Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Universal point spread function engineering for 3D optical information processing

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A trained stack of thin passive surfaces can synthesize an arbitrary set of 3D point-spread functions for incoherent light, enabling snapshot 3D and multispectral imaging without digital reconstruction.

desk verdict Clean numerical extension of 2D universal diffractive transformations to 3D PSF engineering, but the universality claim floats on an empirical DOF threshold with no error bars or code. read the letter →

arxiv 2502.06025 v1 pith:ICM7OY27 submitted 2025-02-09 physics.optics cs.NE

Md Sadman Sakib Rahman , Aydogan Ozcan This is my paper · ORCID
classification physics.opticscs.NE
keywords pointspreadfunctionengineeringdiffractiveopticalprocessor3Dimagingmultispectralspatiallyincoherentlightinformationprocessingcomputationalphase-onlyoptics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a stack of thin, passive, phase-only optical surfaces can be trained so that every voxel of a 3D input volume produces a prescribed diffraction-limited intensity pattern at an output plane. Because the patterns can be chosen independently for each input voxel and each wavelength, the same all-optical stack can implement essentially any nonnegative linear transformation between 3D intensity volumes. The authors demonstrate numerically that such a stack can image a 3D scene from a single detector snapshot, with axial planes separated by about 2.67 wavelengths, and can separate three emission wavelengths at once without spectral filters, axial scanning, or digital reconstruction. The significance, if true, is that complex 3D and multispectral imaging tasks could be performed passively and instantaneously by a fabricated optical element.

What carries the argument

The central object is the intensity transformation matrix A: its columns are the target 3D PSFs, so an input intensity vector i over the input voxels produces an output intensity o = A i, and the processor is trained to realize A-hat approximately equal to A. The processor itself is a cascade of K thin phase-only diffractive surfaces whose N feature values are optimized by backpropagation through a digital forward model. The load-bearing design rule is that roughly N approximately 2 Ni No optimizable features distributed over K at least 4 surfaces suffice to approximate any nonnegative target A with negligible error, with depth suppressing the larger errors seen for K = 2; the diffraction limit enters through the chosen voxel spacing and plane-to-plane distances, which set the axial resolution of the achievable PSFs.

What would settle it

Fabricate the K=4, N approximately 2 Ni No processor from the numerical study and measure its full intensity transformation on random nonnegative test inputs; if the measured outputs deviate from A i beyond the simulation's negligible error, the claimed universal synthesis does not hold. A second decisive test is to put two scatterers in the input volume and check whether the output is the sum of their individual PSFs, since any shadowing-induced deviation would violate the linear model.

Watch

Extended reading notes

Core claim

The central claim is that a spatially incoherent diffractive processor, built from K cascaded transmissive surfaces with N optimizable phase-only features, can synthesize an arbitrarily defined set of 3D diffraction-limited PSFs mapping Ni input voxels to No output voxels, provided the number of features is large enough. In numerical experiments with random nonnegative target matrices A, the approximation error between the all-optical transformation A-hat and A becomes negligible as N approaches 2 Ni No when K is at least 4, and the factor 2 is attributed to the phase-only nature of the features. The authors also show spectrally engineered 3D PSFs, using them to demonstrate snapshot 3D imaging and snapshot 3D multispectral imaging from a single output frame, with demultiplexing accomplished simply by rearranging detector pixel values according to the assigned input planes and wavelengths.

Load-bearing premise

The load-bearing premise is that each emitter in the input volume radiates independently and does not block, shadow, or re-excite its neighbors, so the output is exactly the sum of the individual point-spread functions; any emitter-to-emitter coupling makes the transformation nonlinear and object-dependent, which would void the claimed universality.

