Berry-Esseen Type Bounds

• Berry-Esseen type bounds are quantitative measures for the rate of convergence in central limit theorems, providing explicit distance bounds between normalized sums and the Gaussian distribution.
• The methodology employs tools like Stein's method, randomized concentration, and Malliavin calculus to extend classical bounds to dependent and high-dimensional settings.
• Applications span statistical estimators, random matrix theory, and nonstandard limit distributions, with concrete rates such as O(n⁻¹ᐟ²) under third moment conditions.

A Berry-Esseen type bound gives a quantitative rate of convergence in central limit theorems, typically bounding the Kolmogorov or other distances between the distribution function of a normalized statistic and the Gaussian distribution, in terms determined by higher moments or concentration functionals. Classic theorems treat independent sums; recent advances provide optimal rates under dependency, self-normalization, and for diverse classes of statistics including U-statistics, random matrices, operator-valued noncommutative laws, and nonstandard limit distributions.

1. Classical and Generalized Berry-Esseen Bounds

In the classical case, for independent real random variables $X_1,\ldots,X_n$ with mean zero and variances $\sigma_k^2$, letting $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$ and $B_n = \sum_{k=1}^n \sigma_k^2$, the bound reads

$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$

where $C$ is universal and $L_3$ is the Lyapunov ratio (Bobkov et al., 2011). This rate $O(n^{-1/2})$ is optimal under the third moment condition.

This paradigm extends across multiple structural settings:

• Self-normalized martingales: The optimal rate $N_n^{1/(2p+1)}$ is attained, where $$N_n=\sum \mathbb{E}|X_i|^{2p}+\mathbb{E}|\langle S \rangle_n-1|^p$$ for $\sigma_k^2$0 (Fan et al., 2017).
• Dependent graphs: For sums with a sparse dependency graph (maximum degree $\sigma_k^2$1), the rate is degraded to $\sigma_k^2$2 for variables bounded in $\sigma_k^2$3, $\sigma_k^2$4 (Janisch et al., 2022).
• Random matrix theory and noncommutative probability: For matrix coefficients and spectral radius in random walks on $\sigma_k^2$5, rates are $\sigma_k^2$6 under exponential moment and $\sigma_k^2$7 under a polynomial moment of order $\sigma_k^2$8 (Cuny et al., 2021). For operator-valued free CLT and operator-valued matrices, explicit Berry-Esseen bounds of $\sigma_k^2$9 for Cauchy transforms and $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$0 for Levy distance are achieved (Banna et al., 2021).

2. Methods and Structural Innovations

a. Stein's Method and Extensions

Stein's method is central in modern proofs, allowing adaptation to dependency, heterogeneous variance profiles, nonlinear statistics, and self-normalizing denominators. The general scheme uses an identity such as: $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$1 with $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$2 tailored to the problem (e.g., via Palm theory, zero-bias couplings, exchangeable pairs) (Chen et al., 2020).

b. Randomized Concentration and Smoothing

Sharp bounds for nonlinear and high-dimensional statistics introduce randomized concentration inequalities, with smoothing over convex sets or high-dimensional flows, controlling small-ball probabilities via leave-one-out or selective conditioning (Shao et al., 2021).

c. Malliavin Calculus and Chaos Expansions

For functionals of independent or Gaussian sequences, discrete and continuous Malliavin calculus with chaos expansion yields precise Berry-Esseen inequalities, especially for U-statistics and quadratic forms (Privault et al., 2020, Bourguin et al., 2010). For complex and operator-valued chaoses, the norm of contractions fully characterizes the rate (Chen et al., 2024, Banna et al., 2021).

3. Extension to Structured Dependence and General Statistics

a. Local and Graph Dependence

Under local dependence specified by neighborhoods or dependency graphs, optimizing the Berry-Esseen rate requires only third-moment control for self-normalized sums, achieving $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$3 up to dependence-degree prefactors (Zhang, 2021, Janisch et al., 2022).

b. Permutation, Sampling, and Poisson Models

For stratified permutational statistics (survey sampling, experimental design), recent results yield $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$4 Berry-Esseen bounds as long as no stratum dominates variance, using Stein's method with explicit exchangeable-pair and zero-bias machinery (Tian et al., 18 Mar 2025).

For weighted sums of Poisson avoidance functionals in geometric probability, the approach yields $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$5 Berry-Esseen-type rates in Wasserstein or Kolmogorov distance for union-of-balls volume or quantization error (Barrio, 2015).

