Demixing Signals from Nonlinear Observations

19 concepts17 papers2 proposals

Most classical signal processing assumes that the world is additive. In the standard model of Blind Source Separation (BSS), we observe a mixture $x$ that is simply a linear combination of independent sources $s$, expressed as $x = As$. However, as Tobias Blaschke et al. point out, real-world sensors—from brain imaging hardware to digital cameras—rarely behave so simply. They introduce nonlinear distortions like saturation, quantization, and complex physical interactions that break the fundamental assumptions of linear Independent Component Analysis (ICA).

The Challenge of Nonlinearity

When a mixture is nonlinear, the geometry of the data changes. In a linear mixture, the sources are merely rotated and scaled; in a nonlinear one, the coordinate system itself is warped. Harmeling et al. explain that this creates a "curse of dimensionality" where the data lives on a complex submanifold in feature space. Standard linear algorithms cannot "see" the original sources because they only look for straight-line projections.

A common way to model this complexity is the post-nonlinear (PNL) mixture, where sources are linearly mixed and then passed through an unknown, element-wise nonlinear function $f$, such that $x = f(As)$. Lyu et al. note that this model is particularly relevant for sensors that have a nonlinear response to physical stimuli. Recovering the original signals requires not just finding the unmixing matrix $A^{-1}$, but also identifying and inverting the unknown distortion $f$.

Strategies for Unmixing

To solve these problems, researchers leverage different properties of the underlying signals:

  • Information Theory: One powerful approach is to maximize the independence of the output. Li et al. use deep learning to minimize "total correlation," a measure of redundancy, to force the network to find the most independent underlying factors. Similarly, Solé‐Casals et al. show that maximizing the entropy of the output can effectively invert monotonic nonlinearities, decoupling the distortion from the source separation.
  • Kernel Feature Spaces: Instead of working in the warped input space, Harmeling et al. propose mapping the data into a high-dimensional kernel space. By constructing an orthonormal basis for the data's specific submanifold, they can perform a version of ICA that accounts for nonlinear interactions.
  • Structured Sparsity: If we know the signals are sparse (e.g., they only "fire" occasionally), we can recover them even from non-invertible observations. Soltani et al. developed algorithms like ONESHOT and DHT that can demix signals from nonlinear "mashups" by exploiting the fact that the signals have a specific structure, such as appearing in blocks.

From Exact Recovery to Statistical Matching

In some cases, perfectly reconstructing the original signal is impossible due to noise or extreme distortion. Régaldo-Saint Blancard et al. argue that we should instead focus on "statistical component separation." By using non-linear representations to capture the distinct statistical signatures of signal and noise, we can recover the properties of the source even when the exact waveform is lost.

This transition from linear to nonlinear models is not just a theoretical hurdle; it has profound practical implications. Bourached et al. demonstrate this in art history, using deep neural networks to disentangle the non-linear overlap of paint layers in X-ray images, effectively "seeing" hidden underdrawings. Even in biology, Jochen Triesch shows that individual neurons may use a combination of internal and synaptic plasticity to solve these exact types of non-linear separation problems, allowing them to tune into rare, independent features in their environment.

Ultimately, the goal of this research direction is to bridge the gap between idealized linear models and the messy, nonlinear reality of physical measurements. As Mergny et al. have shown, even complex nonlinear estimation problems often have fundamental theoretical limits that can be mapped back to simpler models, provided we use the right mathematical transformations.

Key Foundational Work

Kernel Feature Spaces and Nonlinear Blind Souce Separation (2001) — contributed 1 concept elements to this direction.
Maximum Likelihood Blind Source Separation: A Context-Sensitive Generalization of ICA (1996) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_nonlinear_signal_mixing_704f0265,
  title = {nonlinear signal mixing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/nonlinear-signal-mixing/rdr_704f0265002d4709ab7abd144509b8cc},
  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_nonlinear_signal_mixing_704f0265_v1,
  title = {nonlinear signal mixing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/nonlinear-signal-mixing/rdr_704f0265002d4709ab7abd144509b8cc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake