pith. sign in

arxiv: 2604.06974 · v1 · submitted 2026-04-08 · 📡 eess.SP · cs.IT· math.IT

The Gaussian data assumption does not always lead to the largest CRB

Jean-Pierre Delmas , Habti Abeida This is my paper

Pith reviewed 2026-05-10 17:44 UTC · model grok-4.3

classification 📡 eess.SP cs.ITmath.IT
keywords Cramér-Rao boundGaussian distributionFisher information matrixparameter estimationnon-Gaussian distributionsnuisance parameterscounterexamples
0
0 comments X

The pith

The Gaussian distribution does not always yield the largest Cramér-Rao bound.

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

The paper challenges the assumption that Gaussian data always produces the maximum Cramér-Rao bound for estimating parameters. It proves this property requires three conditions to hold simultaneously: the mean and covariance must be decoupled in the Fisher information matrix, the parameter of interest must lie in the mean vector, and no additive nuisance parameters can be present. When any of these conditions fail, the authors supply counterexamples in which non-Gaussian distributions produce strictly larger bounds. A reader should care because many practical estimators invoke the Gaussian case to claim optimality or tightness, yet the bound can be exceeded outside the narrow regime.

Core claim

The paper establishes that the Gaussian data model yields the largest CRB only when the mean and covariance parameters are decoupled in the FIM, the parameter of interest resides in the mean vector, and additive nuisance parameters are absent. Outside this regime, non-Gaussian distributions can produce larger CRBs, as demonstrated by explicit counterexamples.

What carries the argument

The three conditions on the Fisher information matrix structure: decoupling of mean and covariance, location of the parameter of interest in the mean vector, and absence of additive nuisance parameters.

If this is right

  • In problems with coupled mean and covariance parameters the CRB obtained under Gaussianity is no longer guaranteed to be the largest possible.
  • Estimation tasks that include additive nuisance parameters can admit non-Gaussian models with higher CRBs than the Gaussian model.
  • Bounds derived under the Gaussian assumption must be re-checked whenever the FIM structure violates any of the three conditions.

Where Pith is reading between the lines

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

  • Signal-processing designers who jointly estimate means and covariances should evaluate bounds for non-Gaussian models rather than defaulting to the Gaussian case.
  • The counterexamples may apply directly to array processing or communications scenarios where parameters are coupled through the data model.
  • A natural next step is to characterize the distribution that actually maximizes the CRB once the decoupling condition is dropped.

Load-bearing premise

The provided counterexamples are constructed correctly and the three listed conditions are necessary and sufficient for the Gaussian-maximizes-CRB property.

What would settle it

A calculation or simulation that reproduces one of the counterexample setups, computes the CRB for both Gaussian and the chosen non-Gaussian distribution, and finds that the Gaussian CRB remains strictly larger would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.06974 by Habti Abeida, Jean-Pierre Delmas.

Figure 1
Figure 1. Figure 1: Fig.1 [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
read the original abstract

This lecture note addresses the common misconception that the Gaussian distribution always yields the largest Cram\'er-Rao Bound (CRB). We show that this property only holds under restrictive conditions: specifically, when the mean and covariance parameters are decoupled in the Fisher Information Matrix (FIM), when the parameter of interest lies in the mean vector and when there are no additive nuisance parameters. Beyond this framework, we provide counterexamples demonstrating that non-Gaussian distributions can produce larger CRB.

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 / 2 minor

Summary. The paper claims that the Gaussian distribution does not always produce the largest Cramér-Rao Bound (CRB). It asserts that this property holds only under three restrictive conditions: decoupling of mean and covariance parameters in the Fisher Information Matrix (FIM), location of the parameter of interest within the mean vector, and absence of additive nuisance parameters. Beyond this setting, the manuscript supplies counterexamples in which non-Gaussian distributions yield strictly larger CRBs.

Significance. If the counterexamples are algebraically correct and the three conditions are shown to be necessary, the note supplies a useful clarification for estimation theory in signal processing. It prevents over-reliance on the Gaussian CRB in settings where the FIM structure or nuisance parameters differ, and it underscores that the Gaussian maximizer property is not distributionally universal once first- and second-order moments are fixed.

major comments (2)
  1. [Counterexamples (likely §3)] The load-bearing element is the set of counterexamples. The manuscript must verify that each non-Gaussian law is constructed with exactly the same first- and second-order moments as its Gaussian counterpart (so that any CRB increase is attributable solely to the distributional assumption and the violated condition). Any algebraic error in the score-function derivation of the FIM, or any implicit re-coupling of mean-covariance blocks, would invalidate the necessity claim.
  2. [Conditions for the Gaussian-max-CRB property (likely §2)] The paper should state explicitly whether the three listed conditions are jointly necessary and sufficient, or whether additional restrictions (e.g., on the support of the distribution or on the form of the nuisance parameters) are required. The abstract presents them as the complete framework, but the derivations must confirm that no other hidden assumptions are used when the Gaussian CRB is shown to be maximal.
minor comments (2)
  1. [Notation and definitions] Notation for the FIM blocks (mean-mean, mean-covariance, covariance-covariance) should be introduced once and used consistently; currently the decoupling condition is described in prose without an accompanying block-matrix equation.
  2. [Summary table] A short table summarizing the three conditions, the corresponding FIM structure, and the outcome for each counterexample would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our lecture note. The feedback has prompted us to strengthen the explicit verification of moment matching and to clarify the scope of the three conditions. We address each major comment below and indicate the corresponding revisions.

read point-by-point responses

  1. Referee: [Counterexamples (likely §3)] The load-bearing element is the set of counterexamples. The manuscript must verify that each non-Gaussian law is constructed with exactly the same first- and second-order moments as its Gaussian counterpart (so that any CRB increase is attributable solely to the distributional assumption and the violated condition). Any algebraic error in the score-function derivation of the FIM, or any implicit re-coupling of mean-covariance blocks, would invalidate the necessity claim.

    Authors: We agree that explicit verification of moment matching is essential. In the original manuscript the counterexamples were constructed to share identical means and covariances with the Gaussian case, but we acknowledge that this was not stated with sufficient detail. In the revised version we have added, in the new Section 3.1, direct calculations of the first- and second-order moments for every non-Gaussian distribution, together with the full score-function derivations and the resulting FIM blocks. These additions confirm that no algebraic errors exist and that mean-covariance decoupling (or coupling) is preserved exactly as required by each counterexample. We believe this removes any ambiguity regarding the source of the CRB increase. revision: yes

  2. Referee: [Conditions for the Gaussian-max-CRB property (likely §2)] The paper should state explicitly whether the three listed conditions are jointly necessary and sufficient, or whether additional restrictions (e.g., on the support of the distribution or on the form of the nuisance parameters) are required. The abstract presents them as the complete framework, but the derivations must confirm that no other hidden assumptions are used when the Gaussian CRB is shown to be maximal.

    Authors: We thank the referee for requesting this clarification. The three conditions are jointly sufficient for the Gaussian distribution to yield the largest CRB when the first two moments are fixed, under the standard regularity conditions that guarantee the existence of the FIM and CRB. The counterexamples demonstrate that violating any one of them can produce a strictly larger CRB for a non-Gaussian law. In the revision we have (i) rephrased the abstract to avoid implying completeness, (ii) added a dedicated paragraph in Section 2 that states the sufficiency result and notes that necessity is established only within the class of distributions sharing the same first two moments, and (iii) included a short remark acknowledging that further restrictions on support or nuisance structure could arise in more general settings. No additional hidden assumptions were employed in the derivations beyond those required for the FIM to be well-defined. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation relies on explicit FIM properties and independent counterexamples

full rationale

The paper establishes the restrictive conditions for the Gaussian-maximizes-CRB property by direct reference to the block structure of the Fisher Information Matrix and then constructs explicit counterexamples that violate at least one condition while preserving first- and second-order moments. No equation is defined in terms of its own output, no fitted parameter is relabeled as a prediction, and no load-bearing premise rests on a self-citation chain. The central demonstration therefore remains self-contained against the standard FIM definition and the supplied algebraic counterexamples.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard properties of the Fisher Information Matrix for parameterized distributions and on the usual regularity conditions required for the CRB to exist; no free parameters, new entities, or ad-hoc axioms are introduced in the abstract.

