Adaptive Local Iterative Filtering (ALIF)

Updated 13 May 2026

Adaptive Local Iterative Filtering (ALIF) is an adaptive, fully local decomposition technique that processes nonstationary and nonlinear signals by using data-driven, spatially variable filters.
It iteratively extracts intrinsic mode functions via local moving averages and a sifting process, enabling separation of components with rapid intra-component frequency variations.
Stabilized and resampled variants like SALIF and FRIF improve convergence and computational efficiency while providing robust time–frequency representations for applications such as chirp decomposition and noise reduction.

Adaptive Local Iterative Filtering (ALIF) is an adaptive, fully local decomposition technique designed for the processing of nonstationary and nonlinear signals. It extends the iterative filtering (IF) framework by employing adaptive, data-driven filters whose support and profile can vary with the signal, enabling time–frequency analysis capable of separating components with rapid intra-component frequency variation and handling strong amplitude/frequency modulation. The ALIF methodology has been rigorously examined for spectral properties, convergence, and has inspired stabilized and resampled variants that address its theoretical and practical limitations, particularly regarding convergence and computational complexity (Cicone et al., 2014, Barbarino et al., 2021, Cicone et al., 2020, Barbarino et al., 2020).

1. Mathematical Foundations of ALIF

The ALIF approach generalizes the IF method by introducing a spatially variable filter. For a real signal $g:[0,1] \to \mathbb{R}$ , the algorithm constructs at each position $x$ a local, compactly supported filter $k_x$ with support $[-\ell(x),\ell(x)]$ , where the length function $\ell(x)$ is extracted in a data-driven fashion from the local spacing of extrema. The normalized filter is given by: $k_x(z) = \frac{1}{\ell(x)} k\left( \frac{z}{\ell(x)} \right),$ where $k$ is even, nonnegative, and unit-mass.

The iterative sifting process involves:

Computing the local moving average

$L(g)(x) = \int_{-\ell(x)}^{\ell(x)} g(x+t) \frac{1}{\ell(x)} k\left( \frac{t}{\ell(x)} \right) dt,$

Subtracting the average to extract oscillatory content: $S(g)(x) = g(x) - L(g)(x),$
Iteratively applying $S$ : $x$ 0, $x$ 1, with the first intrinsic mode function (IMF) defined as $x$ 2.

Subsequent IMFs are extracted recursively from residuals until only a monotonic trend remains (Cicone et al., 2014, Barbarino et al., 2021, Barbarino et al., 2020).

In discrete settings, sampling at $x$ 3 leads to a filter matrix $x$ 4 with entries: $x$ 5 and the inner sifting loop becomes the iteration $x$ 6. Stopping criteria are based on convergence tolerances or maximum iterations.

2. Adaptive Filter Construction and Local Parameterization

The choice of the variable filter length $x$ 7 is central. It is derived by:

Locating extrema of the current signal,
Constructing an interpolant $x$ 8 to the local distances between extrema,
Optionally refining $x$ 9 by subtracting high-frequency components from $k_x$ 0 (using standard IF) to enforce slow variation and frequency selectivity (Cicone et al., 2014).

The actual filter kernel may be a Fokker–Planck (FP) filter, constructed as the steady-state solution of a 1D Fokker–Planck PDE, leading to smooth, compactly supported and positive-distribution functions. This allows ALIF to meet sufficient conditions for filter admissibility and convergence in the underlying IF theory (Cicone et al., 2014).

3. Theoretical Analysis and Spectral Properties

The principal challenge in ALIF analysis stems from the nonstationarity and spatial variability of its filtering operator: the convolution becomes a non-translation-invariant, non-Hermitian, and non-Toeplitz operation. The induced filter matrices $k_x$ 1 are best described as generalized locally Toeplitz (GLT) matrices.

Spectral analysis has established that for convergence of the inner iteration, a necessary condition is that the spectrum of $k_x$ 2 lies within the unit disk, i.e., for $k_x$ 3, $k_x$ 4, or equivalently $k_x$ 5. However, counterexamples demonstrate that this is not a sufficient condition for convergence: negative or complex eigenvalues may arise when $k_x$ 6 varies rapidly, potentially causing divergence or oscillation in the sifting process (Barbarino et al., 2020, Barbarino et al., 2021).

