Edgeworth–Cornish–Fisher Expansions

• Edgeworth–Cornish–Fisher Expansions are higher-order asymptotic series that refine the approximation of distribution functions and quantiles by incorporating cumulants such as skewness and kurtosis.
• They improve finite-sample inference by systematically correcting the normal approximation, yielding more accurate confidence intervals and hypothesis tests via quantile inversion.
• These expansions are versatile, extending to i.i.d., non-i.i.d., dependent, and weighted samples, with applications in finance, biostatistics, and network analysis under appropriate regularity conditions.

Edgeworth–Cornish–Fisher Expansions are higher-order asymptotic series that provide refined approximations to the distribution and quantiles of sample statistics, surpassing the accuracy of the classical central limit theorem (CLT). These expansions systematically incorporate standardized cumulants beyond variance (notably, skewness and kurtosis), leading to corrections which enhance the normal approximation for finite-sample inferential procedures. The Edgeworth expansion addresses the distribution function, while the Cornish–Fisher expansion is its quantile inversion, producing expansions for percentiles or confidence bounds. The formalism applies broadly: i.i.d., non-i.i.d., weighted, dependent, or even randomly-sized samples, provided certain regularity conditions on cumulant growth and smoothness are met (Withers et al., 2012, Withers et al., 2010, Ulyanov et al., 2016, Bertrand et al., 2018).

1. Edgeworth Expansion: Fundamentals and Formulation

The Edgeworth expansion approximates the CDF (and, if desired, density) of a standardized statistic whose cumulants admit power series in the inverse sample size. Let $\hat\theta$ be a standard estimate so that

$$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$

and define the normalized variable

$Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$

whose scaled cumulants $\ell_r$ satisfy $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$, $\ell_r = O(n^{1-r/2})$ (Withers et al., 2012). The $R$-term one-sided Edgeworth expansion is

$$P_n(x) = \mathbb{P}\{Y_n \leq x\} = P(x) - p(x) \sum_{r=1}^R n^{-r/2} h_r(x) + O(n^{-(R+1)/2})$$

where $P(x), p(x)$ are the CDF and PDF of a selected reference law (usually $N(0,1)$), and each $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$0 is a polynomial in $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$1 and the scaled cumulants, with a representation in terms of generalized Hermite polynomials: $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$2 with $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$3 combinatorially defined by Bell polynomials in $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$4 (Withers et al., 2012).

The expansion increases the accuracy of finite-sample approximations, with the order of the error term decreasing as $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$5 increases.

2. Cornish–Fisher Expansion for Quantiles

The Cornish–Fisher expansion inverts the Edgeworth series to obtain asymptotic expansions of quantiles (percentiles) for the statistic. Given $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$6, the $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$7-quantile of the reference law, the expansion for the $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$8-quantile $$\kappa_r(\hat\theta) = a_{r,0} n^{1-r} + a_{r,1} n^{-r} + \cdots, \quad r \geq 1, \quad a_{1,0}=0, \quad a_{2,0}>0$$9 of $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$0 is

$Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$1

where $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$2, like $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$3, are polynomials in $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$4 and the scaled cumulants, and can be calculated using formulas involving Hermite polynomials (Withers et al., 2012, Bertrand et al., 2018). For example, to $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$5 with normal reference: $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$6

$Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$7

These expansions serve to correct coverage of confidence intervals and significance thresholds under non-negligible skewness or kurtosis (Withers et al., 2012, Bertrand et al., 2018).

3. Generalizations: Weighted, Dependent, and Non-i.i.d. Cases

Edgeworth–Cornish–Fisher expansions have been generalized for weighted empirical distributions and for samples with non-identically distributed components. For independent $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$8 with predetermined weights $Y_n = (n/a_{2,0})^{1/2}(\hat\theta - \theta_0)$9, cumulants of smooth functionals $\ell_r$0 are expanded using higher-order von Mises derivatives (Withers et al., 2010). This yields

$\ell_r$1

with explicit formulas for cumulant coefficients up to $\ell_r$2, enabling third-order expansion for both the distribution and quantiles. The expansions are valid provided weights are uniformly bounded and the power sums $\ell_r$3 remain $\ell_r$4. Applications include nonparametric estimates, rank statistics, and regression functionals (Withers et al., 2010).

For dependent sequences, valid Edgeworth and Cornish–Fisher expansions require additional terms reflecting the bias and stochastic variability of variance estimators. For studentized statistics under strong mixing, the expansion involves three series: powers of $\ell_r$5, $\ell_r$6 (where $\ell_r$7 is the number of lag-covariances estimated), and the bias $\ell_r$8 of the studentizing factor. Each enters into the distributional and quantile expansions as specific correction polynomials (Lahiri, 2010).

