 H 0 are k max = 2 k_{\max}=2 k m a x = 2 and 4, but a second-order expansion does not fit the data: its evidence is 97 nats lower. By contrast, a fourth-order expansion fits the data well and is the most probable order. Even so, we cannot confidently say that the homogeneous DES IDL confirms the Hubble tension. This is because k max = 5 k_{\max}=5 k m a x = 5 gives H 0 = 68.52 ± 4.99 k m s − 1 M p c − 1 H_0 = 68.52 \pm4.99\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} H 0 = 68.52 ± 4.99 km s − 1 Mpc − 1 , which is compatible with both SPA and SH0ES. Importantly, it is also consistent with k max = 4 k_{\max}=4 k m a x = 4 at the 1.36 σ 1.36\sigma 1.36 σ C.L. The k max = 6 k_{\max}=6 k m a x = 6 result continues the trend towards lower and less precise H 0 H_0 H 0 , which is now 63.30 ± 7.75 k m s − 1 M p c − 1 63.30 \pm7.75\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} 63.30 ± 7.75 km s − 1 Mpc − 1 . Increasing the polynomial order makes the model more flexible, so the uncertainty grows rapidly. The evidence weighs this flexibility against the improvement in the fit, thereby assigning 20% of the probability to k max = 5 k_{\max}=5 k m a x = 5 and < 1 % <1\% < 1% to k max = 6 k_{\max}=6 k m a x = 6 . This means that the fifth order cannot be dismissed, which is why we average over the orders in Appendix B2. We note that while the specific weights assigned to k max = 4 k_{\max}=4 k m a x = 4 and 5 in the crucial DES-only case depend on the specific priors that we adopt (Table 1 ), χ min 2 \chi^2_{\min} χ m i n 2 improves by only 2.44 when going from the former to the latter. This is a very modest reduction given there are 1555 data points without the legacy SNIa, suggesting that any reasonable way of quantifying the Occam penalty would prefer k max = 4 k_{\max}=4 k m a x = 4 when using DES alone. The situation is less clear when using DES + legacy because then the improvement in χ min 2 \chi^2_{\min} χ m i n 2 is 9.26. Fortunately, the much better coverage of the legacy SNIa at low redshifts makes the choice of polynomial order far less relevant to the inferred cosmological parameters than for DES alone.
Table B2. As Table B1, but for configuration B in which the sound horizon is inferred from the CC data with no CMB input.
If we replace our model-independent cosmographic expansion with flat Λ \Lambda Λ CDM, the resulting χ min 2 \chi^2_{\min} χ m i n 2 is only 2.3 (1.3) above that at k max = 4 k_{\max}=4 k m a x = 4 for DES (DES + legacy). Since Λ \Lambda Λ CDM has three fewer shape parameters than a quartic, this might suggest that Λ \Lambda Λ CDM is the preferred description of the data. However, it is important to stress that the cosmographic expansion should not be seen as a rival model to Λ \Lambda Λ CDM, but rather as a Taylor series which can describe any expansion history over a given redshift range. For this reason, we do not include Λ \Lambda Λ CDM in the average over orders, none of which can exactly match Λ \Lambda Λ CDM for finite k max k_{\max} k m a x . We instead compare Λ \Lambda Λ CDM to specific alternatives using the Bayesian evidence in § 4.5 . An important caveat to our results is that the intrinsic scatter term in the SNIa covariance is calibrated so that Λ \Lambda Λ CDM provides an acceptable fit to the Hubble diagram, since the intrinsic scatter is difficult to obtain any other way. This means a reduced χ 2 \chi^2 χ 2 close to unity is not necessarily a success of the Λ \Lambda Λ CDM model, justifying our focus on relative rather than absolute likelihoods.