4. Beyond Gaussian: Nonstandard and Local Berry-Esseen Theorems

a. Nonstandard Limit Distributions

In nonregular settings (e.g., isotonic regression), a Berry-Esseen bound is derived for convergence to Chernoff-type distributions, with rates $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$6 (up to logs), matching the best rates of oracle local average estimators (Han et al., 2019).

b. Local Limit and Regularity-Dependent Rates

Quantitative bounds for the convergence of densities in local limit theorems depend additionally on density bounds and higher Lyapunov coefficients. For sums of independent random vectors with maximal coordinate density $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$7 and Lyapunov $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$8: $S_n = (X_1+\cdots+X_n)/\sqrt{B_n}$9 with tail corrections, these decay exponentially once $B_n = \sum_{k=1}^n \sigma_k^2$0 (Bobkov et al., 2024).

5. Applications and Implications

a. Statistical Estimators and Algorithms

Berry-Esseen bounds now underpin distributional approximations for $B_n = \sum_{k=1}^n \sigma_k^2$1-estimators, Z-estimators, stochastic gradient descent (SGD), and moment estimators in fractional Ornstein-Uhlenbeck models, often allowing non-asymptotic inference with explicit rates such as $B_n = \sum_{k=1}^n \sigma_k^2$2 or $B_n = \sum_{k=1}^n \sigma_k^2$3 for fOU under discrete fixed-step sampling (Shao et al., 2021, Tang et al., 3 Apr 2025).

b. Multivariate, Functional, and High-Dimensional Statistics

Extending to the multivariate setting, bounds are available for nonlinear $B_n = \sum_{k=1}^n \sigma_k^2$4-dimensional statistics, random matrices, and operator-valued variables. These depend polynomially on dimension and moments, with sharpness up to known lower bounds (Cuny et al., 2021, Shao et al., 2021, Banna et al., 2021).

6. Optimality, Sharpness, and Limitations

In many settings, such as self-normalized and permutation statistics, constructed examples demonstrate the optimality of the exponents in the Berry-Esseen bound (Fan et al., 2017). For operator-valued and functional convergence, loss in the exponent (e.g., move from $B_n = \sum_{k=1}^n \sigma_k^2$5 to $B_n = \sum_{k=1}^n \sigma_k^2$6 in operator-valued CLT under Levy distance) is a consequence of current smoothing technology and lack of additional regularity (Banna et al., 2021, Bobkov et al., 2024).

A notable limitation is that for non-Gaussian and strongly dependent models (e.g., long-memory moving averages), the rate of approach can be arbitrarily slower than $B_n = \sum_{k=1}^n \sigma_k^2$7, fully explicit as $B_n = \sum_{k=1}^n \sigma_k^2$8 where $B_n = \sum_{k=1}^n \sigma_k^2$9 is the memory exponent (Bourguin et al., 2010).

7. Summary Table of Key Asymptotic Rates

Structure / Model	Distance	Condition	Rate	Source
Independent, finite $$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$0	Kolmogorov	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$1	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$2	(Bobkov et al., 2011)
Self-normalized martingale	Kolmogorov	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$3	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$4	(Fan et al., 2017)
Sparse dependency graph	Kolmogorov	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$5	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$6	(Janisch et al., 2022)
Random walk, GL$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$7 coefficients	Kolmogorov	exp./poly moment	$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$8/$$\sup_x | \mathbb{P}(S_n \leq x) - \Phi(x) | \leq C\, L_3, \quad L_3 = \frac{1}{B_n^{3/2}} \sum_{k=1}^n \mathbb{E}|X_k|^3$$9	(Cuny et al., 2021)
Operator-valued free CLT	Cauchy/Levy	Operator moments	$C$0/ $C$1	(Banna et al., 2021)
Stratified permutation statistics	Kolmogorov	3rd moment, balance	$C$2	(Tian et al., 18 Mar 2025)
Fractional OU, $C$3	Kolmogorov	fBm, discrete obs.	$C$4	(Tang et al., 3 Apr 2025)
Fractional OU, $C$5	Kolmogorov	fBm, discrete obs.	$C$6	(Tang et al., 3 Apr 2025)
Chernoff-type limits (cube root)	Kolmogorov	Smoothness $C$7	$C$8 (log factors)	(Han et al., 2019)
Long-memory moving average	Kolmogorov	MA($C$9), $L_3$0	$L_3$1	(Bourguin et al., 2010)

These representative rates and the associated structural dependence of constants summarize the state-of-the-art landscape for Berry-Esseen type bounds in contemporary probability and mathematical statistics.