Esseen’s Smoothing Inequality

• Esseen's smoothing inequality is a fundamental result that quantitatively connects differences between distribution functions with the low-frequency behavior of their characteristic functions.
• It is pivotal in establishing convergence rates in the central limit theorem, enhancing the classical Berry–Esseen bounds with explicit, near-optimal constants.
• Modern extensions utilize optimized smoothing kernels and Fourier-analytic generalizations to control tail errors and extend the inequality to other probabilistic metrics.

Esseen's smoothing inequality is a fundamental result in probability theory and Fourier analysis, providing a quantitative link between the distribution functions of random variables and the behavior of their characteristic functions. This inequality forms the analytic backbone of numerous results on the rate of convergence in the central limit theorem, most notably the Berry–Esseen theorem. Modern developments have extended its scope, sharpened its constants, introduced innovations in smoothing kernels, and generalized it to metrics beyond the classical Kolmogorov distance.

1. Formal Statement and Classical Framework

Let $X$ and $Y$ be real random variables with distribution functions $F_X(a) = \Pr\{X \le a\}$ and $F_Y(a) = \Pr\{Y \le a\}$ and corresponding characteristic functions $\phi_X(t) = \mathbb{E}[e^{itX}]$ and $\phi_Y(t) = \mathbb{E}[e^{itY}]$. Assume $Y$ admits a density $g_Y$ bounded by $M$, $\sup_{y \in \mathbb{R}} g_Y(y) \le M < \infty$.

Esseen's (Prawitz–Bohman–Vaaler) smoothing inequality asserts that for every $\varepsilon > 0$,

$$\sup_{a \in \mathbb{R}} \bigl| F_X(a) - F_Y(a) \bigr| \le 2\int_{-1/\varepsilon}^{1/\varepsilon} \left| \frac{\phi_X(t) - \phi_Y(t)}{t} \right| dt + C M \varepsilon,$$

where $C$ is an absolute constant (Vershynin, 5 Feb 2026). Using the alternative normalization, it is presented as

$$\sup_{x \in \mathbb{R}} \bigl| F_X(x) - F_Y(x) \bigr| \le \frac{1}{\pi} \int_{-T}^{T} \frac{|\phi_X(t) - \phi_Y(t)|}{|t|} dt + \frac{24 M}{T}, \quad T > 0,$$

with the $1/\pi$ factor known to be sharp (Gabdullin et al., 2018).

This inequality quantifies how closely two distributions match in supremum norm in terms of the "low-frequency" distance between their characteristic functions, plus a controlled tail error dependent on the smoothness of the comparison distribution.

2. Smoothing Kernels and Modern Fourier-Analytic Generalizations

The classical Esseen smoothing method introduces a "smoothing multiplier" or filter $M: \mathbb{R} \to \mathbb{C}$ subject to

• $\Re M(t)$ even, $\Im M(t)$ odd,
• $M(t) = 0$ for $|t| > 1$.

Given a truncation parameter $T > 0$, the rescaled kernel is $M_T(t) := M(t/T)$. The generalized inversion operator is

$$\Lambda(h)(x) := \mathrm{p.v.} \int_{-\infty}^\infty e^{-itx} \frac{h(t)}{t} \frac{dt}{2\pi i}$$

for suitable smooth $h$. The Prawitz–Bohman–Vaaler inequalities (a variant of Esseen's inequality) are

$$\Lambda(M_T(-\cdot)\varphi_X(\cdot))(x) \le F_X(x) - \tfrac{1}{2} \le \Lambda(M_T(+\cdot)\varphi_X(\cdot))(x),$$

quantifying the proximity of $F_X$ to a reference distribution through its low-frequency characteristic function weighted by a smooth, compactly-supported kernel (Pinelis, 2013).

Modern approaches replace the classical tilt $x \mapsto x$ with a more general odd function $G$ of bounded variation, whose Fourier–Stieltjes transform vanishes above a chosen frequency $y > 0$. Kernels of the form

$M(t) = \widehat{p}(t) + iK \widehat{pG}(t)$

with $p$ a smooth symmetric pdf and an optimal constant $K$ deliver improved tail estimates and greater flexibility, especially in applications involving higher moments or nonuniform bounds (Pinelis, 2013).

3. Applications to Berry–Esseen and Nonuniform Bounds

The smoothing inequality is central to quantitative limit theorems. For independent, zero-mean random variables $X_1, \dots, X_n$ with $\mathbb{E}|X_k|^3 < \infty$, the normalized sum $S_n = \sum_{k=1}^n X_k$ is compared to the standard normal $Z \sim N(0,1)$.