Bayesian model average over the polynomial order
Unless stated otherwise, values quoted in the main text use our baseline order of k max = 4 k_{\max}=4 k m a x = 4 . However, Appendix B1 shows that we cannot confidently confirm the Hubble tension with the DES IDL because the predicted H 0 H_0 H 0 depends on the order. Rather than selecting one order, we therefore treat k max k_{\max} k m a x as a discrete parameter and marginalise over it. The posterior of any quantity θ \theta θ is then the Bayesian model average
P ( θ ∣ D ) = ∑ k max = 2 6 w ( k max ) P ( θ ∣ D , k max ) , (27) P(\theta\mid D) = \sum_{k_{\max}=2}^{6} w(k_{\max})\,P(\theta\mid D,k_{\max}), \tag*{(27)} P ( θ ∣ D ) = k m a x = 2 ∑ 6 w ( k m a x ) P ( θ ∣ D , k m a x ) , (27)
where P ( θ ∣ D , k max ) P(\theta\mid D,k_{\max}) P ( θ ∣ D , k m a x ) is the posterior at order k max k_{\max} k m a x and w ( k max ) w(k_{\max}) w ( k m a x ) is the weighted posterior probability of that order (Eq. (10) ). We quote the mean and standard deviation of Eq. (27) in the BMA rows in Table B1 and Table B2 . Its variance is the probability-weighted average of the variances at each order plus the variance of their central values, so the average carries the uncertainty in the order as well as the uncertainty at each order. The only exception is q 0 q_0 q 0 for DES + legacy, whose posterior is bimodal because k max = 4 k_{\max}=4 k m a x = 4 and 5 place it on opposite sides of zero. For these cases, we report the two intervals that together comprise the 1 σ 1\sigma 1 σ highest posterior density region.
The first result to highlight is that the BMA is dominated by k max = 4 k_{\max}=4 k m a x = 4 and 5 in every case. The lower orders are excluded by at least 7.9 nats, while k max = 6 k_{\max}=6 k m a x = 6 never carries more than 2.1% of the probability. For DES, k max = 4 k_{\max}=4 k m a x = 4 carries 79.5% of the probability and k max = 5 k_{\max}=5 k m a x = 5 carries 19.8%. The weighted average gives H 0 = 74.77 ± 4.83 k m s − 1 M p c − 1 H_0 = 74.77 \pm4.83\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} H 0 = 74.77 ± 4.83 km s − 1 Mpc − 1 , compared with 76.43 ± 3.01 k m s − 1 M p c − 1 76.43 \pm3.01\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} 76.43 ± 3.01 km s − 1 Mpc − 1 for k max = 4 k_{\max}=4 k m a x = 4 . The central value hardly moves, but the uncertainty grows 1.6 × 1.6\times 1.6 × because the k max = 5 k_{\max}=5 k m a x = 5 posterior is centred 7.9 k m s − 1 M p c − 1 \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} km s − 1 Mpc − 1 lower. When allowing recalibration of the legacy SNIa, k max = 4 k_{\max}=4 k m a x = 4 and 5 carry 45.7% and 52.3% of the probability, respectively, so their weighted average gives H 0 = 69.50 ± 2.43 k m s − 1 M p c − 1 H_0 = 69.50 \pm2.43\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} H 0 = 69.50 ± 2.43 km s − 1 Mpc − 1 and Δ M l e g a c y = + 0.053 ± 0.040 \Delta M_{\mathrm{legacy}} = +0.053 \pm0.040 Δ M legacy = + 0.053 ± 0.040 mag. Configuration B ranks the orders in the same way, with k max = 4 k_{\max}=4 k m a x = 4 preferred for DES and k max = 5 k_{\max}=5 k m a x = 5 marginally preferred for DES + legacy (Table B2 ).
The BMA changes the interpretation of the DES result more than that of any other sample. For DES alone, the H 0 H_0 H 0 tension with SPA falls from 3.13 σ 3.13\sigma 3.13 σ at k max = 4 k_{\max}=4 k m a x = 4 to 1.56 σ 1.56\sigma 1.56 σ , though the result remains consistent with SH0ES at the 0.33 σ 0.33\sigma 0.33 σ C.L. This means that once we propagate the uncertainty in the order, DES alone does not determine H 0 H_0 H 0 precisely enough to discriminate between SH0ES and SPA, even though its central value lies much closer to SH0ES. On the other hand, conclusions on the legacy SNIa are robust. The averaged H 0 H_0 H 0 for DES + legacy is 4.50 σ 4.50\sigma 4.50 σ below SH0ES, compared with 4.69 σ 4.69\sigma 4.69 σ at k max = 4 k_{\max}=4 k m a x = 4 , the small reduction arising from the larger uncertainty. Once the legacy SNIa are recalibrated, the average is consistent with SPA at the 0.94 σ 0.94\sigma 0.94 σ C.L. and with SH0ES at the 1.43 σ 1.43\sigma 1.43 σ C.L., which is only slightly worse than the 0.99 σ 0.99\sigma 0.99 σ tension with SH0ES for k max = 4 k_{\max}=4 k m a x = 4 . The recalibration also raises the averaged H 0 H_0 H 0 by 2.9 k m s − 1 M p c − 1 2.9\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} 2.9 km s − 1 Mpc − 1 , which is similar to the ∼ 3 k m s − 1 M p c − 1 \sim3\ \mathrm{km}\,\mathrm{s}^{-1}\,\mathrm{Mpc}^{-1} ∼ 3 km s − 1 Mpc − 1 shift at k max = 4 k_{\max}=4 k m a x = 4 (§ 4.8 ). Thus, the averaged results confirm that a late-time solution to the Hubble tension remains possible insofar as we exclude or recalibrate the legacy SNIa.
Unlike a least-squares fit, the evidence depends on the prior, so it is important to check how much the weights rely on it. Our prior is uniform over all expansion histories with 0 ≤ E ( a ~ ) ≤ 2 + 7 a ~ 0 \le E(\tilde{a}) \le2 + 7\tilde{a} 0 ≤ E ( a ~ ) ≤ 2 + 7 a ~ (§ 3.1 ). Halving the width of this band about the line E = 1 + 3.5 a ~ E = 1 + 3.5\tilde{a} E = 1 + 3.5 a ~ is a linear contraction in the C k C_k C k , so it divides the prior volume at order k k k by exactly 2 k 2^k 2 k . Provided the posterior lies inside the narrower band, ln Z \ln Z ln Z then rises by ln 2 \ln2 ln 2 , so the evidence ratio between consecutive orders shifts by ln 2 ≈ 0.69 \ln2 \approx0.69 ln 2 ≈ 0.69 in favour of the higher order (the Occam penalty is lower due to the narrower prior range). We find that at least 99% of the posterior lies inside the halved width band for every sample at k max ≥ 3 k_{\max} \ge3 k m a x ≥ 3 , so this shift applies almost exactly. Halving the band moves the DES weights at k max = 4 k_{\max}=4 k m a x = 4 and 5 from 0.80 and 0.20 to 0.65 and 0.33, respectively, while the DES + legacy weights shift from 0.25 and 0.73 to 0.14 and 0.82. However, the averaged H 0 H_0 H 0 moves by at most 1.2 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 and is now 73.56 ± \pm ± 5.60 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 for DES, 66.41 ± \pm ± 1.10 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 for DES + legacy, and 69.05 ± \pm ± 2.38 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 with the legacy SNIa recalibrated. These lie 1.13σ \sigma σ above SPA, 4.85σ \sigma σ below SH0ES, and 1.63σ \sigma σ below SH0ES, respectively, so our conclusions are not driven by the width of the prior. This highlights the advantage of a prior on the normalised expansion history itself over independent priors on each C k C_k C k , whose widths are difficult to set but would nonetheless set the Occam penalty of every order separately.
Full posteriors
The results we obtain with the IDL differ significantly depending on whether we include the legacy SNIa. We emphasise this in Figure C1 , which shows a triangle plot for the main parameters in our baseline analysis (fourth-order cosmography in configuration A). If we include the legacy SNIa but allow them to be recalibrated relative to the main DES sample by Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy , we obtain results intermediate between the DES-only and DES + legacy cases. The bottom right panel shows that this is driven by a preference for Δ M l e g a c y > 0 \Delta M_{\mathrm{legacy}} > 0 Δ M legacy > 0 , though it is probably ≲ 0.1 \lesssim0.1 ≲ 0.1 mag.
Figure C1. Triangle plot showing the main parameters in our baseline analysis, which uses a fourth-order cosmographic expansion with the Planck 2020 prior on r d r_d r d (configuration A). The IDL is constrained using SNIa from DES (black), DES + legacy (red), or DES + legacy with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (blue; Eq. (22) ). The results in the latter case are intermediate between the other two. The dotted orange lines mark H 0 H_0 H 0 and q 0 q_0 q 0 in flat Λ \Lambda Λ CDM with the SPA parameters, whose small uncertainties are neglected here (SPT-3G Collaboration [107 ] ).
