Poisson Wavefront Imaging (PWI)

• PWI is an optimization-based framework that models photon shot noise as a Poisson process in low-light environments.
• It employs multiple SLM phase patterns with a total variation smoothness prior to enhance wavefront reconstruction accuracy.
• Experimental results show up to 1.6× lower RMSE and 1.8× improved resolution compared to classical methods like the Gerchberg-Saxton algorithm.

Poisson Wavefront Imaging (PWI) is an optimization-based framework for quantitative phase imaging optimized for photon-starved scenarios. By explicitly modeling photon shot noise as a Poisson process and integrating a smoothness prior, PWI achieves accurate and high-resolution wavefront reconstruction under severe signal constraints. The method leverages multiple spatial light modulator (SLM) phase patterns to maximize information throughput and regularity, resulting in superior performance relative to classical approaches such as the Gerchberg-Saxton (GS) algorithm, particularly in regimes with few detected photons per pixel (Choi et al., 13 Dec 2025).

1. Physical Forward Model and Poisson Noise Characterization

The central object of estimation in PWI is the pupil-plane phase $\phi \in \mathbb{R}^n$, discretized on $n$ pixels. For each measurement, an SLM applies one of $M$ distinct phase patterns $p_i$ ($i=1,\dots, M$), introducing coded diversity to the illumination. The subsequent propagation through the optical system can be modeled as

$u_i(\phi) = A_i\,A_0\,\exp\left(j\,\phi\right),$

where $A_0$ and $A_i$ are system operators describing propagation to and from the SLM, respectively. The detector records light intensity at pixel $k$,

$$\mu_{ik}(\phi) = \left|\; u_{ik}(\phi) \;\right|^2.$$

In photon-starved settings, the detected photon counts $y_{ik}$ are modeled as independent Poisson random variables: $$y_{ik} \sim \mathrm{Poisson}\left(\mu_{ik}(\phi) + b\right),$$ with $b \geq 0$ representing a fixed background. Compactly, for each pattern,

$$y_i \sim \mathrm{Poisson}\left(H(\phi, p_i) + b\right),$$

where $H(\phi, p_i) = \mu_i$ gives the vector of expected intensities.

2. MAP Reconstruction Formalism and Optimization Procedure

PWI estimates $\phi$ by maximizing the posterior distribution using a Poisson-likelihood fidelity term and a total variation (TV) smoothness prior. The optimization problem is: $$\min_\phi\;-\log\mathcal{L}(\phi; y, p) + \lambda\,R(\phi),$$ with

$$-\log\mathcal{L}(\phi; y, p) = \sum_{i=1}^{M}\sum_{k=1}^{K} \Bigl[\mu_{ik}(\phi) - y_{ik}\,\log\,\mu_{ik}(\phi)\Bigr]$$

and $$R(\phi) = \mathrm{TV}(\phi) = \sum_p \|(\nabla \phi)_p\|_2$$. The TV prior regularizes the phase, promoting smoothness while preserving phase discontinuities. The optimization is solved using the alternating direction method of multipliers (ADMM), with auxiliary variables and closed-form updates for the detector-plane fields, the pupil-plane phase, and TV splits.

3. Multiple SLM Patterns and Fisher Information Enhancement

Utilizing $M$ diverse SLM phase patterns $\{ p_i \}$ substantially increases sensitivity to wavefront features by modulating system response. The Fisher information matrix for parameter vector $\phi$ is: $$[I_F(\phi)]_{pq} = \sum_{i=1}^M \sum_{k=1}^K \frac{1}{\mu_{ik}(\phi)} \frac{\partial \mu_{ik}(\phi)}{\partial \phi_p} \frac{\partial \mu_{ik}(\phi)}{\partial \phi_q}.$$ Maximizing either the trace or minimum eigenvalue of $I_F$ via careful design (e.g., random but smooth patterns) improves the conditioning and theoretical estimation accuracy compared to single-pattern or Shack-Hartmann approaches.

4. Theoretical Error Bound and Empirical Performance

The Cramér–Rao lower bound (CRLB) governs the minimum achievable variance for unbiased estimators: $$\mathrm{Var}(\hat{\phi}_p) \geq \left[ I_F(\phi)^{-1} \right]_{pp}.$$ The mean-per-pixel standard deviation for a total photon count $N$ (fixing the global piston) is defined by

$$\sigma(N) = \sqrt{ \frac{1}{N(P-1)} \sum_{p=1}^{P-1} [I_F(\phi)^{-1}]_{pp} }.$$

Simulation results demonstrate that the root-mean-square error (RMSE) of PWI closely approaches this theoretical bound. Introducing the TV prior introduces bias such that the RMSE can fall below the unbiased CRLB. Simulations with a $40\times 40$ “P”-shaped phase using $M = 10$ patterns yield RMSEs near the CRLB and notably below those of SHWFS or flat-pattern imaging at all tested photon levels ($10^2$–$10^4$ per pixel).

5. Experimental Comparison and Quantitative Results

Experimental results with a USAF target at the image plane and $M=24$ SLM patterns show PWI with TV prior reconstructing phase accurately with only $7.5$ photons/pixel, matching GS quality at $66$ photons/pixel (an $8.7\times$ reduction in photon budget). Across all photon levels, PWI with TV achieves up to $1.6\times$ lower phase RMSE than GS. Spatial resolution, as measured by the maximum resolved line pairs per millimeter (lp/mm), increases from $71.8$ (GS) to $128$ (PWI)—a $1.8\times$ enhancement.

Photon budget (photons/pix)	GS RMSE	PWI+TV RMSE	RMSE ratio	Resolved lp/mm
66	0.18	0.11	1.6× lower	128 vs 71.8

6. Computational Aspects and Application Domains

Each PWI iteration requires $\mathcal{O}(MK \log K)$ operations, dominated by forward and inverse FFT-based propagations. The ADMM solver converges in approximately $300$–$700$ iterations and is roughly $1.7\times$ faster per iteration than Adam-based optimization for the Poisson loss.

Application areas include:

• Astronomy: Real-time correction of wavefronts for faint guide stars under photon-limited conditions.
• Semiconductor metrology: Inspection of high-numerical-aperture (NA) wafers using few-photon EUV/X-ray illumination.
• Biological imaging: Quantitative phase microscopy of live cells under minimal light exposure to mitigate sample damage.

This suggests that PWI’s integration of statistically accurate noise modeling, variational regularization, and coded SLM diversity offers a generalized and effective solution to photon-limited phase retrieval tasks across diverse fields (Choi et al., 13 Dec 2025).