Editorial extensions

If this is right

  • A single passive diffractive processor can perform any prescribed nonnegative linear intensity transformation between 3D volumes at the speed of light, with no digital processing step.
  • One output snapshot from a single detector array can recover the full 3D distribution of independent emitters, with axial planes separated by about 2.67 wavelengths, by pixel rearrangement alone.
  • Spectral information can be encoded in the same snapshot: different wavelengths are routed to different output pixels, so multispectral 3D imaging needs no filters, no scanning, and no reconstruction.
  • The optimized layer designs transfer directly to other wavelengths by rescaling physical dimensions, since the spectral engineering relies on free-space dispersion rather than material dispersion.
  • Shallow processors with only two surfaces are not sufficient for accurate 3D PSF synthesis, so structural depth is a necessary part of the design.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's rule that N approximately 2 Ni No features suffice is inferred from random target matrices at modest sizes; scaling it to large volumes is an extrapolation, and a proof or a larger-scale test would be needed before treating the universality claim as quantitative.
  • Because the output pixels are multiplexed across input planes and wavelengths, the achievable field of view shrinks as the number of axial planes and colors grows; the paper does not address this information-capacity trade-off explicitly.
  • A direct experimental falsification would be to fabricate one of the optimized K=4 designs and measure the full transformation on random test inputs, since the paper's demonstrations are entirely numerical.
  • The same intensity-linear architecture suggests a route to depth-resolved fluorescence lifetime or polarimetric imaging if the programmed PSFs encode those additional degrees of freedom, but that extension is not explored here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a method for universal point spread function (PSF) engineering in three dimensions using spatially incoherent diffractive optical processors composed of cascaded transmissive surfaces. The central claim is that, given sufficient optimizable phase-only features N distributed over K≥4 surfaces, such a processor can approximate, with negligible error, any arbitrarily defined set of spatially varying 3D PSFs, corresponding to an arbitrary nonnegative linear intensity transformation between input and output volumes. The authors numerically demonstrate this by training diffractive networks to approximate random nonnegative target matrices, reporting an empirical threshold of N≈2N_iN_o for low error. They also present numerical demonstrations of snapshot 3D imaging and snapshot multispectral 3D imaging using the engineered PSFs. The physical assumptions are explicitly stated: the input emitters are independent and non-interacting, so the system is linear in intensity.

Significance. If the central claim holds, the work would substantially extend PSF engineering from 2D pupil-plane masks to arbitrary 3D spatially varying PSFs, enabling all-optical 3D and multispectral imaging without axial scanning, spectral filters, or digital reconstruction. This would be a meaningful advance for computational imaging, microscopy, and optical information processing. The numerical demonstrations in Figs. 4 and 5 show that the proposed approach can realize these applications at least at a small scale. However, the universality claim rests on an empirical, unproven design rule, and the paper provides no code, data, or repeated-trial statistics to support the generality of the results. The significance is therefore conditional on strengthening the numerical evidence and carefully qualifying the scope of the universality claim.

major comments (3)
  1. [Results (Fig. 2b)] The manuscript reports only numerical simulations and contains no code, data, or data-availability statement. Because the universality claim rests entirely on the numerical results in Figs. 2–5, the absence of reproducible artifacts hinders independent verification of the convergence threshold and the reported error values. I request that the authors release the simulation code and the trained network parameters, or at minimum provide a detailed specification of the forward model and optimization hyperparameters in the main text or supplementary material.
  2. [Discussion] The linearity assumption—independent, non-interacting emitters with no shadowing or re-excitation—is explicitly acknowledged in the Discussion, and the authors correctly note that if this assumption is violated the transformation becomes nonlinear and object-dependent. However, the abstract and introduction present the method as 'universal PSF engineering' without prominently stating this restriction. The universality claim should be qualified in the abstract and introduction to make clear that it applies only to spatially incoherent, non-interacting emitters, which is a meaningful constraint for many imaging scenarios.
  3. [Figs. 4 and 5] The application demonstrations use very small volumes: Fig. 4 uses 4 input planes each discretized into 6×6 pixels, and Fig. 5 uses 3 input planes at 3 wavelengths with 6×6 pixels per plane. The required number of features N≈2N_iN_o grows quadratically with the product of input and output voxel counts, which would become impractically large for realistic imaging volumes (e.g., 10^6 input and output voxels would require on the order of 2×10^12 features). The paper does not discuss this scalability constraint, yet it is central to the practical significance of the claimed universality. I ask the authors to address the scalability of the approach and to clarify the range of problem sizes for which the design rule is practically feasible.
minor comments (4)
  1. [Discussion (Fig. 5 paragraph)] There is a typo in the sentence 'see Fig. 5c) )'—an extra closing parenthesis appears after 'Fig. 5c'.
  2. [Abstract] The phrase 'rigorously analyze' overstates the nature of the evidence: the paper provides numerical demonstrations, not a formal mathematical proof. Suggest rewording to 'numerically analyze' or 'characterize'.
  3. [Throughout] The main text references 'Methods' for the forward model and optimization details, but the Methods section is placed in the Supplementary Information. The main text should explicitly direct readers to the Supplementary Methods, especially since the Methods are not included in the main manuscript.
  4. [Fig. 3] For the diffraction-limit study, it would be informative to overlay the theoretical axial resolution limit (e.g., 2λ/NA²) on the error curves in Fig. 3a and 3b to allow a direct quantitative comparison between the numerical error onset and the expected diffraction-limited resolution.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: the random-target optimization benchmark is externally specified, and the 3D demonstrations are independent numerical experiments.