axioms (1)
  • standard math The Fisher Information Matrix is well-defined and the regularity conditions for the CRB hold for the distributions considered.
    Invoked implicitly when stating that the CRB exists and can be compared across distributions.

pith-pipeline@v0.9.0 · 5368 in / 1337 out tokens · 31818 ms · 2026-05-10T17:44:09.928704+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

11 extracted references · 11 canonical work pages

  1. [1]

    Stoica and R

    P. Stoica and R. Moses,Introduction to spectral analysis, Prentice-Hall, 1997

    work page 1997

  2. [2]

    Stoica and P

    P. Stoica and P. Babu, ”The Gaussian data assumption leads to the largest CRB”, Lecture note,IEEE Signal Process. Magazine, May 2011

    work page 2011

  3. [3]

    S. Park, E. Serpedin, and K. Qarage, ”Gaussian assumption: the least favorable but the most useful”, Lecture note,IEEE Signal Process. Magazine, May 2013

    work page 2013

  4. [4]

    Stein, A

    M. Stein, A. Mezghani, and J.O. Nossek, ”A lower bound for the Fisher information measure”,IEEE Signal Process. Letters, vol. 21, no. 7, July 2014

    work page 2014

  5. [5]

    Porat,Digital processing of random signalsPrentice-Hall, 1994

    B. Porat,Digital processing of random signalsPrentice-Hall, 1994

    work page 1994

  6. [6]

    Delmas, Chapter ”Background on real and complex elliptically symmetric distributions”, in Elliptical Distributions in Signal Processing and Machine Learning, Springer 2024

    J.-P. Delmas, Chapter ”Background on real and complex elliptically symmetric distributions”, in Elliptical Distributions in Signal Processing and Machine Learning, Springer 2024. April 9, 2026 DRAFT 9

    work page 2024

  7. [7]

    Abramowitz and I

    M. Abramowitz and I. A. Stegun,Handbook of mathematical functions with formulas, graphs, and mathematical tables, National Bureau of Standards, Applied Mathematics, 1972

    work page 1972

  8. [8]

    On the Fisher information matrix for multivariate elliptically contoured distributions,

    O. Besson and Y . I. Abramovich, “On the Fisher information matrix for multivariate elliptically contoured distributions,”IEEE Signal Process. Lett., vol. 20, no. 11, pp. 1130-1133, Nov. 2013

    work page 2013

  9. [9]

    Slepian-Bangs formula and Cram ´er Rao bound for circular and non-circular complex elliptical symmetric distributions,

    H. Abeida and J.-P. Delmas, “Slepian-Bangs formula and Cram ´er Rao bound for circular and non-circular complex elliptical symmetric distributions,”IEEE Signal Process. Letters, vol. 26, no. 10, pp. 1561-1565, Oct. 2019

    work page 2019

  10. [10]

    Abeida and J.-P

    H. Abeida and J.-P. Delmas, ”Efficiency of subspace-based estimators for elliptical symmetric distributions,”Signal Processing, vol. 174, Oct. 2020

    work page 2020

  11. [11]

    Abeida and J.-P

    H. Abeida and J.-P. Delmas, ”Slepian-Bangs formulas for parameterized density generator of elliptically symmetric distributions,” Signal Processing, vol. 205, Jan. 2023. April 9, 2026 DRAFT

    work page 2023