Hierarchical Gaussian-process test of DESI's dynamical dark-energy preference

Enkun Li; Yu-Hao Mu; Zhihuan Zhou

Hierarchical Gaussian-process test of DESI's dynamical dark-energy preference

T0 review · reviewed 2026-07-30 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read With compressed CMB priors, a hierarchical Gaussian process finds late-time dark energy only about 0.8σ from a cosmological constant, and the same data in CPL improve ΛCDM by only ~1σ.

desk verdict Careful hierarchical-GP + ablation study on DESI DR2 with compressed Planck; mild ~1σ result is real for that pipeline but cannot adjudicate DESI’s full-likelihood claim. read the letter →

arxiv 2607.23593 v1 pith:SERH4K3T submitted 2026-07-26 astro-ph.CO

Yu-Hao Mu , Zhihuan Zhou , Enkun Li This is my paper · ORCID

classification astro-ph.CO

keywords darkenergyGaussianprocessDESIBAOcosmicchronometershierarchicalBayesianinferenceCPLparameterisationcompressedCMBpriors

topics Dark Energy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

DESI’s baryon acoustic oscillation data, when combined with CMB and supernovae in a simple two-parameter dark-energy model, have been reported to prefer evolving dark energy at several sigma. This paper asks whether that preference survives a more flexible, non-parametric reconstruction of the late-time expansion history under a carefully controlled analysis. The authors condition a hierarchical Gaussian process on cosmic-chronometer H(z) points, co-sample the kernel hyperparameters with the usual cosmological parameters, and couple DESI BAO, compressed Planck distance priors, and Pantheon+ supernovae through a Monte-Carlo effective likelihood. The reconstructed equation of state at low redshift sits only about 0.8σ from w = −1, a same-data CPL fit improves nested ΛCDM by roughly one in chi-squared, and a battery of ablations leaves the median still within about 1σ of a cosmological constant. The mild result is therefore tied to the compressed-CMB pipeline and does not speak to DESI’s full Planck-likelihood claim; a sympathetic reader cares because it isolates how much of the reported dynamical preference is analysis choice versus data.

What carries the argument

Hierarchical Gaussian process with co-sampled kernel hyperparameters (σ_f, l): the latent H(z) is conditioned on cosmic chronometers, then BAO, compressed CMB, and supernova likelihoods are averaged by Monte-Carlo draws from that GP posterior so hyperparameter uncertainty and non-linear distance functionals enter the joint posterior together.

What would settle it

Repeat the identical hierarchical GP and same-data CPL comparison after replacing the compressed (R, ℓ_A, ω_b h²) priors with the full Planck temperature-and-polarization likelihood; if the low-redshift w offset and Δχ² then rise to several sigma, the mild-preference claim fails to transfer.

Watch

Extended reading notes

Core claim

Under a hierarchical Gaussian process that co-samples kernel hyperparameters with cosmological parameters, conditioned on 37 cosmic-chronometer H(z) points and coupled to DESI DR2 BAO, compressed Planck distance priors, and Pantheon+ via a Monte-Carlo effective likelihood, the baseline posterior is w(z ≃ 0) = −0.80^{+0.26}_{-0.23} (about 0.8σ from −1). A CPL fit on the identical compressed pipeline improves nested ΛCDM by only Δχ² ∼ 1 (~1σ), and ablations that fix hyperparameters, drop supernovae or LRG1/2 BAO, change the kernel, or tighten the length-scale prior leave the median w(0) within ≲1σ of −1.

Load-bearing premise

Compressed Planck distance priors plus freezing the late-time expansion model above redshift 2.5 are treated as an adequate stand-in for the CMB information that drives the stronger DESI claim.

