Point Spread Function Method

• Point Spread Function Method is a framework that models an instrument’s impulse response to accurately recover the true object via convolution.
• It is applied in XFEL scintillating-screen diagnostics, astronomical imaging, and scattering deconvolution to enhance resolution and measurement precision.
• The approach leverages PSF-aware forward fitting to overcome limitations of naive Gaussian fitting and achieve improved inference of beam and object profiles.

Searching arXiv for the core paper and closely related PSF-method papers to ground the article in current literature. The point spread function method is a class of imaging and inference methods that explicitly model an instrument’s response to a point source and then use that response in forward modeling, fitting, interpolation, deconvolution, subtraction, or operator approximation. In the European XFEL scintillating-screen context, the term denotes a specific workflow: accurately simulate $PSF_{\text{aber}}$ and $PSF_{\text{scint}}$ for the actual screen diagnostic, and incorporate these PSFs into the beam-profile fitting function in order to recover smaller beam sizes than are accessible with naive Gaussian fitting (Novokshonov, 2024). Across astronomy, scattering imaging, computational optics, AO PSF prediction, and inverse problems, the same conceptual core recurs: the observed image is modeled as a convolution of the true object with a PSF, and PSF knowledge is then exploited either to reconstruct the object or to improve measurement fidelity (Hartung, 2013, Xu et al., 2017, Bergé et al., 2011, Kuznetsov et al., 24 Mar 2026).

1. Definition and formal structure

In imaging, the PSF is the impulse response of the system. In the XFEL scintillating-screen formulation, it is defined as the intensity distribution on the camera sensor produced by an infinitesimally small beam spot on the scintillator, incorporating the scintillator, the objective lens, and the imaging geometry (Novokshonov, 2024). The generic forward model is a convolution,

$$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$

and closely related forms appear in astronomical imaging, scattering through thin diffusers, and wide-field telescope modeling (Novokshonov, 2024, Hartung, 2013, Xu et al., 2017, Jia et al., 2020).

In the European XFEL study, the total PSF is explicitly split into an aberrational component $PSF_{\text{aber}}$ and a scintillator-geometry component $PSF_{\text{scint}}$. In the plane without observation angle, only $PSF_{\text{aber}}$ enters. In the angle plane, the image is modeled as the source convolved with $PSF_{\text{aber}} * PSF_{\text{scint}}$ (Novokshonov, 2024). This decomposition is central because the observation angle affects only one transverse plane, producing a strongly anisotropic response.

A broader implication is that “point spread function method” is not a single algorithm but a methodological family. In astronomy, it can denote PSF matching and PSF-shaped filtering before image subtraction (Hartung, 2013), interpolation of spatially varying stellar PSFs across a focal plane (Bergé et al., 2011), or data-calibrated AO PSF prediction from reduced telemetry (Kuznetsov et al., 24 Mar 2026). In scattering imaging, it can mean retrieving a speckle-like PSF from a known reference object and then using that PSF for reconstruction (Xu et al., 2017). In PDE-constrained inversion, it can denote approximation of a high-rank Hessian by sampling impulse responses and interpolating kernel entries (Alger et al., 2023). This suggests that the unifying criterion is not the application domain but the explicit use of the system’s point response as the organizing object.

2. The European XFEL scintillating-screen formulation

At the European XFEL, transverse electron beam profiles are measured with scintillating screen monitors rather than with optical transition radiation screens, because OTR images can be badly distorted by coherent OTR due to micro-bunching, whereas scintillators do not suffer from these coherence effects and yield more robust, reproducible images (Novokshonov, 2024). The scintillating material is GAGG:Ce, the typical thickness is $200\ \mu\text{m}$, the screen is perpendicular to the electron beam, and emitted scintillation light is viewed at a backward angle with the camera tilted to obey the Scheimpflug condition (Novokshonov, 2024).

Resolution degradation arises from two physical effects. The first is finite scintillator thickness combined with the observation angle: because the screen is $200\ \mu\text{m}$ thick, each electron produces a track in the scintillator thickness, and light from different depths is projected onto different camera positions. This creates geometrical broadening in the plane containing the observation angle, while the orthogonal plane is unaffected by this specific effect (Novokshonov, 2024). The second is optical aberration of the objective, including spherical aberration, coma, and astigmatism (Novokshonov, 2024).

The anisotropy is therefore intrinsic. In the no-angle plane, the PSF is essentially the lens PSF and is close to isotropic and well described by a Gaussian with $\sigma_{\text{aber}}$ of order $PSF_{\text{scint}}$0. In the angle plane, the PSF becomes strongly anisotropic and non-Gaussian because of projection of the $PSF_{\text{scint}}$1 scintillator thickness and multiple internal reflections in the scintillator and optics (Novokshonov, 2024). This is the immediate reason that naive Gaussian fitting fails for small beams in the angle plane.

The study treats the beam profile itself as Gaussian and fits the measured profile with a Gaussian convolved with $PSF_{\text{scint}}$2 and, where needed, $PSF_{\text{scint}}$3. A plausible implication is that the method is best understood as PSF-aware forward fitting rather than deconvolution in the strict inverse-problem sense, because the analysis is posed directly in terms of a parameterized source model convolved with a known response (Novokshonov, 2024).

