Functional uniqueness and stability of Gaussian priors in optimal L1 estimation

Alex Dytso; Leighton Barnes

REVIEW 3 major objections 5 minor 16 references

Functional uniqueness and stability of Gaussian priors in optimal L1 estimation

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read If an optimal estimator is nearly linear under Gaussian noise, the prior must be nearly Gaussian.

desk verdict The L1 stability framework is genuinely new, but Theorem 2 rests on a false convolution identity as printed; repairable, but not citable in this form. read the letter →

arxiv 2511.16864 v2 pith:BEGYAJFB submitted 2025-11-21 cs.IT math.IT

Leighton Barnes , Alex Dytso This is my paper

classification cs.ITmath.IT

keywords GaussianpriorslinearestimatorsconditionalmedianL1estimationHermiteexpansionsstabilityLévymetricBayesian

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

This paper proves a stability theorem for the classical Gaussian-noise model Y = X + Z, where Z is standard normal and independent of X. Earlier work showed that the Gaussian prior is the unique distribution making the optimal estimator linear, both for squared-error loss (conditional mean) and for absolute-error loss (conditional median). The paper shows that uniqueness is stable: if the optimal estimator is only approximately linear, the prior must be approximately Gaussian. For L2 loss it derives an explicit rate of closeness to Gaussian in the Lévy metric; for L1 loss it develops a Hermite-expansion and adjoint-operator framework, showing that near-linearity of the conditional median forces all non-Gaussian Hermite coefficients of the prior density to be small. This gives the first quantitative stability statement for L1-optimal linearity in this setting.

What carries the argument

The engine of the L1 proof is the operator T_a[f](y) = ∫ f(x) sign(x−ay) e^{−(x−y)^2/2} dx, which vanishes for all y exactly when the conditional median equals ay. The paper rewrites T_a as a convolution with a sign kernel, analyzes a tilted density f̃(x)=e^{x^2(1−a)/(2a)}f(x), and derives a growth estimate on local averages of f̃. It then constructs, for each Hermite function H_n, a function φ_n such that the adjoint operator satisfies T_a^* φ_n = H_n, with the Fourier norm of e^{−(1−a)y^2/2}φ_n growing like n^{1/4}. This lets the authors convert an L1 bound on the tilted T_a[f] into a bound on the Hermite coefficient ⟨f, H_n⟩.

What would settle it

Substitute x = a u into the defining integral (21) for T_a[f] and compare the result with the printed identity (75) for a test density such as a Gaussian; a mismatch would invalidate the proof of Theorem 2 and require either a corrected identity or a new argument.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 2: under two decay assumptions on the difference between the conditional median and the line ay — the difference weighted by a Gaussian factor is integrable, and the inverse difference is uniformly bounded by a step function whose total integral is at most ε — the inner product ⟨f, H_n⟩ between the prior density f and the nth Hermite function is at most L_n ε for every n≥1, where L_n grows like n^{1/4}. Because the Hermite functions form a complete orthonormal basis, this means f is almost orthogonal to every basis element except the Gaussian H_0, so it is close to Gaussian in a functional sense. With fast Hermite decay, Corollary 1 upgrades this to L2 cl

Load-bearing premise

The proof relies on the convolution identity at equation (75) that re-expresses the conditional-median operator as a single integral; if this identity is not valid, the growth estimate, the L1 bound, and Theorem 2 all collapse.

Editorial extensions

If this is right

If the conditional median is pointwise close to a line with Gaussian-tail decay, the prior density is nearly orthogonal to all Hermite functions except the Gaussian, so it is weakly close to Gaussian.
Corollary 1: if the prior's Hermite coefficients decay fast enough, near-linearity of the conditional median upgrades to L2 closeness to the Gaussian density.
Theorem 1 gives a quantitative Lévy-metric rate for L2 stability, improving the qualitative stability known for conditional means.
Setting ε=0 in Theorem 2 recovers the uniqueness of the Gaussian prior for linear conditional medians via a purely functional proof that avoids tempered distributions.
The inverse-function assumption (63) shows that L1 proximity of the estimator to a line is equivalent, up to constants, to the same proximity in the quantile domain.

Reading between the lines

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

The adjoint-operator construction could generalize to other estimation problems where optimality is characterized by a sign-kernel integral equation, not just linear conditional medians.
The explicit L2 rate in Theorem 1 is likely not optimal; the paper itself leaves the tightness of the ε-order open.
If the convolution identity in the L1 proof is repaired, the same growth-estimate machinery might yield stability under weaker tail assumptions.
The result has a practical reading: if an empirical estimator in Gaussian noise looks nearly linear, any Bayesian explanation is constrained to use a nearly Gaussian prior.

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

What you should know: the L1 stability result (Theorem 2) is a real step beyond the authors' earlier uniqueness paper, and the Hermite-adjoint machinery is the right sort of tool for the problem. But the proof has a load-bearing algebraic error: equation (75) does not follow from the definitions. Substituting x=au into (21) with their tilde-f gives an extra factor a and an extra Gaussian factor; the printed identity is simply false. Since (75) feeds directly into (77), (80), and the final L1 bound (90), the main theorem is unsupported as written. This is not a cosmetic slip — it's the bridge between the operator T_a and the convolution form that the whole argument depends on.

That said, the paper earns credit where it is due. The L2 section is a clean, self-contained re-derivation of known stability with explicit constants; the Esseen-step bound in Theorem 1 looks off by a log factor or exponent, but that part is incremental anyway. Proposition 5's construction of the adjoint preimages phi_n is clever, and the growth estimate (80) would be a solid route to the theorem if (75) can be repaired. The authors also deserve credit for not hiding the strong assumptions: they explicitly state the decay condition (62) and the step-envelope condition (63), and the corollary makes clear that functional stability in L2 requires extra Hermite decay. There is no fitted-parameter circularity here; epsilon and a are problem inputs.

On the secondary issues: the Stirling estimate in Proposition 5 has a reciprocal inconsistency (C_a sqrt(n!)/2^{n/2} vs 2^{n/2}/sqrt(n!)), so the O(n^{1/4}) rate is not proven as written either. That one looks easily fixable. The reliance on the authors' own [1] is heavy, but it is natural — they are extending their own theorem.

The bottom line: the paper deserves a serious referee, but the referee should be told to focus on (75) and the Stirling step. I would not cite the L1 theorem in its current form, and I would not assign it to a reading group until a corrected version appears. If the identity can be fixed (and the missing factor only changes constants, so it probably can), the theorem is likely true and the framework is worth publishing.

Referee Report

3 major / 5 minor

Summary. The paper studies stability of Gaussian priors under approximate linearity of Bayesian estimators in the additive Gaussian noise model Y=X+Z. Theorem 1 gives an explicit Lévy-metric bound on the prior when the conditional mean is close to a linear function in mean-squared error. Theorem 2 states that if the conditional median is ε-close to a line in a strong pointwise/integrated sense (conditions (62)-(63)), then all non-constant Hermite coefficients of the prior density are bounded by L_n ε with L_n = O(n^{1/4}), implying that the prior is close to Gaussian in an L^2 sense when the Hermite coefficients decay sufficiently fast. The L1 proof uses a Hermite expansion framework and an adjoint operator T_a^*, building on the authors' earlier uniqueness result.

Significance. If the results are correct, the L1 stability theorem is a meaningful new contribution: it provides the first quantitative stability analogue of Gaussian uniqueness for conditional medians under Gaussian noise, a setting where standard orthogonality tools fail. The Hermite/adjoint framework is well chosen and is a strength of the paper. The L2 proof is also clean and gives explicit rates. No constants are fitted to make the conclusion true; ε and a are inputs, which is a further strength. However, the L1 proof as printed contains load-bearing algebraic errors, and the displayed L2 bound has a typo that makes it vacuous as ε→0. These issues are repairable in a revision, but the theorems are not proven as stated.

major comments (3)

[§3.2, Eq. (75)] Equation (75) is false as printed. Substituting x = a u in (21), multiplying by e^{(1-a)y^2/2}, and using \tilde f(a u)=e^{(1-a)a u^2/2}f(a u) gives e^{(1-a)y^2/2} T_a[f](y) = a ∫ \tilde f(a u) sign(u-y) e^{-a(u-y)^2/2} du, not the printed identity. The factor a is missing. This identity is used in (76)-(77) to derive the local bound (80) and again in (82)-(88) to derive the key estimate (90); as printed, that chain is invalid. The error is repairable by inserting a and tracking constants, but it is load-bearing.
[Appendix, Proposition 5, Eqs. (43)-(45)] There is a reciprocal inconsistency in the Stirling estimate. Since K_n = π^{1/4} 2^{n/2} √(n!) σ^{n-1/2}, one has 1/K_n ∝ 2^{-n/2}/√(n!), so the coefficient in (43) should be C_a 2^{n/2}/√(n!), not C_a √(n!)/2^{n/2}. Consequently (44) bounds the L1 norm by √(n!)·Γ((n+2)/2)/2^{n/2}, which grows far faster than n^{1/4}; (45) computes the reciprocal ratio 2^{n/2}Γ/√(n!) = O(n^{1/4}). Since Theorem 2's L_n = O(n^{1/4}) rests on this proposition, Proposition 5 is not established as written.
[§3.1, Theorem 1, Eq. (48) vs (59)-(61)] The displayed L2 bound in (48) is inconsistent with the proof. With T = √(log(1/ε)/(1+δ)), the second term of (60) becomes (24/π)√(2πσ_X^2)√((1+δ)/log(1/ε)), which tends to 0 as ε→0. The theorem statement instead has (24/π)√(2πσ_X^2/(1+δ))log(1/ε), which tends to ∞, making the bound vacuous. The first term in (48) also has an extra denominator. The proof's own estimate (61) also appears to misstate the second term. This needs correction; the claimed L2 rate is not established as printed.

