Traditional deep learning models treat time series data as discrete sequences or grids—ordered arrays of values indexed by fixed time steps ($x_1, x_2, \dots, x_T$). While this formulation aligns well with standard architectures like Recurrent Neural Networks (RNNs) or Transformers, it struggles with the messiness of real-world data. When observations are irregularly sampled, sensors drop out, or forecasting horizons change, discrete models require complex preprocessing, alignment, or imputation steps that often degrade the underlying signal.
A growing body of research addresses these limitations by shifting the paradigm: treating time series not as discrete sequences, but as continuous functions of time. By mapping temporal coordinates directly to signal values, this functional approach provides a unified, mathematically elegant framework that natively handles irregular sampling, missing data, and arbitrary resolutions.
The Core Intuition: From Grids to Curves
To understand why this works, consider how a standard model processes a missing value in a multivariate time series. If a sensor fails to record a value at $t=3$, a discrete model faces a broken grid. It must either fill the gap using a heuristic (like forward-filling) or deploy a separate imputation model before it can proceed with forecasting.
In contrast, the functional paradigm models the underlying data-generating process as a continuous curve, $f(t)$. Under this view, observing a time series means sampling from $f(t)$ at a specific set of timestamps. Imputation and forecasting cease to be distinct algorithmic challenges; instead, they collapse into a single task: querying the function at arbitrary coordinates.
Discrete View: [x_1] ---> [x_2] ---> [???] ---> [x_4] (Broken Grid)
Functional View: f(t) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ (Continuous Curve)
^ ^ ^
t=1 t=2 t=3 (Query anywhere!)
Evaluating the function inside the range of observed timestamps performs imputation; evaluating it beyond the observed range performs forecasting. Because the model parameterizes the curve itself rather than the transitions between steps, gaps in the data do not disrupt the temporal flow.
Parameterizing the Continuum
How do we represent an infinite-dimensional continuous function using a finite set of parameters?
Historically, Functional Data Analysis (FDA) relied on linear combinations of fixed basis functions, such as B-splines or Fourier series. For instance, in speech and behavioral analysis, the authors of Analyzing Temporal Dynamics of Dyadic Synchrony in Affective Interactions used B-spline interpolation to transform discrete interaction metrics into continuous curves, allowing them to study the shape of behavioral variability over time. Similarly, Kohei Hayashi incorporated Fourier and wavelet bases into tensor decompositions in Tensor Decomposition with Smoothness to inherently enforce spatial and temporal continuity.
Modern deep learning scales this concept by using Implicit Neural Representations (INRs). An INR parameterizes a continuous signal using a Multi-Layer Perceptron (MLP) that maps a continuous coordinate (such as time $t$ or space-time $[x, y, z, t]$) directly to its corresponding value:
$$f_\theta(t) = \text{MLP}(t; \theta)$$
To generalize across thousands of different time series, researchers leverage meta-learning or hypernetworks. In frameworks like TimeFlow, developed by Etienne Le Naour et al., and iHyperTime, developed by Elizabeth Fons et al., a context encoder first processes the available sparse observations of a sequence to produce a latent code $z$. This code $z$ then parameterizes or conditions the INR, yielding an instance-specific continuous function $f(t; z)$.
Why It Works: Native Handling of Irregularity
By decoupling the model's internal representation from the sampling rate of the data, functional models bypass the constraints of fixed look-ahead windows and rigid grids.
Constant Memory and Computational Efficiency
In Learning Deep Time-index Models for Time Series Forecasting, Gerald Woo et al. introduced DeepTime, a meta-optimization framework that treats forecasting as a continuous mapping task. By pairing deep time-index representations with an efficient ridge-regression inner loop, DeepTime achieves state-of-the-art forecasting accuracy. Crucially, because it queries time indices directly rather than autoregressively stepping through a sequence, it operates with constant memory complexity across arbitrary time horizons, vastly outperforming Transformer-based alternatives on long-context tasks.
Zero-Shot Generalization Across Frequencies
Because time is treated as a continuous coordinate, a model trained on hourly data can be queried at minutely intervals or across irregular gaps without architectural modifications. Luke Nicholas Darlow et al. leveraged this property in DAM (Deep Data-dependant Approximate analytical Model), demonstrating a foundation model for forecasting. By framing forecasting as continuous function approximation, a single DAM model can generalize zero-shot across diverse datasets, remaining completely robust to missing values and varying sampling rates.
Robustness to Extreme Sparsity
When data is highly incomplete, discrete models fail because they cannot reliably learn transition matrices. Functional representations overcome this by optimizing for global curve shapes. Christian Klötergens et al. designed an attention-based encoder that maps irregularly sampled observations directly to the parameters of a continuous latent function, focusing entirely on estimating the global coefficients of the underlying temporal curve.
This global optimization allows models to withstand extreme data degradation. For example, Mengxuan Li 0003 et al. introduced ImputeINR, demonstrating that modeling time series as continuous functions preserves structural integrity and diagnostic utility even when up to 90% of the data points are missing.
Preserving Continuity Throughout Deep Networks
The benefits of functional representations extend beyond input processing. Recent work demonstrates that maintaining data in a continuous state throughout a neural network's hidden layers prevents the information loss associated with discretization.
Aniruddha Rajendra Rao et al. developed architectures with continuous hidden layers, allowing networks to map functions directly to other functions (function-on-function regression) without flattening curves into scalar vectors (Aniruddha Rajendra Rao et al.).
This preservation of mathematical continuity is vital for capturing higher-order dynamics. As Boxin Zhao demonstrated in Direct Estimation of Differential Functional Graphical Models, modeling time series as continuous functional objects allows for the extraction of derivatives (such as velocity and acceleration of change). These continuous dynamics make it possible to recover complex structural dependencies and graph differences in high-dimensional biological and neural systems, capturing subtle temporal variations that discrete models completely smooth away.