3. Simulation, parameterization, and validation

All PSFs in the XFEL study are modeled using Ansys Zemax OpticStudio®. Sequential mode is used for aberration modeling, and non-sequential mode is used for bulk-media interactions, internal reflections, and scintillator geometry (Novokshonov, 2024). For the 1:2 optics with the Schneider-Kreuznach Makro-Symmar 5.9/120 lens, the source–objective distance is 360 mm, the objective–camera distance is 180 mm, and the camera tilt is about $PSF_{\text{scint}}$4. For the 1:1 optics with the Schneider-Kreuznach 5.6/180 lens, both distances are 360 mm and the camera tilt is $PSF_{\text{scint}}$5 (Novokshonov, 2024).

The aberrational PSF is computed on-axis and at off-axis positions. Owing to Scheimpflug alignment, on-axis and off-axis PSFs match very well, indicating that defocus is well corrected (Novokshonov, 2024). The central aberration PSF for the 1:2 optics is fitted with a Gaussian yielding $PSF_{\text{scint}}$6, while the 1:1 optics gives $PSF_{\text{scint}}$7 (Novokshonov, 2024).

For $PSF_{\text{scint}}$8, the scintillator is modeled as a $PSF_{\text{scint}}$9-thick slab of GAGG:Ce with a line of discrete point sources placed within the bulk, each emitting isotropically in $$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$0. Rays propagate through the slab and into the lens and camera, undergoing internal reflections (Novokshonov, 2024). The resulting PSF is strongly non-Gaussian, shows several steps or lobes due to internal reflections, and is much wider than $$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$1 (Novokshonov, 2024). The central profile is parameterized as a sum of four high-order Gaussians with exponent 10 rather than 2,

$$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$2

The parameters are obtained from fits to the OpticStudio simulation and then used in data analysis (Novokshonov, 2024).

Validation proceeds through the modulation transfer function. The optical transfer function is the Fourier transform of the PSF,

$$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$3

with

$$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$4

Measured MTF from a USAF 1951 target agrees well with OpticStudio simulations after adjusting the image plane position to account for slight defocus (Novokshonov, 2024). The scintillator PSF is validated indirectly by agreement between predicted and experimentally reported resolutions: for the 1:1 station, the plane without observation angle gives measured resolution $$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$5 while the model gives $$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$6; in the angle plane, other studies report $$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$7 while the model predicts $$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$8 (Novokshonov, 2024).

4. PSF-aware fitting as a resolution-recovery method

The practical core of the XFEL method is to replace a pure Gaussian fit with a Gaussian convolved with the known PSFs (Novokshonov, 2024). In the no-angle plane,

$$I_{\text{meas}}(x,y) = (I_{\text{true}} * \mathrm{PSF})(x,y),$$9

and because both distributions are approximately Gaussian,

$PSF_{\text{aber}}$0

In the angle plane,

$PSF_{\text{aber}}$1

and the convolution is evaluated numerically or through pre-tabulated PSFs (Novokshonov, 2024).

The quantified performance gain is substantial. For 1:2 optics in the angle plane, simulated beam sizes of 5–60 $PSF_{\text{aber}}$2 show that a naive Gaussian fit drops below 10% deviation from true size only when the true beam size is $PSF_{\text{aber}}$3, whereas the PSF-convolved fit crosses the 10% deviation threshold at around 10–20 $PSF_{\text{aber}}$4 (Novokshonov, 2024). The study quotes an effective resolution of roughly $PSF_{\text{aber}}$5 for the angle plane in this geometry (Novokshonov, 2024).

In the no-angle plane, where only $PSF_{\text{aber}}$6 matters, the effective resolution is about $PSF_{\text{aber}}$7 for 1:2 optics and about $PSF_{\text{aber}}$8 for 1:1 optics (Novokshonov, 2024). In the 1:1 angle plane, the convolved Gaussian deviates by more than 10% only at $PSF_{\text{aber}}$9, so the effective resolution is comparable to or slightly better than $PSF_{\text{scint}}$0 (Novokshonov, 2024). The analysis also explains why 1:1 optics gives slightly better effective resolution in the no-angle plane despite a slightly larger raw $PSF_{\text{scint}}$1: after image rescaling by magnification, aberrations are effectively magnified in 1:2 processing (Novokshonov, 2024).

The practical workflow is explicit: simulate or measure $PSF_{\text{scint}}$2 and $PSF_{\text{scint}}$3; parameterize them; acquire beam images in the characterized geometry; extract 1D profiles; fit the no-angle plane with Gaussian plus quadrature subtraction or $PSF_{\text{scint}}$4, and fit the angle plane with $PSF_{\text{scint}}$5; then use the fitted $PSF_{\text{scint}}$6 as the true beam size (Novokshonov, 2024). The stated limitations are equally specific: the method assumes a Gaussian beam shape, depends on PSF accuracy, is more computationally expensive than pure Gaussian fitting, and should be re-simulated or re-measured if the geometry drifts (Novokshonov, 2024).

