Depth Projection Methods

• Depth projection is a framework that quantifies data centrality, outlyingness, and geometric relationships by projecting high-dimensional data onto lower dimensions.
• Efficient computation using methods like Monte Carlo approximations and GPU-parallelized algorithms enables practical analysis in high-dimensional and tensor data contexts.
• Extensions into tensor, functional, and kernel settings broaden its applicability to robust statistics, imaging, and machine learning, ensuring invariance and high breakdown robustness.

Depth projection refers to a family of mathematical, statistical, and computational frameworks in which the centrality, outlyingness, and geometric relationships of data points are quantified or leveraged using their projections along low-dimensional directions. This general principle is foundational in robust statistics (e.g., data depth functions), high-dimensional data analysis, tensor and functional data processing, geometric optics, and computer vision. Depth projection also arises in modern imaging and 3D scene understanding, where the transformation between two- and three-dimensional representations is governed by projection models, often with ambiguities and invariances.

1. Statistical Depth Functions and Projection Depth

A statistical depth function assigns to each point $x$ in $\mathbb{R}^d$ a measure $D(x;F)$ of centrality with respect to a distribution $F$. The projection depth (PD) is a canonical example, defined by projecting the data and the query point onto all possible unit directions $u \in S^{d-1}$ and measuring maximal standardized deviation: $$O(x;F) = \sup_{\|u\|=1} \frac{|u^\top x - \mu(F_u)|}{\sigma(F_u)}$$ where $F_u$ is the distribution of $u^\top X$, and $\mu$, $\sigma$ are robust univariate location and scale functionals (e.g., median, MAD). The projection depth is $\mathbb{R}^d$0 (Ramsay et al., 2023).

This framework provides:

• Affine invariance: $\mathbb{R}^d$1 for invertible $\mathbb{R}^d$2.
• Maximality at center: Attains a maximum at the center of symmetry.
• Monotonicity and vanishing at infinity: $\mathbb{R}^d$3 decreases radially from the center and vanishes as $\mathbb{R}^d$4.
• Breakdown robustness: The estimator induced by maximizing PD has breakdown point $\mathbb{R}^d$5 (Ramsay et al., 2023).

For fuzzy random variables, depth projection generalizes by taking all support directions and $\mathbb{R}^d$6-levels, preserving affine invariance and robustness (Fuente et al., 2023).

2. High-Dimensional and Computational Aspects

Exact computation of projection depth is challenging in high dimensions due to the need for optimization over the unit sphere $\mathbb{R}^d$7. For bivariate data, the supremum reduces to a maximum over a finite set of $\mathbb{R}^d$8 data-driven directions, enabling exact, polynomial-time computation of projection depth, contours, and medians using linear-fractional programming (Liu et al., 2011).

In higher dimensions, Monte Carlo (random projection) approximation is widely used. Uniform error bounds scale as $\mathbb{R}^d$9, where $D(x;F)$0 is the number of sampled directions (Nagy et al., 2019). GPU-parallelized algorithms (e.g., Refined Random Search, RRS) achieve speedups up to $D(x;F)$1, allowing interactive computation of projection-based depths for $D(x;F)$2–$D(x;F)$3 (Leone et al., 9 Jun 2025). The RRS algorithm alternates between global random exploration and spherical-cap local refinement, with univariate depth computations parallelized across massive numbers of directions.

Projection-depth based medians and contours may also be computed in regression settings (projection regression depth, PRD), by considering projected residuals along random directions. Exact and approximate algorithms, as well as computationally efficient "fast" depth medians with minor loss in robustness, have been developed (Zuo, 2019).

3. Extensions to Multilinear, Functional, and Kernel Settings

Tensors and Multilinear Data

When data are naturally multidimensional arrays (matrices or tensors as in neuroimaging, images, video), flattening destroys structure and computational tractability. Tensor-based Projection Depth (TPD) generalizes PD by considering scalar multi-mode projections: $D(x;F)$4 and defines $D(x;F)$5 (Hu et al., 2012). An alternating maximization over each mode provides tractable computation. TPD is strictly more powerful than vector-based PD whenever the data possess natural multi-mode structure, outperforming PD for classification on tensor-structured data.

Functional Data

Classical projection-depth degenerates in infinite dimensions; arbitrary directions can have vanishing projected scale, causing the depth to collapse to zero almost surely. Regularized Projection Depth (RPD) restricts to directions $D(x;F)$6 with median absolute deviation above a regularization threshold $D(x;F)$7. Efficient approximate computation uses random projections filtered by their sample MAD (Bočinec et al., 23 Dec 2025, Bočinec et al., 26 Feb 2026). RPD preserves affine invariance, maximality, convexity of depth regions, and Lipschitz continuity, with a robust median possessing breakdown point $D(x;F)$8.

Kernel and Nonlinear Data

Random projection depth (RPD) is extended via kernel methods to handle multimodal or non-convex data via Kernel Random Projection Depth (KRPD). Data are embedded in a Reproducing Kernel Hilbert Space (RKHS), with KPCA providing a low-dimensional representation. Robustness and breakdown properties are inherited from the use of median/MAD in the projected space (Tamamori, 2023).

4. Geometric Depth Projection in Imaging and Computer Vision

In geometric computer vision, depth projection describes the mapping between scene depth and image measurements under the pinhole camera model: $D(x;F)$9 Ambiguity arises: any scaled 3D point projects to the same $F$0; this is the "projection ambiguity" underlying modern monocular depth algorithms (Wang et al., 10 Jul 2025).

Novel monocular or depth-completion techniques exploit this structure:

• Projection ambiguity and consistency (PacGDC): Depth-completion pipelines synthesize pseudo-geometries with varying shape and position/scale but identical image projections, augmenting training diversity while enforcing position (via sparse points) and shape (via semantic image cues) consistency (Wang et al., 10 Jul 2025).
• Forward vs. inverse projection: In dynamic scenes, forward projection provides correct geometric mapping under independent object and camera motion, leading to sharper reconstructions and lower errors compared to inverse warping (Lee et al., 2021).

360-degree (omnidirectional) depth estimation combines complementary projections—equirectangular and cubemap or icosahedral—and bi-projection fusion modules to aggregate both local and global FoVs, overcoming distortions and limited receptive field (Wang et al., 2022, Ai et al., 2024).

In structured light, depth projection can refer to inferring depth from the analysis of projected patterns. For rapidly moving objects where patterns are blurred, depth can be reliably recovered by analyzing the blur width ratios of two projected stripe patterns (light flow) and their known projection Jacobians (Furukawa et al., 2017).

5. Specialized Applications: Robustness, Privacy, and Beyond

Robustness and Privacy

Projection-depth based medians are robust to outliers, with finite-sample breakdown points near $F$1 and bounded influence functions, making them critical in contaminated or adversarial environments (Ramsay et al., 2023). Differentially private depth medians can be constructed using the propose-test-release (PTR) and exponential mechanisms, maintaining robust breakdown and optimal or near-optimal finite-sample accuracy bounds (Ramsay et al., 2023).

Nonconvex and Fuzzy Settings

In problems without natural total order (e.g., fuzzy random variables), projection depth generalizes using support functions of fuzzy $F$2-cuts, with depth and monotonicity properties analogous to the classical case. Metric-based $F$3-depths offer complementary approaches for fuzzy data, at the expense of affine invariance and robustness (Fuente et al., 2023).

6. Role in High-Dimensional, Multimodal, and Multivariate Analysis

Depth projection methods avoid the curse of dimensionality and remain interpretable and robust in high dimensions, tensors, and functional spaces. Regularization or structure-preserving projections (tensor-PCA, random filtering) are essential for tractable computation and avoidance of degeneracy (Hu et al., 2012, Bočinec et al., 26 Feb 2026, Bočinec et al., 23 Dec 2025). Kernelization yields nonparametric adaptability to nonlinear structure and multimodal distributions (Tamamori, 2023). Integration with deep architectures enables simultaneous recovery of depth and spectral information from structured illumination or color-coded projections (Li et al., 2022).

7. Summary Table: Principal Depth Projection Variants

Variant	Key Domain	Core Principle	Robustness
PD	$F$4 vectors	Max outlyingness over $F$5	High
TPD	matrices/tensors	Bilinear/multilinear projection per mode	High
RPD	functional (Hilbert space)	Directional filtering by spread (MAD $F$6)	High
KRPD	RKHS / kernelized	Random projections in KPCA embedding	High / kernel-dependent
Fuzzy PD	fuzzy sets	Outlyingness of support function statistics	High
Geometric	Computer vision, imaging	Pinhole/camera projection, depth–image mapping	Model-dependent

By leveraging depth projection in multivariate analysis, computer vision, high-dimensional statistics, and robust estimation, researchers obtain principled, invariant, and computationally tractable frameworks for quantifying outlyingness, structure, and ambiguity in both classical and modern data regimes (Hu et al., 2012, Leone et al., 9 Jun 2025, Bočinec et al., 26 Feb 2026, Ramsay et al., 2023, Wang et al., 2022, Wang et al., 10 Jul 2025).