Dimensionality Reduction for Functional Data

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2151 papers175 conceptsfrozen 2026-10-04

In classical statistics, an observation is typically a point or a vector—a fixed set of numbers like a person’s height, weight, and age. But in many modern fields, an "observation" is actually a continuous entity: a heart rate monitor’s trace over 24 hours, the 3D trajectory of a robotic arm, or a series of brain activity maps from an fMRI.

These are examples of functional data. Mathematically, these objects exist in infinite-dimensional space. If you tried to treat every millisecond of a sensor log as a separate variable, your model would immediately collapse under the "curse of dimensionality." To make this data tractable, researchers use a suite of techniques to project these infinite-dimensional curves into a compact, finite-dimensional space of "scores" that capture the essence of the signal.

The Curve as a Single Unit

The fundamental shift in this research direction is treating a time-series or a spatial scan not as a sequence of independent snapshots, but as a single functional unit. As Giuseppe Averta and Edoardo Battaglia demonstrate in Unvealing the Principal Modes of Human Upper Limb Movements through Functional Analysis, analyzing movement as a continuous function allows us to see coordination patterns—"synergies"—that are invisible when looking at discrete time points.

To work with these functions, we need a way to represent them using a few numbers. This is typically achieved through basis expansion. As explained by Sylvain Calinon in Mixture Models for the Analysis, Edition, and Synthesis of Continuous Time Series, we can view a complex signal as a weighted sum of simpler "building block" functions, such as Fourier series for periodic data or B-splines for smooth trajectories.

Functional Principal Component Analysis

While fixed bases (like wavelets) are useful, the most powerful tool in this area is Functional Principal Component Analysis (FPCA). Unlike standard PCA, which finds the directions of greatest variance in a cloud of points, FPCA finds the "modes of variation" in a collection of curves.

Think of it this way: if you are looking at the growth curves of children, the first "functional principal component" might represent the overall average height (a vertical shift), while the second might represent the timing of the growth spurt (a horizontal stretch or "warping").

Ulrich Stadtmüller shows in Generalized functional linear models that by representing a predictor function using a truncated expansion of its most significant eigenfunctions, we can perform regression on infinite-dimensional data. This allows us to predict a single outcome (like a disease diagnosis) from an entire curve (like a patient's glucose history).

Bridging the Gap Between Discrete and Continuous

In practice, we never actually observe a "continuous curve." We observe discrete, often noisy, and irregularly timed data points. A major contribution of this field is the ability to reconstruct the underlying smooth function from these "sparse and noisy" observations.

Jie Peng addresses this in Distance-based clustering of sparsely observed stochastic processes, with applications to online auctions. By using Mercer’s theorem to decompose trajectories into a mean function and eigenfunctions, researchers can cluster behaviors (like bidding strategies) even when they only have a few data points per subject. Similarly, Ci‐Ren Jiang, Jane-Ling Wang, and John A. D. Aston demonstrate in A functional approach to deconvolve dynamic neuroimaging data that FPCA can act as a powerful denoiser, filtering out localized noise by focusing on the global "basis shapes" of brain activity.

Handling Complexity: Warping and Multi-level Data

Real-world functional data often contains variations in both amplitude (how intense a signal is) and phase (when a signal happens). If two people perform the same gesture but at different speeds, standard PCA might see them as completely different.

To solve this, researchers like Rushil Anirudh and Pavan K. Turaga use "elastic" functional analysis. In Elastic functional coding of human actions: From vector-fields to latent variables, they map trajectories onto Riemannian manifolds to achieve a 250x reduction in dimensionality that is invariant to timing differences.

Furthermore, data is often "repeated"—for example, a patient might have a 24-hour heart rate curve recorded every day for a month. Poorbita Kundu and colleagues in fdarep: Two-Dimensional FPCA, Marginal FPCA, and Product FPCA for Repeated Functional Data provide frameworks to decompose this variation into "within-day" and "between-day" components, allowing for a multi-level understanding of temporal dynamics.

Why It Works: The Low-Dimensional Hypothesis

The reason these methods are so successful across diverse fields—from predicting brain states to modeling COVID-19 trajectories—is the observation that complex natural phenomena often follow a low-dimensional structure.

As Maurice Berk notes in Functional modelling of microarray time series with covariate curves, even the expression of thousands of genes over time can often be explained by a few fundamental shapes, like linear trends or quadratic humps. By identifying these characteristic temporal signatures, researchers can compress massive datasets into a handful of interpretable scores without losing the essential "story" told by the data.

Papers in this version

Showing 60 of 151 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 116 papers
  2. Version 1 · 2026-10-04 · 151 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_functional_principal_component_analysis_f0e789be,
  title = {functional principal component analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/functional-principal-component-analysis/rdr_f0e789be35fd455f8c2aa86f3a72d436},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dimensionality_reduction_for_functional__f0e789be_v1,
  title = {Dimensionality Reduction for Functional Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dimensionality-reduction-for-functional-data/rdr_f0e789be35fd455f8c2aa86f3a72d436/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}