Editorial extensions

If this is right

In this compressed-CMB setting a DESI-like quintessence-to-phantom transition is not required by the late-time data.
Quoted dynamical-dark-energy significances remain tied to CMB treatment and statistical setup rather than to BAO alone.
Co-sampling kernel hyperparameters changes the high-z error budget at the several-percent level and shifts the low-z median by ~0.1 in w relative to fixing them at chronometer maximum likelihood.
Probe, tracer, kernel, and length-prior switches inside one pipeline leave w = −1 viable at ~1σ, so those modelling choices do not restore a strong transition here.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If full-likelihood re-analyses continue to show several-sigma preference while compressed-prior pipelines stay near ΛCDM, the discrepancy itself becomes a diagnostic of which CMB modes or early-universe assumptions are driving dynamical dark energy.
Freezing the GP above z = 2.5 when integrating to recombination means any genuine high-redshift dark-energy evolution would be absorbed into the sound-horizon and distance-prior calibration; joint early-plus-late non-parametric models would test that leakage.
The large posterior length scale (l ~ 3.8) that prefers smooth H(z) may systematically down-weight rapid low-z transitions that parametric CPL can still fit, so length-scale priors are themselves a hidden model choice when comparing non-parametric and parametric claims.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The punchline is simple: under compressed Planck distance priors, both their hierarchical GP and a same-pipeline CPL fit sit only ~0.8–1σ from ΛCDM on DESI DR2 + CC + Pantheon+. They say so clearly and do not pretend otherwise.

What is new is the controlled package, not the GP idea itself. They co-sample (σ_f, l) with the cosmology, build a Monte-Carlo effective likelihood that averages BAO/CMB/SN over GP draws conditioned on the 37 chronometers, and then run a clean ablation table (fixed hypers, no SN, no LRG1/2, Matérn-5/2, short-l prior) plus a same-data CPL Δχ²~1. That table is useful. The math is standard and consistently applied: GP conditioning, f_DE inversion to w(z), flat priors, emcee. Citations cover the right prior GP and DESI literature, including their own earlier work, without looking padded. Public data, private code — mid reproducibility, not a red flag for this subfield.

The soft spot is exactly the one they flag and the stress-test repeats. Compressed (R, ℓ_A, ω_b h²) plus freezing the GP above z=2.5 means the CMB enters as three numbers and a fitting-formula r_s. Their CPL check shares that compression, so it corroborates the GP inside the pipeline; it does not independently test whether DESI’s 2.8–4.2σ lives in the discarded modes. That is a scope limit, not a hidden flaw — the abstract and Sec. 5 state it. Secondary and smaller: N_s=20–30 draws per L_eff and modest N_eff on the ablations add some pseudo-marginal noise that can widen posteriors toward the null; the deterministic CPL cross-check blunts the worry.

This is for people who care about how DESI claims move under CMB compression and non-parametric reconstruction. It will not settle the dynamical-DE debate, and it does not claim to. I would send it to referees. A serious editor should; the methods are grounded and the disclaimer is honest. Engage if you are writing on DESI systematics or GP cosmography; otherwise skim the table and the caveat.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: hierarchical GP reconstruction and same-pipeline CPL are independent analyses of external data, not inputs redefined as outputs.

full rationale

The paper conditions a hierarchical GP on 37 public cosmic-chronometer H(z) points, co-samples kernel hyperparameters with cosmological parameters, and evaluates BAO, compressed Planck distance priors, and Pantheon+ through a Monte-Carlo effective likelihood. w(z) is obtained from the reconstructed expansion history via the standard continuity relation (Eq. 5), not by fitting a target that is then relabeled a prediction. The same-data CPL comparison (Δχ² ~ 1) is a separate parametric fit on the identical compressed pipeline, used as a consistency check rather than a quantity forced by construction. Ablations (fixed hypers, no SN, no LRG1/2, Matérn-5/2, short-l prior) further probe sensitivity without closing a definitional loop. Self-citations to the group’s earlier GP cosmography are background methodology references and do not supply a uniqueness theorem or load-bearing premise that forces the mild-preference conclusion. The paper’s own disclaimer that compressed CMB is not equivalent to DESI’s full Planck likelihood is a scope limitation, not circular reasoning. No step reduces a claimed prediction to its fitted input or to a self-citation chain.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The claim rests on standard FLRW distance relations and energy conservation, a zero-mean GP prior on H(z) conditioned only on 37 chronometers, flat priors on six sampled parameters, and the modeling choice that compressed Planck distance priors plus a high-z freeze of f_DE adequately represent CMB constraints for testing DESI’s dynamical preference. No new physical entities are introduced; free parameters are the usual cosmological set plus GP hyperparameters. The load-bearing domain assumption is inequivalence-tolerant use of compressed CMB.

