arxiv: 2604.24050 · v1 · submitted 2026-04-27 · 🌌 astro-ph.CO

Recognition: unknown

A sound-horizon-free measurement of the Hubble constant from DESI DR2 baryon acoustic oscillations using artificial neural networks

Gaurav N. Gadbail , Kazuharu Bamba

Authors on Pith no claims yet

Pith reviewed 2026-05-08 02:00 UTC · model grok-4.3

classification 🌌 astro-ph.CO

keywords Hubble constantbaryon acoustic oscillationsartificial neural networksHubble tensiondistance duality relationmodel-independentDESI DR2

0 comments

The pith

Artificial neural networks reconstruct distances from DESI BAO data to measure H0 at 71.5 km/s/Mpc without the sound horizon scale.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how to derive the Hubble constant from late-universe observations alone by using artificial neural networks to reconstruct redshift-distance functions in a fully data-driven way. It combines baryon acoustic oscillation measurements from DESI DR2 with type Ia supernovae and cosmic chronometer data through the distance duality relation. This avoids any input from the early-universe sound horizon or assumptions about supernova absolute brightness. The joint result is H0 = 71.5 plus or minus 2.2 km/s/Mpc, which sits closer to local distance-ladder values than to the lower Planck CMB inference.

Core claim

Artificial neural networks are trained to reconstruct the necessary distance functions directly from the data. These reconstructions are then combined via the distance duality relation across DESI DR2 baryon acoustic oscillations, supernovae, and cosmic chronometers to produce a model-independent Hubble constant of 71.5 plus or minus 2.2 km s^{-1} Mpc^{-1} at 68 percent . The value is consistent with local measurements within 0.6 sigma and with Planck within 2 sigma, thereby favoring a higher H0.

What carries the argument

Artificial neural network reconstruction of the distance-redshift relations, applied to the three probes and linked by the distance duality relation to remove dependence on the sound horizon scale r_d.

If this is right

The method supplies an independent, late-universe-only anchor for H0 that does not rely on early-universe physics.
It shows that DESI BAO data can constrain H0 once combined with supernovae and chronometers through duality.
The higher central value adds support for the Hubble tension being present in the data rather than an artifact of modeling.
Larger future BAO samples can be processed the same way to reduce the uncertainty without introducing new priors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the neural network outputs remain stable under changes in network size or training procedure, the technique could be applied to other cosmological parameters that currently depend on the sound horizon.
Persistent tension after more data would shift attention toward possible breakdowns in the distance duality relation itself.
The approach opens a path to test whether the tension appears in purely geometric, model-independent reconstructions at higher redshifts.

Load-bearing premise

The distance duality relation holds exactly between luminosity distances and angular diameter distances across the datasets.

What would settle it

A direct measurement showing that the product of reconstructed luminosity distance and angular diameter distance deviates from (1 plus z) squared by more than the reported uncertainty at overlapping redshifts would falsify the H0 result.

Figures

Figures reproduced from arXiv: 2604.24050 by Gaurav N. Gadbail, Kazuharu Bamba.

**Figure 1.** Figure 1: FIG. 1: The left panel shows the reconstruction of the Hubble parameter view at source ↗

**Figure 2.** Figure 2: FIG. 2: Posterior probability distribution functions view at source ↗

**Figure 3.** Figure 3: FIG. 3: Sound-horizon-free BAO combination view at source ↗

**Figure 5.** Figure 5: FIG. 5: Comparison of the joint view at source ↗

read the original abstract

We present a model-independent, sound-horizon-free measurement of the Hubble constant $H_0$ using baryon acoustic oscillation tracers from the Dark Energy Spectroscopic Instrument Data Release 2. The function reconstructions are performed using the artificial neural network method, which is a completely data-driven approach that avoids the mild $\Lambda$CDM prior dependence. Our approach is based on the distance duality relation and combines three complementary observational probes, such as Type Ia supernovae, cosmic chronometer, and DESI DR2 BAO -- without requiring any knowledge of the sound horizon scale $r_d$ or any assumption about the absolute luminosity of SNe Ia. We obtain a joint constraint of $H_0 = 71.5\pm2.2$ km s$^{-1}$ Mpc$^{-1}$ at 68\% confidence for 1000 bootstrap realisations and 4096 neurons, which is consistent with the TRGB result and the SH0ES measurement within $0.6\sigma$, consistent with the Planck 2020 result within $2\sigma$. Our results favor a higher value of $H_0$ compared to the Planck CMB inference, adding independent support for the reality of the Hubble tension.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a sound-horizon-free H0 of 71.5 from DESI DR2 BAO via ANN reconstruction and distance duality, landing between Planck and local values.