Figure C2 shows the triangle plots for the main analyses in this paper, namely the cosmographic (k max = 4 k_{\max}=4 k m a x = 4 ) and Λ \Lambda Λ CDM versions of the IDL constrained by DES and DES + legacy in configurations A and B. We overlay the DES and DES + legacy results, showing the former in black and the latter in red. The top panels show configuration A, while the bottom panels show configuration B. The left panels show results using our fourth-order cosmographic expansion, while the right panels use Λ \Lambda Λ CDM. The legacy SNIa do not much shrink the Λ \Lambda Λ CDM contours because Λ \Lambda Λ CDM already imposes a rigid shape to the expansion history, minimising the uncertainty associated with extrapolating across the DES coverage gap at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 . In our more flexible cosmographic expansion, the legacy SNIa substantially shrink the parameter uncertainties. However, the preferred region is not expected on the basis of DES alone. This emphasises that the legacy SNIa are likely subject to systematics, at least when recalibrated onto the DES filter system in order to enable a joint analysis.
Figure C2. Triangle plots for the main analyses presented in this paper, showing the 1 σ 1\sigma 1 σ and 2 σ 2\sigma 2 σ contours for DES (black) and DES + legacy (red). The left panels use our fourth-order cosmographic expansion (Eq. (7) ), which we express through the expansion rate H k ≡ H ( a ˉ e m a x k / k max ) H_k \equiv H(\bar{a}_{\mathrm{emax}} k/k_{\max}) H k ≡ H ( a ˉ emax k / k m a x ) at four points equally spaced in a ˉ \bar{a} a ˉ , which lie at z = 0.21 z = 0.21 z = 0.21 , 0.54 0.54 0.54 , 1.11 1.11 1.11 , and 2.33 2.33 2.33 for k = 1 … 4 k = 1 \ldots4 k = 1 … 4 . All H k H_k H k are given in km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . We also show q 0 = C 1 − 1 q_0 = C_1 - 1 q 0 = C 1 − 1 , though it is fully determined by the H k H_k H k . Over the H k H_k H k ranges shown here, the priors are essentially flat. The right panels assume flat Λ \Lambda Λ CDM (Eq. (9) ). The dotted orange lines on some panels show predictions for the corresponding quantities in flat Λ \Lambda Λ CDM with the SPA parameters, neglecting their small uncertainties [107 ] . The top panels show results for configuration A with a tight prior on r d r_d r d (Planck 2020). This is relaxed in the bottom panels, which show our configuration B results where r d r_d r d is inferred from CC, making it far less precise. As a result, the r d r_d r d axis range differs between the top and bottom panels. Notice how the DES + legacy parameters in the cosmographic configuration A analysis largely fall outside the range expected from DES alone.
Late-time solutions
In Section 4.5, we tested whether two proposed phenomenological adjustments to Λ \Lambda Λ CDM can alleviate the Hubble tension, bearing in mind their greater model flexibility. In what follows, we only consider the exp model (Eq. (12) ) because the CPL model predicts a low H 0 H_0 H 0 and thus cannot be combined with local H 0 H_0 H 0 anchors (Table 3 ). We first show the triangle plots for the analyses unconstrained by any local H 0 H_0 H 0 anchor (Figure C3 ). As might be expected without any local anchor, the amplitude term A A A is consistent with zero in all cases, though there is an interesting 1.3 σ 1.3\sigma 1.3 σ preference for A < 0 A < 0 A < 0 using DES + legacy. This is associated with a preference for lower H 0 H_0 H 0 , making it even more difficult to solve the Hubble tension. If we allow recalibration of the legacy SNIa, the result is very similar to that using DES alone. Given the lack of preference for any excess H 0 H_0 H 0 at all, the redshift scale of the excess (z 1 z_1 z 1 ) is largely unconstrained.
Figure C3. Triangle plot for the exp model (Eq. (12) ), proposed as a late-time solution to the Hubble tension. Results are shown using SNIa from DES (black), DES + legacy (red), and DES + legacy with free Δ M l e g a c y \Delta M_{\mathrm{legacy}} Δ M legacy (blue). No H 0 H_0 H 0 anchors are considered here. As a result, the excess amplitude A A A is consistent with zero and the redshift scale z 1 z_1 z 1 is unconstrained. The dotted orange lines mark Ω m \Omega_{\mathrm{m}} Ω m and H 0 H_0 H 0 in flat Λ \Lambda Λ CDM with the SPA parameters, whose small uncertainties we neglect (SPT-3G Collaboration [107 ] ).
