pith. sign in

arxiv: 2606.18446 · v1 · pith:CLPVCKFFnew · submitted 2026-06-16 · 🌌 astro-ph.CO · astro-ph.IM· stat.ME

Covariance shrinkage for cosmological inference with Sellentin-Heavens-type likelihoods

Mattera Raffaele This is my paper

Pith reviewed 2026-06-26 22:52 UTC · model grok-4.3

classification 🌌 astro-ph.CO astro-ph.IMstat.ME
keywords covariance shrinkageSellentin-Heavens likelihoodcosmological inferencesimulation-estimated covarianceregularizationmarginalizationHartlap correctioncovariance uncertainty
0
0 comments X

The pith

Covariance scaling corrections depend on the likelihood form, and shrinkage with marginalisation over its intensity improves conditioning without needing extra Hartlap scaling under Sellentin-Heavens inference.

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

The paper establishes that scalar covariance corrections are likelihood-dependent: Hartlap scaling emerges as optimal for the Gaussian plug-in likelihood but the unscaled sample covariance is optimal once covariance uncertainty is marginalised via the Sellentin-Heavens likelihood. It then introduces a shrinkage procedure that regularises the sample covariance toward a spherical target while treating the shrinkage intensity itself as an auxiliary inferential quantity with its own prior. The final parameter posterior is obtained by marginalising over the posterior distribution of that intensity, which both improves matrix conditioning and carries uncertainty about the regularisation amount into the inference. Monte Carlo experiments confirm these effects for simulation-estimated covariances in cosmological settings.

Core claim

Scalar covariance corrections are likelihood-dependent and an additional Hartlap-type global scaling is not favoured once covariance uncertainty is marginalised through the Sellentin-Heavens likelihood. A shrinkage formulation in which the sample covariance is regularised towards a spherical target and the shrinkage intensity is treated as an auxiliary inferential quantity yields improved conditioning while propagating uncertainty about the amount of regularisation into posterior inference when the intensity is marginalised over.

What carries the argument

Shrinkage intensity treated as auxiliary inferential quantity with assigned prior whose posterior is induced by the likelihood, then marginalised to obtain the parameter posterior.

If this is right

  • Hartlap-type global scaling is not needed when covariance uncertainty is already marginalised through the Sellentin-Heavens likelihood.
  • Shrinkage substantially improves covariance conditioning.
  • Marginalisation over shrinkage intensity propagates uncertainty about the regularisation amount into the parameter posterior.
  • The approach combines covariance-marginalised likelihood inference with structural regularisation of noisy simulation-estimated covariance matrices.

Where Pith is reading between the lines

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

  • The method could be applied to other inference problems that rely on simulation-estimated covariances, such as in particle physics or financial risk modelling.
  • Alternative target matrices beyond the spherical choice could be substituted while retaining the marginalisation step.
  • The framework might be extended to non-Gaussian likelihoods or to joint inference of multiple covariance-related hyperparameters.

Load-bearing premise

A spherical target matrix is a suitable regularisation anchor and the chosen prior on shrinkage intensity does not introduce bias.

What would settle it

Monte Carlo experiments that compare posterior calibration and parameter constraints obtained with and without the proposed shrinkage marginalisation, using known true covariances from simulations, would show whether the conditioning and uncertainty-propagation benefits hold.

Figures

Figures reproduced from arXiv: 2606.18446 by Mattera Raffaele.

Figure 1
Figure 1. Figure 1: Empirical excess negative log-likelihood loss for scalar covariance corrections in a rep￾resentative autoregressive scenario with p = 50 and ns/p = 5. The dashed lines indicate the theoretical scalar optima. The two panels use separate vertical scales, and each loss is shifted by its minimum. Under the Gaussian plug-in likelihood, the empirical minimum is close to the Hartlap covariance-side scaling c ⋆ G.… view at source ↗
Figure 2
Figure 2. Figure 2: Nonlinear example with simulated data. Points denote the observed data vector, the dashed curve is the true mean function µi(θ) = exp(θ0 + θ1zi), and the solid curve is the posterior mean fit obtained by sampling the joint posterior of (θ, α) with Stan. B Technical results We collect here the standard multivariate distributional results used in the proofs. Let W ∼ Wp(Σ, ν) denote a p-dimensional Wishart ra… view at source ↗
read the original abstract

