Paper: arXiv 2509.13923

Authors: Lamia Lamrani, Benoît Collins, Jean-Philippe Bouchaud

Abstract

Cross-validation is one of the most widely used methods for model selection and evaluation; its efficiency for large covariance matrix estimation appears robust in practice, but little is known about the theoretical behavior of its error. In this paper, we derive the expected Frobenius error of the holdout method, a particular cross-validation procedure that involves a single train and test split, for a generic rotationally invariant multiplicative noise model, therefore extending previous results to non-Gaussian data distributions. Our approach involves using the Weingarten calculus and the Ledoit-Péché formula to derive the oracle eigenvalues in the high-dimensional limit. When the population covariance matrix follows an inverse Wishart distribution, we approximate the expected holdout error, first with a linear shrinkage, then with a quadratic shrinkage to approximate the oracle eigenvalues. Under the linear approximation, we find that the optimal train-test split ratio is proportional to the square root of the matrix dimension. Then we compute Monte Carlo simulations of the holdout error for different distributions of the norm of the noise, such as the Gaussian, Student, and Laplace distributions and observe that the quadratic approximation yields a substantial improvement, especially around the optimal train-test split ratio. We also observe that a higher fourth-order moment of the Euclidean norm of the noise vector sharpens the holdout error curve near the optimal split and lowers the ideal train-test ratio, making the choice of the train-test ratio more important when performing the holdout method.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced random matrix theory, Weingarten calculus, and detailed asymptotic derivations for error estimation and optimal split ratios, indicating high mathematical complexity. However, it relies solely on Monte Carlo simulations for validation without presenting backtests, real financial datasets, or implementation-heavy metrics, resulting in lower empirical rigor.

Research Flowchart

  flowchart TD
  Goal["Research Goal:<br>Assess theoretical behavior of<br>holdout cross-validation error<br>for large non-Gaussian<br>covariance matrix estimation"]

  Methodology["Key Methodology:<br>Weingarten calculus &<br>Ledoit-Péché formula<br>for oracle eigenvalue derivation"]

  Inputs["Data/Inputs:<br>Rotationally invariant<br>multiplicative noise model<br>(Gaussian, Student, Laplace)"]

  Process["Computational Process:<br>1. Derive expected Frobenius error<br>2. Apply linear/quadratic shrinkage<br>3. Monte Carlo simulations"]

  Outcomes["Key Findings:<br>• Optimal split ratio ∝ √dimension<br>• Quadratic approx. improves accuracy<br>• Higher 4th moment sharpens error curve<br>• Changes optimal split ratio"]

  Goal --> Methodology
  Methodology --> Inputs
  Inputs --> Process
  Process --> Outcomes

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper derives the expected Frobenius holdout error for rotationally invariant multiplicative noise models using Weingarten calculus and extends the analysis to non‑Gaussian noise distributions.

Source sentence
In this paper, we derive the expected Frobenius error of the holdout method, a particular cross-validation procedure that involves a single train and test split, for a generic rotationally invariant multiplicative noise model, therefore extending previous results to non-Gaussian data distributions. Our approach involves using the Weingarten calculus and the Ledoit-Péché formula to derive the oracle eigenvalues in

Assumes 3

  • The population covariance matrix Σ follows an inverse Wishart distribution. model
    Source sentence
    However, when the population covariance matrix Σ is inverse Wishart and the data are Gaussian, the computations reduce to the linear shrinkage described in Proposition (1). In the following section, we will approximate the oracle eigenvalues using first a linear, then a quadratic shrinkage and will consider an inverse Wishart population covariance matrix.
  • The noise vectors are rotationally invariant (multiplicative noise model). data
    Source sentence
    Although the data are not assumed Gaussian but rotationally invariant, we will approximate the oracle eigenvalues by the eigenvalues of the optimal linear shrinkage given in proposition (1), let us define Diag(V t inΣVin)L as the diagonal matrix containing the optimal linear shrinkage eigenvalues, i.e. for i an integer between 1 and n (Diag(V t inΣVin)L)ii = rλEin i + (1 −r),
  • Finite sixth and eighth moments of the noise norm are assumed (E‖ξ‖⁶<∞ and E‖ξ‖⁸<∞). statistical
    Source sentence
    When Σ is an inverse Wishart of parameters n and p such that p is negligible in front of n, under the assumptions of proposition (3) and by assuming that E  ∥ξ∥6 < ∞ (83) and E  ∥ξ∥8 < ∞, (84) the expected holdout Frobenius error written in Eq. (82) in the high dimension limit equals E �∥ΞH −Σ∥2 Q  = " k t 3E  Tr((ξξT )2)  n(n + 2) −1 ! −1 # E + 1 + p,

Data

  • Monte Carlo simulations of rotationally invariant noise synthetic unknown
    Source sentence
    (87) 4 Applications to rotationally invariant noise distributions In this section, we compute the expected holdout error for different rotationally invariant noise distributions. For stability, the Monte Carlo is averaged over 20000 simulations for different distributions of the norm of the noise, namely Gaussian, Student and Laplace as defined in 4.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Under the linear shrinkage approximation, the optimal train‑test split ratio is proportional to the square root of the matrix dimension. theoretical
    Source sentence
    Under the linear approximation, we find that the optimal train-test split ratio is proportional to the square root of the matrix dimension.
  • The quadratic shrinkage approximation yields a substantial improvement in holdout error, especially around the optimal train‑test split ratio. theoretical
    Source sentence
    Then we compute Monte Carlo simulations of the holdout error for different distributions of the norm of the noise, such as the Gaussian, Student, and Laplace distributions and observe that the quadratic approximation yields a substantial improvement, especially around the optimal train-test split ratio.
  • A higher fourth‑order moment of the noise norm sharpens the holdout error curve near the optimal split and lowers the ideal train‑test ratio. theoretical
    Source sentence
    We also observe that a higher fourth-order moment of the Euclidean norm of the noise vector sharpens the holdout error curve near the optimal split and lowers the ideal train-test ratio, making the choice of the train-test ratio more important when performing the holdout method.
  • Monte Carlo simulations confirm that the quadratic approximation reduces the error compared to the linear approximation for Gaussian, Student, and Laplace noise norms. performance
    Source sentence
    For stability, the Monte Carlo is averaged over 20000 simulations for different distributions of the norm of the noise, namely Gaussian, Student and Laplace as defined in 4. We observe a small negative bias in the simulations which becomes more visible for higher values of γ.
  • A small negative bias is observed in simulations, becoming more pronounced for higher values of the noise‑norm factor γ. statistical
    Source sentence
    We observe a small negative bias in the simulations which becomes more visible for higher values of γ. Furthermore, the higher the value of γ, the higher the dispersion of the error around its mean making the choice of the optimal k more important.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 14 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.