---
title: Non-Uniform Edgeworth Expansions
url: https://www.emergentmind.com/topics/non-uniform-edgeworth-expansions
type: topic
---

# Non-Uniform Edgeworth Expansions

Non-uniform Edgeworth expansions are asymptotic series that refine the normal approximation to sums (or functionals) of random variables, providing explicit correction terms whose accuracy can be sharply tracked as a function of both the sample size and the point at which the approximation is evaluated. The non-uniformity refers to remainder estimates that decay faster as one moves away from the center, typically featuring weights of the form $(1+|x|)^{-m}$. This stands in contrast to classical (uniform) expansions, which bound the global sup-norm of the error. The modern theory rigorously establishes non-uniform Edgeworth expansions for a broad class of dependent and inhomogeneous sequences, Markov chains, dynamical systems, products of random matrices, non-i.i.d. arrays, and even self-normalized or weighted statistics, using tools of spectral analysis, analytic perturbation, characteristic function techniques, and cumulant control.

## 1. General Formulation and Basic Principles

Consider a sequence or triangular array of random variables $X_1, X_2, \dots, X_n$ and their normalized sum
$$
W_n = \frac{S_n - A_n}{B_n}, \qquad S_n = \sum_{k=1}^n X_k, \quad B_n > 0, \quad A_n = O(B_n^{-1}).
$$
The distribution function under study is $F_n(x) = \mathbb{P}(W_n \leq x)$. The Edgeworth expansion provides an asymptotic series:
$$
F_n(x) = \Phi(x) + \sum_{j=1}^r B_n^{-j} H_{j,n}(x)\varphi(x) + R_{n,r}(x),
$$
where $\Phi$ and $\varphi$ are the standard normal cdf and pdf, $H_{j,n}$ are Edgeworth polynomials (in general depending on $n$ and $x$), and $R_{n,r}(x)$ is a remainder term. Non-uniformity enters through bounds of the form
$$
|F_n(x) - F_{n,r}(x)| \leq C B_n^{-(r+1)} (1+|x|)^{-m},
$$
for some integer $m \geq r+2$, providing polynomial decay in the tails. When $B_n \sim n^{1/2}$ and $r$ is fixed, this ensures strong tail control and precise rates at all $x \in \mathbb{R}$ [2210.07204, 2511.06414].

The Edgeworth polynomials $H_{j,n}(x)$ (or $P_j(x)$ in some conventions) are universal combinatorial expressions in the normalized cumulants or moments up to order $m+1$, and can be written as
$$
P_p(x) = \sum_{k: \sum \ell k_\ell = p} C_k \prod_{\ell} \kappa_{\ell+2}^{k_\ell} H_{p-1}(x),
$$
where the constants $C_k$ arise from the combinatorial expansion of the limiting log-characteristic function [2210.07204, 1104.3759].

## 2. Hypotheses and Analytical Conditions

Sharp non-uniform Edgeworth expansions rely on two classes of analytic hypotheses:

- **(A) Local Control ("small-$t$" regime):** Bounding higher derivatives of the log-characteristic function,
  $$
  \phi_n(t) = \mathbb{E} e^{itW_n}, \quad A_n(t) = \log \phi_n(t) + t^2/2,
  $$
  there exists $m \geq 3$ such that for $|t| \leq \epsilon B_n$,
  $$
  |A_n^{(j)}(t)| \leq C_j B_n^{-(j-2)}, \quad \forall\; 3 \leq j \leq m+1.
  $$

- **(B) Cumulant-Growth/Weak Stationarity and Tail Control ("large-$t$" regime):** Ensuring cumulants $\kappa_k(S_n)=n\kappa_k+O(\rho^n)$ for $k = 2,\ldots,m+1$, $0<\rho<1$, so that cumulants exist and grow linearly with $n$. For the tails, integrals of the characteristic function satisfy, for large $B$,
  $$
  \int_{c B_n}^{B B_n^{(m-3)/2}} |\phi_n^{(m)}(t)||t|^{-1} dt = o(B_n^{-(m-2)}).
  $$
  In applications, these are typically established through spectral gap and analytic perturbation theory for suitably defined transfer operators on Banach spaces [2210.07204, 2511.06414].