4. Distributional Reference Laws and Matched-Skewness Expansions

While the standard normal is a frequent reference law, expansions about alternative distributions yield improved efficiency for skewed statistics. Expanding about a gamma law with matched skewness (i.e., choosing the gamma parameter $\ell_r$9 so $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$0 of the standardized statistic matches $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$1 of the gamma) nullifies leading-order skewness corrections, drastically reducing the polynomial complexity required for a fixed accuracy (Withers et al., 2012). This "matched-skewness gamma" approach can cut the number of necessary nonzero correction terms by up to 90% for $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$2 accuracy relative to the standard-normal expansion.

5. Explicit Formulas, Implementation, and Non-Asymptotic Error

Edgeworth correction polynomials are built from cumulant-derived coefficients and generalized Hermite polynomials. For the standard normal, $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$3, $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$4, $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$5, $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$6, $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$7. For a regular smooth function of the mean, cumulants up to order 4 can be computed by symbolic differentiation and substitution of population moments, with implementation facilitated via automated tools (e.g., Maple sheets and R code) for complex parametric or nonparametric estimators (Bertrand et al., 2018).

The Cornish–Fisher coefficients for quantiles are derived via formal Taylor expansion or via explicit back-substitution, with non-asymptotic error bounds available when the Edgeworth remainder is controlled. The minimum required for second-order accuracy is typically explicit control of the Edgeworth remainder and the possibility of a monotone Bartlett correction transformation (Ulyanov et al., 2016). Bounds hold uniformly only on central probability intervals.

6. Applications and Impact

Edgeworth–Cornish–Fisher expansions have broad applicability in contemporary statistics:

• Confidence intervals and hypothesis tests for sample means, variances, medians (including under random sample size), U-statistics, M-estimators, and network moments (Zhang et al., 2020, Christoph et al., 2019).
• Correcting critical values for heavy-tailed, skew, or dependent samples, with applications in finance (VaR), biostatistics, insurance, and network inference (Ulyanov et al., 2016, Zhang et al., 2020).
• Improved approximations for non-normal limit laws (Student's $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$8, Laplace, Gamma) in scenarios with random or overdispersed sample sizes (Christoph et al., 2019, Withers et al., 2012).
• Accurate quantile estimation for high-dimensional or weighted empirical functionals (Withers et al., 2010).
• Enhanced coverage accuracy for bootstrap confidence intervals and bias-corrected and accelerated (BCA) resampling (Bertrand et al., 2018).

The expansions provide significant reductions in coverage error compared to first-order (CLT-based) inference, with quantitative benefits persisting in moderate $\kappa_r(Y_n) = r!\,\ell_r + O(n^{-1/2})$9 regimes when higher cumulants are non-negligible. The practical implementation is enabled by explicit coefficient expressions and algorithmic differentiation, with code templates available for symbolic and numerical computation (Bertrand et al., 2018).

7. Limitations, Regularity, and Error Rates

The validity of these expansions is contingent on regularity conditions: existence of a cumulant expansion (i.i.d. or weakly dependent data), nonlattice structure or sufficient "self-smoothing" (particularly for U-statistics and network moments), boundedness of higher moments, and adequate smoothness of the statistic's functional representation (Withers et al., 2012, Zhang et al., 2020). For non-i.i.d. or weighted cases, von Mises derivatives must remain uniformly bounded. In certain settings (e.g., extreme tails or excessively small sample size), the expansions do not converge uniformly, and explicit error bounds require knowledge of specific Edgeworth remainder terms (Ulyanov et al., 2016).

The accuracy is characterized by $\ell_r = O(n^{1-r/2})$0 error for $\ell_r = O(n^{1-r/2})$1-term expansions; for non-i.i.d. or weighted cases, achieving $\ell_r = O(n^{1-r/2})$2 error is standard. For dependent or studentized statistics, the error in quantile estimation accumulates from normal, estimation-variance, and bias contributions, whose coefficients must be balanced according to bias-variance tradeoffs (e.g., choice of lag-length in long-run variance estimation) (Lahiri, 2010).

In applied contexts, these expansions permit computation of accurate, higher order-corrected critical values and confidence intervals for a wide variety of statistics, over a flexible range of sample regimes and underlying data structures (Withers et al., 2012, Withers et al., 2010, Bertrand et al., 2018).