---
title: 'h* Statistic: Definition and Applications'
url: https://www.emergentmind.com/topics/h-statistic
type: topic
---

# h* Statistic: Definition and Applications

Searching arXiv for papers on “h* statistic”, Hajek projection/U-statistics, and Coxeter-group h* usage to support a precise, up-to-date article.
The notation \(h^*\) is used in several technically distinct ways across the arXiv literature. In recent statistics, it denotes a parametric, frequentist test statistic for evaluating global outliers without a normality assumption [2508.06792]. In the literature on differentially private U-statistics, the empirical local Hájek projection \(\hat h(i)\) is sometimes denoted \(h^*\), although that paper explicitly uses the term “local Hajek projection” [2407.04945]. In algebraic combinatorics, \(h^*\) appears as a statistic on symmetric groups whose type-\(B\) analogue is the statistic \(L\) on hyperoctahedral groups [1303.0990]. This suggests that \(h^*\) is not a single universal object, but a notation whose meaning is field-specific.

## 1. Distinct uses of the notation \(h^*\)

Three usages are explicitly represented in the cited literature.

| Context | Meaning of \(h^*\) | Source |
|---|---|---|
| Outlier analysis | A test statistic comparing outlier-to-group distance with inlier pairwise spread | [2508.06792] |
| Private U-statistics | A notation sometimes used for the local Hájek projection \(\hat h(i)\) | [2407.04945] |
| Coxeter combinatorics | A statistic on symmetric groups generalized to type \(B\) by \(L\) | [1303.0990] |

The most explicit standalone definition of an \(h^*\) statistic is the outlier-analysis construction of 2025, where \(h^*\) is designed for one-dimensional homogeneous data and is positioned as analogous to the role of Student’s \(t\) in comparing means [2508.06792]. By contrast, in the U-statistics setting the notation is ancillary: the principal object is the local Hájek projection, and the statement that it is “sometimes denoted \(h^*\) in the literature” is primarily terminological rather than definitional [2407.04945]. In Coxeter theory, the available paper does not redefine \(h^*\) directly, but states that the restriction of \(L\) to \(S_n\) coincides with the “\(h^*\) statistic” introduced earlier by Klopsch and Voll, so the cited paper is best read as a type-\(B\) generalization rather than a primary exposition of \(h^*\) itself [1303.0990].

A common misconception is therefore to treat \(h^*\) as denoting one canonical statistic across fields. The cited material instead supports a contextual reading.

## 2. \(h^*\) as an outlier statistic

For a set of \(n\) real-valued, one-dimensional, homogeneous data points
\[
\mathbf{X}=(X_1,X_2,\ldots,X_n)^\intercal,
\]
let \(X^*\) denote a candidate outlier, typically the maximum or minimum value. The \(h^*\) test statistic is defined by
\[
h^*=\sqrt{
\frac{
\displaystyle\sum_{k=1}^n (X_k-X^*)^2/(n-1)
}{
\displaystyle\sum_{\substack{i>j\\ X_i,X_j\neq X^*}} (X_i-X_j)^2/\binom{n-1}{2}
}
}.
\]
Its numerator is the mean squared distance from the candidate outlier to every data point, while its denominator is the mean squared pairwise distance among all other data points, excluding \(X^*\) [2508.06792].

The paper also gives an alternative algebraic form and introduces a scaled form \(\tilde h^*\), with the stated range \(\tilde h^* \in [1/\sqrt{n-2},\infty)\) [2508.06792]. Operationally, the calculation proceeds by summing squared distances from the candidate outlier to the rest of the sample, computing the pairwise inlier squared distances among the remaining \(n-1\) points, and taking the ratio of the corresponding root-mean-square quantities [2508.06792].

The interpretation is explicitly geometric. High \(h^*\) means that the candidate outlier is much farther from the rest of the dataset than the typical pairwise spread among inliers; low \(h^*\) means that the candidate “outlier” may not be unusually extreme relative to the group [2508.06792]. Both \(h^*\) and \(\tilde h^*\) are stated to be invariant under linear transformations such as shifting or scaling all data [2508.06792].

The paper’s worked example uses \(n=4\) with \(\mathbf X=[3,4,5,8]\) and \(X^*=8\). The numerator is based on \(25+16+9=50\), the denominator on \(1+4+1=6\), giving
\[
h^*=\sqrt{(50/3)/2}\approx 3.54
\]
[2508.06792].

## 3. Statistical inference, null distribution, and interpretation

The outlier-analysis \(h^*\) statistic is used as a one-tailed test for whether an observed value is exceptionally large or small under a chosen distributional assumption for the ordinary data [2508.06792]. The null hypothesis is that the candidate outlier does not differ sufficiently from ordinary data to be considered statistically exceptional [2508.06792].