The order of the expansion $r$ (highest $B_n^{-j}$ term) is constrained by the available moments: $1 \leq r \leq m-2$, and $m$ is typically taken as the greatest integer less than the available (possibly fractional) moment order [1104.3759, 2207.14402].

## 3. Main Results: Expansion Statements and Remainder Bounds

The primary theorem can be formulated as follows:  
For $1 \leq r \leq m-2$,
$$
F_n(x) = \Phi(x) + \sum_{j=1}^r B_n^{-j} H_{j,n}(x)\varphi(x) + R_{n,r}(x),
$$
with the sharp remainder estimate,
$$
|R_{n,r}(x)| \leq C B_n^{-(r+1)} (1+|x|)^{-m},
$$
where $C$ depends only on moments and analytic constants up to order $m+1$ [2210.07204, 2511.06414, 1104.3759].

For i.i.d. summands with $E|X_1|^s<\infty$ ($s \geq 2$ possibly non-integer, $m = [s]$), the non-uniform density approximation (local limit) is
$$
P_n(x) = \varphi(x) \left[ 1 + \sum_{k=1}^{m-2} n^{-k/2} P_k(x) \right] + R_n(x),
$$
and for $s=m$ integer, the remainder is $O(n^{-(m-2)/2} (1+|x|)^{-m})$; for fractional moments $2<s<3$, the remainder is $O(n^{-(s-2)/2}(1+|x|^s))$. Here, $P_k(x)$ are Hermite-type polynomials determined by the cumulants of $X_1$ up to order $m+2$ [1104.3759].

For general weakly dependent or non-i.i.d. cases—including weighted arrays, Markov chains, dynamical systems, and functionals of weighted empirical distributions—analogous expansions hold using generalizations of von Mises calculus to signed measures [2511.06414, 1002.4338].

A key feature is the explicit decay rate of the remainder with $x$, improving substantially upon uniform error bounds:
$$
\sup_{x \in \mathbb{R}} (1+|x|)^m |F_n(x) - \Phi(x)| \leq C n^{-1/2},
$$
illustrating that the Gaussian approximation's error is polynomially small even for large deviations, provided enough moments exist [2210.07204].

## 4. Methodology: Construction and Proof Strategies

The analytical core involves characteristic function analysis and Fourier inversion, with key steps as follows:

- **Taylor Expansion and Cumulant Control:** Write the log-characteristic function as a Taylor series at $t=0$, matching terms with cumulants up to degree $m+1$ [2210.07204, 1104.3759].
- **Splitting Small and Large $t$:** For $|t| \leq cB_n$, use the expansion and analytic bounds; for large $t$, decay estimates from moment control and analytic perturbation yield exponentially small contributions. Esseen-type smoothing inequalities are used to balance these regimes optimally [2511.06414].
- **Hermite Polynomial Representation:** Edgeworth polynomials are constructed to ensure the term-by-term cancellation of the error to each desired order, with explicit combinatorial formulas involving Hermite polynomials and cumulants [2210.07204, 1104.3759, 1002.4338].
- **Control of the Remainder:** For each $m$ available moment, the non-uniformity in $x$ is achieved by repeated integration by parts and careful tracking of the way polynomial weights in $x$ emerge from the Fourier inversion. For dependent structures, analytic perturbation of the transfer operator supplies the necessary regularity [2210.07204, 2511.06414].