free parameters (7)

H_0 = 70.1^{+3.6}_{-3.1} km s⁻¹ Mpc⁻¹
Sampled freely in [55,80] km s⁻¹ Mpc⁻¹ with no SH0ES or Planck H_0 prior; posterior median enters all distance predictions.
Ω_m = 0.317^{+0.032}_{-0.029}
Matter density sampled in [0.05,0.60]; controls Friedmann equation and sound-horizon inputs.
Ω_k = −0.0070^{+0.0013}_{-0.0014}
Curvature sampled in [−0.05,0.05]; affects transverse distances used in BAO and CMB compressed priors.
ω_b h² = 0.0224^{+0.0001}_{-0.0002}
Baryon density sampled in [0.018,0.026]; enters r_s(z_*) and compressed CMB vector.
σ_f (GP amplitude) = 284.6^{+135.9}_{-72.2} km s⁻¹ Mpc⁻¹
Kernel amplitude co-sampled via ln σ_f; sets allowed H(z) variance around the GP mean.
l (GP length scale) = 3.79^{+0.81}_{-1.12} h⁻¹ Gpc
Correlation length co-sampled via ln l; large posterior l smooths w(z) and is central to the mild low-z deviation.
N_s (MC draws per L_eff) = 20–30
Hand-chosen 20–30 GP draws per likelihood evaluation; controls Monte-Carlo noise in the effective BAO+CMB+SN likelihood.

assumptions (6)

domain assumption Late-time cosmology is described by FLRW with H²(z) as in Eq. (1) and distances from the line-of-sight integral r(z).
Section 2.1; standard background assumption shared with DESI CPL analyses.
domain assumption Dark-energy equation of state follows from energy conservation via w(z)=−1+(1+z)/3 f'_DE/f_DE after inverting the Friedmann equation for f_DE.
Eqs. (4)–(5); used as the reconstruction pipeline identity.
domain assumption H(z) is a latent draw from a zero-mean GP with RBF (baseline) or Matérn-5/2 kernel, conditioned only on 37 cosmic-chronometer points with diagonal noise.
Secs. 3.2–3.3; non-parametric prior that replaces a fixed w(z) family.
ad hoc to paper Compressed Planck (R, ℓ_A, ω_b h²) priors plus freezing f_DE above z=2.5 suffice to test whether DESI-like dynamical preference appears in this data combination.
Sec. 3.1 and Sec. 5 explicitly note inequivalence to full Planck likelihood; this is the pivotal modeling choice for the headline contrast.
domain assumption BAO, CMB compressed, and SN likelihoods can be replaced by a Monte-Carlo average of χ² over finite GP posterior draws (L_eff).
Eqs. (12)–(13); standard hierarchical approximation with acknowledged finite-N_s noise.
domain assumption Flat priors on the six parameters in Table 1 and analytic marginalization of SN absolute magnitude are non-informative enough not to drive w(0).
Sec. 3.5 and SN treatment in Sec. 3.1; standard but prior bounds (e.g. l≤5, short-l ablation l≤1.5) are shown to matter at the ~0.2–0.4 level in w.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hierarchical Gaussian-process test of DESI's dynamical dark-energy preference." pith.science (2026). https://pith.science/paper/SERH4K3T

