Iterative Ptychography Methods

• Iterative ptychography is a reconstruction approach that uses overlapping diffraction measurements to recover both object and probe with phase retrieval.
• It leverages scan overlap-induced redundancy to enforce object–probe consistency, thereby improving convergence and mitigating local minima.
• The method integrates classical projection techniques with modern optimization, multiscale, and machine-learning enhancements to boost reconstruction quality and speed.

Iterative ptychography is the class of reconstruction methods that recover a complex object, and often the illumination probe, from overlapping diffraction measurements recorded as a localized coherent beam is scanned across a specimen. In the standard far-field formulation, the exit wave at scan position $j$ is $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$, and the measured intensity is $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$; reconstruction proceeds by repeatedly enforcing measured detector amplitudes together with object–probe consistency across overlapping positions (2207.13760). From early projection and optimization formulations to modern regularized, multiscale, autodiff, live, and machine-learning-augmented schemes, the field has remained centered on the same structural fact: overlap-induced redundancy couples neighboring measurements and makes the phase retrieval problem tractable, but convergence and correctness depend strongly on the illumination structure, initialization, scan geometry, update rules, and noise regime (Yang et al., 2011, Edo, 2015).

1. Measurement model, redundancy, and the role of overlap

In iterative ptychography, a localized probe is translated across an extended object so that adjacent illuminated regions overlap. Each diffraction pattern is therefore not an isolated measurement but one member of a coupled set in which the same object pixels are observed under multiple shifted illuminations. This redundancy is the principal source of stability relative to single-shot coherent diffractive imaging, because it ties neighboring patches together and constrains the missing phases globally rather than locally (2207.13760, Parada et al., 2024).

A common overlap estimate is the average linear overlap

$\text{overlap}=1-\frac{R}{D},$

where $R$ is the average scan step size and $D$ the probe diameter (2207.13760). In practical X-ray ptychography, one paper states that the illumination overlap is usually chosen between $50\%$ and $60\%$, while another describes ePIE as commonly recommended around $60\%-70\%$ spatial overlap (2207.13760, Nakahata et al., 2024). The practical reason is that when overlap is reduced, neighboring measurements become weakly coupled, convergence slows, and reconstructions become more susceptible to local minima, phase wraps, vortices, and low-frequency artifacts (2207.13760).

The dependence on redundancy is not captured fully by scan overlap alone. A vector-space analysis showed that ptychography can be viewed as projecting the object onto a family of translated and Fourier-modulated probe functions, and that the illumination Wigner distribution function

$$\chi_a(\mathbf u,\mathbf R)=\int d\mathbf r\; a(\mathbf r)\,a^*(\mathbf r+\mathbf R)\, e^{-2\pi i\,\mathbf u\cdot\mathbf r}$$

is the natural kernel governing how information is shared between measurements (Edo, 2015). In that view, oversampling is only a coarse counting metric; the actual distribution of redundancy depends essentially on probe shape and phase structure. This implies that two experiments with similar nominal overlap can have materially different convergence behavior if their probes induce different coupling patterns in $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$0-space (Edo, 2015).

2. Algorithmic foundations and classical iterative schemes

By 2011, ptychographic phase retrieval had already been recast explicitly as a nonlinear optimization problem as well as a projection problem (Yang et al., 2011). For scan operators $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$1, frame Fourier transform $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$2, and measured amplitudes $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$3, two standard objectives are

$$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$4

which support steepest descent, nonlinear conjugate gradient, Gauss–Newton, and Newton-type iterations (Yang et al., 2011). That same analysis showed that classical error-reduction or alternating-projection updates can be interpreted as preconditioned steepest descent on the amplitude objective, thereby unifying numerical optimization and projection heuristics within a single framework (Yang et al., 2011).

Projection-based methods remain central. In the ptychograph formulation of far-field known-probe phase retrieval, one alternates an amplitude projection

$$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$5

with a consistency projection $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$6, giving the alternating-projections update

$$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$7

Difference Map modifies this by combining the two projectors through a relaxation parameter $$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$8, and is widely used as a stronger projection-based baseline than plain alternating projections (Welker et al., 2022). A live extension of this family later reformulated ER and DM around a fixed-size moving buffer of exit waves, showing that projection-based ptychography can be performed during acquisition rather than only after the full scan is complete (Welker et al., 2023).

The PIE family implements the same physical logic in an object–probe update form. In ePIE, one forms the exit wave, propagates it to the detector, replaces the calculated modulus with the measured modulus, inverse propagates, and updates object and probe from the exit-wave discrepancy (Dzhigaev et al., 2013, Weber et al., 2023). A standard form is

$$\psi_j(\mathbf r)=P(\mathbf r-\mathbf r_j)O(\mathbf r)$$9

$$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$0

with $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$1 the corrected-minus-predicted exit wave (Weber et al., 2023). Regularized variants such as rPIE alter the denominators to improve stability, while related blind formulations jointly recover object and probe and can incorporate partial coherence or multiple probe modes (Ekmekci et al., 2024, Zheng et al., 1 May 2026).

3. Convergence, initialization, and parameter sensitivity

A persistent theme in iterative ptychography is that convergence depends not only on overlap and data quality but also on the starting point. An experimental X-ray study using the refPIE framework compared conventional flat initialization with a moment-based object initialization

$$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$2

where absorption and phase-gradient information are extracted from diffraction moments before iterative refinement (2207.13760). On a full scan with $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$3 average overlap, a half scan with $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$4, and a quarter scan with $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$5, flat initialization yielded FRC resolutions of $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$6, $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$7, and $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$8, whereas object initialization gave $$I_j(\mathbf q)=\left|\mathcal F\{\psi_j(\mathbf r)\}\right|^2$$9, $\text{overlap}=1-\frac{R}{D},$0, and $\text{overlap}=1-\frac{R}{D},$1 (2207.13760). The half-scan result agreed closely with the resolution expected from inverse fourth-root photon-count scaling, leading to the interpretation that reducing overlap from $\text{overlap}=1-\frac{R}{D},$2 to $\text{overlap}=1-\frac{R}{D},$3 caused essentially no reconstruction penalty beyond lower photon count, provided the solver was initialized with a physically informed low-resolution object (2207.13760).

Regularization provides a second route to improved low-overlap behavior. In electron ptychography from 4-D STEM, LoRePIE augments ePIE with transform-domain hard-thresholding of amplitude and phase,

$\text{overlap}=1-\frac{R}{D},$4

where $\text{overlap}=1-\frac{R}{D},$5 and $\text{overlap}=1-\frac{R}{D},$6 retain only the $\text{overlap}=1-\frac{R}{D},$7 largest coefficients after an orthogonal transform (Moshtaghpour et al., 2024). On experimental Rotavirus data originally acquired at $\text{overlap}=1-\frac{R}{D},$8 overlap, synthetic downsampling experiments showed that high-quality reconstruction could still be obtained from data with overlap as low as $\text{overlap}=1-\frac{R}{D},$9 (Moshtaghpour et al., 2024). This supports the broader view that low-overlap failure is not purely a property of the measurement physics; it is also shaped by the prior embodied in the reconstruction method.

A third convergence issue is the choice of update strengths. A 2025 study of low-dose atomic-resolution electron ptychography demonstrated that carefully chosen object and probe update coefficients are essential, and that the optimal values can be much smaller than those commonly used in the literature (Chennit et al., 9 Jul 2025). For the tested FAPbBr$R$0 datasets, useful ePIE reconstructions appeared around $R$1, while rPIE and WASP performed best around $R$2, noting that the interpretation of the coefficient reverses in rPIE because it appears in the denominator (Chennit et al., 9 Jul 2025). Crucially, the paper showed that a decreasing sum squared error does not guarantee a correct solution: under low dose, reconstructions could display apparently convergent error curves while still failing to recover the correct object and probe (Chennit et al., 9 Jul 2025). This directly challenges a common misconception that numerical convergence alone certifies physical correctness.

4. Multiscale, autodiff, and real-time acceleration

Because ptychography is FFT-dominated and often data-intensive, substantial effort has gone into accelerating classical iterative solvers without discarding their physics. One route is multilevel optimization. A multigrid optimization framework built around the distance-to-feasibility objective used coarse object grids together with low-pass-filtered diffraction data, and in numerical experiments a 5-level MG/OPT scheme outperformed both single-level LBFGS and PIE in objective reduction, reconstruction error, and noise robustness (Fung et al., 2018). The central idea is that coarse errors should be removed on coarse representations, reserving the finest grid for fine detail (Fung et al., 2018).

Another route is low-frequency warm starting. A 2024 method introduced fast Partial Fourier Transforms that compute only centered low-frequency Fourier coefficients at cost

$R$3

with $R$4, rather than the full FFT cost $R$5 (Parada et al., 2024). The PFT stage is used only in early PIE or ePIE iterations, after which the reconstruction switches back to the usual FFT-based solver. This plug-and-play warm start reduces the computational budget while preserving reconstruction quality, especially in large-scale regimes where even applying the FFT is considered computationally taxing (Parada et al., 2024).

Automatic differentiation has supplied a more general optimization substrate. A TensorFlow/Keras framework formulated ptychography as minimizing an intensity-domain mean-squared error over a differentiable forward model, using Adam rather than hand-derived PIE updates (Seifert et al., 2020). On the tested simulations it performed comparably to mPIE in speed and quality at suitable learning rate, and its flexibility was illustrated by treating the propagation distance $R$6 as a trainable parameter and recovering it in about $R$7 epochs (Seifert et al., 2020). This suggests that iterative ptychography can be recast as generic differentiable inverse optimization without abandoning the underlying physics.

Real-time operation is a separate but related acceleration objective. A live ePIE implementation at the PtyNAMi hard X-ray microscope started reconstruction from a small subset of the data and extended it incrementally as measurements arrived, giving interpretable results already with a small subset and clear specimen recognition by about a quarter of a long scan (Weber et al., 2023). A more formal live framework later adapted ER and DM using an RTISI-style moving buffer and partial consistency projector, showing that live DM could reconstruct arbitrary-sized objects with fixed computational resources and, under comparable effective computational load, could even achieve higher-quality reconstructions than classical offline DM (Welker et al., 2023).

5. Learned and Bayesian hybrids

Learned methods have entered iterative ptychography chiefly as augmentations of classical solvers rather than replacements of the measurement model. Deep Iterative Projections inserted a residual CNN after each alternating-projection step, preserving both the amplitude constraint and the consistency projection while learning to subtract the residual ptychograph error (Welker et al., 2022). On simulated MNIST ptychography, the phase-only-trained DIP$R$8 reached the target $R$9 in $D$0 iterations, compared with $D$1 for Difference Map and more than $D$2 for plain alternating projections, and also reduced runtime to $D$3 s from more than $D$4 s for AP on the reported GPU benchmark (Welker et al., 2022). The same study also found that DIP alone could stagnate in PSNR at longer runs, which motivated using it as an initialization stage before DM rather than as a full replacement (Welker et al., 2022).

Deep unrolling has followed a similar physics-preserving pattern. PtychoDV combines a vision transformer initializer with a short unrolled sequence of Wirtinger-flow data-consistency steps and learned CNN priors, using only $D$5 unrolled stages (Gan et al., 2023). On simulated sparse scans, it matched or outperformed PMACE and AWF while reducing inference time by orders of magnitude; at scan pattern $D$6, for example, PtychoDV reported NRMSE $D$7 in $D$8 s, compared with $D$9 in $50\%$0 s for PMACE (Gan et al., 2023). The same work showed that using PtychoDV as an initializer could make a $50\%$1-iteration PMACE reconstruction outperform a $50\%$2-iteration PMACE reconstruction in some settings (Gan et al., 2023).

Transformer-based one-shot models have also been combined with ePIE. PtychoFormer processes subsets of diffraction patterns and stitches the resulting local inferences, while extended-PtychoFormer uses that stitched reconstruction to initialize ePIE (Nakahata et al., 2024). In the reported tests, PtychoFormer was $50\%$3 to $50\%$4 faster than ePIE, and the ePF hybrid reduced NRMSE compared with ePIE by $50\%$5 for amplitude and $50\%$6 for phase (Nakahata et al., 2024). The same paper reported useful performance at $50\%$7 spatial overlap, well below the overlap commonly recommended for conventional iterative reconstruction (Nakahata et al., 2024). At the same time, the paper also states that this speed comes at the expense of quality, and that ePF outperforms pure PtychoFormer, which underscores the continuing value of iterative refinement (Nakahata et al., 2024).