Covariance matrices used in astronomical and cosmological parameter inference are often estimated from a finite number of simulations, so covariance uncertainty can affect posterior calibration and parameter constraints. We study covariance regularisation from the perspective of likelihood-based inference with simulation-estimated covariance matrices. First, we analyse scalar covariance scaling under the Gaussian plug-in likelihood and the covariance-marginalised Sellentin--Heavens likelihood. Using an expected negative log-likelihood loss, we show that Hartlap covariance-side scaling is recovered as the optimum under the Gaussian plug-in likelihood, whereas the unscaled sample covariance is optimal under the Sellentin--Heavens likelihood. This shows that scalar covariance corrections are likelihood-dependent and that an additional Hartlap-type global scaling is not favoured once covariance uncertainty is marginalised through the Sellentin--Heavens likelihood. We then introduce a shrinkage formulation in which the sample covariance is regularised towards a spherical target and the shrinkage intensity is treated as an auxiliary inferential quantity. A prior is assigned to the shrinkage intensity, the likelihood induces its posterior distribution, and the final parameter posterior is obtained by marginalising over it. Monte Carlo experiments show that shrinkage substantially improves covariance conditioning, while marginalisation over the shrinkage intensity propagates uncertainty about the amount of regularisation into posterior inference. The proposed approach provides a simple way to combine covariance-marginalised likelihood inference with structural regularisation of noisy simulation-estimated covariance matrices.

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.

Referee Report

2 major / 1 minor

Summary. The manuscript claims that scalar covariance corrections are likelihood-dependent: Hartlap-type global scaling is optimal under the Gaussian plug-in likelihood but the unscaled sample covariance is optimal under the Sellentin-Heavens likelihood, as derived from an expected negative log-likelihood loss. It further proposes shrinking the sample covariance toward a spherical target while treating the shrinkage intensity as an auxiliary parameter with an assigned prior; the final posterior is obtained by marginalization over this intensity. Monte Carlo experiments are reported to show improved covariance conditioning and propagation of regularization uncertainty into the parameter posterior.

Significance. If the results hold, the work usefully demonstrates the likelihood dependence of simple covariance corrections and supplies a concrete procedure for combining Sellentin-Heavens marginalization with structural regularization. The Monte Carlo experiments constitute reproducible evidence for gains in conditioning and uncertainty propagation; these are concrete strengths that would be valuable for cosmological analyses limited by simulation-based covariance estimates.

major comments (2)
  1. [Shrinkage formulation] Shrinkage formulation: the adoption of a spherical target matrix together with a prior on the shrinkage intensity is introduced as a modeling choice whose mismatch with the strong off-diagonal structure of cosmological covariance matrices is not obviously corrected by marginalization; the manuscript does not provide a derivation or test showing that the marginal posterior on cosmological parameters remains unbiased under target misspecification.
  2. [Monte Carlo experiments] Monte Carlo experiments: while conditioning improvement is demonstrated, the experiments do not include a controlled test in which the spherical target is deliberately misspecified relative to the true covariance structure, leaving open whether the propagated uncertainty fully mitigates bias in the recovered parameter constraints.
minor comments (1)
  1. [scalar covariance scaling analysis] The expected-negative-log-likelihood derivation for optimality is summarized but the explicit loss function and its minimization steps are not shown; adding these would allow direct verification that the optimum is independent of the fitted parameters.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive report and for recognizing the manuscript's contributions on likelihood-dependent covariance corrections and the combination of Sellentin-Heavens marginalization with shrinkage. We address each major comment below.

read point-by-point responses

  1. Referee: [Shrinkage formulation] Shrinkage formulation: the adoption of a spherical target matrix together with a prior on the shrinkage intensity is introduced as a modeling choice whose mismatch with the strong off-diagonal structure of cosmological covariance matrices is not obviously corrected by marginalization; the manuscript does not provide a derivation or test showing that the marginal posterior on cosmological parameters remains unbiased under target misspecification.

    Authors: The spherical target is adopted for its analytical simplicity and precedent in the shrinkage literature. Marginalization over the shrinkage intensity is designed to let the data inform the effective amount of regularization rather than to guarantee unbiasedness under arbitrary target misspecification. The manuscript does not contain a derivation or dedicated test of unbiasedness for misspecified targets; we will add an explicit limitations paragraph discussing this assumption and its implications for cosmological covariances with strong off-diagonal structure. revision: partial

  2. Referee: [Monte Carlo experiments] Monte Carlo experiments: while conditioning improvement is demonstrated, the experiments do not include a controlled test in which the spherical target is deliberately misspecified relative to the true covariance structure, leaving open whether the propagated uncertainty fully mitigates bias in the recovered parameter constraints.