@misc{pith2026260723593,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Gaussian-process test of DESI's dynamical dark-energy preference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SERH4K3T}},
  note         = {Machine review of arXiv:2607.23593}
}

abstract

DESI DR2 BAO combined with CMB and Type~Ia supernovae in the Chevallier--Polarski--Linder (CPL) parameterisation prefers evolving dark energy over $\Lambda$CDM at roughly $2.8$--$4.2\sigma$. We reconstruct the same late-time expansion history with a hierarchical Gaussian process (GP) that co-samples the kernel hyperparameters $(\sigma_f,l)$ together with $(H_0, \Omega_m, \Omega_k, \omega_b h^2)$, coupling BAO, compressed Planck distance priors, and Pantheon+ through a Monte-Carlo effective likelihood conditioned on 37 cosmic-chronometer $H(z)$ points. The baseline posterior gives $w(z\simeq 0)=-0.80^{+0.26}_{-0.23}$ (68\%~C.L.), about $0.8\sigma$ from $w=-1$, with $l=3.79^{+0.81}_{-1.12}$. A CPL fit on the identical compressed Planck pipeline improves nested $\Lambda$CDM by only $\Delta\chi^2\sim 1$ ($\sim 1\sigma$). Ablations that fix $(\sigma_f,l)$, drop SN or LRG1/2 BAO, replace the radial-basis kernel by Mat\'ern-$5/2$, or tighten the prior on $l$ leave the median $w(0)$ within $\lesssim 1\sigma$ of $-1$. The mild preference reported here is therefore specific to this compressed-CMB analysis and does not address DESI's full Planck-likelihood result.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references

[1]

Riess, et al., Astron

A.G. Riess, et al., Astron. J.116, 1009 (1998)

1998
[2]

Perlmutter, et al., Astrophys

S. Perlmutter, et al., Astrophys. J.517, 565 (1999)

1999
[3]

Schmidt, et al., Astrophys

B.P. Schmidt, et al., Astrophys. J.507, 46 (1998)

1998
[4]

Weinberg, Rev

S. Weinberg, Rev. Mod. Phys.61, 1 (1989)

1989
[5]

Peebles, B

P.J.E. Peebles, B. Ratra, Rev. Mod. Phys.75, 559 (2003)

2003
[6]

Copeland, M.S

E.J. Copeland, M.S. Sahni, S. Tsujikawa, Int. J. Mod. Phys. D15, 1753 (2006)

2006
[7]

M. Li, X.D. Li, S. Wang, Y. Wang, Commun. Theor. Phys.56, 525 (2011)

2011
[8]

Chevallier, D

M. Chevallier, D. Polarski, Int. J. Mod. Phys. D10, 213 (2001)

2001

Show all 66 references

[9]

Linder, Phys

E.V. Linder, Phys. Rev. Lett.90, 091301 (2003)

2003
[10]

Jassal, J.S

H.K. Jassal, J.S. Bagla, T. Padmanabhan, Mon. Not. Roy. Astron. Soc.356, L11 (2005)

2005
[11]

Barboza, J.S

E.M. Barboza, J.S. Alcaniz, Phys. Lett. B666, 415 (2008)

2008
[12]

Huterer, G

D. Huterer, G. Starkman, Phys. Rev. Lett.90, 031301 (2003)

2003
[13]

Albrecht, et al., (2006)

A. Albrecht, et al., (2006)

2006
[14]

Crittenden, L

R.G. Crittenden, L. Pogosian, G.B. Zhao, JCAP12, 025 (2009)

2009
[15]

Holsclaw, U

T. Holsclaw, U. Alam, B. Sanso, H. Lee, K. Heitmann, S. Habib, D. Higdon, Phys. Rev. Lett.105, 241302 (2010)

2010
[17]

Shafieloo, A.G

A. Shafieloo, A.G. Kim, E.V. Linder, Phys. Rev. D85, 123530 (2012)

2012
[18]

Zhang, H

M.J. Zhang, H. Li, Eur. Phys. J. C78(6), 460 (2018)

2018
[19]