minor comments (5)

[§3.2, Theorem 2 statement] The proof uses the L2 Hermite expansion f = Σ c_i H_i, but the theorem only assumes f(x) ≤ M and does not explicitly state f ∈ L^2. The statement should include square-integrability of f (or otherwise justify the expansion). This is a missing hypothesis rather than a technical gap.
[§3.2, Eq. (77)] The dominated convergence argument is asserted without the truncation needed because \tilde f(ax) is not globally integrable. Please provide details on the limiting argument and the boundary term absorbed into C_1.
[§3.2, Eq. (65)] The equality ∫|ψ(y)-ay|dy = ∫|ψ^{-1}(x)-x/a|dx is true for monotone ψ (both sides equal the area between the two graphs), but it should be stated and justified, especially because ψ is only assumed nondecreasing and may have flat pieces.
[§3.2, Eq. (88)] The notation ∆(x)=b_i for x∈[i,i+1) requires b_i to be nonnegative and summable; please state this explicitly so that ∫∆ = Σ b_i is unambiguous.
[§3.1, Eq. (60)] The phrase 'boundedness of the Dawson integral' in the step leading to (60) is inaccurate; the relevant bound is just ∫_{-T}^T e^{ω^2/2} dω ≤ 2e^{T^2/2}. The reference to Dawson's integral is unnecessary here.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 2 derives Hermite-coefficient bounds from ε-linearity assumptions via operator estimates; the Gaussian target is not used as an input or fit.

full rationale

I walked the derivation chain in Theorem 2. The input is the ε-close conditional median assumptions (62)–(63), and the output is the Hermite-coefficient bound |⟨f,H_n⟩| ≤ L_n ε (64). The proof introduces a rescaled density f̃, establishes local L1 growth estimates (68)–(80), and then uses the adjoint construction of Proposition 5 to transfer T_a[f] bounds to Hermite coefficients (91)–(94). None of these steps defines ε, a, or the Hermite coefficients in terms of the conclusion: ε and a are problem data, H_n are fixed basis functions, and L_n is obtained from a Stirling estimate rather than fitted to force (64). The appeals to the authors' prior work [1] are for the convolution identity (75) and the kernel lower bound (79); these are technical lemmas about the fixed operator T_a and are not restatements of near-Gaussianity of f, so they do not make the argument circular. The skeptical observation that (75) is missing a factor a is a correctness concern about the proof as printed, not a self-referential reduction; repairing the constant would not change the logical structure. No fitted-input-called-prediction, self-definitional, or uniqueness-imported-by-self-citation pattern occurs in the main theorem. The uniqueness result from [1] is used as background and as a source of technical tricks, but the stability result is not obtained by assuming the conclusion.

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

No parameters fitted to data. The quantities epsilon, a, sigma^2 are problem inputs; delta in Theorem 1 is optimized over, not fitted. The main premises are standard analytic tools plus two imports from prior work, one of which (75) is algebraically suspect.

assumptions (6)

standard math Hermite functions form a complete orthonormal basis of L^2(R).
Used to expand f in (67) and to interpret small Hermite coefficients as closeness to Gaussian; cited to [12].
standard math Esseen's inequality converts characteristic-function closeness into CDF/Levy closeness.
Invoked in (58) for the L2 stability theorem; cited to [16].
standard math FKG covariance inequality for monotone functions.
Used in Lemma 1 to show the conditional median is non-decreasing, which is later used to turn an L1 bound into a sup bound at (73).
domain assumption The prior density f is bounded and square-integrable.
Theorem 2 assumes f(x)<=M and uses L^2 Hermite expansions; not every probability density satisfies this, so the theorem does not cover all priors.
ad hoc to paper The antiderivative trick from [1]: g(x) >= c * 1_{[-0.5,0.5]}(x).
Imported without proof from the authors' prior work [1]; used at (79)-(80) to derive the crucial local growth estimate for \tilde f.
ad hoc to paper Convolution identity (75): e^{(1-a)y^2/2} T_a[f](y) = integral of \tilde f(ax) sign(x-y) exp(-a(x-y)^2/2) dx.
This is a load-bearing algebraic premise. A direct change of variables with the stated \tilde f introduces an extra factor a*exp(-(1-a)^2 x^2/2), so the identity as printed does not hold; the proof of Theorem 2 depends on it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Functional uniqueness and stability of Gaussian priors in optimal L1 estimation." pith.science (2026). https://pith.science/paper/BEGYAJFB

@misc{pith2026251116864,
  author       = {Pith},
  title        = {Pith review of: Functional uniqueness and stability of Gaussian priors in optimal L1 estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEGYAJFB}},
  note         = {Machine review of arXiv:2511.16864}
}

abstract

We study when optimal Bayesian estimators under Gaussian noise are approximately linear, and what this implies about the underlying prior distribution. Consider the classical model \(Y = X + Z\), where \(Z\) is Gaussian and independent of \(X\). It is well known that under squared-error loss, the conditional mean \(\mathbb{E}[X|Y]\) is a linear function of \(Y\) if and only if the prior is Gaussian. Much less is understood under absolute-error loss, where the optimal estimator is the conditional median and standard orthogonality-based tools no longer apply. Recent work has established that, in the Gaussian noise model, the Gaussian prior is also the unique distribution that induces an exactly linear conditional median. In this paper, we move beyond exact characterizations and develop a quantitative stability theory: if the optimal estimator is approximately linear, must the prior be close to Gaussian? For the \(L_2\) setting, we derive explicit rates showing that near-linearity of the conditional mean forces the prior to be close to Gaussian in the Levy metric. For the \(L_1\) setting, we develop a functional-analytic framework based on Hermite expansions and adjoint operators, establishing that approximate linearity of the conditional median implies proximity to the Gaussian family.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 1 linked inside Pith

[1]

L1 estimation: On the optimality of linear estimators,

L. P. Barnes, A. Dytso, J. Liu, and H. Vincent Poor, “L1 estimation: On the optimality of linear estimators,”IEEE Transactions on Information Theory, vol. 70, no. 11, pp. 8026–8039, 2024

2024
[2]

Conjugate priors for exponential families,

P. Diaconis and D. Ylvisaker, “Conjugate priors for exponential families,”The Annals of Statistics, vol. 7, no. 2, pp. 269–281, 1979

1979
[3]

On conditions for linearity of optimal estimation,

E. Akyol, K. Viswanatha, and K. Rose, “On conditions for linearity of optimal estimation,”IEEE Transactions on Information Theory, vol. 58, no. 6, pp. 3497–3508, 2012

2012
[4]

Kailath, A

T. Kailath, A. H. Sayed, and B. Hassibi,Linear Estimation. Prentice Hall, 2000

2000
[5]

Conditional mean estimation in gaussian noise: A meta derivative identity with applications,

A. Dytso, H. V. Poor, and S. Shamai Shitz, “Conditional mean estimation in gaussian noise: A meta derivative identity with applications,”IEEE Transactions on Information Theory, vol. 69, no. 3, pp. 1883–1898, 2023

2023
[6]

Multivariate priors and the linearity of optimal bayesian estimators under gaussian noise,

L. P. Barnes, A. Dytso, J. Liu, and H. V. Poor, “Multivariate priors and the linearity of optimal bayesian estimators under gaussian noise,” inProc. IEEE International Symposium on Information Theory (ISIT), 2024

2024
[7]

Completely monotone functions and characterizations of the gamma distri- bution,

K.-T. Chou and I. Olkin, “Completely monotone functions and characterizations of the gamma distri- bution,”Annals of the Institute of Statistical Mathematics, vol. 21, pp. 601–617, 1969

1969
[8]

Optimal estimation for the poisson noise model: A character- ization of linear conditional means,

A. Dytso, H. V. Poor, and S. Shamai, “Optimal estimation for the poisson noise model: A character- ization of linear conditional means,”IEEE Transactions on Information Theory, vol. 66, no. 10, pp. 6306–6329, 2020

2020

Show all 16 references

[9]

Bounds for the difference between median and mean of gamma and Poisson distributions,

J. Chen and H. Rubin, “Bounds for the difference between median and mean of gamma and Poisson distributions,”Statistics & Probability letters, vol. 4, no. 6, pp. 281–283, 1986

1986
[10]

Linearity-inducing priors for poisson parameter estimation underl 1 loss,

L. P. Barnes, A. Dytso, and H. V. Poor, “Linearity-inducing priors for poisson parameter estimation underl 1 loss,” 2025. [Online]. Available: https://arxiv.org/abs/2505.21102

2025 arXiv
[11]

On the minimum meanp th error in Gaussian noise channels and its applications,

A. Dytso, R. Bustin, D. Tuninetti, N. Devroye, H. V. Poor, and S. S. Shitz, “On the minimum meanp th error in Gaussian noise channels and its applications,”IEEE Transactions on Information Theory, vol. 64, no. 3, pp. 2012–2037, 2017

2012
[12]

Hermite expansions for spaces of functions with nearly optimal time- frequency decay,

L. Neyt, J. Toft, and J. Vindas, “Hermite expansions for spaces of functions with nearly optimal time- frequency decay,”Journal of Functional Analysis, vol. 288, no. 3, p. 110706, 2025

2025
[13]

Bounds on Dawson’s integral occurring in the analysis of a line distribution network for electric vehicles,

A. Janssen, “Bounds on Dawson’s integral occurring in the analysis of a line distribution network for electric vehicles,”EURANDOM Preprint Series, vol. 14, 2021

2021
[14]

R. M. Dudley,Real Analysis and Probability, 2nd ed., ser. Cambridge Studies in Advanced Mathematics. Cambridge: Cambridge University Press, 2002, vol. 74

2002
[15]

Strong data processing inequalities for input constrained additive noise channels,

F. du Pin Calmon, Y. Polyanskiy, and Y. Wu, “Strong data processing inequalities for input constrained additive noise channels,”IEEE Transactions on Information Theory, vol. 64, no. 3, pp. 1879–1892, 2017

2017
[16]

Feller,An Introduction to Probability Theory and Its Applications, Volume II, 2nd ed

W. Feller,An Introduction to Probability Theory and Its Applications, Volume II, 2nd ed. New York: John Wiley & Sons, 1971. A Auxiliary Results Lemma 2.Forτ∈R ϕX (τ)−e − σ2 X τ 2 2 = e− σ2 X τ 2 2 Z τ 0 e σ2 X t2 2 ϕ′ X (t) +σ 2 X tϕX (t) dt 12 Proof.Next we use the following ...

1971

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

L1 estimation: On the optimality of linear estimators,

L. P. Barnes, A. Dytso, J. Liu, and H. Vincent Poor, “L1 estimation: On the optimality of linear estimators,”IEEE Transactions on Information Theory, vol. 70, no. 11, pp. 8026–8039, 2024

2024

[2] [2]

Conjugate priors for exponential families,

P. Diaconis and D. Ylvisaker, “Conjugate priors for exponential families,”The Annals of Statistics, vol. 7, no. 2, pp. 269–281, 1979

1979

[3] [3]

On conditions for linearity of optimal estimation,

E. Akyol, K. Viswanatha, and K. Rose, “On conditions for linearity of optimal estimation,”IEEE Transactions on Information Theory, vol. 58, no. 6, pp. 3497–3508, 2012

2012

[4] [4]

Kailath, A

T. Kailath, A. H. Sayed, and B. Hassibi,Linear Estimation. Prentice Hall, 2000

2000

[5] [5]

Conditional mean estimation in gaussian noise: A meta derivative identity with applications,

A. Dytso, H. V. Poor, and S. Shamai Shitz, “Conditional mean estimation in gaussian noise: A meta derivative identity with applications,”IEEE Transactions on Information Theory, vol. 69, no. 3, pp. 1883–1898, 2023

2023

[6] [6]

Multivariate priors and the linearity of optimal bayesian estimators under gaussian noise,

L. P. Barnes, A. Dytso, J. Liu, and H. V. Poor, “Multivariate priors and the linearity of optimal bayesian estimators under gaussian noise,” inProc. IEEE International Symposium on Information Theory (ISIT), 2024

2024

[7] [7]

Completely monotone functions and characterizations of the gamma distri- bution,

K.-T. Chou and I. Olkin, “Completely monotone functions and characterizations of the gamma distri- bution,”Annals of the Institute of Statistical Mathematics, vol. 21, pp. 601–617, 1969

1969

[8] [8]

Optimal estimation for the poisson noise model: A character- ization of linear conditional means,

A. Dytso, H. V. Poor, and S. Shamai, “Optimal estimation for the poisson noise model: A character- ization of linear conditional means,”IEEE Transactions on Information Theory, vol. 66, no. 10, pp. 6306–6329, 2020

2020