We next show the triangle plots when adding a local H 0 H_0 H 0 anchor (Figure C4 ). The top panel considers three anchors that do not involve SNIa, while the bottom panel uses the SH0ES anchor (Breuval et al. 2024). These anchors are in principle independent but are consistent
Figure C4. Similar to Figure C3 , but including a local H 0 H_0 H 0 anchor (§ 3.4.1 ). The top panel jointly uses the three anchors not including SNIa (Jensen et al. 2025; Vogl et al. 2025; Mukherjee et al. 2021). The bottom panel uses the SH0ES anchor (Breuvall et al. 2024).
with each other, allowing us to combine all four anchors. We refrain from doing so because the statistical power would be dominated by SH0ES. Our results show that z 1 z_1 z 1 is not well constrained if we include the legacy SNIa, even if we allow them to be recalibrated. Excluding them leads to a tight upper limit on z 1 z_1 z 1 . With the non-SNIa anchors, the z 1 z_1 z 1 posterior is truncated at low values only by our prior floor of 0.001. However, when using DES + SH0ES without the legacy SNIa, the z 1 z_1 z 1 posterior becomes sharply peaked and is not truncated by the prior on z 1 z_1 z 1 . This peak does not measure z 1 z_1 z 1 on its own, however. The SH0ES anchor constrains the product A z 1 A z_1 A z 1 , and the peak sits where this ridge meets the upper edge of our prior on A A A (§ 4.6 ). We therefore do not interpret z 1 ≈ 0.03 z_1 \approx0.03 z 1 ≈ 0.03 as a property of a local void.
Redshift-floor sweep in LCDM mocks
A central result of this paper is the difference in H 0 H_0 H 0 and q 0 q_0 q 0 depending on whether we consider DES or DES + legacy (Figure 4 ). Our result for H 0 D E S H_0^{\mathrm{DES}} H 0 DES involves extrapolating over the poorly covered regime at z ≲ 0.2 z \lesssim0.2 z ≲ 0.2 to reach z → 0 z \to0 z → 0 . The resulting extrapolation uncertainty can be mitigated by adding legacy SNIa, which reach down to z = 0.025 z = 0.025 z = 0.025 and thereby yield a far more precise H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy . However, the result differs from H 0 D E S H_0^{\mathrm{DES}} H 0 DES . To determine how plausible it is that H 0 D E S H_0^{\mathrm{DES}} H 0 DES exceeds H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy by the observed Δ H 0 = 8.27 k m s − 1 M p c − 1 \Delta H_0 = 8.27\ \mathrm{km}\ \mathrm{s}^{-1}\ \mathrm{Mpc}^{-1} Δ H 0 = 8.27 km s − 1 Mpc − 1 , we set up mock datasets assuming flat Λ \Lambda Λ CDM with the SPA parameters based on the latest CMB observations (H 0 = 67.19 k m s − 1 M p c − 1 H_0 = 67.19\ \mathrm{km}\ \mathrm{s}^{-1}\ \mathrm{Mpc}^{-1} H 0 = 67.19 km s − 1 Mpc − 1 and Ω m = 0.3175 \Omega_{\mathrm{m}} = 0.3175 Ω m = 0.3175 ; SPT-3G Collaboration [107 ] ) and the Planck 2020 sound horizon of r d = 147.05 c M p c r_{\mathrm{d}} = 147.05\ \mathrm{cMpc} r d = 147.05 cMpc . We neglect uncertainties in these parameters because we expect larger uncertainties from the
SNIa, especially for DES alone due to the need to extrapolate to the present epoch. We generate mock BAO and SNIa data at the same redshifts as the real data, ensuring that every mock realisation of Λ \Lambda Λ CDM with the SPA parameters has the same redshift coverage as the real data.
