Berry–Esseen Inequalities Overview

Updated 27 January 2026

Berry–Esseen type inequalities are quantitative refinements of the CLT that provide explicit convergence rates based on the third absolute moment and characteristic function analysis.
They employ smoothing inequalities, truncation, and exponential tilts to extend classical results to dependent structures, bootstrapping, and high-dimensional settings.
Recent advancements using Zolotarev metrics and nonuniform corrections have sharpened these bounds, impacting applications in martingale theory, nonlinear functionals, and large deviations.

A Berry–Esseen type inequality is a quantitative refinement of the central limit theorem (CLT) that provides explicit rates of convergence in probability metrics—typically the Kolmogorov metric—between the distribution of a normalized sum of independent random variables and the standard normal law. These inequalities are foundational in probability theory and statistical inference, precisely quantifying how quickly the distribution of partial sums converges to normality in terms of moments of the summands. Beyond the i.i.d. scalar case, the theory encompasses dependent structures, nonlinear functionals, bootstrapping, high-dimensional and non-standard (e.g., non-Gaussian) limit regimes.

1. Classical Berry–Esseen Inequality and Constants

The classical Berry–Esseen inequality applies to normalized sums $S_n = X_1 + \cdots + X_n$ of real i.i.d. random variables with mean zero, unit variance, and finite third absolute moment $\beta_3 = \mathbb{E}|X_1|^3$ . The inequality is stated as

$\Delta_n := \sup_x |F_n(x) - \Phi(x)| \leq C_0\,\frac{\beta_3}{\sqrt{n}},$

where $F_n$ is the cdf of $S_n/\sqrt{n}$ and $\Phi$ is the standard normal cdf. The optimal (smallest possible) constant $C_0$ has been the subject of extensive study:

Shevtsova (Shevtsova, 2011) sharpened the best upper estimate to $C_0 < 0.4748$ by proving

$\Delta_n \leq 0.3328 \frac{\beta_3 + 0.429}{\sqrt{n}} \quad \text{and} \quad \Delta_n \leq 0.33554 \frac{\beta_3 + 0.415}{\sqrt{n}}.$

For all $\beta_3 \geq 1$ , both forms yield $C_0<0.4756$ , and the second is sharper for $\beta_3 < 1.2854$ .

Earlier landmarks:
- Esseen originally established $C_0 \leq 0.56$ .
- Korolev–Shevtsova (2010): $C_0 < 0.4784$ by optimization over Lyapunov fractions.
- Tyurin and others progressively improved explicit numerical bounds.
Zolotukhin–Nagaev–Chebotarev (Zolotukhin et al., 2018) provided a finite- $n$ computational approach for Bernoulli summands, proving $C_{02}\leq 0.409954$ , strictly smaller than the Esseen constant for $1 \leq n \leq 5\cdot 10^5$ , with similar asymptotics for non-Bernoulli two-point laws.

The dependence on the third absolute moment is optimal in the classical i.i.d. setting.

2. Methodological Foundations

The proof techniques for Berry–Esseen bounds are based on smoothing inequalities and characteristic function analysis. Key elements include:

Smoothing–inequality (Prawitz [Zolotarev 1965–67]–type) arguments:

$\Delta_n \leq \text{(integrals of remainders of characteristic functions with smooth cutoffs)}.$

Shevtsova (Shevtsova, 2011) used enhanced remainder bounds, analysis of the Lyapunov fraction, and explicit numerical optimization over parameters.

Truncation and exponential tilts are used in non-uniform and moderate/large deviation regimes (Pinelis, 2013, Pinelis, 2013).
For sums of non-identically distributed or dependent summands (e.g., $m$ -dependent, Markov, or martingale differences), adaptations include replacement of classical expansions with local limit results and anti-concentration bounds (Bong et al., 2022, Guo et al., 2017, Fan et al., 2016).
In smooth-function bounds, Zolotarev metrics $\zeta_s$ and the Wasserstein-1/Kantorovich distance are employed, allowing Berry–Esseen type results in $L^1$ , Lipschitz, or Zolotarev norms (Mattner et al., 2017, Mattner, 2022).

3. Scope: Extensions and Generalizations

3.1 Non-i.i.d. Sums and Dependence Structures

Non-identical summands: Nonuniform and uniform BE bounds have forms similar to the classical, with Lyapunov ratios reflecting the pooled variance and third-moment data (Pinelis, 2013, Mattner, 2022).
Martingales: Under a conditional Bernstein condition,

$|P(S_n \leq x) - \Phi(x)| \leq C(1 + x^2)D(\epsilon, \delta; x)\exp\{-\frac{\widehat{x}^2}{2}\}$

producing bounds with optimal exponential tails, unifying normal approximation with Cramér–type large deviations (Fan et al., 2016).

Regenerative and mixing processes: Provided weak moment and mixing assumptions, Berry–Esseen rates of $O(n^{-\delta/2})$ are obtainable, generalizing Bolthausen’s results to lower-moment or weaker mixing regimes (Guo et al., 2017).
Conditioned structures: Quantitative BE bounds exist for sums conditional on combinatorial constraints (e.g., occupancy, hashing, random forests), retaining $O(N^{-1/2})$ rates under uniform control of moments and characteristic function decay (Klein et al., 2019).