The cited paper states that the distribution of \(h^*\), and hence significance thresholds, depends on both \(n\) and the distribution of the inlier data [2508.06792]. Exact null distributions are described as nontrivial except in special cases, so \(p\)-values are obtained from simulated or sampled empirical distributions for the chosen distribution and sample size, or by Monte Carlo simulation [2508.06792]. The stated degrees of freedom are \(\nu=n-2\) [2508.06792].

The appendix provides simulated critical values for normal and log-normal data. One example reported in the paper is that, for \(n=10\), the 95% threshold is approximately \(2.59\) under normality and approximately \(9.19\) for log-normal data [2508.06792]. This makes clear that the procedure does not eliminate distributional assumptions altogether; rather, it removes the requirement of normality and replaces it with simulation-based calibration under the selected inlier model.

The paper emphasizes that \(h^*\) is not a tool for automatic outlier detection or removal. Instead, it provides an objective assessment of an individual point’s extremeness, intended for interpretative and substantive analysis [2508.06792]. In that sense, the statistic quantifies a “degree” of outlierness rather than merely producing a binary inlier/outlier classification.

## 4. Comparison with classical outlier tests and proposed extensions

The paper contrasts \(h^*\) with Grubbs’ test and Dixon’s \(Q\) test. Grubbs’ test is described as based on the \(Z\)-score of the maximum or minimum versus the mean, assumes normality, considers all data, does not handle clustering, is not robust to distribution, and is framed as discarding outliers as errors. Dixon’s \(Q\) is described as a ratio of gap to range, also assuming normality, focusing on nearest neighbors and the extremal gap [2508.06792]. By comparison, \(h^*\) is described as requiring no normality assumption, considering all distances except those involving the removed candidate, handling clustering, being robust to distribution, and quantifying the degree of outlierness [2508.06792].

The practical scope given in the paper is one-dimensional, homogeneous, interval- or ratio-scale data, with caveats for certain rating scales [2508.06792]. The method is described as empirical and proximity-based, able to handle heterogeneity, multimodality, and local clustering, and as linearly invariant [2508.06792].

Several extensions are proposed. A Bayesian formulation treats \(h^*_{\mathrm{obs},j}\) as entering a posterior probability of outlierness \(P(L_j=1\mid h^*_{\mathrm{obs},j})\), with further marginalization and posterior updating described in Equations 8–13, and with MCMC-based simulation used to obtain posterior probabilities for candidate outliers [2508.06792]. A paired-analysis procedure is also described: identify outliers pre-intervention, compute \(h^*\) pre and post for the same individuals, and use the Wilcoxon signed-rank or paired \(t\)-test to assess whether \(h^*\) drops significantly among extreme cases [2508.06792].

The paper further introduces a weighted statistic
\[
h^*_{\mathrm w}=
\sqrt{
\frac{
\left.\sum_{k=1}^{n} w_{k,*}(X_k-X^*)^2 \middle/ \sum_{k=1}^n w_{k,*}\right.
}{
\left.\sum_{\substack{i>j\\X_i,X_j\neq X^*}} w_{ij}(X_i-X_j)^2 \middle/ \sum_{\substack{i>j\\X_i,X_j\neq X^*}} w_{ij}\right.
}
}
\]
and a sensitivity-generalized form
\[
h^*_\eta=
\left(
\frac{
\left.\sum_{k=1}^{n} w_{k,*}(X_k-X^*)^\eta \middle/ \sum_{k=1}^n w_{k,*}\right.
}{
\left.\sum_{\substack{i>j\\X_i,X_j\neq X^*}} w_{ij}(X_i-X_j)^\eta \middle/ \sum_{\substack{i>j\\X_i,X_j\neq X^*}} w_{ij}\right.
}
\right)^{1/\eta},
\]
where higher \(\eta\) increases sensitivity to large differences [2508.06792]. The paper presents these as mechanisms for contextual weighting and for moving from the root-mean-square case \(\eta=2\) to a Hölder-mean family.

## 5. \(h^*\) as local Hájek projection in private U-statistics

A distinct usage arises in the theory of U-statistics under differential privacy. The paper studies estimation of
\[
\theta=\mathbb E[h(X_1,\ldots,X_k)],
\]
where \(X_1,\ldots,X_k\) are i.i.d. and \(h\) is a permutation-invariant kernel [2407.04945]. The corresponding non-private U-statistic is
\[
U_n=\frac{1}{\binom{n}{k}}\sum_{S\in I_{n,k}} h(X_S),
\]
which the paper describes as the minimum variance unbiased estimator for \(\theta\) under mild conditions [2407.04945].