5. Variants of the PSF method in other research domains

The same phrase denotes distinct but structurally related procedures in other fields. In astronomical image subtraction, PSF matching is used to transform a reference-image PSF into a science-image PSF before subtraction, and the specific contribution of "Image Subtraction Noise Reduction Using Point Spread Function Cross-correlation" is to apply PSF cross-correlation as a pre-processing step before fitting a spatially varying Dirac delta basis kernel (Hartung, 2013). There the PSF acts as a matched filter for unresolved point sources, reducing high-frequency pixel-level noise and improving the signal-to-noise ratio of variable objects in the difference image by approximately $PSF_{\text{scint}}$7 (Hartung, 2013).

In imaging through scattering layers and around corners, the PSF method treats the diffuser or rough reflector as a shift-invariant linear system with a speckle-like PSF, retrieves that PSF from a reference object of known shape, and reconstructs hidden objects through a non-iterative Fourier-domain deconvolution (Xu et al., 2017). The central transmission formula is

$PSF_{\text{scint}}$8

and the method supports single-shot reconstruction, dynamic scenes, and reflection geometry (Xu et al., 2017). This suggests that, in scattering imaging, “PSF method” emphasizes calibration by a known object rather than instrumental design modeling.

For wide-field small-aperture telescopes, PSF methods frequently mean PSF estimation or interpolation across the field of view. One line of work uses deep networks trained on calibration data and telescope misalignment states so that a sparse “PSF-Cube” can be mapped to a dense field of PSFs at arbitrary positions (Jia et al., 2020). Another compares interpolation schemes and finds that Kriging gives the most reliable interpolation of PCA-based PSF coefficients, significantly better than polynomial interpolation for weak-lensing systematics control (Bergé et al., 2011). These are PSF methods in the sense of field-dependent forward modeling rather than image restoration.

Recent AO work uses a compact, physics-informed, and data-calibrated Fourier PSF model corrected by a lightweight neural network trained end to end on on-sky data (Kuznetsov et al., 24 Mar 2026). In that setting, the PSF is factored into static, turbulence, and tip–tilt OTF components, while the network corrects a compact set of physically meaningful parameters retrievable from the archive (Kuznetsov et al., 24 Mar 2026). A plausible implication is that modern PSF methods increasingly combine analytical optics with learned calibration, especially where full telemetry is unavailable.

6. Broader methodological significance

Across domains, the PSF method supports three recurring operations. The first is forward fitting: use a parameterized source model convolved with a known PSF and fit in measurement space, as in the XFEL beam-size recovery workflow (Novokshonov, 2024). The second is correction or inversion: deconvolution of images or subtraction of modeled PSF structure, as in scattering imaging (Xu et al., 2017), solar or partially occulted calibration methods (Hofmeister et al., 2022), and EUV instrumental PSF correction (Hofmeister et al., 2 May 2026). The third is interpolation or approximation of operators: use impulse responses at sampled positions to approximate the action of a spatially varying system, as in high-rank Hessian approximation with locally supported non-negative kernels (Alger et al., 2023).

A technical distinction runs through the literature between empirical PSFs, physics-based PSFs, and hybrid PSFs. The XFEL scintillating-screen study is strongly physics-based: Zemax simulations, explicit geometry, and MTF validation (Novokshonov, 2024). Astronomy interpolation papers emphasize empirical star-based estimation and spatial statistics (Bergé et al., 2011). AO PSF prediction and some telescope PSF estimation studies are hybrid, combining physical forward models with neural or statistical calibration (Jia et al., 2020, Kuznetsov et al., 24 Mar 2026). This suggests that the method class is less defined by the optimization algorithm than by the decision to make the PSF an explicit inferential object.

A second distinction concerns whether the PSF is treated as isotropic, anisotropic, Gaussian-like, or highly structured. The XFEL case is exemplary because anisotropy is not a secondary complication but the principal effect in the angle plane (Novokshonov, 2024). Scattering media produce speckle-like PSFs (Xu et al., 2017); AO systems produce wavelength- and field-dependent PSFs (Kuznetsov et al., 24 Mar 2026); partially occulted-image methods explicitly avoid assuming a predefined functional form (Hofmeister et al., 2022). The common lesson is that incorrect structural assumptions about the PSF directly limit recoverable resolution, photometric accuracy, or source separation.

Finally, the method’s epistemic role is often calibration rather than mere correction. In the XFEL study, incorporating the PSF into the fitting function substantially improves the reliable resolution for beam sizes below $PSF_{\text{scint}}$9, especially in the plane affected by scintillator thickness and observation angle (Novokshonov, 2024). In other settings, the same logic improves transient detection (Hartung, 2013), hidden-object reconstruction (Xu et al., 2017), weak-lensing systematics control (Bergé et al., 2011), or AO-assisted photometry and de-blending (Kuznetsov et al., 24 Mar 2026). The point spread function method is therefore best regarded as a general research strategy: estimate the system’s impulse response with enough fidelity that downstream inference can be posed in terms of the actual imaging operator rather than an idealized surrogate.