full rationale

The paper's derivation chain is not circular. The central claim is that a spatially incoherent diffractive processor can approximate any prescribed nonnegative linear intensity transformation A, whose columns are treated as the desired 3D PSFs. The targets in Fig. 2a are random nonnegative matrices generated independently of the network, and the optimizer is evaluated against that externally specified target, so the benchmark is not derived from the network itself. The reported low values of ||A - Ahat|| are training losses for separately optimized devices, one per (K,N) point, but for an inverse-design feasibility claim this is the appropriate validation: the task is to synthesize a specified transformation, not to generalize to unseen transformations. The DOF threshold N approximately 2NiNo is an empirical observation from Fig. 2b rather than a parameter fitted to the claim, and the factor of 2 is given a heuristic degrees-of-freedom rationale; it is an extrapolation risk but not a tautology. The linearity assumption o = Ai is explicitly stated and physically motivated by spatial incoherence and non-interacting emitters, and the Discussion flags where it fails. Prior self-citations (refs 27-30) supply context and 2D results, but the 3D random-target simulations and imaging demonstrations are new and independent of those citations. No equation or definition is shown to reduce to its own input, so no circular step is identified.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim requires the linear incoherent intensity model, diffraction-limited voxel discretization, and an empirical DOF threshold; none of these are derived in the visible text. No new entities are postulated. The main free parameters are the hand-chosen simulation geometry and the empirical factor 2 in the DOF rule, which is read from Fig. 2b rather than proven.

free parameters (3)
  • Empirical DOF threshold factor (2 in N approximately 2N_iN_o) = 2
    The factor of 2 is asserted as a convergence threshold and read from Fig. 2b, not derived from information-theoretic bounds; it acts as a design rule for all subsequent simulations.
  • Volume discretization parameters (d_pp, d_i=d_o, wavelengths, pixel counts) = d_pp = 2.67 to 3 lambda; d_i = d_o = 10.5 or 21 lambda; 6x6 pixels per plane; 3-4 planes
    These are hand-picked simulation geometries; the diffraction-limit analysis shows errors grow when d_pp is below the axial resolution, so the reported success depends on these choices.
  • Target matrix A entries = Uniform random in [0,1]
    A is randomly sampled to represent 'any' transformation; results may not generalize to structured or adversarial target matrices.
assumptions (5)
  • domain assumption Spatially incoherent emitters are independent and non-interacting, so the optical system is linear in intensity: o = A i.
    Stated in the Introduction and Discussion; if emitters shadow, scatter, or excite each other, the model becomes nonlinear and object-dependent.
  • domain assumption Input and output volumes can be discretized into diffraction-limited voxels at approximately lambda/2 sampling, with plane spacing d_pp and distances d_i, d_o defining the NA.
    Used throughout Figs. 2-5; the diffraction limit is invoked as the resolution limit for independent channels.
  • ad hoc to paper A phase-only diffractive processor with N approximately 2N_iN_o optimizable features distributed over K >= 4 surfaces can approximate arbitrary nonnegative linear transformations.
    This is the core capability claim; supported only by numerical experiments in Fig. 2, not by a theorem, and prior 2D results are cited as justification.
  • standard math Angular-spectrum (Rayleigh-Sommerfeld) free-space diffraction plus incoherent intensity addition is the correct forward model.
    Standard Fourier optics; referenced via Goodman and prior diffractive network papers, not derived in this manuscript.
  • standard math Axial resolution limit approximately 2 lambda / NA^2 bounds distinguishable input planes.
    Used in Fig. 3b to explain error onset; standard diffraction theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universal point spread function engineering for 3D optical information processing." pith.science (2026). https://pith.science/paper/ICM7OY27

@misc{pith2026250206025,
  author       = {Pith},
  title        = {Pith review of: Universal point spread function engineering for 3D optical information processing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICM7OY27}},
  note         = {Machine review of arXiv:2502.06025}
}
read the original abstract

