Stability Collapse in High Dimensional Phase Retrieval

18 concepts6 papers2 proposals

In many physical imaging and signal processing applications, sensors can only record the intensity or magnitude of a signal, completely losing its phase. Reconstructing the original signal from these magnitude-only measurements is known as phase retrieval—a foundational challenge spanning ptychography, speech recognition, and quantum mechanics (Lukas Liehr).

While phase retrieval is often perfectly tractable in low-dimensional settings, a severe mathematical obstacle emerges as we push toward high-resolution imaging: the stability of the reconstruction process collapses. Understanding why this stability collapse occurs, and how we can bypass it, is central to modern high-dimensional signal recovery.

The Geometry of Stability Collapse

Why does increasing a signal's resolution make its phase so much harder to recover? The root of the problem lies in how the "stability constant"—the mathematical factor that dictates how much measurement noise is amplified during reconstruction—scales with dimension.

In finite-dimensional spaces, the stability constant degrades exponentially as the signal's dimension increases (Sarah Koppensteiner et al.). By the time we reach true infinite-dimensional signal spaces, uniform stability collapses entirely (Ingrid Daubechies et al.). This means the problem becomes fundamentally ill-posed: an arbitrarily small perturbation or numerical error in the magnitude measurements can result in a massive, completely incorrect deviation in the reconstructed signal.

This high-dimensional instability is a geometric property of complex reconstruction spaces. A parallel phenomenon can be observed in generative AI: Han Zhang et al. demonstrated that when reconstructing high-resolution images using diffusion Probability-Flow ODEs, the probability of encountering severe numerical instability converges to one as the pixel count grows. In both domains, scaling up dimensionality forces the inversion process to navigate regions where error amplification is nearly unavoidable.

Why More Data Cannot Save Us

A standard engineering intuition for combating noise is oversampling—simply taking more measurements to average out the errors. However, phase retrieval in infinite dimensions defies this intuition.

Rima Alaifari proved that even if we use highly redundant measurement schemes, such as continuous frames in Banach spaces, uniform stability cannot be rescued. Moving to an infinite, continuous family of measurements fails to prevent arbitrary error amplification because the condition number of the phase retrieval operator itself becomes infinite. The instability is not caused by a shortage of data points; it is baked into the geometry of the phase-loss transformation itself.

To understand why the math breaks down so spectacularly, consider what happens when a signal's energy is concentrated on disjoint or poorly connected sets. As the authors of Stable Phase Retrieval in Infinite Dimensions point out, if a signal consists of two distinct components separated by a wide gap of zero intensity, their relative phase becomes completely unanchored.

Signal Component A          Zero-Energy Gap          Signal Component B
     | * * |                                              | * * |
     | * * |               ................               | * * |
  (Phase: θ₁)              (No Interference)           (Phase: θ₂)

Because the components do not overlap or interfere with one another, you can shift the phase of Component A by $\theta_1$ and Component B by $\theta_2$ without altering the combined magnitude profile. In infinite dimensions, signals can easily "hide" these independent phase variations across disjoint regions, making global phase alignment highly sensitive to tiny measurement fluctuations.

Strategy 1: Redefining the Goal via Local Phases

If recovering the global phase of an infinite-dimensional signal is inherently unstable, how can we make practical imaging systems work? One powerful solution is to change the objective.

Instead of trying to recover a signal up to a single, global phase factor, Ingrid Daubechies et al. introduced a paradigm shift: recovering the signal up to local or component-wise phase shifts. If we accept that distant, disconnected parts of a signal can have independent phase ambiguities, we can focus on stably reconstructing the local structure. The authors demonstrated that by utilizing localized measurement frameworks—such as Gabor frames and wavelets—stable reconstruction up to these local phase modifications becomes entirely feasible.

Strategy 2: Propagating Phase Through Graph Connectivity

If we must recover the global phase, we have to restrict the types of signals we attempt to reconstruct. The fundamental vulnerability is the presence of "gaps" or "holes" that isolate parts of the signal from one another.

To exploit this, Cheng Cheng et al. developed a framework based on Locally Stable and Conditionally Connected (LSCC) measurements. The core intuition is to treat the measurement space as a connected graph. If local neighborhoods are stable (meaning we can resolve relative phases locally), and these neighborhoods overlap sufficiently to form a well-connected graph without large gaps, we can reliably propagate phase information across the entire signal.

[ Local Region 1 ] <--- Overlap ---> [ Local Region 2 ] <--- Overlap ---> [ Local Region 3 ]
  Phase resolved                       Phase chained                        Phase chained
  locally                              across boundary                      across boundary

By ensuring the signal lacks large structural holes, this connectivity prevents the relative phases of different regions from decoupling. Under the LSCC framework, the stability constant bypasses the exponential curse of dimensionality entirely, scaling polynomially instead. This makes high-dimensional recovery robust and practically feasible, showing that while infinite dimensions naturally invite instability, structural connectivity provides a reliable path home.

Key Foundational Work

Phase Retrieval In The General Setting Of Continuous Frames For Banach Spaces (2017) — contributed 7 concept elements to this direction.
The Mathematics of Phase Retrieval (2019) — contributed 3 concept elements to this direction.
Stable Phase Retrieval in Infinite Dimensions (2019) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 7 papers
  2. Version 1 · 2026-10-04 · 13 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_infinite_dimensional_instability_16d40029,
  title = {infinite-dimensional instability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-dimensional-instability/rdr_16d4002915ba4b2b94a23565671c424f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_infinite_dimensional_instability_16d40029_v2,
  title = {infinite-dimensional instability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-dimensional-instability/rdr_16d4002915ba4b2b94a23565671c424f/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake