Classical Berry–Esseen Inequality

• The classical Berry–Esseen inequality is a key result in probability that quantifies the rate at which the distribution of normalized sums converges to the normal law.
• It employs smoothing techniques and characteristic function analysis to bound the Kolmogorov distance in terms of the third absolute moment and sample size.
• Recent analytic and computer-assisted refinements have improved estimates for the constant C₀, impacting statistical inference, simulation accuracy, and theoretical research.

The classical Berry–Esseen inequality is a fundamental result in quantitative probability theory that provides a non-asymptotic, uniform bound on the rate of convergence in the central limit theorem (CLT) for sums of independent, identically distributed random variables. Specifically, it estimates the Kolmogorov distance between the distribution of the normalized sum and the standard normal distribution in terms of the third absolute moment and the sample size. Determining sharp upper and lower bounds for the absolute constant associated with this inequality remains a major focus of mathematical research, given its implications for statistical applications, numerical analysis, and theoretical advances in probability.

1. Formal Statement of the Inequality

Let $X_1,X_2,\dots,X_n$ be independent, identically distributed real random variables with mean zero, unit variance, and finite third absolute moment $\beta_3 = \mathbb{E}|X_1|^3$. Define the normalized sum and associated distribution function as

$$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$

and let $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$ denote the standard normal distribution function. The Kolmogorov distance between $F_n$ and $\Phi$ is

$$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$

The classical Berry–Esseen inequality then asserts the existence of an absolute constant $C_0 > 0$ such that

$\Delta_n \leq C_0 \frac{\beta_3}{\sqrt{n}},$

for all $n \geq 2$ and all distributions meeting the conditions above (Shevtsova, 2011).

2. Historical Development and Determination of $\beta_3 = \mathbb{E}|X_1|^3$0

Since its original formulation in the mid-twentieth century, the value of the optimal absolute constant $\beta_3 = \mathbb{E}|X_1|^3$1 has been a subject of extensive research. Esseen (1956) proved a lower bound $\beta_3 = \mathbb{E}|X_1|^3$2, showing that no smaller constant suffices in full generality. Successive improvements have been achieved primarily via analytic refinements and, in certain cases, computer-assisted proofs:

• Korolev and Shevtsova (2011) improved the best known upper bound on $\beta_3 = \mathbb{E}|X_1|^3$3, proving $\beta_3 = \mathbb{E}|X_1|^3$4 and $\beta_3 = \mathbb{E}|X_1|^3$5, yielding $\beta_3 = \mathbb{E}|X_1|^3$6 and $\beta_3 = \mathbb{E}|X_1|^3$7, respectively (Shevtsova, 2011).
• In the i.i.d. Bernoulli case, Zolotukhin, Nagaev, and Chebotarev used supercomputer-assisted interval analysis to demonstrate that $\beta_3 = \mathbb{E}|X_1|^3$8 is strictly less than the Esseen constant $\beta_3 = \mathbb{E}|X_1|^3$9 up to $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$0, with an explicit bound $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$1 (Zolotukhin et al., 2018).
• In the general i.i.d. framework, the best currently published $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$2 value is $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$3 (Mattner, 2022).
• Despite efforts, it remains unresolved whether the best constant for arbitrary lattice distributions coincides with $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$4.

3. Proof Methodology: Smoothing and Characteristic Functions

The modern proof approach incorporates smoothing inequalities and careful analysis of characteristic functions, following lines initiated by Zolotarev and developed by Prawitz, Bhattacharya–Ranga Rao, and others. The technical steps involve:

1. Smoothing inequality (Esseen–Prawitz type): Relating $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$5 to integrals of the modulus of the difference between the characteristic function of $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$6 and the Gaussian, plus a Gaussian tail term. Formally, for a suitable kernel $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$7,

$$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$8

where $$S_n = \sum_{k=1}^n X_k, \quad F_n(x) = \mathbb{P}(S_n < x\sqrt{n}),$$9.

2. Taylor expansion near the origin: Controlling $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$0 by a third-moment term, utilizing

$$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$1

and bounding the remainder in terms of $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$2.