Point spread function (PSF) engineering has been pivotal in the remarkable progress made in high-resolution imaging in the last decades. However, the diversity in PSF structures attainable through existing engineering methods is limited. Here, we report universal PSF engineering, demonstrating a method to synthesize an arbitrary set of spatially varying 3D PSFs between the input and output volumes of a spatially incoherent diffractive processor composed of cascaded transmissive surfaces. We rigorously analyze the PSF engineering capabilities of such diffractive processors within the diffraction limit of light and provide numerical demonstrations of unique imaging capabilities, such as snapshot 3D multispectral imaging without involving any spectral filters, axial scanning or digital reconstruction steps, which is enabled by the spatial and spectral engineering of 3D PSFs. Our framework and analysis would be important for future advancements in computational imaging, sensing and diffractive processing of 3D optical information.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 29 canonical work pages

  1. [1]

    Goodman, J. W. Introduction to Fourier Optics. (Roberts and Company Publishers, 2005)

    work page 2005

  2. [2]

    & Rajora, S

    Khare, K., Butola, M. & Rajora, S. Fourier Optics and Computational Imaging. (Springer International Publishing, Cham, 2023). doi:10.1007/978-3-031-18353-9. 7

    work page doi:10.1007/978-3-031-18353-9 2023

  3. [3]

    & Díaz, A

    Ojeda-Castañeda, J., Andrés, P. & Díaz, A. Annular apodizers for low sensitivity to defocus and to spherical aberration. Opt. Lett. 11, 487–489 (1986)

    work page 1986

  4. [4]

    & Berriel-Valdos, L

    Ojeda-Castaneda, J. & Berriel-Valdos, L. R. Zone plate for arbitrarily high focal depth. Appl. Opt. 29, 994–997 (1990)

    work page 1990

  5. [5]

    Dowski, E. R. & Cathey, W. T. Extended depth of field through wave-front coding. Appl. Opt. 34, 1859–1866 (1995)

    work page 1995

  6. [6]

    Hell, S. W. Increasing the Resolution of Far-Field Fluorescence Light Microscopy by Point-Spread- Function Engineering. in Topics in Fluorescence Spectroscopy: Volume 5: Nonlinear and Two-Photon- Induced Fluorescence (ed. Lakowicz, J. R.) 361–426 (Springer US, Boston, MA, 2002). doi:10.1007/0- 306-47070-5_9

    doi:10.1007/0- 2002

  7. [7]

    & Zhuang, X

    Huang, B., Wang, W., Bates, M. & Zhuang, X. Three-Dimensional Super-Resolution Imaging by Stochastic Optical Reconstruction Microscopy. Science 319, 810–813 (2008)

    work page 2008

  8. [8]

    E., Backer, A

    Shechtman, Y., Weiss, L. E., Backer, A. S., Lee, M. Y. & Moerner, W. E. Multicolour localization microscopy by point-spread-function engineering. Nat. Photonics 10, 590–594 (2016)

    work page 2016

Show all 37 references
  1. [9]

    Recent advances in point spread function engineering and related computational microscopy approaches: from one viewpoint

    Shechtman, Y. Recent advances in point spread function engineering and related computational microscopy approaches: from one viewpoint. Biophys. Rev. 12, 1303–1309 (2020)

    2020

  2. [10]

    Opatovski, N. et al. Multiplexed PSF Engineering for Three-Dimensional Multicolor Particle Tracking. Nano Lett. 21, 5888–5895 (2021)

    2021

  3. [11]

    Lee, S.-H. et al. Increasing the storage density of a page-based holographic data storage system by image upscaling using the PSF of the Nyquist aperture. Opt. Express 19, 12053–12065 (2011)

    2011

  4. [12]

    Kim, D. et al. Advances in optical engineering for future telescopes. Opto-Electron. Adv. 4, 210040– 24 (2021)

    2021

  5. [13]

    E., Michaeli, T

    Hershko, E., Weiss, L. E., Michaeli, T. & Shechtman, Y. Multicolor localization microscopy and point- spread-function engineering by deep learning. Opt. Express 27, 6158–6183 (2019). 8

    2019

  6. [14]

    Nehme, E. et al. DeepSTORM3D: dense 3D localization microscopy and PSF design by deep learning. Nat. Methods 17, 734–740 (2020)

    2020

  7. [15]

    Bai, B. et al. To image, or not to image: class-specific diffractive cameras with all-optical erasure of undesired objects. eLight 2, 14 (2022)