A more explicitly probabilistic alternative is Bayesian latent-space inversion. In a simulated low-overlap framework, a deep generative prior $50\%$8 was combined with the ptychographic Poisson likelihood and sampled using the unadjusted Langevin algorithm, producing posterior samples rather than a single point estimate (Ekmekci et al., 2024). This method consistently outperformed rPIE in reduced-overlap conditions and provided uncertainty maps whose pixelwise standard deviation correlated positively with true reconstruction error over 100 test samples (Ekmekci et al., 2024). A plausible implication is that some of the perceived overlap requirement of standard iterative solvers reflects the weakness of their implicit priors rather than a hard physical boundary.

A more operationally mature learned augmentation is the 2026 fast-forward operator inserted once into a standard CDTools reconstruction after a short warm-up (Zheng et al., 1 May 2026). Acting only on the object estimate, it preserves subsequent iterative data-fidelity minimization while reducing time-to-solution; on held-out experimental samples it achieved more than a two-fold wall-clock speedup relative to the standard solver and was deployed in production at a synchrotron beamline (Zheng et al., 1 May 2026).

6. Extensions, applications, and scope

Iterative ptychography is no longer limited to recovering a single two-dimensional object from a monochromatic scan. In hard X-ray beam metrology, ePIE has been used to jointly reconstruct a Siemens-star test object and a nanofocused probe containing phase singularities, yielding probe widths of $50\%$9 nm horizontally and $60\%$0 nm vertically and revealing a vortex pair with topological charges $60\%$1 and $60\%$2 (Dzhigaev et al., 2013). Numerical propagation of the recovered probe then tracked the vortices through focus, demonstrating that iterative ptychography can function as wavefront metrology rather than only sample imaging (Dzhigaev et al., 2013).

In spectroscopic X-ray ptychography, SPA formulated multi-energy blind reconstruction as a joint Poisson maximum-likelihood problem constrained by a chemical dictionary and solved it with ADMM-style splitting (Chang et al., 2019). By coupling the redundancy across energies, SPA was designed to recover chemical maps directly rather than first reconstructing independent ptychographic images at each energy and only then performing spectral analysis (Chang et al., 2019). The same work reported enhanced robustness when reconstructing reduced-redundancy data with large scan step sizes, indicating that spectral coupling can play a role analogous to spatial overlap in stabilizing iterative inversion (Chang et al., 2019).

A related generalization is joint ptychography–tomography. An ADMM-based formulation solved for a 3D object directly from diffraction data across all scan positions and rotation angles, with 3D total variation regularization (Chang et al., 2019). In synthetic tests with large scan step size $60\%$3 and $60\%$4 angles, the standard two-step pipeline failed, whereas the joint methods reconstructed the sample; the TV-regularized variant improved SNR from $60\%$5 to $60\%$6 over the unregularized joint method in that setting (Chang et al., 2019). This suggests that iterative ptychography can be embedded as one block in larger coupled inverse problems without losing its characteristic detector-domain amplitude correction structure.

The scope of the field also now includes nonstandard measurement physics. Ptychographical intensity interferometry replaced coherent diffraction magnitudes with second-order intensity correlations from a pseudothermal source and used overlapping probe positions plus loose supports to recover the object (Wang et al., 2017). Although the measurement modality differs from conventional coherent ptychography, the reconstruction engine still follows the same forward-propagate, modulus-replace, backpropagate, and overlap-consistency logic (Wang et al., 2017).

Across these extensions, several unresolved issues recur. Overlap remains important even when initialization, sparsity regularization, or learned priors allow it to be reduced; low error curves do not eliminate the risk of local minima; and many learned methods remain validated only on simulations or restricted experimental settings (2207.13760, Chennit et al., 9 Jul 2025, Ekmekci et al., 2024). Iterative ptychography therefore continues to evolve along two linked axes: richer forward models and priors, and more reliable ways of navigating the nonconvex optimization landscape they induce.