1. Optimizing cutoff parameters and offsets: Selecting parameters such as $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$3 and Lyapunov fractions for optimal constants, with some steps numerical/computer-assisted.
2. Majorant construction: Constructing two-term majorants $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$4 for $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$5 for improved local control near the origin.

This methodology yields the sharp $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$6 convergence rate of the CLT in a non-asymptotic fashion and allows precise estimation of constants (Vershynin, 5 Feb 2026, Shevtsova, 2011).

4. Structural Refinements and Recent Advances

Structural refinements to the classical inequality have achieved further reductions in $$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$7:

• Korolev and Shevtsova (2011) obtained

$$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$8

and

$$\Phi(x) = \frac1{\sqrt{2\pi}} \int_{-\infty}^x e^{-t^2/2}\,dt$$9

which, for $F_n$0, yield $F_n$1 and $F_n$2 (Shevtsova, 2011). These refinements emerged from improved estimates of $F_n$3 near $F_n$4 and tighter control of the remainder $F_n$5, as well as computer-assisted parameter optimization.

• For discrete two-point (Bernoulli) distributions, the combination of analytic and computational methods enables pinning down $F_n$6 below the Esseen lower bound for practical ranges of $F_n$7, with error within $F_n$8 of $F_n$9 (Zolotukhin et al., 2018).
• Replacing the third absolute moment by distances such as the Wasserstein ($\Phi$0) or Zolotarev's $\Phi$1 norms yields sharper results, particularly when the underlying distributions are “Zolotarev-close” to normal (Mattner, 2022).

5. Probability Metrics, Generalizations, and Extensions

Several metrics underpin the quantitative comparison in the Berry–Esseen context:

• Kolmogorov distance:

$\Phi$2

• Wasserstein ($\Phi$3) distance:

$\Phi$4

• Zolotarev's $\Phi$5 distances:

Defined for $\Phi$6 via supremum over classes of $\Phi$7 test functions with variational normalization.

Recent research shows that under enhanced conditions (e.g., summands Zolotarev-close to normal), the classical $\Phi$8-based bound can be strictly improved. The refined Berry–Esseen bound in (Mattner, 2022) replaces the $\Phi$9 moment by

$$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$0

Such refinements are particularly notable in applications where $$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$1 or $$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$2 is small, for instance, discretized or winsorized Gaussian distributions.

The extension to non-identically distributed or weakly dependent summands is also immediate in the framework of smoothing inequalities, with analogous $$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$3 rates and dependence on aggregated third moments (Vershynin, 5 Feb 2026).

6. Applications and Impact

The classical Berry–Esseen inequality underpins the quantitative theory of the CLT, enabling:

• Explicit, uniform error bounds in goodness-of-fit testing and robust sample size requirements in statistical inference.
• Construction of non-asymptotic confidence bands for empirical distributions.
• Precise analysis of the rate of convergence in randomized algorithms and Monte Carlo simulations.
• Improvements in classical and modern CLT settings for sums, generalizations to finite population sampling, and discrete structure analysis (Mattner, 2022).

Structural advances that reduce $$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$4 by even a small margin translate to tangible gains in these areas, yielding tighter error guarantees and more efficient statistical designs (Shevtsova, 2011).

7. Ongoing Directions and Open Problems

Determining the exact best possible constant $$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$5 for the classical Berry–Esseen inequality remains unresolved. The Zolotarev conjecture posits that $$\Delta_n = \sup_{x \in \mathbb{R}} |F_n(x) - \Phi(x)|.$$6 is optimal, but for arbitrary i.i.d. (and in particular, lattice) distributions this is unproven. The interplay between analytic improvements and large-scale computational verification—especially for discrete or lattice cases—continues to drive progress (Shevtsova, 2011, Zolotukhin et al., 2018). Parallel research refines the bounds when distributional closeness is measured by Wasserstein or Zolotarev norms rather than moment conditions, suggesting further potential for sharp quantitative convergence in probabilistic and statistical analysis (Mattner, 2022).