Traditional signal processing and time-series analysis rely heavily on a comforting assumption: stationarity. In a stationary world, the fundamental rhythms of a signal remain constant over time. The classic Fourier transform, for instance, decomposes a signal into a collection of infinite, unchanging sine waves. This mathematical framework works beautifully for a steady tuning fork, but it stumbles when applied to the real world.
Real-world signals—from the fluctuating pulse of a patient (Pilz et al. in Local Group Invariance for Heart Rate Estimation From Face Videos in the Wild) to the shifting cycles of financial markets (Liu et al. in T-WaveNet: A Tree-Structured Wavelet Neural Network for Time Series Signal Analysis) or the warped accretion discs surrounding black holes (An et al. in Temporal evolution of long-timescale periodicities in ULX NGC 5408 X-1)—rarely keep a steady beat. Instead, their periodicities drift, accelerate, decay, and evolve.
When a signal's frequency changes within an observation window, static frequency analysis smears its energy across multiple spectral bins. This creates an illusion of high-frequency noise or multiple concurrent frequencies, failing to capture the true underlying trajectory of the system. To solve this, a growing body of research focuses on explicitly modeling drifting periodicity—treating frequency not as a static property, but as a dynamic, time-varying state.
Why Stationary Frameworks Stumble
To understand why new architectures are necessary, we must look at where traditional probabilistic and deep learning models break down when encountering non-stationary frequencies.
Consider Gaussian Processes (GPs), a staple for non-parametric signal estimation. Standard GP kernels assume stationarity, meaning the correlation between any two points depends solely on their relative displacement. As Wilson et al. explain in Gaussian Process Covariance Kernels for Pattern Discovery and Extrapolation, this structural constraint prevents the model from naturally adapting when the underlying frequency or structure evolves over time.
Furthermore, stationary processes mathematically dictate that different frequency components must be uncorrelated. As Tobar demonstrates in Bayesian Nonparametric Spectral Estimation and Altamirano et al. expand upon in Nonstationary multi-output Gaussian processes via harmonizable spectral mixtures, this restriction blinds models to the continuous transitions of a drifting frequency, making it impossible to capture spectral evolution over long durations.
Stationary Assumption:
Fixed Frequency ───► [ Constant Correlation Window ] ───► Sharp Spectrum
Real-World Drift:
Drifting Frequency ──► [ Constant Correlation Window ] ───► Smeared / Blurred Spectrum
In deep learning and time-series forecasting, a similar issue arises from the "fixed-period assumption." Generative models and temporal retrieval mechanisms often utilize a fixed cycle length to look back at historical data or synthesize future steps. If a patient's heart rate changes or an environmental cycle shifts, these rigid windows fall out of sync, leading to tracking errors or synthesis collapse (Lee et al. in Conditional Generation of Periodic Signals with Fourier-Based Decoder; Cao et al. in Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval).
Strategies for Tracking Spectral Evolution
To overcome these limitations, recent methodologies shift the objective from finding a global frequency to tracking the trajectory of a changing frequency. They achieve this through three primary mechanisms: instance-level normalization, flexible coordinate transforms, and state-space tracking.
Instance-Level Normalization and Residuals
Instead of forcing a non-stationary signal into a rigid global frequency grid, frameworks can adaptively normalize signals to isolate the drift. The authors of Ye et al. in Frequency Adaptive Normalization For Non-stationary Time Series Forecasting handle this by adjusting for frequency shifts at the individual instance level.
Complementing this, Lin et al. in CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns decouple the problem by isolating "stationary residuals" from the shifting cyclic patterns. By separating the baseline distribution shifts from the rhythmic changes, the model can track how the underlying period evolves without getting confused by changes in signal intensity or baseline drift.
Flexible and Fractional Transforms
Traditional windowed transforms face an inherent trade-off: a wide window provides precise frequency resolution but blurs temporal changes, while a narrow window tracks time well but loses frequency precision.
When frequencies modulate rapidly, standard transforms break down. To track paths that change faster than a standard window can accommodate, Kakutani et al. in Improvement of wavelet-synchrosqueezing transform with time shifted angular frequency introduced time-domain modifications to the Wavelet-Synchrosqueezing Transform, bypassing complex frequency derivatives to maintain a sharp lock on rapid modulations.
For complex network data, such as traffic or neural signals on a graph, Li in Sampling of Graph Signals Based on Joint Time-Vertex Fractional Fourier Transform introduces fractional-order flexibility into joint time-vertex transforms. By allowing the mathematical coordinate system itself to rotate fractionally, the transform can compactly represent "chirp-like" signals—patterns where the frequency accelerates or decelerates linearly over time.
Standard Transform: Fixed Window ──► Cannot capture rapid acceleration
Fractional Transform: Rotated Basis ──► Aligns with the acceleration path (Chirp)
Evolving Frequencies as State Variables
Another powerful intuition is to treat periodic components as dynamic systems rather than static functions. By representing periodic covariance functions as a series of stochastic resonators, the authors of Solin et al. in Explicit Link Between Periodic Covariance Functions and State Space Models show how to maintain probabilistic tracking of quasi-periodic signals.
This formulation allows a model to continuously update its estimate of the fundamental frequency as new data streams in. Instead of computing a massive, static kernel over the entire history, the model treats frequency as an evolving state variable within a state-space model, allowing it to adapt linearly to non-stationary drifts.
Turning Noise into Trajectory
Ultimately, these approaches succeed because they change how models perceive frequency variations. In a rigid, stationary framework, a drifting frequency looks like noise, chaos, or structural breakdown. By implementing learnable, invertible neural networks that dynamically adjust frequency subbands (Liu et al. in T-WaveNet) or recurrent memory states that explicitly track frequency components over time (Qi et al. in State-Frequency Memory Recurrent Neural Networks), modern architectures embrace the fluidity of real-world signals.
By modeling how frequencies drift rather than assuming they stay still, these frameworks turn non-stationarity from a source of error into a predictable, informative trajectory.
Key Foundational Work
Insurance claims modulated by a hidden Brownian marked point process (2009) — contributed 1 concept elements to this direction.
Online Detection of Fire in Video (2007) — contributed 1 concept elements to this direction.