📊 statistics

Depth-based estimation for multivariate functional data with phase variability

This paper proposes a robust depth-based approach for estimating the central pattern of multivariate functional data subject to both individual phase variation and cross-component time warping, establishing the necessary conditions for consistency and demonstrating performance through simulations and real-world applications.

Original authors: Ana Arribas-Gil, Sara López-Pintado

Published 2026-02-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ana Arribas-Gil, Sara López-Pintado

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding the "True Shape" in a Messy World

Imagine you are a conductor trying to figure out the perfect melody of a song. You ask 50 different musicians to play the same piece. However, there are two problems:

Timing Issues (Phase Variation): Some musicians start a little early, some drag their feet, and some speed up in the middle. They are all playing the right notes, but at different speeds.
Multiple Instruments (Multivariate Data): Instead of just one violin, each musician is playing a whole orchestra (violins, drums, flutes) all at once. The timing issues might affect the drums differently than the flutes, but the person playing them is the same.

The goal of this paper is to find the original, perfect melody (the "common pattern") without getting confused by the musicians' timing mistakes or the fact that they are playing multiple instruments at once.

The Old Way vs. The New Way

The Old Way (Registration):
Traditionally, statisticians tried to fix this by "aligning" the curves first. Imagine taking a piece of elastic tape with the music drawn on it and physically stretching or shrinking it until it matches a master template. You have to do this for every single musician and every single instrument.

The Problem: This is like trying to stretch a rubber band perfectly while someone is throwing rocks at it. If one musician is playing a weird, distorted tune (an outlier), the whole stretching process gets messed up. It's also very slow and computationally heavy.

The New Way (Depth-Based Estimation):
The authors propose a smarter, more robust method. Instead of stretching the rubber bands, they use a concept called "Data Depth."

Think of "Depth" like a crowd-sourced center. If you have a pile of rocks, the "deepest" rock is the one right in the middle, surrounded by others. The "shallowest" rocks are the ones on the very edge.

The Analogy: Imagine looking at a crowd of people. The person standing in the exact center of the group is the "deepest." If you want to find the "average" person, you don't measure everyone's height and average it (which gets ruined if one person is a giant). Instead, you just pick the person standing in the middle.
The Innovation: The authors show that if you pick the "deepest" curve from your messy data, it is actually the best guess for the true, underlying pattern, even if the data is messy, distorted, or has weird outliers.

The Two Main Challenges They Solved

1. The "One Person, Many Instruments" Problem
In many real-world scenarios (like tracking human growth), one person has multiple measurements (height, weight, arm length). The paper assumes that if you are running late, you are late for your height measurement, your weight measurement, and your arm length measurement all at the same time.

The Solution: The authors developed a way to look at all these different "instruments" together. They proved that if you find the "deepest" curve across all these different measurements, you can mathematically separate the "true shape" from the "timing errors."

2. The "Outlier" Problem
What if one musician is playing a completely different song?

The Solution: Because their method relies on finding the "middle" (the median) rather than the "average," it is naturally resistant to outliers. If 49 musicians play the same song and 1 plays a different one, the "average" method gets confused. The "deepest" method simply ignores the weird one and finds the center of the 49 who are in sync.

The "WHyRA" Tool: The Detective's Magnifying Glass

The authors also created a visual tool called the WHyRA plot (Warping Hypograph Ranking Agreement).

The Analogy: Imagine you have two different maps of the same city drawn by the same person, but one is stretched horizontally and the other vertically. You want to know: "Did the person stretch both maps in the same way?"
How it works: The plot ranks the timing errors of every musician for every instrument. If the rankings match up perfectly (e.g., the musician who was slowest on the drums was also the slowest on the flute), the dots on the graph form a straight line. If the dots are scattered everywhere, it means the assumption that "one person = one timing error" is broken. This helps researchers know if their math is valid before they trust the results.

What the Simulations Showed

The authors tested their method with computer-generated data (simulations) to see how it held up against the old method (called the "CM method").

Speed: Their method was much faster. It didn't need to solve complex stretching puzzles for every single curve.
Accuracy with Noise: When the data was clean and simple, the old method was slightly better.
Accuracy with Chaos: When the data was messy, had lots of timing variations, or included "weird" outliers, the new Depth-Based method was significantly more accurate. It didn't break down when the data got messy.
Robustness: When they intentionally added "bad" data (outliers) to the mix, the old method's results got distorted, but the new method stayed steady, correctly identifying the true pattern.

Summary

This paper introduces a new, faster, and tougher way to find the "true shape" behind messy, multi-dimensional data. Instead of trying to force messy curves to line up perfectly (which fails when data is weird), it uses a "center-finding" technique (Depth) to naturally filter out the noise and find the common pattern. It works especially well when you have multiple types of measurements per person and when your data contains some strange, unusual observations.

Technical Summary: Depth-based estimation for multivariate functional data with phase variability

Problem Statement
The paper addresses the statistical challenge of estimating a common underlying pattern (template function) in multivariate functional data when the data exhibits both individual phase variation (time warping) and cross-component time distortion. In many scientific fields, such as growth analysis or distributional data analysis, individuals are observed across multiple functional components (e.g., different physiological structures or marginal distributions). These observations often suffer from two distinct sources of phase variability:

Individual variability: Each subject has a unique time-warping function.
Component variability: Different components of the multivariate process may be distorted by different deterministic functions.

Existing methods, such as those by Carroll and Müller (2023), typically rely on a registration-based approach: first estimating warping functions by solving minimization problems to align curves, and then estimating the common pattern from the registered sample. However, this sequential process can be computationally intensive and may lack robustness against atypical observations (outliers). Furthermore, standard registration methods often assume specific parametric forms or rely on mean-based estimators, which are sensitive to outliers.

Methodology
The authors propose a robust, registration-free estimation framework based on functional data depth. The approach is grounded in the Latent Deformation Model (LDM) introduced by Carroll and Müller (2023), where an observed multivariate curve $X_{ij}(t)$ for individual $i$ and component $j$ is modeled as:
$X_{ij}(t) = a_{ij} \lambda(\psi_j(h_i(t)))$
Here, $\lambda$ is the unknown common amplitude function (latent curve), $\psi_j$ represents component-based deterministic distortions, $h_i$ represents individual time-warping functions, and $a_{ij}$ are scale factors.

The core of the proposed methodology involves the following steps:

Depth-Based Median Estimation: Instead of minimizing a registration loss, the method identifies the "deepest" curve in the sample as the estimator for the common pattern. The authors establish that if a functional depth measure ($FD$) satisfies two key properties—invariance under strictly monotone transformations and consistency of the sample median to the population median—the deepest curve consistently estimates the target function $\lambda$ (or $\lambda \circ \psi_j$ for specific components) even in the presence of time warping.
Generalization to Multivariate Settings: The method extends the univariate depth-based approach (Arribas-Gil and Romo, 2012) to the multivariate LDM. It allows for the direct estimation of the common pattern $\lambda$ by pooling observations across components, provided the composition of warping processes satisfies specific symmetry conditions (Proposition 3).
Handling Non-Monotone Targets: For cases where the target function $\lambda$ is not strictly monotone, the authors incorporate a "monotonizing operator" $T$ (as proposed by Dupuy et al., 2011) to transform the sample curves before applying the depth-based median, ensuring consistency.
Diagnostic Tool (WHyRA): To validate the critical LDM assumption that individual warping functions $h_i$ are identical across all components, the authors introduce the Warping Hypograph Ranking Agreement (WHyRA) plot. This graphical tool visualizes the rank correspondence of warping function estimates across components using the Modified Hypograph Index (MHI). It helps detect violations of the shared warping assumption.

Key Contributions

Registration-Free Estimation: The paper proposes an alternative to the standard two-step registration-then-estimation pipeline. By using functional depth, the common pattern is estimated directly from the observed, unregistered data, avoiding the need to solve multiple complex minimization problems for warping functions first.
Theoretical Conditions for Consistency: The authors derive necessary conditions for a functional depth measure to provide consistent estimation in time-warping models. They prove that any depth measure satisfying monotonicity invariance and median consistency (such as Modified Band Depth, Integrated Depths, or Quantile Integrated Depths) can be used to consistently estimate the latent curve under the LDM.
Robustness: The depth-based approach is inherently robust to outliers. Unlike mean-based estimators used in registration methods, the depth median is less influenced by atypical trajectories.
Diagnostic Framework: The introduction of the WHyRA plot provides a practical, visual method to assess the validity of the separability assumption (that individual warping is consistent across components), which is crucial for the model's applicability.

Results
The authors evaluate the method through extensive simulations and real data applications:

Simulation Performance: In simulations comparing the depth-based method to the Carroll and Müller (2023) registration method (CM), the depth-based approach demonstrated:
- Superiority under high variability: As time-warping variability and cross-component distortion increased, the depth-based method yielded lower Integrated Square Errors (LISE) for the estimated target function compared to the CM method.
- Computational Efficiency: The depth-based method was significantly faster (e.g., ~0.17s vs. ~11s for $p=4, n=50$ ) because it bypasses the iterative registration process.
- Sensitivity to Noise: The depth-based estimates of the target function and individual curves were found to be more sensitive to additive Gaussian noise than the mean-based CM estimates, a trade-off inherent to robust median-based methods.
- Robustness to Outliers: In simulations with contaminated data (outliers generated from a different amplitude function), the depth-based method maintained better performance in estimating the true common pattern compared to the CM method.
Real Data Applications: The method was applied to two datasets:
1. Arctic Ocean Ice Extent: Analyzing multivariate functional data over time.
2. Maternity-Age Distributions in Europe: Examining distributions across different countries.
  In both cases, the method successfully identified the common patterns and detected atypical observations or deviations from the model assumptions, demonstrating its utility in exploratory data analysis.

Significance and Claims
The paper claims that its primary significance lies in providing a robust, efficient, and theoretically grounded alternative to traditional registration-based methods for multivariate functional data with phase variability. By shifting the focus from minimizing warping errors to maximizing data depth, the authors offer a procedure that:

Avoids the computational burden of solving multiple registration problems.
Provides consistent estimation under broader conditions regarding the warping processes (specifically requiring central symmetry rather than just mean identity).
Offers a built-in mechanism (WHyRA) to diagnose model validity, which is often overlooked in standard registration pipelines.

The authors modestly note that while their method excels in robustness and speed, it may be less efficient than mean-based methods in scenarios dominated by high additive Gaussian noise. They emphasize that their contribution is the extension of depth-based consistency results to the multivariate LDM and the development of a diagnostic tool to support the model's assumptions, rather than a universal replacement for all registration techniques.