Wang, X.H

D. Wang, X.H. Meng, Phys. Rev. D95(2), 023508 (2017)

2017
[20]

G´ omez-Valent, L

A. G´ omez-Valent, L. Amendola, JCAP04, 051 (2018)

2018
[21]

J. Niu, K. Jiao, P. He, T.J. Zhang, Astrophys. J.972, 14 (2024)

2024
[22]

Y. Mu, E.K. Li, L. Xu, JCAP06, 022 (2023)

2023
[23]

E.K. Li, M. Du, Z.H. Zhou, H. Zhang, L. Xu, Mon. Not. Roy. Astron. Soc.501(4), 4452 (2021)

2021
[24]

E.K. Li, H. Zhang, M. Du, Z.H. Zhou, L. Xu, JCAP08, 042 (2018)

2018
[25]

Zhang, E.K

W. Zhang, E.K. Li, M. Du, Y. Mu, L. Xu, (2019)

2019
[28]

Abdul-Karim, et al., (2025)

M. Abdul-Karim, et al., (2025)

2025
[29]

Adame, et al., JCAP02, 021 (2025)

A.G. Adame, et al., JCAP02, 021 (2025)

2025
[30]

Adame, et al., JCAP04, 012 (2025)

A.G. Adame, et al., JCAP04, 012 (2025)

2025
[31]

Lodha, et al., (2025)

K. Lodha, et al., (2025)

2025
[32]

G. Gu, X. Wang, Y. Wang, G.B. Zhao, L. Pogosian, K. Koyama, et al., Nature Astron. (2025)

2025
[34]

Cortˆ es, A.R

M. Cortˆ es, A.R. Liddle, JCAP12, 007 (2024)

2024
[35]

Wolf, P.G

W.J. Wolf, P.G. Ferreira, Phys. Rev. D108(10), 103519 (2023)

2023
[36]

Efstathiou, Mon

G. Efstathiou, Mon. Not. Roy. Astron. Soc.538(2), 875 (2025)

2025
[37]

Calder´ on, K

R. Calder´ on, K. Lodha, A. Shafieloo, E. Linder, W. Sohn, A. de Mattia, J.L. Cervantes-Cota, R. Crittenden, T.M. Davis, et al., JCAP10, 048 (2024)

2024
[38]

Lodha, A

K. Lodha, A. Shafieloo, R. Calder´ on, E. Linder, W. Sohn, J.L. Cervantes-Cota, A. de Mattia, J. Garc ´ ıa-Garc ´ ıa, et al., Phys. Rev. D111(2), 023532 (2025)

2025
[40]

Y. Wang, K. Freese, (2025)

2025
[41]

Loubser, et al., (2025)

S.I. Loubser, et al., (2025)

2025
[42]

Cortˆ es, A.R

M. Cortˆ es, A.R. Liddle, (2025)

2025
[43]

D. Wang, D. Mota, (2025)

2025
[44]

G. Ye, S.J. Lin, (2025)

2025
[45]

de Souza, A

R. de Souza, A. Sousa-Neto, J.E. Gonz´ alez, J. Alcaniz, (2025)

2025
[46]

Zhang, Y.H

X. Zhang, Y.H. Xu, Y. Sang, (2025)

2025
[47]

J.X. Li, S. Wang, (2025)

2025
[48]

Rasmussen, C.K.I

C.E. Rasmussen, C.K.I. Williams,Gaussian Processes for Machine Learning(MIT Press, Cambridge, MA, 2006)

2006
[49]

Seikel, C

M. Seikel, C. Clarkson, M. Smith, JCAP06, 036 (2012)

2012
[50]

Benisty, R

D. Benisty, R. D’Agostino, R.C. Nunes, E. Di Valentino, JCAP01, 018 (2023)

2023
[51]

Holanda, J.C

R.F.L. Holanda, J.C. Carvalho, J.S. Alcaniz, JCAP01, 060 (2017)

2017
[52]