We calculate ‘true’ SNIa apparent magnitudes in the assumed cosmology as follows:
m S N I a Λ C D M = M D E S + 5 log 10 [ ( 1 + z h e l ) D M , Λ C D M ( z H D ) ] + 25 , (28) m_{\mathrm{SNIa}}^{\Lambda\mathrm{CDM}} = M_{\mathrm{DES}} + 5\log_{10}\left[(1+z_{\mathrm{hel}})D_{\mathrm{M},\Lambda\mathrm{CDM}}(z_{\mathrm{HD}})\right] + 25, \tag*{(28)} m SNIa Λ CDM = M DES + 5 log 10 [ ( 1 + z hel ) D M , Λ CDM ( z HD ) ] + 25 , (28)
where Λ \Lambda Λ CDM superscripts indicate that the quantity does not include observational errors, the redshifts are the same as observed, and D M D_{\mathrm{M}} D M is the comoving distance to z H D z_{\mathrm{HD}} z HD in the SPA cosmology. Since the absolute calibration of the SNIa is irrelevant to the inferred cosmological parameters, we adopt M D E S = − 19.40 M_{\mathrm{DES}}=-19.40 M DES = − 19.40 and do not report M D E S M_{\mathrm{DES}} M DES results from our mocks. We then add observational scatter based on the total covariance matrix C S N I a \mathbf{C}_{\mathrm{SNIa}} C SNIa of the real data in the DES-Dovekie catalogue ((Popovic et al. 2026)). The apparent magnitude is therefore as follows:
m S N I a ∼ N ( m S N I a Λ C D M , C S N I a ) . (29) m_{\mathrm{SNIa}} \sim\mathcal{N}\left(m_{\mathrm{SNIa}}^{\Lambda\mathrm{CDM}},\mathbf{C}_{\mathrm{SNIa}}\right). \tag*{(29)} m SNIa ∼ N ( m SNIa Λ CDM , C SNIa ) . (29)
We follow a similar approach with the BAO. We generate mock BAO data at the same redshifts as those with real data from DESI DR2 BAO, propagating errors using their published covariance matrix C B A O \mathbf{C}_{\mathrm{BAO}} C BAO at each redshift (except at the lowest redshift where only D V D_{\mathrm{V}} D V is available; see DESI Collaboration 2025). The mock BAO measurements of D H , M , V / r d D_{\mathrm{H,M,V}}/r_{\mathrm{d}} D H , M , V / r d are therefore as follows:
D M , H , V r d ∼ N ( D M , H , V Λ C D M r d , C B A O ) . (30) \frac{D_{\mathrm{M,H,V}}}{r_{\mathrm{d}}} \sim\mathcal{N}\left(\frac{D_{\mathrm{M,H,V}}^{\Lambda\mathrm{CDM}}}{r_{\mathrm{d}}},\mathbf{C}_{\mathrm{BAO}}\right). \tag*{(30)} r d D M , H , V ∼ N ( r d D M , H , V Λ CDM , C BAO ) . (30)
For both the SNIa and BAO, we set up a single noiseless mock in which we do not add observational errors, so m S N I a = m S N I a Λ C D M m_{\mathrm{SNIa}}=m_{\mathrm{SNIa}}^{\Lambda\mathrm{CDM}} m SNIa = m SNIa Λ CDM and D M , H , V B A O / r d = D M , H , V Λ C D M / r d D_{\mathrm{M,H,V}}^{\mathrm{BAO}}/r_{\mathrm{d}}=D_{\mathrm{M,H,V}}^{\Lambda\mathrm{CDM}}/r_{\mathrm{d}} D M , H , V BAO / r d = D M , H , V Λ CDM / r d . We use the noiseless mock to separate the impact of observational errors and deficiencies in our fourth-order cosmographic expansion.
We conduct 10 4 10^4 1 0 4 realisations of the above setup. In each case, we extract H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy using the least-squares approach to reduce computational costs. The top panel of Figure D1 shows the joint distribution of these two H 0 H_0 H 0 estimates as a scatter plot, with the 1 σ 1\sigma 1 σ and 2 σ 2\sigma 2 σ contours overlaid due to the large number of points in the central region. The solid red star shows the result for the real data using NUTS, corresponding to the main part of this paper. The open blue diamond shows the result using instead the least-squares method, which gives an almost identical result but is much faster to evaluate. The observed combination of H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy is quite unlikely in the SPA cosmology, for which the true H 0 H_0 H 0 is shown using the solid lines. The dashed grey line shows the line of equality, from which the dotted red line is offset upwards by the observed Δ H 0 \Delta H_0 Δ H 0 . It is immediately apparent that in our mocks, both recovered H 0 H_0 H 0 values are somewhat higher than the truth. This remains the case even in our noiseless mock, which we indicate using a green plus symbol. The bias is thus an inherent