read the letter

The main thing to know is that the authors extract H0 without any input from the sound horizon scale or supernova absolute magnitude. They reconstruct the distance and expansion functions from DESI DR2 BAO using an artificial neural network, then combine those with supernovae and cosmic chronometers through the distance duality relation to set the absolute scale at z=0. The output is 71.5 ± 2.2 km/s/Mpc from 1000 bootstrap realizations with 4096 neurons fixed. This is new in its direct application to the latest DESI release with this fully data-driven setup. It does a clean job avoiding the usual priors that enter other analyses, and the derivation itself shows no internal contradictions: duality links DA and DL at shared redshifts, chronometers supply the scale, and H0 follows from the joint fit. The bootstrap approach gives a straightforward uncertainty estimate. The soft spots sit in the reconstruction details. The abstract and available description do not show explicit tests against mocks or checks for bias in how the network recovers the functions from the combined dataset. With a large fixed neuron count, sensitivity to that choice or to redshift matching across probes could matter, even if the central math holds. Readers working on the Hubble tension or model-independent expansion history would find this useful as one more independent data point. It deserves a serious referee to verify the implementation and any systematics in the ANN step. I would recommend sending it to peer review rather than desk rejecting it.

Referee Report

2 major / 2 minor

Summary. The manuscript claims a model-independent, sound-horizon-free determination of the Hubble constant using artificial neural networks to reconstruct distance and expansion functions from DESI DR2 BAO data. These are combined with Type Ia supernovae and cosmic chronometers via the distance-duality relation, without input values for the sound horizon rd or SN absolute magnitudes. The reported result is H0 = 71.5 ± 2.2 km s^{-1} Mpc^{-1} (68% CL) from 1000 bootstrap realizations with a 4096-neuron network, stated to be consistent with SH0ES/TRGB within 0.6σ and Planck within 2σ.

Significance. If the result holds, the work supplies an independent H0 constraint that sidesteps the sound-horizon scale and SN luminosity assumptions common in other analyses. The data-driven ANN reconstruction with fixed hyperparameters and bootstrap uncertainty estimation is a clear methodological strength, providing a flexible, non-parametric route to combine the three probes and yielding a falsifiable numerical prediction that can be tested against future data releases.

major comments (2)

[Methods / ANN reconstruction and error estimation] The central error budget (quoted ±2.2 km s^{-1} Mpc^{-1}) rests on 1000 bootstrap resamples of the combined dataset, yet the manuscript provides insufficient detail on how the joint DESI DR2 BAO + SNe + CC sample is constructed and resampled (e.g., whether probe-specific covariances are preserved). This directly affects the reliability of the final H0 uncertainty and must be expanded with explicit pseudocode or equations for the bootstrap procedure.
[ANN architecture and training] The choice of 4096 neurons is presented as fixed, but no justification, hyperparameter scan, or validation against mocks is shown to demonstrate that reconstruction biases remain sub-dominant to the statistical error. Without such tests, it is difficult to confirm that the quoted H0 value is free of method-induced systematics at the claimed precision.

minor comments (2)

[Abstract] The abstract states the final numerical result but omits the redshift range spanned by the DESI DR2 BAO measurements used in the joint fit; adding this would improve clarity.
[Introduction / Methods] Notation for the reconstructed functions (e.g., DA(z), H(z)) should be defined once at first use with explicit reference to the distance-duality relation employed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of our work and the recommendation for minor revision. We address each major comment below and will update the manuscript accordingly to improve clarity and reproducibility.

read point-by-point responses

Referee: The central error budget (quoted ±2.2 km s^{-1} Mpc^{-1}) rests on 1000 bootstrap resamples of the combined dataset, yet the manuscript provides insufficient detail on how the joint DESI DR2 BAO + SNe + CC sample is constructed and resampled (e.g., whether probe-specific covariances are preserved). This directly affects the reliability of the final H0 uncertainty and must be expanded with explicit pseudocode or equations for the bootstrap procedure.

Authors: We agree that additional detail on the bootstrap procedure is required for full reproducibility and to confirm the robustness of the quoted uncertainty. In the revised manuscript we will expand the Methods section with an explicit description of how the joint DESI DR2 BAO + SNe Ia + cosmic chronometer dataset is assembled, including the treatment of probe-specific covariances. We will also insert pseudocode (or equivalent equations) that outlines the resampling steps, making clear that each probe's covariance structure is preserved during the bootstrap. revision: yes
Referee: The choice of 4096 neurons is presented as fixed, but no justification, hyperparameter scan, or validation against mocks is shown to demonstrate that reconstruction biases remain sub-dominant to the statistical error. Without such tests, it is difficult to confirm that the quoted H0 value is free of method-induced systematics at the claimed precision.

Authors: We acknowledge that the manuscript would benefit from a clearer justification of the fixed 4096-neuron architecture. While internal convergence checks guided our choice, these were not documented. In the revision we will add a concise rationale based on reconstruction stability and note that the bootstrap ensemble already samples reconstruction variability. A full hyperparameter scan and dedicated mock validation campaign, however, would require substantial new computational work beyond the scope of a minor revision; we therefore provide only the existing checks and a statement that any residual bias is expected to be sub-dominant to the reported statistical error. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The derivation reconstructs expansion and distance functions via ANN applied directly to DESI DR2 BAO data, then links the probes through the standard distance-duality relation to SNe Ia and cosmic chronometers without any input value for rd or SN absolute magnitude. H0 is obtained at z=0 from the absolute scale supplied by the chronometers. No equation reduces a fitted parameter to a renamed prediction, no load-bearing premise rests on a self-citation chain, and the ANN is treated as a non-parametric interpolator whose hyperparameters are fixed externally to the target result. The bootstrap procedure estimates uncertainties from the data itself, leaving the central claim self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The measurement depends on the neural network being able to accurately reconstruct the functions in a model-independent way and on the validity of the distance duality relation. Additional details on training and potential free parameters in the network are not available from the abstract.

free parameters (2)

4096 neurons
Hyperparameter for the artificial neural network reconstruction
1000 bootstrap realizations
Number of resamples for uncertainty estimation

axioms (1)

domain assumption The distance duality relation holds exactly
Invoked to combine BAO, SNe, and chronometer data without additional assumptions on sound horizon or luminosity

pith-pipeline@v0.9.0 · 5528 in / 1479 out tokens · 90849 ms · 2026-05-08T02:00:54.554179+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

37 extracted references · 6 canonical work pages · 3 internal anchors

[1]

The function reconstructions are performed using the artificial neural network method, which is a completely data-driven approach that avoids the mildΛCDM prior dependence. Our approach is based on the distance duality relation and combines three complementary observational probes, such as Type Ia supernovae, cosmic chronometer, and DESI DR2 BAO – without...

work page internal anchor Pith review Pith/arXiv arXiv 2020
[2]

• One hidden layer with 4096 neurons and an Expo- nential Linear Unit (ELU) activation function [31]: f(x) = ( x, forx>0 α(ex −1), forx≤0 (8) withα=1

Network Architecture and T raining Procedure The network consists of: 4 • An input layer accepting the scalar 2 redshiftz. • One hidden layer with 4096 neurons and an Expo- nential Linear Unit (ELU) activation function [31]: f(x) = ( x, forx>0 α(ex −1), forx≤0 (8) withα=1. • An output layer with 2 neurons returning the re- constructed function value and i...
[3]

Log-Space Reconstruction for SN Ia SinceΞ SNIa spans several orders of magnitude across the redshift rangez∈[0.01, 2.26], we perform the ANN 2 A scalar is a single numerical value. 3 Batch normalisation standardises the distribution of layer activa- tions within each training batch, stabilising the optimisation land- scape and reducing sensitivity to the ...