Crittenden, G.B

R.G. Crittenden, G.B. Zhao, L. Pogosian, L. Samushia, X. Zhang, JCAP02, 048 (2012)

2012
[53]

Zhao, et al., Nature Astron.1(9), 627 (2017)

G.B. Zhao, et al., Nature Astron.1(9), 627 (2017)

2017
[54]

Y. Wang, L. Pogosian, G.B. Zhao, A. Zucca, Astrophys. J. Lett.869, L8 (2018)

2018
[55]

Jimenez, A

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

2002
[56]

Eisenstein, et al., Astrophys

D.J. Eisenstein, et al., Astrophys. J.633, 560 (2005)

2005
[58]

Brout, et al., Astrophys

D. Brout, et al., Astrophys. J.938(2), 110 (2022)

2022
[60]

Vagnozzi, Universe9(9), 393 (2023)

S. Vagnozzi, Universe9(9), 393 (2023)

2023
[61]

Foreman-Mackey, D.W

D. Foreman-Mackey, D.W. Hogg, D. Lang, J. Goodman, Publ. Astron. Soc. Pac.125, 306 (2013)

2013
[62]

Goodman, J

J. Goodman, J. Weare, Commun. Appl. Math. Comput. Sci.5(1), 65 (2010)

2010
[63]

S¨ arkk¨ a, A

S. S¨ arkk¨ a, A. Solin,Applied Stochastic Differential Equa- tions(Cambridge University Press, Cambridge, 2019)

2019
[64]

Moresco, et al., Living Rev

M. Moresco, et al., Living Rev. Rel.25(1), 6 (2022)

2022
[65]

Huang, K

Q.G. Huang, K. Wang, JCAP05, 035 (2020)

2020
[66]

Aghanim, et al., Astron

N. Aghanim, et al., Astron. Astrophys.641, A6 (2020)

2020
[67]

Goliath, R

M. Goliath, R. Amanullah, P. Astier, A. Goobar, R. Pain, Astron. Astrophys.380, 6 (2001)

2001
[68]

Conley, et al., Astrophys

A. Conley, et al., Astrophys. J. Suppl.192, 1 (2011)

2011
[69]

W. Hu, N. Sugiyama, Astrophys. J.471, 542 (1996)

1996
[70]

Eisenstein, W

D.J. Eisenstein, W. Hu, Astrophys. J.496, 605 (1998)

1998
[71]

Riess, et al., Astrophys

A.G. Riess, et al., Astrophys. J. Lett.934(1), L7 (2022)

2022
[72]

Keeley, A

R.E. Keeley, A. Shafieloo, W.L. Matthewson, (2025)

2025
[73]

Busti, C

V.C. Busti, C. Clarkson, M. Seikel, Mon. Not. Roy. As- tron. Soc.441, 11 (2014)

2014

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.

[1] [1]

Riess, et al., Astron

A.G. Riess, et al., Astron. J.116, 1009 (1998)

1998

[2] [2]

Perlmutter, et al., Astrophys

S. Perlmutter, et al., Astrophys. J.517, 565 (1999)

1999

[3] [3]

Schmidt, et al., Astrophys

B.P. Schmidt, et al., Astrophys. J.507, 46 (1998)

1998

[4] [4]

Weinberg, Rev

S. Weinberg, Rev. Mod. Phys.61, 1 (1989)

1989

[5] [5]

Peebles, B

P.J.E. Peebles, B. Ratra, Rev. Mod. Phys.75, 559 (2003)

2003

[6] [6]

Copeland, M.S

E.J. Copeland, M.S. Sahni, S. Tsujikawa, Int. J. Mod. Phys. D15, 1753 (2006)

2006

[7] [7]

M. Li, X.D. Li, S. Wang, Y. Wang, Commun. Theor. Phys.56, 525 (2011)

2011

[8] [8]

Chevallier, D

M. Chevallier, D. Polarski, Int. J. Mod. Phys. D10, 213 (2001)

2001