    Authors: The existing Monte Carlo suite demonstrates conditioning gains and uncertainty propagation when the target is a reasonable (if simplified) choice. We agree that a controlled misspecification test is absent and would strengthen the claims. In the revised version we will add such an experiment, deliberately using a mismatched spherical target and reporting the resulting shifts in the marginal posterior on cosmological parameters. revision: yes

Circularity Check

0 steps flagged

No circularity: optimality results derived from independent expected-loss functional; shrinkage is an explicit modeling proposal with assigned prior

full rationale

The paper derives that Hartlap scaling is optimal under the Gaussian plug-in likelihood but not under the Sellentin-Heavens likelihood by minimizing an expected negative log-likelihood loss that does not depend on any parameter fitted in the final result. The shrinkage construction treats intensity as an auxiliary variable with an explicit prior and performs marginalization; this is a modeling choice rather than a reduction of the target posterior to its own inputs by construction. No self-citation load-bearing steps, no fitted-input-called-prediction, and no self-definitional relations appear in the provided derivation chain. The spherical-target assumption is an unvalidated modeling premise but does not create circularity in the reported results.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on the Sellentin-Heavens likelihood (prior literature) and on the modeling choice of a spherical target plus a prior for the shrinkage intensity; no new physical entities are introduced.

free parameters (1)
  • prior on shrinkage intensity
    A prior must be assigned to the auxiliary shrinkage intensity; its functional form and hyperparameters are modeling choices not fixed by the data.
axioms (2)
  • domain assumption Covariance matrices are estimated from a finite number of simulations
    This is the starting point of the entire analysis.
  • domain assumption Sellentin-Heavens likelihood correctly marginalises covariance uncertainty
    Invoked when comparing optimality under the two likelihoods.

pith-pipeline@v0.9.1-grok · 5776 in / 1295 out tokens · 31817 ms · 2026-06-26T22:52:47.084477+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

  1. [1]

    Unbiased estimation of the inverse covariance matrix , author=

    Why your model parameter confidences might be too optimistic. Unbiased estimation of the inverse covariance matrix , author=. Astronomy & Astrophysics , volume=. 2007 , publisher=

    2007

  2. [2]

    Monthly Notices of the Royal Astronomical Society: Letters , volume=

    Parameter inference with estimated covariance matrices , author=. Monthly Notices of the Royal Astronomical Society: Letters , volume=. 2015 , publisher=

    2015

  3. [3]

    1982 , publisher=

    Aspects of multivariate statistical theory , author=. 1982 , publisher=

    1982

  4. [4]

    Monthly Notices of the Royal Astronomical Society: Letters , volume=

    Non-linear shrinkage estimation of large-scale structure covariance , author=. Monthly Notices of the Royal Astronomical Society: Letters , volume=. 2017 , publisher=

    2017

  5. [5]

    Journal of Multivariate Analysis , volume=

    A well-conditioned estimator for large-dimensional covariance matrices , author=. Journal of Multivariate Analysis , volume=. 2004 , publisher=

    2004

  6. [6]

    The Annals of Statistics , volume=

    Analytical nonlinear shrinkage of large-dimensional covariance matrices , author=. The Annals of Statistics , volume=. 2020 , publisher=

    2020

  7. [7]

    Statistics , volume=

    Improved precision matrix estimation for mean-variance portfolio selection , author=. Statistics , volume=. 2025 , publisher=

    2025

  8. [8]

    Journal of Empirical Finance , volume=

    Improved estimation of the covariance matrix of stock returns with an application to portfolio selection , author=. Journal of Empirical Finance , volume=. 2003 , publisher=

    2003

  9. [9]

    Intelligent Systems with Applications , volume=

    Shrinkage estimation with reinforcement learning of large variance matrices for portfolio selection , author=. Intelligent Systems with Applications , volume=. 2023 , publisher=

    2023

  10. [10]

    Journal of Financial Econometrics , volume=

    Learning the Shrinkage Intensity: A Data-Driven Approach for Risk-Optimized Portfolios , author=. Journal of Financial Econometrics , volume=. 2026 , publisher=

    2026

  11. [11]