## 5. Classes of Examples and Applications

Non-uniform Edgeworth expansions have been established for a wide array of settings, including but not limited to:

| Class of Processes                         | Mechanism for Assumption Verification           | Notable Features                    |
|--------------------------------------------|-------------------------------------------------|-------------------------------------|
| I.i.d. or non-i.i.d. sums                  | Moment conditions, characteristic function      | Petrov–Götze bounds, fractional moments, local moderate deviation rates [1104.3759] |
| Inhomogeneous Markov chains                | Spectral gap, transfer operator quasi-compactness| Stationary expansions, slow variance growth [2210.07204, 2511.06414] |
| Products of random matrices                | Projective operator, Furstenberg–Kesten theory  | All orders, exponential moment [2210.07204, 2511.06414]         |
| (Partially) expanding/hyperbolic dynamics  | Ruelle–Perron–Frobenius operator, Young towers  | Exponential tails, stationary and non-stationary cases [2210.07204, 2511.06414] |
| Weighted empirical measures / functionals  | von Mises calculus for signed measures          | Cumulant expansions for smooth functionals, Cornish–Fisher quantiles [1002.4338] |
| Self-normalized statistics (e.g., Student $T$) | Conditioning/deconvolution, non-normalized CFs | Edgeworth expansions for $S_n/V_n$ with strong tail accuracy [2207.14402] |

Key applications include:  
- Edgeworth expansions for moments and smooth functionals $h(W_n)$ with asymptotic series for $\mathbb{E}[h(W_n)]$ where $|h'(x)|(1+|x|)^m$ is integrable,
- Strong Berry–Esseen theorems and Wasserstein metric expansions: $W_p(\mathrm{Law}(W_n), N(0,1)) = O(B_n^{-1})$, with higher-order expansions in $W_p$ [2511.06414],
- Enhanced local large deviation and moderate deviation results due to tail-accuracy,
- Improved entropic central limit theorems and total variation rates for self-normalized sums [2207.14402],
- Explicit asymptotic corrections for quantiles in nonparametric statistics via Cornish–Fisher expansions [1002.4338].

## 6. Optimality, Range of Validity, and Comparison with Uniform Expansions

The attainable order $r$ in the expansion is sharply determined by the maximal integer $m$ such that moments up to order $m+1$ (or cumulants up to $m+1$) can be controlled. The exponent in the non-uniform polynomial weight can, in regular examples, be taken arbitrarily large provided enough moments exist. For fractional moments $s>2$, the degree of non-uniformity interpolates continuously between the classical Petrov regime (integer $m$) and the lower-moment case [1104.3759].

In stationary or weakly stationary scenarios—in particular, for Markov chains and dynamical systems with transfer operator spectral gaps—the Edgeworth polynomials do not accumulate $n$-dependence, and one recovers "stationary" polynomials as in the classical independent case [2210.07204, 2511.06414].

The non-uniform expansions are strictly stronger than uniform expansions:
- Uniform expansions control $\sup_{x}|F_n(x)-\Phi(x)|$ at a slower rate and do not capture the rapid decay of the error in the tails.
- Non-uniform expansions serve as the essential technical input for fine-grained Berry–Esseen results, $L^p$-approximations, and optimal transport error rates.

The results are valid under minimal regularity, requiring only moment conditions and analytic control available from spectral gap or quasi-compactness methods, operator-analytic perturbations, or, for weighted and non-i.i.d. arrays, from suitably bounded von Mises derivatives [2210.07204, 1002.4338].

## 7. Limitations, Extensions, and Modern Directions

The necessity of controlling cumulants and moments up to order $m+1$ constrains the maximal order of expansion and the polynomial decay exponent. In certain pathological cases, such as lacking spectral gaps or exponential tails, the expansions may not be achievable to arbitrary order. However, the methods extend beyond independence and stationary settings—encompassing weak dependence, non-identically distributed arrays, sequential or random dynamical systems, and broad classes of functionals, indicating the robustness of the approach [2511.06414, 2210.07204, 1002.4338]. Applications to self-normalized processes, entropic CLTs, and Wasserstein metrics continue to expand the relevance of non-uniform Edgeworth asymptotics, with explicit formulas for moments, quantiles, and functionals becoming computationally feasible under the explicit bracket-and-cumulant calculus established in the referenced papers.

Source: https://www.emergentmind.com/topics/non-uniform-edgeworth-expansions