    2022

  8. [16]

    Li, J. et al. Rapid sensing of hidden objects and defects using a single-pixel diffractive terahertz sensor. Nat. Commun. 14, 6791 (2023)

    2023

  9. [17]

    Bai, B. et al. Information-hiding cameras: Optical concealment of object information into ordinary images. Sci. Adv. 10, eadn9420 (2024)

    2024

  10. [18]

    Sakib Rahman, M. S. & Ozcan, A. Integration of Programmable Diffraction with Digital Neural Networks. ACS Photonics 11, 2906–2922 (2024)

    2024

  11. [19]

    Lin, X. et al. All-optical machine learning using diffractive deep neural networks. Science 361, 1004– 1008 (2018)

    2018

  12. [20]

    Hu, J. et al. Diffractive optical computing in free space. Nat. Commun. 15, 1525 (2024)

    2024

  13. [21]

    Sakib Rahman, M. S. & Ozcan, A. Computer-Free, All-Optical Reconstruction of Holograms Using Diffractive Networks. ACS Photonics 8, 3375–3384 (2021)

    2021

  14. [22]

    & Ozcan, A

    Mengu, D. & Ozcan, A. All-Optical Phase Recovery: Diffractive Computing for Quantitative Phase Imaging. Adv. Opt. Mater. 10, 2200281 (2022)

    2022

  15. [23]

    Işıl, Ç. et al. Super-resolution image display using diffractive decoders. Sci. Adv. 8, eadd3433 (2022)

    2022

  16. [24]

    Bai, B. et al. Data-Class-Specific All-Optical Transformations and Encryption. Adv. Mater. 35, 2212091 (2023)

    2023

  17. [25]

    Rahman, M. S. S. et al. Learning diffractive optical communication around arbitrary opaque occlusions. Nat. Commun. 14, 6830 (2023)

    2023

  18. [26]

    Li, J. et al. All-optical complex field imaging using diffractive processors. Light Sci. Appl. 13, 120 (2024). 9

    2024

  19. [27]

    & Ozcan, A

    Kulce, O., Mengu, D., Rivenson, Y. & Ozcan, A. All-optical information-processing capacity of diffractive surfaces. Light Sci. Appl. 10, 25 (2021)

    2021

  20. [28]

    & Ozcan, A

    Kulce, O., Mengu, D., Rivenson, Y. & Ozcan, A. All-optical synthesis of an arbitrary linear transformation using diffractive surfaces. Light Sci. Appl. 10, 196 (2021)

    2021

  21. [29]

    Li, J. et al. Massively parallel universal linear transformations using a wavelength-multiplexed diffractive optical network. Adv. Photonics 5, 016003 (2023)

    2023

  22. [30]

    Rahman, M. S. S., Yang, X., Li, J., Bai, B. & Ozcan, A. Universal linear intensity transformations using spatially incoherent diffractive processors. Light Sci. Appl. 12, 195 (2023)

    2023

  23. [31]

    Wang, H. et al. Deep learning enables cross-modality super-resolution in fluorescence microscopy. Nat. Methods 16, 103–110 (2019)

    2019

  24. [32]

    Wu, Y. et al. Three-dimensional virtual refocusing of fluorescence microscopy images using deep learning. Nat. Methods 16, 1323–1331 (2019)

    2019

  25. [33]

    & Ozcan, A

    Li, Y., Li, J. & Ozcan, A. Nonlinear encoding in diffractive information processing using linear optical materials. Light Sci. Appl. 13, 173 (2024)

    2024

  26. [34]

    Yang, X., Rahman, M. S. S., Bai, B., Li, J. & Ozcan, A. Complex-valued universal linear transformations and image encryption using spatially incoherent diffractive networks. Adv. Photonics Nexus 3, 016010 (2024)

    2024

  27. [35]

    Ma, G. et al. Unidirectional imaging with partially coherent light. Adv. Photonics Nexus 3, 066008 (2024)

    2024

  28. [36]

    Saleh, B. E. A. & Teich, M. C. Fundamentals of Photonics. (John Wiley & Sons, 2019)

    2019

  29. [37]

    Paszke, A. et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. in Advances in Neural Information Processing Systems vol. 32 (Curran Associates, Inc., 2019). 10 Figures Fig. 1 3D PSF engineering using a spatially incoherent diffractive optical processor...

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.