---
title: Scatter Halfspace Depth in Multivariate Analysis
url: https://www.emergentmind.com/topics/scatter-halfspace-depth
type: topic
---

# Scatter Halfspace Depth in Multivariate Analysis

Scatter halfspace depth (sHD) is a robust, nonparametric measure quantifying the fit of a positive-definite scatter (or covariance) matrix to a multivariate data cloud. Extending Tukey's location-based halfspace depth, sHD generalizes the concept of data "centrality" from points in $\mathbb{R}^d$ to the space of symmetric positive-definite matrices, providing a principled approach for multivariate scatter estimation with strong robustness and minimax-optimality properties. Recent work has developed both exact and approximate computational methods, clarified the geometry of depth regions, and explored its performance in both classical elliptically symmetric and more general $\alpha$-symmetric models [2208.05173][1704.06160][2512.07634].

## 1. Definition of Scatter Halfspace Depth

Let $X_1, \dots, X_n \in \mathbb{R}^d$ be a sample, and fix a location center $\mu \in \mathbb{R}^d$ (typically the Tukey median). For any symmetric positive-definite matrix $\Sigma \succeq 0$, the scatter halfspace depth is defined as
\[
\mathrm{sHD}(\Sigma; \mu) = \min_{u \in S^{d-1}} \min \left\{ \#\left\{i: |u^\top(X_i-\mu)| \le \sqrt{u^\top \Sigma u} \right\}, \; \#\left\{i: |u^\top(X_i-\mu)| \ge \sqrt{u^\top \Sigma u} \right\} \right\}.
\]
Conceptually, sHD measures the smallest mass (number or probability) of data points in any slab delimited by two parallel hyperplanes orthogonal to $u$ and at Mahalanobis (ellipsoidal) distance $\sqrt{u^\top \Sigma u}$ from $\mu$, across all directions $u$. This definition extends to the population version using probabilities:
\[
\mathrm{H}\!D^{\mathrm{sc}}_{P,T}(\Sigma) = \inf_{u \in S^{d-1}} \min \left\{ P\left( |u^\top (X - T_P)| \le \sqrt{u^\top \Sigma u} \right), \; P\left(|u^\top (X - T_P)| \ge \sqrt{u^\top \Sigma u} \right) \right\}
\]
where $P$ is the distribution of $X$ and $T_P$ is an affine-equivariant location functional [2208.05173][1704.06160].

When $d=1$, sHD coincides with the univariate median absolute deviation depth ranking; for general $d$, sHD can be understood as employing Tukey depth for the projections $u$ of the centered data cloud [2208.05173].

## 2. Statistical Properties and Theory

### Core Properties

- **Affine Invariance**: For any invertible transformation $A$ and shift $b$, $\mathrm{sHD}(A \Sigma A^\top; b + A\mu)$ computed from transformed data $\{AX_i+b\}$ equals $\mathrm{sHD}(\Sigma; \mu)$ on $\{X_i\}$ [2208.05173][1704.06160][2512.07634].
- **Robustness**: sHD downweights the influence of outlying observations; only scatter matrices aligned with the main data pattern achieve high depth [2208.05173].
- **Continuity and Quasiconcavity**: For generic data in general position, $\Sigma \mapsto \mathrm{sHD}(\Sigma)$ is upper semicontinuous and sHD contours are quasiconcave/nested as depth varies. The population version exhibits upper semicontinuity under both the Frobenius and geodesic (Riemannian) metrics, and full continuity under smoothness assumptions on $P$ [1704.06160].
- **Monotonicity and Boundedness**: sHD decreases ("vanishes") at the boundary of the scatter parameter space as $\|\Sigma\|_F \to \infty$ or smallest eigenvalue $\lambda_{\min}(\Sigma) \to 0$ [1704.06160].

### Structural Insights

The deepest scatter matrix (sHD-median) exists for smooth distributions, and for elliptical models it is uniquely maximized at the true scatter [1704.06160][2512.07634]. For elliptical $P$, the maximum depth is $1/2$; for other distributions, the supremum may be less. Depth regions (super-level sets) are convex and nested [1704.06160][2512.07634].

### Extensions to $\alpha$-Symmetric Distributions

Recent developments introduce an $\alpha$-symmetric scatter depth for distributions where the characteristic function has the form $\varphi_X(t) = \phi(\|t\|_\alpha)$. The definition generalizes the Mahalanobis ellipsoid to the $\ell_\alpha$ norm:
\[
\mathrm{SD}_\alpha(\Sigma; P) = \inf_{u \in S^{d-1}} \min \left\{ P\left( \langle X - T(P), u \rangle \le \| \Sigma^{1/2} u \|_\alpha \right), P\left( \langle X - T(P), u \rangle \ge \| \Sigma^{1/2} u \|_\alpha \right) \right\}
\]
Retaining affine equivariance and convexity, this version achieves minimax-optimality in broader distributional classes [2512.07634].

## 3. Robustness, Minimax Optimality, and Breakdown Properties

For $\mu$ taken as the Tukey median, the sHD-median $\Sigma^*$ defined by maximizing sHD is a robust scatter estimator. Under the Huber $\epsilon$-contamination model—where data is a mixture of a “clean” distribution and arbitrary outliers—$\Sigma^*$ achieves minimax rate $n^{-1/2}$ (up to constants) for estimation error, and exhibits a breakdown point approaching $1/2$. Thus, sHD attains both high resistance to gross errors and optimal error rates under contamination [2208.05173][2512.07634].

The optimality at the true scatter persists in elliptical models, with explicit characterizations and guarantees of existence and uniqueness. In $\alpha$-symmetric families, the unique sHD-median is always isotropic: $\Sigma = \sigma^2 I_d$ for a scale $\sigma$ determined by the marginal distribution, and minimax-optimal finite-sample error bounds have been established [2512.07634].

## 4. Computation: Exact and Approximate Algorithms

### Exact sHD Computation

Exact sHD computation is combinatorially expensive but theoretically tractable for low dimensions. The key geometric concepts include:

- **Spherical Circles and Shells**: Each point with $\|X_i\| > 1$ defines the pair of spheres $C_i = \{u \in S^{d-1}: u^\top X_i = 1\}$ and its antipode, partitioning the sphere into regions where point counts are constant.
- **Maximal Tangent Hyperplanes**: Only directions $u$ corresponding to maximal supporting hyperplanes require evaluation.
- **Algorithm Outline**: (i) Affine-standardize the data, (ii) enumerate k-subsets corresponding to candidate tangent planes, (iii) detect those outside the unit ball, (iv) update sHD via point counts, (v) mark visited subsets to avoid redundancy.

This enumeration has worst-case $O(n^d d^2)$ complexity, but is practical for $d\leq5$ and moderate $n$ [2208.05173].

### Fast Approximations

For larger $d$ or $n$, approximate algorithms are necessary:

- **Random-Directions ("rdirections")**: Sample $N$ random $u$’s on $S^{d-1}$, evaluate the minimal point count slab for each, and take the minimum. Complexity is $O(Nnd)$.
- **Random-Points ("rpoints")**: Randomly sample $M$ $(d-1)$-tuples of points, focusing only on the most likely critical slabs. More accurate at fixed computational budget, with complexity $O(Mnd^2)$.

Empirical results show rpoints has near-exact accuracy $\sim$90% for reasonable $M$ (e.g., $M=10^4$), while rdirections is less reliable (50% accuracy for comparable parameter settings) [2208.05173].

Implementation is available via C++ (for $d\leq5$) and R (arbitrary $d$), with attention to numerical stability (projections differing by less than $10^{-14}$ are carefully excluded) [2208.05173].

## 5. Geometry and Topology of Depth Regions

The set of all $\Sigma$ with sHD above a given threshold forms a depth region, denoted $R^{\mathrm{sc}}_{P,T}(\alpha) = \{\Sigma: \mathrm{H}\!D^{\mathrm{sc}}_{P,T}(\Sigma) \ge \alpha\}$. These regions are nested and convex in the Frobenius geometry, closed under mild regularity, and may be unbounded at the boundary of the space of scatter matrices.

Considering the geodesic Riemannian metric on the symmetric positive-definite cone confers additional structure: for smooth $P$, geodesically deep regions are always compact for all but the minimal nonzero depth value. In elliptical models, depth is geodesically quasiconcave, and each region is geodesically convex [1704.06160].

## 6. Applications and Empirical Behavior

Scatter halfspace depth underpins robust scatter estimation in high dimensions and in contaminated or heavy-tailed settings. In finance, practical utility is demonstrated via diagnostic analysis of scatter and shape outlyingness in high-frequency returns: sHD-based measures align with major financial events and outperform Euclidean distances in identifying both scale and shape outliers [1704.06160].

For sample estimation, empirical sHD uniformly converges to its population counterpart under the empirical measure and affine-equivariant location functional, under mild conditions. The DepthDescent algorithm—an SPD-matrix descent scheme—handles dimensions up to $k=10$ and sample sizes of several thousand [1704.06160].

In the context of $\alpha$-symmetric distributions, the unique sHD-median is isotropic and computable by solving a one-dimensional root-finding problem, with each step costing $O(n)$. The estimator enjoys optimal concentration under Huber contamination, with estimation error matching minimax rates [2512.07634].

## 7. Historical Perspective and Relation to Other Depths

The sHD formalism generalizes ideas from classic Tukey depth (location) to scatter estimation and builds on early work by Zhang (2002) and Chen–Gao–Ren (2017), who proposed depth functionals for scatter with varying assumptions and properties. sHD distinguishes itself by its precise geometric definition, robust optimality, structural analysis of depth regions under multiple topologies, and generalization to $\alpha$-symmetric laws [1704.06160][2512.07634].

Alternative centering strategies (such as pairwise differences to avoid estimating a location parameter) can yield related depth notions, but fail to achieve Fisher consistency under general elliptical laws, highlighting the necessity of affine-equivariant centering [1704.06160].

## Table: Summary of Key Properties

| Property                | sHD (Elliptical)           | sHD ($\alpha$-Symmetric)               |
|-------------------------|----------------------------|-----------------------------------------|
| Affine Invariance       | Yes                        | Yes                                    |
| Uniqueness of Median    | Yes (elliptical)           | Yes (isotropic)                        |
| Robustness              | High, breakdown $\sim1/2$  | High, breakdown $\sim1/2$              |
| Minimax Optimality      | Yes (Huber contamination)  | Yes (by construction)                  |
| Region Convexity        | Frobenius and geodesic     | Convex (Frobenius), holds for depth     |
| Efficient Computation   | Low $d$ via enumeration;   | Isotropic, 1D root-finding per sample   |
|                         | Large $d$: stochastic      |                                         |

Scatter halfspace depth constitutes a foundational methodology for robust, nonparametric, affine-invariant scatter estimation in multivariate analysis, with rigorous guarantees and practical computational tools [2208.05173][1704.06160][2512.07634].

Source: https://www.emergentmind.com/topics/scatter-halfspace-depth