    Physical Review D , volume=

    Dimensionality reduction techniques for statistical inference in cosmology , author=. Physical Review D , volume=. 2025 , publisher=

    2025

  12. [12]

    Monthly Notices of the Royal Astronomical Society , volume=

    Large covariance matrices: smooth models from the two-point correlation function , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2016 , publisher=

    2016

  13. [13]

    Monthly Notices of the Royal Astronomical Society , volume=

    Large covariance matrices: accurate models without mocks , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2019 , publisher=

    2019

  14. [14]

    Monthly Notices of the Royal Astronomical Society , volume=

    Putting the Precision in Precision Cosmology: How accurate should your data covariance matrix be? , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2013 , publisher=

    2013

  15. [15]

    arXiv preprint arXiv:2604.19463 , year=

    On combining estimated and analytic covariance matrices , author=. arXiv preprint arXiv:2604.19463 , year=

    Pith/arXiv arXiv

  16. [16]

    Physical Review D , volume=

    Fewer mocks and less noise: Reducing the dimensionality of cosmological observables with subspace projections , author=. Physical Review D , volume=. 2021 , publisher=

    2021

  17. [17]

    The Annals of Statistics , volume=

    Empirical Bayes estimation of the multivariate normal covariance matrix , author=. The Annals of Statistics , volume=. 1980 , publisher=

    1980

  18. [18]

    Journal of Multivariate Analysis , volume=

    Minimax estimators for a multinormal precision matrix , author=. Journal of Multivariate Analysis , volume=. 1977 , publisher=

    1977

  19. [19]

    Computational Statistics & Data Analysis , volume=

    Improved Stein-type shrinkage estimators for the high-dimensional multivariate normal covariance matrix , author=. Computational Statistics & Data Analysis , volume=. 2011 , publisher=

    2011

  20. [20]

    The Annals of Statistics , pages=

    Estimation of a covariance matrix under Stein's loss , author=. The Annals of Statistics , pages=. 1985 , publisher=

    1985

  21. [21]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Large covariance estimation by thresholding principal orthogonal complements , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2013 , publisher=

    2013

  22. [22]

    Physical Review D—Particles, Fields, Gravitation, and Cosmology , volume=

    The effect of covariance estimator error on cosmological parameter constraints , author=. Physical Review D—Particles, Fields, Gravitation, and Cosmology , volume=. 2013 , publisher=

    2013

  23. [23]

    Astronomy & Astrophysics , volume=

    Simulation-based inference has its own Dodelson--Schneider effect (but it knows that it does) , author=. Astronomy & Astrophysics , volume=. 2025 , publisher=

    2025

  24. [24]

    Monthly Notices of the Royal Astronomical Society , volume=

    Precision matrix expansion--efficient use of numerical simulations in estimating errors on cosmological parameters , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2018 , publisher=

    2018

  25. [25]

    Physical Review D , volume=

    Data compression with noise suppression for inference under noisy covariance , author=. Physical Review D , volume=. 2025 , publisher=

    2025

  26. [26]

    Monthly Notices of the Royal Astronomical Society , volume=

    Dark Energy Survey year 3 results: covariance modelling and its impact on parameter estimation and quality of fit , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2021 , publisher=

    2021

  27. [27]

    Monthly Notices of the Royal Astronomical Society , volume=

    CARPool covariance: fast, unbiased covariance estimation for large-scale structure observables , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2022 , publisher=

    2022

  28. [28]

    Monthly Notices of the Royal Astronomical Society , volume=

    CARPool: fast, accurate computation of large-scale structure statistics by pairing costly and cheap cosmological simulations , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2021 , publisher=

    2021

  29. [29]

    Monthly Notices of the Royal Astronomical Society , volume=

    Bayesian control variates for optimal covariance estimation with pairs of simulations and surrogates , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2022 , publisher=

    2022

  30. [30]

    Monthly Notices of the Royal Astronomical Society , volume=

    Matching Bayesian and frequentist coverage probabilities when using an approximate data covariance matrix , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2022 , publisher=

    2022

  31. [31]

    Monthly Notices of the Royal Astronomical Society , volume=

    Dark energy survey year 3 results: likelihood-free, simulation-based w cdm inference with neural compression of weak-lensing map statistics , author=. Monthly Notices of the Royal Astronomical Society , volume=. 2025 , publisher=

    2025