The relevance of \(h^*\) enters through the Hájek projection. The paper defines
\[
\hat S_n=\sum_{i=1}^n \mathbb E[U_n\mid X_i]-(n-1)\mathbb E[U_n]
=\frac{k}{n}\sum_{i=1}^n \mathbb E[h(X_S)\mid X_i]-(n-1)\theta,
\]
and its empirical local version
\[
\hat h(i):=\frac{1}{\binom{n-1}{k-1}}\sum_{S\in I^{(i)}_{n,k}} h(X_S),
\]
where \(I^{(i)}_{n,k}\) denotes the \(k\)-subsets containing \(i\) [2407.04945]. The paper states that \(\hat h(i)\) is sometimes denoted \(h^*\) in the literature, while using the term “local Hajek projection” itself [2407.04945].

This quantity is central to the paper’s privacy mechanism. U-statistics are difficult to privatize because of overlapping dependence, elevated global sensitivity relative to sample means, and degeneracy, where the variance can decay as \(O(1/n^2)\) while standard privacy noise can remain too large [2407.04945]. The proposed remedy is a thresholding-based approach using local Hájek projections to identify “good” and “bad” points, reweight tuples, and then add carefully calibrated smooth-sensitivity noise [2407.04945].

The stated significance of the local projection is twofold. First, it identifies and reweights outlier influence so that sensitivity is low on most datasets. Second, in degenerate cases where \(\zeta_1=0\), it is sharply concentrated around \(\theta\), allowing the private estimator to remain close to the non-private U-statistic while adding much less privacy noise [2407.04945]. The paper reports nearly optimal private error for non-degenerate U-statistics and strong evidence of near-optimality for degenerate U-statistics [2407.04945]. In this literature, then, “\(h^*\)” is not an outlier score but a projection-based linearization tool for statistical and algorithmic analysis.

## 6. \(h^*\) in Coxeter combinatorics and its type-\(B\) generalization

In algebraic combinatorics, the available arXiv source does not redefine \(h^*\) directly, but it does locate the statistic within the theory of Coxeter groups. The paper introduces a new statistic \(L\) on the hyperoctahedral group \(B_n\),
\[
L(w):=\#\{(i,j)\in[\pm n]^2\mid i<j,\ w(i)>w(j),\ i+j\equiv 1 \pmod 2\},
\]
and describes it as analogous to the classical Coxeter length function, with a crucial parity condition [1303.0990]. It also gives the decomposition
\[
L(w)=a(w)+b(w)+2c(w)
\]
in terms of refined combinatorial statistics on signed permutation matrices [1303.0990].

The direct connection to \(h^*\) is explicit: the restriction of \(L\) to \(S_n\) coincides with the “\(h^*\) statistic” introduced by Klopsch and Voll, and the paper presents \(L\) as a type-\(B\) analogue of that earlier statistic [1303.0990]. The parity restriction \(i+j\equiv 1 \pmod 2\) is singled out as the defining feature that distinguishes \(L\) from ordinary Coxeter length [1303.0990].

The paper’s main conjecture concerns the signed generating function of \(L\) over descent classes:
\[
\sum_{w\in B_n^I} (-1)^{\ell(w)} X^{L(w)} = f_{n,I}(X),
\]
with
\[
f_{n,I}(X)=\frac{(2n)!_X}{\prod_{r=1}^l (2(i_{r+1}-i_r))!_X}.
\]
For singleton descent classes, the conjectured formula yields the Poincaré polynomials of varieties of symmetric matrices of fixed rank; for the full group,
\[
\sum_{w\in B_n} (-1)^{\ell(w)} X^{L(w)}=(2n)!_X
\]
[1303.0990]. The proof strategy for several cases uses supporting sets and sign-reversing involutions to show pairwise cancellation outside those sets [1303.0990].

Within this literature, \(h^*\) is therefore a combinatorial statistic tied to generating functions, descent classes, Poincaré polynomials, and representation-theoretic structures, rather than to statistical inference in the ordinary probabilistic sense.

## 7. Conceptual synthesis

Across the cited literature, \(h^*\) names objects that share notation but not domain, null model, or inferential target. In outlier analysis, it is a root-mean-square distance ratio for testing whether a candidate datum is exceptionally far from its peers [2508.06792]. In differentially private U-statistics, it is a notational alias sometimes attached to the local Hájek projection used to control sensitivity and error [2407.04945]. In Coxeter combinatorics, it denotes a statistic on symmetric groups that is generalized by a parity-sensitive type-\(B\) statistic \(L\) [1303.0990].

This multiplicity matters methodologically. The outlier \(h^*\) is calibrated by empirical or simulated null distributions tied to the inlier model and sample size [2508.06792]. The U-statistics \(h^*\) usage is embedded in a privacy-utility analysis involving degeneracy, smooth sensitivity, and projection-based thresholding [2407.04945]. The combinatorial \(h^*\) is studied through signed generating functions, supporting sets, and sign-reversing involutions [1303.0990]. These are not interchangeable constructs.

A plausible implication is that references to an “\(h^*\) statistic” should always be read with explicit disciplinary qualification. Without that qualification, the notation is underdetermined.

Source: https://www.emergentmind.com/topics/h-statistic