2020
[4]

Aghanimet al.(Planck Collaboration), Astron

N. Aghanimet al.(Planck Collaboration), Astron. Astro- phys.641, A6 (2020)

2020
[5]

A. G. Riesset al.(SH0ES Collaboration), Astrophys. J. Lett. 934, L7 (2022). 9

2022
[6]

W. L. Freedmanet al., Astrophys. J.891, 57 (2020)

2020
[7]

W. L. Freedman, Astrophys. J.919, 16 (2021)

2021
[8]

The DESI Experiment, a whitepaper for Snowmass 2013

M. Leviet al., 2013, arXiv:1308.0847

work page Pith review arXiv 2013
[9]

Abdul Karimet al.(DESI Collaboration), Phys

M. Abdul Karimet al.(DESI Collaboration), Phys. Rev. D 112, 083514 (2025)

2025
[10]

Abdul Karimet al.(DESI Collaboration), Phys

M. Abdul Karimet al.(DESI Collaboration), Phys. Rev. D 112, 083515 (2025)

2025
[11]

Renzi, and A

F. Renzi, and A. Silvestri, Phys. Rev. D107, 023520 (2023)

2023
[12]

Guoet al., Astrophys

W. Guoet al., Astrophys. J. Lett.978,L33 (2025)

2025
[13]

Jiang, D

J.Q. Jiang, D. Pedrotti, S.S. da Costa, and S. Vagnozzi, Phys. Rev. D110, 123519 (2024)

2024
[14]

Wang, X.-J

G.-J. Wang, X.-J. Ma, S.-Y. Li, and J.-Q. Xia, Astrophys. J. Suppl. Ser.246, 13 (2020)

2020
[15]

Fluriet al., Phys

J. Fluriet al., Phys. Rev. D100, 063514 (2019)

2019
[16]

George, and E.A

D. George, and E.A. Huerta, Phys. Rev. D 97, 044039 (2018)

2018
[17]

Escamilla-Rivera, M.A.C

C. Escamilla-Rivera, M.A.C. Quintero and S. Capozziello, JCAP03, 008 (2020)

2020
[18]

Colgain and M.M

E.O. Colgain and M.M. Sheikh–Jabbari, (2021), arXiv:2101.08565

work page arXiv 2021
[19]

Cao and N

S. Cao and N. Liang, Research in Astron. Astrophys. 11(10), 1199-1208 (2011)

2011
[20]

S. Cao, M. Biesiada, X. Zheng and Z.-H. Zhu, MNRAS 457, 281-287 (2016)

2016
[21]

Qiet al., Phys

J.-Z. Qiet al., Phys. Rev. D99, 063507 (2019)

2019
[22]

Zhenget al., Astrophys

X. Zhenget al., Astrophys. J.892, 103 (2020)

2020
[23]

Liu,et al., Phys

T. Liu,et al., Phys. Lett. B838, 137687 (2023)

2023
[24]

Broutet al., Astrophys

D. Broutet al., Astrophys. J.938, 110 (2022)

2022
[25]

Riess,et al.Astrophys

A.G. Riess,et al.Astrophys. J.826, 56 (2016)

2016
[26]

Jimenez and A

R. Jimenez and A. Loeb, Astrophys. J.573, 37 (2002)

2002
[27]

Zhang, Z

C. Zhang, Z. Han, Y. Shuo, L. Siqi, Z. Tong-Jie, and S. Yan- Chun, Res. Astron. Astrophys.14, 1221 (2014)

2014
[28]

Stern, R

D. Stern, R. Jimenez, L. Verde, M. Kamionkowski, and S.A. Stanford, J. Cosmol. Astropart. Phys.02, 008 (2010)

2010
[29]

Morescoet al., J

M. Morescoet al., J. Cosmol. Astropart. Phys.08, 006 (2012)

2012
[30]

Morescoet al., J

M. Morescoet al., J. Cosmol. Astropart. Phys.05, 014 (2016)

2016
[31]

A. L. Ratsimbazafyet al., Mon. Not. Roy. Astron. Soc.467, 3239 (2017)

2017
[32]

Moresco, Mon

M. Moresco, Mon. Not. Roy. Astron. Soc.450, L16 (2015)

2015
[33]

J.-Z. Qi, P . Meng, J.-F. Zhang, and X. Zhang, Phys. Rev. D. 108, 063522 (2023)

2023
[34]

Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

D.-A. Clevert, T. Unterthiner, and S. Hochreiter, 2015, arXiv:1511.07289

work page Pith review arXiv 2015
[35]

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

S. Ioffe, and C. Szegedy, 2015, arXiv:1502.03167

work page internal anchor Pith review arXiv 2015
[36]

Adam: A Method for Stochastic Optimization

D.P . Kingma, and J. Ba, 2014, arXiv:1412.6980

work page internal anchor Pith review arXiv 2014
[37]

Ivezi´cet al., Astrophys

ˇZ. Ivezi´cet al., Astrophys. J.873, 111 (2019)

2019