Must the outcome be a curve?

No. Scalar-on-function regression uses a scalar outcome and a functional predictor.

Are the curves observed continuously?

Usually not; they are measured on discrete grids and represented using suitable methods.

Does a coefficient curve identify causal time windows?

No. It describes the fitted association under the model, with uncertainty and identification limits.

AI Glossary / Statistics and probability / Advanced

Functional regression

Definition

Functional regression models relationships in which an outcome, a predictor or both are functions, such as curves observed over time, wavelength or space. A function is treated as a data object, even though measurements are usually recorded at a finite set of locations. Common forms are scalar-on-function regression, function-on-scalar regression and function-on-function regression. The term concerns functional data analysis; it does not merely mean that a conventional regression uses a mathematical function or that its software follows a functional-programming style.

What it means

A curve can contain useful information about timing, shape and local variation that a single summary loses. Two daily activity profiles may have the same total but different patterns across the day. Functional regression can relate those patterns to an outcome or predict an entire response curve. Observations along a curve are usually dependent and do not constitute independent subjects. The statistical challenge is to learn relationships across multiple curves while controlling the complexity of variation within each curve and accounting for measurement quality.

How it works

Define the functional domain and how curves are observed, aligned and represented. Basis functions, smoothing penalties or functional principal components can reduce dimensionality and regularize estimation. In a scalar-on-function linear model, E[Y given X] = alpha + integral beta(t)X(t) dt, where beta(t) weights the predictor curve across its domain. Function-on-scalar models let regression coefficients vary over the response domain, while function-on-function models relate an input curve to an output curve, often through a coefficient surface. Select complexity using validation appropriate to the independent units. Irregular sampling, noisy measurements, registration and missing curve segments may require specialized treatment.

Three practical examples

Manufacturing: relate a temperature trajectory during a run to a scalar quality measurement for that run.
Energy analysis: predict a daily demand curve from weather and calendar predictors rather than forecasting only a daily total.
Spectral analysis: use a measured spectrum as a functional predictor of a material property while controlling the complexity of wavelength-specific effects.

Think of it this way

Think of comparing melodies rather than only their average loudness. The order and shape of the notes can matter. Functional regression aims to connect that structured pattern with another measurement, while recognizing that many notes belong to one melody rather than to many independent examples.

Worked example

Consider a stipulated scalar-on-function model over t from zero to one with intercept alpha = one and coefficient function beta(t) = 3t. For predictor curve X1(t) = 2t, the weighted integral is integral 6t squared dt = two, so the predicted mean outcome is three. For X2(t) = 2(1 - t), the weighted integral is integral [6t - 6t squared] dt = three minus two = one, giving predicted mean two. Both curves have the same unweighted integral of one, yet the model gives different predictions because beta(t) places greater weight on later parts of the domain. This calculation uses exact hypothetical functions and specified coefficients. It does not show that such a coefficient function can be estimated reliably from a small dataset, nor does it establish a causal effect of changing the curve.

Illustrative case study

Fictional case: Cypress Process Analytics records a sensor curve for each production run. A simple average obscures a late-run pattern associated with the measured output, so analysts compare a regularized functional model with simpler summaries. They split validation data by complete runs rather than scattering points from one run across training and testing. They also check whether sensor alignment creates misleading patterns. This invented case illustrates a possible modeling workflow without claiming that functional regression necessarily outperforms a simpler model.

Common misconceptions and mistakes

Do not count every time point as an independent experimental unit. Dense measurement of a few curves does not replace replication across subjects or runs. Smoothing can remove real features as well as noise, and misaligned curves can mix timing variation with amplitude variation. Coefficient functions may be unstable where predictor curves provide little independent variation. A visually prominent coefficient region is not automatically statistically or causally important. Using portions of a curve that would not be available at prediction time creates leakage. Ordinary pointwise intervals do not automatically provide simultaneous coverage over the entire domain.

Business use

Businesses can use functional regression when the shape of a trajectory or profile is central to the decision. Define the independent unit, prediction time and available portion of each curve before modeling. Compare against meaningful simple summaries and evaluate on new units under realistic collection conditions. Document preprocessing, alignment, basis selection and regularization because these choices shape interpretation. Present predictions and uncertainty in a form that reflects the functional domain. When a model is used prospectively, ensure its inputs are genuinely observable at the decision time rather than relying on later measurements.

Check your understanding

Two predictor curves have equal totals but different shapes. Must a functional linear model predict the same outcome?

Answer: No. A nonconstant coefficient function weights different parts of the curves differently, so equal unweighted totals need not imply equal weighted integrals.

Related terms

Smoothing
Regularization
Functional data analysis
scalar-on-function regression
function-on-scalar regression
function-on-function regression
basis expansion
functional principal components
curve registration

Learn next

Learn how curves are represented with bases and penalties, then compare the three main functional-regression settings. Study validation by independent unit and the distinction between pointwise and simultaneous uncertainty before interpreting coefficient curves.

Sources

Jeffrey S. Morris, Functional Regression: https://arxiv.org/abs/1406.4068 Morris, institutional publication record: https://mdanderson.elsevierpure.com/en/publications/functional-regression/ Reiss and colleagues, Methods for scalar-on-function regression, primary article record: https://pubmed.ncbi.nlm.nih.gov/28919663/

Previous: Functional presence engine Next: Fundamental frequency