Esseen's method yields the uniform Berry–Esseen theorem: $$\sup_x \left| F_{S_n}(x) - \Phi(x) \right| \le C \sum_{k=1}^n \mathbb{E}|X_k|^3$$ with best-known constants currently $C_u \approx 0.4748$ (i.i.d. case) and a provable lower bound $C_u \ge 0.4097$ (Pinelis, 2013). For nonuniform bounds emphasizing accuracy in the tails, the smoothing method is extended as follows: $$|\mathbb{P}\{S_n > \sigma z\} - \mathbb{P}\{Z > z\}| \le C_{\mathrm{nu}} \frac{\rho}{\sigma^3 \sqrt{n}} (1 + z^3)$$ with optimized constants $C_{\mathrm{nu}} \approx 1.0135$ in the i.i.d. case (Pinelis, 2013). Integration by parts within the smoothing operator moves powers of $t$ onto the kernel, resulting in precise control over higher moments (e.g., via derivatives of order three or higher), enabling sharper nonuniform and large deviation results.

In particular, Pinelis showed that with appropriately smooth kernels, constants arbitrarily close to the lower bound can be achieved, essentially closing the gap between upper and lower bounds for nonuniform Berry–Esseen constants (Pinelis, 2013).

4. Smoothing Inequality in Other Metrics and Structures

Recent developments have generalized the smoothing inequality to metrics beyond the Kolmogorov distance, notably to the $p$-Wasserstein metric on compact Lie groups. For two probability measures $\mu$, $\nu$ on a compact group $G$ with geodesic distance $p(x,y)$, one considers a generalized modulus-of-continuity metric $g$. The key result is a smoothing bound of the form

$W_g(\mu, \nu) \le \Phi(M) + C_G \Phi(M) \Psi(M)$

where $\Phi(M)$ is a smoothing error from convolution with a frequency cutoff kernel, and $\Psi(M)$ involves a Fourier-truncation of the difference between the group Fourier transforms of $\mu$ and $\nu$ (Borda, 2020).

In the Abelian or real line case, this recovers the classical Esseen inequality, establishing its optimality and flexibility across a broad range of probabilistic distances and algebraic structures.

5. Analytical Innovations and Explicit Constants

The sharpness and applicability of Esseen's smoothing inequality rest on several analytical refinements:

• Quadratic-tail inequalities crucial for controlling characteristic function decay in high-frequency regimes, replacing previous coarse constants (e.g., Esseen's 40, Rozovskii's 12, improved to the sharp value 4) (Gabdullin et al., 2018).
• Convolution symmetrization for improving bounds on $|\varphi_n(t)|$, leveraging independence and symmetry to yield exponential decay for large $t$.
• Two-piece smoothing decomposes the smoothing integral at a threshold $T_0 \ll T$, applying Taylor-remainder bounds near the origin and Fourier decay in the tails.
• Computation of explicit, often asymptotically exact, constants: The best constants in smoothing inequalities are now known to within $4\%$ of optimality for certain fractions, with for instance $C_E(\infty,1)\le 2.66$ and limiting sharpness $C^*_E(\infty,1) \approx 1.73$ (Gabdullin et al., 2018).

These innovations have led to convergence-rate inequalities surpassing classical Lyapunov and Osipov forms whenever the summands avoid large skewness or heavy tails.

6. Impact, Extensions, and Open Questions

The smoothing inequality's reach extends not only to rates in the central limit theorem but to a spectrum of problems in probability and analysis:

• It allows fine control in nonuniform CLT, large deviations, random walk convergence, and metrics of empirical measure convergence.
• Generalizations encompass high-order moment bounds (via repeated integration by parts), compact group structures, and Wasserstein-type distances.
• An open problem is whether the lower-bound constant for the nonuniform Berry–Esseen bound ($C_{\mathrm{nu}} \ge 1.0135$) is achievable exactly, or whether the optimum is achieved for a one-point law ($C \approx 1$) as conjectured by Bentkus (Pinelis, 2013).
• The choice and optimization of smoothing kernels—especially tempered tilts $G$ of compact spectral support—remain central for minimizing constants and maximizing the method’s power (Pinelis, 2013).

7. Summary Table: Key Variants and Constants

Inequality Variant	Explicit Upper Bound	Limiting Sharpness	Reference
Uniform Berry–Esseen (i.i.d.)	0.4748	0.4097	(Pinelis, 2013)
Nonuniform Berry–Esseen (i.i.d.)	$<1.1$	1.0135	(Pinelis, 2013)
Esseen-style smoothing (Kolmogorov)	2.66	1.73	(Gabdullin et al., 2018)
Wasserstein-type smoothing (G)	$C_G$ group-specific	best known for $\mathbb{R}$	(Borda, 2020)

The theoretical and practical power of Esseen's smoothing inequality lies in its universality—centralizing the role of Fourier analysis in quantitative probability, facilitating explicit rates of convergence, and enabling intricate control over the limiting behavior and tail estimates of sums of independent random variables.