Fast2D Phase Unwrapping

• Fast2D is a quality-guided, path-following phase unwrapping algorithm that recovers continuous phase measurements in adaptive optics.
• It computes pixel reliability from second differences and unwraps edges in order of decreasing reliability to minimize error propagation.
• Comparative studies show Fast2D achieves near-perfect accuracy and the lowest latency among tested methods, with slight sensitivity at aperture boundaries.

Searching arXiv for Fast2D and closely related phase-unwrapping work to ground the article in current literature. arxiv_search(query="Fast2D phase unwrapping adaptive optics", max_results=10) Fast2D is a quality-guided, path-following phase unwrapping algorithm used to recover a continuous phase from wrapped phase measurements in adaptive optics wavefront sensing. In the context of complex field wavefront sensors, which estimate electric-field phase and amplitude through interference or diffraction effects, the sensor output is typically a wrapped phase that must be unwrapped before correction by continuous-surface deformable mirrors. A further systems constraint is latency: for real-time adaptive optics, unwrapping must occur within $1$ millisecond or faster. In a comparative study of four prevalent and mature methods—Fast2D, Zernike Gradient, Discrete Fourier Transform, and Least Squares Principle Value—Fast2D was reported as the lowest-latency method and as essentially numerically exact in the reported ideal simulations, with only slight sensitivity at aperture boundaries (Huerta et al., 19 Aug 2025).

1. Problem setting in adaptive optics

Phase unwrapping is required because the measured phase is usually available only modulo $2\pi$. The wrapped phase $\Phi_w$ is given by

$$\Phi_w = \mathrm{Arg}(e^{i\Phi}) = \tan^{-1}\left(\frac{\mathrm{Im}(e^{i\Phi})}{\mathrm{Re}(e^{i\Phi})}\right) = \Phi - 2\pi \cdot \mathrm{floor}\left(\frac{\Phi+\pi}{2\pi}\right).$$

For adaptive optics, the unwrapping problem is not merely formal. The reconstructed phase must be converted into a continuous surface command for a deformable mirror, and that reconstruction has to occur on millisecond or sub-millisecond timescales. The comparison study in which Fast2D was evaluated used simulations of atmospheric turbulence following a Kolmogorov spectrum and examined performance both with and without a binary circular aperture boundary defining the edge of a monolithic telescope (Huerta et al., 19 Aug 2025).

Within that setting, accuracy was measured by RMS wavefront error (WFE), defined as the root-mean-square of the difference between the piston-removed unwrapped phase and the true phase, expressed in waves. This metric isolates the quality of the unwrapping itself rather than the absolute piston component.

2. Core algorithmic mechanism

Fast2D is a quality-guided, path-following phase unwrapping algorithm. Its central design principle is that adjacent pixels with similar values likely do not contain wraps; these regions are therefore unwrapped first so that error propagation is minimized. The method begins by computing a second difference $D$ at each pixel, from which the pixel reliability $R$ is defined as

$R = \frac{1}{D}.$

Edge reliability is then defined as the sum of the reliabilities of neighboring pixels. All edges are sorted by reliability, and unwrapping proceeds in order of decreasing edge reliability, so that the most trustworthy pixel connections are resolved first (Huerta et al., 19 Aug 2025).

This reliability-ordered traversal distinguishes Fast2D from global modal or transform-based approaches. Rather than solving for coefficients in a basis or enforcing a global periodicity model, it constructs the unwrapped phase incrementally along a reliability-guided path. This suggests that its excellent reported performance in ideal simulations is tied to the quality metric’s ability to identify locally coherent regions before more ambiguous regions are traversed.

3. Boundary behavior and aperture geometry

The principal algorithmic complication for Fast2D lies at boundaries. At edge pixels there are insufficient neighbors to determine reliability in the same way as in the interior, so boundary handling becomes critical. This is especially relevant in adaptive optics because telescope pupils are often circular or otherwise non-rectangular, meaning that the computational grid contains a binary aperture boundary rather than a full rectangular support (Huerta et al., 19 Aug 2025).

In the reported simulations, Fast2D retained exceptional accuracy when a circular aperture boundary was present, except for possibly masked outermost edge pixels. The study further noted that performance is robust except at masked outermost edge pixels and can degrade slightly with complex binary boundaries, such as apertures with central obstructions. The reported interpretation was that Fast2D’s slight boundary sensitivity can be handled through masking and careful edge treatment.

