Separating Phase and Amplitude in Functional Data Analysis

73 concepts52 papers1 proposal

Imagine you have ten recordings of a human heartbeat or a specific gene expression cycle. Every recording captures the exact same underlying physiological event—a sharp spike followed by a gradual decline. However, because the subjects' biological clocks run at slightly different speeds or start at different times, the peaks do not line up perfectly.

If you take the standard mathematical average (the cross-sectional mean) of these curves at each time point, a striking artifact occurs: the sharp peaks get smeared out into a flat, smooth hill. The resulting "average" curve looks nothing like any individual observation in your dataset. As noted by Hans-Georg Müller et al. in Functional Data Analysis, this temporal misalignment—known as phase variation—distorts aggregate representations and frequently leads to misleading statistical conclusions.

The core challenge is that traditional metrics compare data points at fixed time indices. When signals exhibit nonlinear temporal distortions, rigid point-to-point comparisons treat horizontal shifts (changes in timing) as vertical differences (changes in structural shape). To build accurate models, we must separate what the shape is (amplitude variability) from when it occurs (phase variability).

The Geometry of Elastic Registration

To mathematically disentangle phase and amplitude, we need a way to align the time axes of different curves. This process is known as curve registration or elastic alignment. Conceptually, we want to find a "warping function"—a smooth, strictly increasing function $\gamma(t)$ that maps the time axis of one curve to another, stretching and compressing time until their structural features line up.

However, simply minimizing the standard $L^2$ distance between a warped curve and a template introduces a severe mathematical flaw: it is not symmetric. Warping curve $A$ to match curve $B$ yields a different distance than warping curve $B$ to match curve $A$, and it can lead to degenerate solutions where an entire segment of time is compressed into a single point.

To solve this, researchers turned to differential geometry. Sebastian Kurtek et al. in Signal Estimation Under Random Time-Warpings and Nonlinear Signal Alignment utilized the Fisher-Rao Riemannian metric combined with the Square-Root Velocity Function (SRVF) framework. The SRVF transforms a curve $f(t)$ into its velocity-based representation:

$$q(t) = \frac{\dot{f}(t)}{\sqrt{|\dot{f}(t)|}}$$

Under this transformation, the mathematically complex Fisher-Rao metric simplifies beautifully into the standard $L^2$ distance, and the action of time-warping becomes an isometry (preserving distances). This allows researchers to compute a true "structural expectation" or Fréchet mean that preserves sharp features like peaks and valleys, rather than blurring them away.

Alternative non-parametric views also address this challenge. For instance, Jean-François Dupuy et al. in Non parametric estimation of the structural expectation of a stochastic increasing function proposed estimating a structural expectation without relying on an arbitrary reference template, avoiding template-induced bias. From a manifold perspective, Dong Chen in Nonlinear manifold representations for functional data demonstrated that functional data with time-warping lies on a nonlinear manifold. This means that linear averaging fails inherently, and manifold learning is required to "unwrap" the data into a parsimonious Euclidean space where standard statistics can safely apply.

Amortized and Generative Alignment

While geometric frameworks like SRVF provide rigorous mathematical guarantees, traditional optimization via dynamic programming is computationally expensive, especially when scaling to large datasets. Modern machine learning addresses this by embedding phase-amplitude separation directly into generative neural networks and probabilistic models.

Instead of optimizing a warping function iteratively for every single pair of curves, deep networks can learn to predict the warping functions in a single forward pass. For example, Elvis Nunez et al. introduced SrvfNet in SrvfNet: A Generative Network for Unsupervised Multiple Diffeomorphic Functional Alignment, which combines the mathematical rigor of the SRVF framework with a Variational Autoencoder (VAE) architecture. This allows the model to jointly discover an optimal population template and align large collections of functional data up to 150 times faster than traditional methods.

Integrating alignment into the latent space also unlocks powerful unsupervised clustering capabilities. If a model is sensitive to phase shifts, it might group two identical signals into different clusters simply because one started a second later. To prevent this, Kaspar Märtens et al. developed BasisVAE in BasisVAE: Translation-invariant feature-level clustering with Variational Autoencoders, embedding translation and scale invariance directly into the VAE decoder. This ensures that features sharing the same underlying functional shape are clustered together regardless of temporal offsets.

Extending this to complex multi-dimensional functional data, Samuel V. Singh et al. proposed FAEclust in Shape-Informed Clustering of Multi-Dimensional Functional Data via Deep Functional Autoencoders, which uses a shape-informed clustering objective to remain robust against phase variations where traditional linear methods fail.

Why Decoupling Matters in Practice

Why go through the trouble of separating phase and amplitude? When these two sources of variability are confounded, downstream statistical tasks degrade significantly. In classification, a model might misinterpret a simple delay in a heartbeat as a pathological structural deformity.

By explicitly decoupling these components, models achieve higher accuracy and robustness. For instance, James Matuk et al. demonstrated in Geometric Empirical Bayesian Model for Classification of Functional Data Under Diverse Sampling Regimes that separating amplitude and phase variability prevents the misinterpretation of temporal shifts, significantly boosting classification accuracy even when curves are sparsely or irregularly observed.

Whether analyzing biological gene expression, EEG brainwaves, or cardiac motion, treating temporal misalignment not as noise, but as a distinct geometric property to be isolated, is the key to unlocking the true structure hidden within functional data.

Key Foundational Work

Probabilistic Models For Joint Clustering And Time-Warping Of Multidimensional Curves (2003) — contributed 1 concept elements to this direction.
Gene Expression Clustering with Functional Mixture Models (2003) — contributed 1 concept elements to this direction.
Bayesian Detection of Infrequent Differences in Sets of Time Series with Shared Structure (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_misalignment_functional_data_mi_86105ca3,
  title = {temporal misalignment (functional data misalignment)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-misalignment-functional-data-misalignment/rdr_86105ca358ba4fcfb2582d062416a409},
  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_temporal_misalignment_functional_data_mi_86105ca3_v1,
  title = {temporal misalignment (functional data misalignment)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-misalignment-functional-data-misalignment/rdr_86105ca358ba4fcfb2582d062416a409/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake