---
title: Higher-Order Inverse Stochastic Dominance
url: https://www.emergentmind.com/topics/higher-order-inverse-stochastic-dominance
type: topic
---

# Higher-Order Inverse Stochastic Dominance

Higher-order inverse stochastic dominance (n-ISD) is a quantitative framework for comparing probability distributions based on integrated quantile functions. Extending the standard concept of stochastic dominance, n-ISD leverages higher-degree accumulation of quantiles to provide necessary conditions for ordering, particularly relevant in economics, finance, and welfare analysis. The orders of dominance encapsulate distributional features beyond the mean, reflecting the influence of tails and minima, and offer a rigorous platform for both theoretical characterization and empirical hypothesis testing [2601.03541], [2306.12271].

## 1. Definitions and Core Notation

Let $X$ be a real-valued random variable with cumulative distribution function $F_X(x)$ and left-continuous quantile function $F_X^{-1}(p) = \inf\{x : F_X(x) \ge p\}$ for $p \in [0,1]$. The $n$-th integrated quantile, or $n$-th "inverse-CDF," is defined recursively as:
- $F_X^{[-1]}(p) = F_X^{-1}(p)$,
- $F_X^{[-n]}(p) = \int_0^p F_X^{[-(n-1)]}(u)\, du$ for $n \ge 2$.

An equivalent representation for $n \ge 2$:
$$
F_X^{[-n]}(p) = \frac{1}{(n-2)!} \int_0^1 F_X^{-1}(u)\, (p-u)_+^{n-2}\, du.
$$

The n-th order inverse stochastic dominance between random variables $X$ and $Y$, denoted $X \le_n^{-} Y$, holds if $F_X^{[-n]}(p) \le F_Y^{[-n]}(p)$ for all $p \in (0,1)$. The "strict" variant, $X <_n^{-} Y$, further requires strict inequality at some $p$.

For $k\ge 1$, let $X_{1:k} = \min\{X_1, \dots, X_k\}$ for i.i.d. $X_i$ with law $X$, and $\mu_{1:k}^X = \mathbb{E}[X_{1:k}]$. It follows that:
$$
\mu_{1:k}^X = k! F_X^{[-(k+1)]}(1).
$$

## 2. Moment-Inequality Characterizations

Necessary moment-inequality characterizations for n-ISD parallel Fishburn's results for ordinary $n$-SD.

- **Theorem 3.4 (n-ISD moment inequalities):**  
  For $n>2$, $X, Y \in L^1$ with $X \le_n^{-} Y$, let $k$ with $0 \le k < n-2$, and suppose $\mu_{1:(n-1-j)}^X = \mu_{1:(n-1-j)}^Y$ for $j = 0,\dots,k$. Then:
  $$
  (-1)^{k+1} \mu_{1:(n-2-k)}^X \le (-1)^{k+1} \mu_{1:(n-2-k)}^Y
  $$

- **Theorem 3.5 (strong n-ISD):**  
  If $X <_n^{-} Y$ and either  
  (i) $\mu_{1:j}^X = \mu_{1:j}^Y$ for $j=1,\dots,n-1$, then $\mu_{1:n}^X > \mu_{1:n}^Y$,  
  or  
  (ii) $\mu_{1:j}^X = \mu_{1:j}^Y$ for $j=2,\dots,n$, then $(-1)^n \mathbb{E}[X] < (-1)^n \mathbb{E}[Y]$.

These inequalities concern moments of minimum order statistics and provide sharp necessary—though not sufficient—conditions for n-ISD [2601.03541].

## 3. Relation to Integrated Quantiles and Social Welfare

Higher-order ISD accumulates quantiles from below (upward ISD) or above (downward ISD). For $m \ge 3$ and CDFs $F_1$, $F_2$, with quantile functions $Q_j$, Aaberge, Havnes, and Mogstad define:
$$
\Lambda_j^m(p) = \frac{1}{(m-2)!} \int_0^p (p-t)^{m-2} Q_j(t)\, dt, \\
\widetilde{\Lambda}_j^m(p) = \frac{1}{(m-2)!}\left[(1-p)^{m-2}\mu_j - \int_p^1 (t-p)^{m-2} Q_j(t) dt\right]
$$
where $\mu_j = \int_0^1 Q_j(t)dt$ is the mean.

- $F_1$ has $m$-th-degree upward ISD over $F_2$ if $\Lambda_1^m(p) \ge \Lambda_2^m(p)$ for all $p$.
- $F_1$ has $m$-th-degree downward ISD over $F_2$ if $\widetilde{\Lambda}_1^m(p) \ge \widetilde{\Lambda}_2^m(p)$ for all $p$.

This is equivalent to requiring that, for any Gini-type social welfare weight function $u$ with nonnegative $(m-1)$-th derivative,
$$
\int_0^1 u(p) dF_1^{-1}(p) \ge \int_0^1 u(p) dF_2^{-1}(p),
$$
and analogously for Lorenz-type criteria [2306.12271].

## 4. Proof Structure and Historical Connections

Fishburn's (1980b) method for standard stochastic dominance utilizes the asymptotics of the integrated-CDF as $x \to +\infty$ and relates the expansion's polynomial coefficients to moments. For inverse SD, the asymptotic analysis targets $F_X^{[-n]}(p)$ as $p \to 1$ (or $0$), deploying linear combinations such as $A_n^X(p) = F_X^{[-n]}(p) - \widetilde F_X^{[-n]}(p)$ for odd $n$ and $B_n^X(p) = F_X^{[-n]}(p) + \widetilde F_X^{[-n]}(p)$ for even $n$. Repeated integration provides an expansion in powers of $(1-p)$; coefficients are the $\mu_{1:j}^X$, constraining possible violations of n-ISD via leading-term behavior.

Analogous reasoning establishes the necessity—and sharpness—of moment inequalities for inverse orders, highlighting the theoretical symmetry and distinctions between direct and inverse dominance [2601.03541].

## 5. Special Cases and Practical Examples

For small $n$:
- **First-order:** $X \le_1^{-} Y$ iff $F_X^{-1}(p) \le F_Y^{-1}(p)$ $\forall p$, implying $\mathbb{E}[X] \le \mathbb{E}[Y]$.
- **Second-order:** $F_X^{[-2]}(p) \le F_Y^{[-2]}(p)$ for all $p$ implies $\mu_{1:2}^X \le \mu_{1:2}^Y$; for equal means, this entails $\operatorname{Var}(X) \ge \operatorname{Var}(Y)$.
- **Third-order:** $F_X^{[-3]}(p) \le F_Y^{[-3]}(p)$ yields necessary conditions on $(1-u)$-weighted mean minima, further constraining distributional tails.

The strict forms afford strong conclusions about reversal of order in the next moment statistic when all lower minima match [2601.03541].

## 6. Nonparametric Testing and Empirical Evidence

A nonparametric test for $m$-th-degree ISD utilizes empirical process theory. Given independent samples from $F_1, F_2$, empirical CDFs $\hat{F}_j$, quantiles $\hat{Q}_j$, and corresponding $\hat{\Lambda}_j^m, \hat{\widetilde{\Lambda}}_j^m$, the difference processes $\hat\phi_m^{u/d}$ are defined, measuring upward or downward ISD gaps. Test statistics are based on functionals $\mathcal{S}(h) = \sup_{p \in [0,1]} h(p)$ or $\mathcal{I}(h) = \int_0^1 \max\{h(p), 0\}dp$, and asymptotic inference employs weighted (multiplier) bootstrap techniques.

Under standard regularity—CDFs supported on $[0, \infty)$ with positive density, finite higher moments, and mild copula and differentiability assumptions—the test controls size and is consistent. Empirical illustrations with "double-Pareto" distributions confirm strong finite-sample properties. Application to UK income data (1995–2010) reveals that higher-order downward ISD almost totally ranks distributions by upper-tail changes, while upward ISD stresses lower-tail experiences, matching theoretical expectations for welfare analysis [2306.12271].

## 7. Corollaries, Limitations, and Extensions

The moment inequalities are necessary but not sufficient—distinct distributions may satisfy all mean minimum inequalities without obeying the full integrated-quantile ordering. The bounds are nevertheless tight, coinciding with the binomial-moment terms arising in asymptotic expansions.

For ordinary $n$-SD, background risk effects generalize Pomatto–Strack–Tamuz's result: given $X, Y$ distinguishable only in the $n$-th moment (lower moments matched and strict $n$-th moment inequality), there exists an independent background risk $Z$ such that $X+Z >_n Y+Z$, illustrating the amplification of higher-order dominance by additive noise [2601.03541].

Further examples, technical proofs, and tables of empirical results appear in Guan–Zou–Hu [2601.03541] and the nonparametric testing framework of [2306.12271].

Source: https://www.emergentmind.com/topics/higher-order-inverse-stochastic-dominance