GLT-sequence theory rigorously links the large- $k_x$ 7 spectral distribution of the discrete ALIF matrices to parametrized symbols $k_x$ 8, providing a precise description of the limiting eigenvalue distribution for regular enough $k_x$ 9 and $[-\ell(x),\ell(x)]$ 0. However, general, verifiable sufficient conditions for convergence for all signals and parameterizations are still not available (Barbarino et al., 2020).

4. Stabilized and Resampled Variants

Addressing ALIF's convergence limitations, two stabilized variants have been formulated:

Stable Adaptive Local Iterative Filtering (SALIF) implements an iteration of the form

$[-\ell(x),\ell(x)]$ 1

where $[-\ell(x),\ell(x)]$ 2 is positive semidefinite. If $[-\ell(x),\ell(x)]$ 3, SALIF guarantees eigenvalues in $[-\ell(x),\ell(x)]$ 4; thus, every inner sifting sequence converges. This method, however, requires higher-complexity operations (two dense matrix multiples per step) (Barbarino et al., 2021).

Resampled Iterative Filtering (RIF) and its Fast variant (FRIF) use a resampling strategy: they reparametrize the signal, flattening the nonuniform time–frequency structure so that standard (stationary) iterative filtering can be applied. In the FRIF approach, the computations further exploit the circulant structure of the resulting filter matrices in the frequency domain, achieving significant computational acceleration via the FFT. Convergence is guaranteed provided the filtering kernel has a nonnegative Fourier transform (Barbarino et al., 2021).

Method	Convergence Guarantee	Per-IMF Complexity
ALIF	Not in general	$[-\ell(x),\ell(x)]$ 5
SALIF	Always (if $[-\ell(x),\ell(x)]$ 6)	$[-\ell(x),\ell(x)]$ 7
FRIF	Always	$[-\ell(x),\ell(x)]$ 8

5. Practical Performance and Applications

Numerical studies consistently indicate that ALIF and its variants can successfully decompose highly nonstationary signals, particularly separating components with rapidly varying instantaneous frequencies (e.g., chirps, whistles) and addressing mode-mixing issues that challenge EMD and stationary IF (Cicone et al., 2014, Barbarino et al., 2021).

Key empirical findings include:

Chirp Decomposition: FRIF achieves higher accuracy and much faster computation compared to ALIF and SALIF in separating close frequency chirps.
Noise Robustness: FRIF and ALIF allocate noise into the initial IMFs, demonstrating effective denoising properties even at very low SNRs.
Real Data (e.g., bat echolocation): FRIF accurately isolates rapid frequency-modulation events, with extracted IMFs matching instantaneous frequency tracks in the time–frequency plane (Barbarino et al., 2021).

Instantaneous frequency estimation in ALIF is performed using envelope-based local methods, entirely sidestepping the Hilbert transform, which provides faithful, positive-definite frequency estimates for each IMF (Cicone et al., 2014).

6. Further Developments: Convergence Analysis and SIFT Extension

ALIF convergence theory has been rigorously developed for certain function classes. For adaptive harmonic models built from separated intrinsic mode type (IMT) functions, iteration dynamics and error propagation can be explicitly controlled under slow amplitude and phase variation assumptions. Freezing arguments and Gaussian kernel behaviors yield geometric convergence to targeted IMFs for adequately parameterized window lengths (Cicone et al., 2020).

The SIFT (Synchrosqueezing Iterative Filtering Technique) method augments ALIF with short-time synchrosqueezing to provide robust, adaptive local frequency estimation for kernel bandwidth selection—demonstrating enhanced separation and time–frequency representation in challenging low SNR, highly nonstationary regimes (Cicone et al., 2020).

7. Open Problems and Outlook

While ALIF provides a powerful and versatile framework for fully adaptive, local signal decomposition, several key questions remain:

General Convergence: No general, verifiable criterion for convergence of discrete ALIF exists; particularly, strong inhomogeneity in $[-\ell(x),\ell(x)]$ 9 can induce spectral instabilities not apparent from limiting GLT symbols (Barbarino et al., 2020).
Optimal Parameter Selection: Adaptive procedures for $\ell(x)$ 0 remain heuristic in practical deployment; rigorous links between local signal characteristics and parameter choices are needed.
Computational Acceleration: The FRIF method demonstrates that leveraging structure for FFT-based acceleration is both effective and mathematically principled. A plausible implication is that future research may further unify adaptive decomposition schemes with efficient numerical linear algebra, and deeper spectral theory may be required to close the convergence questions (Barbarino et al., 2021, Barbarino et al., 2020).