3.2 Functionals, Bootstrapping, and High-Dimensional Settings

Smooth functionals ( $f(S_n)$ ): For functions $f\in C^2(\mathbb{R})$ with Lipschitz Hessians,

$|\mathbb{E}[f(S_n)] - \mathbb{E}[f(N)]| \leq C\,\frac{\beta_3}{6n} \,\|f''\|_{\mathrm{Lip}}$

with constants $C_{\mathrm{opt}} < 1/6$ , via sharp analysis of the Zolotarev metrics (Mattner et al., 2017).

Multivariate, nonlinear functionals and high-dimensional CLT: For $T = W + D$ with $W$ a sum and $D$ a small nonlinear remainder,

$\sup_{A\text{ convex}} |P\{T \in A\} - P\{Z \in A\}| \leq C d^{1/2} n^{-1/2} + \text{(remainder terms)}$

with $C$ explicit and sharp up to $d^{1/2}$ factor (Shao et al., 2021).

Bootstrap accuracy: Multivariate, higher-order Berry–Esseen bounds (Zhilova, 2016) yield optimal $O(p/\sqrt{n})$ –type rates for bootstrap approximations of smooth functionals and likelihood ratio statistics in high dimensions, provided higher moment-matching is available.

3.3 Non-normal and Conditional Limit Laws

Chernoff-type (non-normal) limits: For isotonic regression, Berry–Esseen bounds of order $O(n^{-1/3}\log^a n)$ reflect the cube-root asymptotics of the underlying estimation problem (Han et al., 2019).
Conditioned-to-stay-positive random walks: The limiting law is Rayleigh, and an order $O(n^{-1/2})$ Berry–Esseen bound in the Kolmogorov norm holds for the properly rescaled sum on the event of staying positive, with an explicit (but potentially non-optimal) exponent-3 dependence on the third-moment (Denisov et al., 2024).

Recent advances have established that classical metrics in Berry–Esseen inequalities can be replaced or refined:

Zolotarev norms $\zeta_1$ , $\zeta_3$ : Mattner (Mattner, 2022) showed

$\|F_n - \Phi\|_\infty \leq 7.2\,\frac{\zeta(F - \Phi)}{\sqrt{n}},\quad \zeta := \zeta_1 \vee \zeta_3,$

so if $F$ is "Zolotarev-close" to normal, the Berry–Esseen bound can be significantly sharper than in terms of $\beta_3$ .

Local limit (density) bounds: For independent vectors in $\mathbb{R}^d$ with sufficient smoothness/density,

$\sup_{x\in\mathbb{R}^d}|p_n(x) - \phi_d(x)| \leq (C\sigma)^d M^2 \frac{B_3}{\sqrt{n}},$

with smoothness penalty $M^2$ arising from the sup-norm of densities (Bobkov et al., 2024).

5. Nonuniform Inequalities and Large Deviations

Nonuniform Berry–Esseen bounds are essential for tail accuracy and moderate/large deviations:

Canonical form: For independent summands,

$|P(S/B > x) - P(Z > x)| \leq C\,\frac{r_3}{1 + x^3},$

with best proven $C_{\mathrm{nu}}\leq 25.80$ in the iid case (Pinelis, 2013), though lower bounds show $C_{\mathrm{nu}}>1$ and smoothing approaches may eventually reduce this constant further.

Martingale analogues: Under conditional Bernstein conditions, nonuniform bounds combine optimal tail decay (exponential in deviation) with polynomial prefactors, unifying Cramér–type expansions and BE theory (Fan et al., 2016).

6. Fast Rates and Moment Matching

Assuming extra structure or smoothness leads to accelerated convergence:

If the first $k$ moments of $X$ match those of the standard normal and there is a local lower bound on the density, then

$\sup_s |P(N^{-1/2}\sum X_i \leq s) - \Phi(s)| \leq C \frac{E|X|^{k+1}}{N^{(k-1)/2}} + \exp(-c h w^3 N / E|X|^{k+1}),$

and for $k=3$ this yields an $O(1/N)$ rate for symmetric variables (Johnston, 2023).

The regime and sharpness of these results have been confirmed via Bernoulli plus small-uniform perturbations.

7. Applications and Statistical Inference

Berry–Esseen type bounds underpin a variety of quantitative statistical analysis:

Small deviation probabilities (e.g., Feige’s conjecture) via an overview of BE and semidefinite-moment inequalities (Guo et al., 2020).
Parameter estimation in stochastic processes: Explicit Berry–Esseen rates for parameter estimation in fractional OU models via Malliavin–Stein techniques elucidate the connection between Hurst index, estimation rate, and the fine structure of the likelihood (Chen et al., 2018).
High-dimensional dependence: $m$ -dependent random vector arrays, bootstrapped sums, and even nonlinear functional estimators (such as $M$ -estimators, SGD averages) fall within the reach of contemporary multivariate and probabilistic BE inequalities (Bong et al., 2022, Zhang, 2021, Shao et al., 2021, Zhilova, 2016).