property of our cosmographic IDL, though it is small compared to the random scatter caused by observational errors. In particular, ⟨ H 0 D E S ⟩ \langle H_0^{\mathrm{DES}}\rangle ⟨ H 0 DES ⟩ is biased by + 2.76 k m s − 1 M p c − 1 +2.76\ \mathrm{km\,s^{-1}\,Mpc^{-1}} + 2.76 km s − 1 Mp c − 1 compared to the true value, but the standard deviation of the best-fitting H 0 D E S H_0^{\mathrm{DES}} H 0 DES between different mocks is 2.86 k m s − 1 M p c − 1 2.86\ \mathrm{km\,s^{-1}\,Mpc^{-1}} 2.86 km s − 1 Mp c − 1 . H 0 D E S H_0^{\mathrm{DES}} H 0 DES is thus biased upwards by only + 0.96 σ +0.96\sigma + 0.96 σ . Due to the better redshift coverage of DES+legacy, it achieves a smaller bias: ⟨ H 0 D E S + l e g a c y ⟩ \langle H_0^{\mathrm{DES+legacy}}\rangle ⟨ H 0 DES + legacy ⟩ exceeds the true value by only + 0.57 k m s − 1 M p c − 1 +0.57\ \mathrm{km\,s^{-1}\,Mpc^{-1}} + 0.57 km s − 1 Mp c − 1 . Since the standard deviation of the best-fitting H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy values is 0.69 k m s − 1 M p c − 1 0.69\ \mathrm{km\,s^{-1}\,Mpc^{-1}} 0.69 km s − 1 Mp c − 1 , this represents a bias of + 0.82 σ +0.82\sigma + 0.82 σ . The results also show a weak positive correlation between H 0 D E S H_0^{\mathrm{DES}} H 0 DES and H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy , with a correlation coefficient of
+ 0.25 +0.25 + 0.25 . Higher H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy is thus generally associated with a correspondingly higher H 0 D E S H_0^{\mathrm{DES}} H 0 DES (the regression of H 0 D E S H_0^{\mathrm{DES}} H 0 DES on H 0 D E S + l e g a c y H_0^{\mathrm{DES+legacy}} H 0 DES + legacy has a slope of 1.04 1.04 1.04 ), albeit with large scatter. This correlation mitigates the impact on Δ H 0 \Delta H_0 Δ H 0 .
The bottom panel of Figure D1 shows the distribution of Δ H 0 \Delta H_0 Δ H 0 . The dotted black line shows the result of our noiseless mock. This sits near the mean of the distribution from our 10 4 10^4 1 0 4 mocks including noise, which we show as a solid black line. These results indicate a small positive bias of around + 2.2 k m s − 1 M p c − 1 +2.2\ \mathrm{km\,s^{-1}\,Mpc^{-1}} + 2.2 km s − 1 Mp c − 1 . The solid red line shows the result for the real data using NUTS. This is very similar to the corresponding least-squares result, which is shown as a dotted red line. Since the mocks are fitted by least squares, the like-for-like statistic is the least-squares Δ H 0 = 8.70 k m s − 1 M p c − 1 \Delta H_0=8.70\ \mathrm{km\,s^{-1}\,Mpc^{-1}} Δ H 0 = 8.70 km s − 1 Mp c − 1 of the real data, which only 85 of the 10 4 10^4 1 0 4 mocks exceed (0.85%, 2.39 σ 2.39\sigma 2.39 σ for a one-tailed test). The NUTS value that we quote is therefore the more conservative of the two. We shade the region in which Δ H 0 \Delta H_0 Δ H 0 exceeds the observed 8.27 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 . This occurs in just 124 of our 10 4 10^4 1 0 4 mocks. The P P P -value of 1.24% corresponds to 2.24 σ 2.24\sigma 2.24 σ for a one-tailed test, which is appropriate given the known sign of the Hubble tension (§ 4.9 ). It is therefore clear that the high observed Δ H 0 \Delta H_0 Δ H 0 is indeed unlikely in mock datasets where the underlying cosmology is Λ \Lambda Λ CDM with the SPA parameters. Part of this signal is the inaccuracy of our cosmographic expansion at finite k max k_{\max} k m a x , which biases H 0 D E S H_0^{\mathrm{DES}} H 0 DES upwards by 2.76 km s− 1 ^{-1} − 1 Mpc− 1 ^{-1} − 1 in our mocks, i.e. by 0.96 times its formal uncertainty. The formal uncertainty matches the scatter between mocks, but it does not absorb this bias, so the ± 1 σ \pm1\sigma ± 1 σ interval contains the input H 0 H_0 H 0 in only 49% of the mocks. These results indicate that H 0 D E S H_0^{\mathrm{DES}} H 0 DES cannot easily be reconciled with the SPA cosmology, despite assuming standard pre-recombination physics to set r d r_d r d and not using the standard distance ladder.