A common misunderstanding is to equate near-perfect interior performance with complete insensitivity to pupil geometry. The reported results do not support that stronger claim. They instead indicate that Fast2D remains highly accurate in the presence of a binary circular aperture, while also identifying boundary handling as the main technical vulnerability of the method.

4. Comparative performance in the reported simulations

The comparative study reported Fast2D, Zernike, DFT, and LSPV performance on $128 \times 128$ pixel arrays. In the unbounded case, Fast2D achieved RMS WFE at the numerical noise floor; with a circular aperture boundary, the error remained at essentially the same scale. In the same experiments, Fast2D was also the fastest CPU implementation among the tested methods (Huerta et al., 19 Aug 2025).

Method	WFE (no bndry)	WFE (w/ bndry)
Fast2D	$(6.80 \pm 0.04) \times 10^{-17}$	$(1.15 \pm 0.03) \times 10^{-16}$
Zernike	$2\pi$0	$2\pi$1
DFT	$2\pi$2	$2\pi$3
LSPV	$2\pi$4	$2\pi$5

Method	Latency [ms]
Fast2D	$2\pi$6
Zernike	$2\pi$7
DFT	$2\pi$8
LSPV	$2\pi$9

These results support several distinctions. Fast2D and LSPV reached essentially perfect accuracy in ideal unbounded cases, but Fast2D remained substantially faster. Zernike Gradient performed well and was described as diffraction-limited, especially for circular apertures, but its accuracy was limited by the number of modes and its latency was significantly higher. DFT was less accurate and more sensitive to edge and boundary artifacts as well as periodicity assumptions. LSPV was highly accurate in unbounded regions, but its boundary performance degraded more strongly and its matrix inversion latency made it slower than Fast2D.

5. Computational profile and real-time relevance

For the reported $\Phi_w$0 arrays, no algorithm met the $\Phi_w$1 millisecond requirement using CPU-only implementations. Fast2D, at $\Phi_w$2 ms, was the closest to that target and was identified as the method with the most immediate promise for satisfying real-time adaptive-optics constraints if further optimization or hardware acceleration is applied (Huerta et al., 19 Aug 2025).

The reported strengths of Fast2D were its fastest observed CPU implementation, numerically perfect accuracy in ideal simulations, conceptual straightforwardness, and implementation across Matlab, Python, and C++. The algorithm is based on simple arithmetic, which supports the view that it is amenable to acceleration. At the same time, its path-following character introduces a nontrivial systems issue: the ordering and interconnectedness of the traversal make naive spatial decomposition difficult.

Specifically, if the array is broken into spatial segments for parallelization, new artificial boundaries are introduced. Because Fast2D’s reliability estimation is weakest near boundaries, segmentation can reduce reliability at those segment edges and thereby affect accuracy unless stitching is handled carefully. A plausible implication is that acceleration strategies preserving global ordering information may be preferable to coarse spatial partitioning.

6. Scope, limitations, and prospective directions

Fast2D was characterized in the comparison study as the most accurate and lowest-latency method among those tested, but that conclusion is bounded by the experimental regime. The evaluation used simulations of atmospheric turbulence following a Kolmogorov spectrum and considered the presence or absence of a binary circular aperture boundary. The method was not explicitly evaluated under strong noise or for highly segmented or obstructed pupils, and future work was identified in more complex pupil geometries, realistic noise, stronger turbulence, and hardware acceleration options (Huerta et al., 19 Aug 2025).

This delimitation matters for interpretation. The reported numerical-noise-floor WFE values establish that Fast2D can behave as an effectively exact unwrapper in ideal simulations, not that all operational wavefront-sensing conditions are equally benign. Likewise, the fact that it was the fastest tested CPU implementation does not imply that real-time deployment is already solved; the measured latency still exceeded the $\Phi_w$3 millisecond requirement.

Within the specific adaptive-optics comparison reported in 2025, Fast2D occupies a well-defined position: a mature phase-unwrapping method whose quality-guided path-following strategy delivers near-perfect unwrapping in idealized conditions, strong robustness in the presence of a circular aperture boundary after appropriate masking, and the best latency among the methods tested, while leaving unresolved the usual boundary, parallelization, and non-ideal-data challenges that dominate practical deployment.