Decoupling Phase and Amplitude for Functional Data Registration

This direction is merged into another direction: temporal misalignment (functional data misalignment). Its last version stays readable and citable here.

Version 1 of 133 papers40 conceptsfrozen 2026-10-04

In traditional statistics, we treat observations as discrete points. However, in many modern domains—from tracking infant growth to monitoring heart rhythms—the data is better understood as continuous curves or "functions." This is the domain of Functional Data Analysis (FDA).

The central challenge in FDA is that curves vary in two distinct ways: how "high" their features are (amplitude) and "when" those features occur (phase). If you simply average a set of misaligned curves, the resulting "mean" curve becomes a distorted, blurry mess where sharp peaks are smoothed out. This phenomenon, often called "mushing," fails to represent any actual individual in the population. To solve this, researchers use registration (or alignment) to decouple timing from magnitude.

The Mushing Problem and the Need for Registration

When analyzing a population of curves, such as the growth trajectories studied by Pantelis Z. Hadjipantelis and Michael S. Kramer, simple cross-sectional averaging fails. As Hans-Georg Müller and Jeng‐Min Chiou explain in Functional Data Analysis, horizontal shifts (phase variation) distort the aggregate representation.

Imagine two heartbeats that are identical in shape but shifted by half a second. A pointwise average would show two small bumps instead of one strong pulse. Chloé Dimeglio and Jean-Michel Loubes argue in Manifold embedding for curve registration that we must instead find a "structural median"—a representative curve that preserves essential features like peaks and valleys by accounting for these temporal shifts.

Decoupling Phase and Amplitude

The core mechanism for fixing this is to mathematically separate the signal into two components:

  1. Amplitude Variability: Differences in the intensity or magnitude of features (e.g., how tall a growth spurt is).
  2. Phase Variability: Differences in the timing or speed of the underlying process (e.g., when the growth spurt occurs).

James Matuk, Sebastian Kurtek, and Oksana Chkrebtii have developed several frameworks, including Geometric Empirical Bayesian Models, to perform this decoupling. By treating time-warping as a nuisance variable, they can align curves to a common template, allowing for more accurate classification and uncertainty quantification.

The Elastic Framework and SRVF

A breakthrough in making registration mathematically rigorous is the Elastic Functional Data Analysis framework. This approach uses the Square-Root Velocity Function (SRVF) to represent curves. As Anuj Srivastava and Sebastian Kurtek demonstrate in their work on AneuRisk65 data, the SRVF transforms the problem of non-linear time warping into a geometry problem.

In this space, the distance between two warped curves can be calculated as a geodesic (the shortest path) on a manifold. This "elastic" metric is invariant to how a curve is parameterized, meaning it doesn't matter if one subject "moves" through the process faster than another; the underlying shape remains the same.

Joint Alignment and Learning

Historically, researchers used a "tandem" approach: first align the curves, then cluster or classify them. However, Andrew Parnell notes in his Review of Clustering Methods that this is often flawed because the alignment depends on the cluster, and the cluster depends on the alignment.

To solve this, Padhraic Smyth and Scott Gaffney pioneered Joint Probabilistic Curve Clustering and Alignment, using the Expectation-Maximization (EM) algorithm to solve both simultaneously. More recently, this "synergistic" approach has moved into deep learning. Chao Chen and Anuj Srivastava introduced SrvfRegNet, which uses neural networks to perform unsupervised registration orders of magnitude faster than traditional dynamic programming. Similarly, Siyuan Jiang developed DeepFRC, the first end-to-end framework to jointly register and classify functional data.

Why It Works: Intuition from Manifolds

The intuition behind the most successful methods is to stop treating curves as a sequence of points and start treating each curve as a single point on a high-dimensional manifold.

In Nonlinear manifold representations for functional data, Dong Chen explains that while curves might look complex and non-linear in Euclidean space, they often lie on a low-dimensional manifold. By "unwrapping" this manifold, we can find the true Fréchet mean—the point on the manifold that is centrally located relative to all others. This mean preserves the structural integrity of the data because it respects the geometric constraints of the curves rather than just averaging their coordinates.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_decoupling_phase_and_amplitude_for_funct_a84167f7,
  title = {Decoupling Phase and Amplitude for Functional Data Registration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decoupling-phase-and-amplitude-for-functional-data-registration/rdr_a84167f7119e4b2e867caee10ee067dc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_decoupling_phase_and_amplitude_for_funct_a84167f7_v1,
  title = {Decoupling Phase and Amplitude for Functional Data Registration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decoupling-phase-and-amplitude-for-functional-data-registration/rdr_a84167f7119e4b2e867caee10ee067dc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}