Our mocks show that the recovered H 0 H_0 H 0 becomes less precise and less accurate (more biased) if we remove the legacy SNIa due to the loss of coverage at low redshifts. To better understand this process, we repeat the redshift-floor sweep of § 4.4 in the Λ \Lambda Λ CDM mock datasets mentioned above. The results are shown in Figure D2 , with the left panel showing H 0 H_0 H 0 and the right panel showing q 0 q_0 q 0 in every 20th mock, so that the same 500 realisations appear in both panels. The thick solid line without points shows the median value at each redshift across the 10 4 10^4 1 0 4 mocks. The thin solid (dotted) lines on either side show the 1 σ 1\sigma 1 σ (2 σ 2\sigma 2 σ ) uncertainty using Gaussian equivalent percentiles. The curves with points show results with the real data. The dotted horizontal line in each panel shows the true value in the mocks. The recovered H 0 H_0 H 0 and q 0 q_0 q 0 are not much biased at low z f l o o r z_{\mathrm{floor}} z floor values due to the minimal extrapolation required to reach the present epoch. As we increase z f l o o r z_{\mathrm{floor}} z floor , the recovered H 0 H_0 H 0 is systematically biased high, while the opposite occurs for q 0 q_0 q 0 . However, this bias is not sufficient to explain the high observed H 0 D E S H_0^{\mathrm{DES}} H 0 DES within the SPA cosmology. Indeed, very few mocks have a curve that rises steeply to match the observed H 0 D E S H_0^{\mathrm{DES}} H 0 DES at its redshift floor of 0.073.
The bottom panels of Figure D2 show that the bias of our analysis is largely within its statistical uncertainty. The uncertainty our analysis returns for a single mock closely matches the scatter of the best-fitting values between mocks, both of which grow steadily with z f l o o r z_{\mathrm{floor}} z floor . Across a broad range of redshift floors, including that of DES, the true H 0 H_0 H 0 lies close to the 1 σ 1\sigma 1 σ lower limit of the median H 0 D E S H_0^{\mathrm{DES}} H 0 DES from the mocks, while the true q 0 q_0 q 0 lies close to the 1 σ 1\sigma 1 σ upper limit. The two mirror each other because a lower q 0 q_0 q 0 implies a faster late-time acceleration, and hence a higher H 0 H_0 H 0 . Our mocks cannot tell us the truncation error for the real cosmology, but in a Λ \Lambda Λ CDM-like one it is comparable to our reported uncertainty, even if DES had no SNIa below z = 0.3 z = 0.3 z = 0.3 . Since DES reaches z = 0.073 z = 0.073 z = 0.073 and has good coverage down to z ≈ 0.2 z \approx0.2 z ≈ 0.2 , we expect our cosmographic expansion to perform well on the real data.
We caution that in a redshift-floor sweep, the different points are all strongly correlated because every point uses a dataset that is strictly a subset of the data used at the preceding point. Even so, it is interesting that the DES-only result remains quite stable to the progressive addition of SNIa at low redshifts by decreasing z f l o o r z_{\mathrm{floor}} z floor . This suggests a lack of redshift-dependent internal tension within DES, making it all the more surprising that adding the legacy SNIa causes such a sharp change in behaviour for both H 0 H_0 H 0 and q 0 q_0 q 0 . The change is so sharp that in the narrow redshift band where DES and legacy samples overlap, there is a visible discrepancy between the results using DES and DES + legacy. This is especially evident in the insets to the top panels of Figure D2 at z ≈ 0.08 z \approx0.08 z ≈ 0.08 . The result of the redshift-floor sweep using the real DES + legacy data is therefore
not expected from our Λ \Lambda Λ CDM mocks. The smoother behaviour in the DES alone case is much more in line with what we might expect, but in this case the values of H 0 H_0 H 0 and q 0 q_0 q 0 differ from those in the SPA cosmology despite assuming a standard r d r_d r d . This issue could presumably be rectified if H 0 r d H_0 r_d H 0 r d and q 0 q_0 q 0 had values different from those in the SPA cosmology. However, the sharp feature in the DES + legacy case at precisely the redshift where legacy SNIa start being included seems unlikely to arise in any realistic cosmology. This coincidence is better understood as a systematic problem with the legacy SNIa measurements, at least when recalibrated onto the DES filter system.
This paper has been typeset from a \TeX/\LaTeX{} file prepared by the authors.