Pith. sign in

REVIEW 3 major objections 5 minor 52 references

What should a linear optical frontend compute? Assessing the role of meta-optics, nonlocality, and coherence in hybrid inference systems

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A linear optical frontend can beat the best trained linear preprocessor only when it is nonlocal and coherent.

desk verdict A clean simulation study that identifies when and why a linear optical frontend helps classification, but the headline super-linear advantage rests on an idealized operator whose physical realizability is left open. read the letter →

arxiv 2608.10304 v1 pith:GO3FIXG7 submitted 2026-08-10 physics.optics

Nicholas Behrens , Yandong Li , Francesco Monticone This is my paper · ORCID
classification physics.optics
keywords hybridoptical-electronicinferencelinearopticalfrontendBhattacharyyadistancenonlocalitycoherenceintensitydetectionmetasurfacequadraticfeatures
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

Hybrid inference systems that send light through an optical frontend before a digital classifier need to know what the optics should compute. The paper shows that the right goal is reshaping the class-conditional statistics of the sensor readout, and that a training-free statistical distance, the Bhattacharyya distance, measures this reshaping and predicts downstream accuracy. Because a camera records intensity, any linear field transformation produces features that are quadratic in the input field; the extra discriminative information lives in interference cross-terms that only nonlocal, coherent optics can access. The paper's central result is that such systems can surpass the best trained linear preprocessor of the same dimension, with an idealized unconstrained linear operator reaching 93% accuracy at a 2x2 sensor against an 87% linear ceiling.

What carries the argument

The central object is the sensor readout equation $I_j = \sum_k |M_{jk}|^2 |u_k|^2 + \sum_{k\neq l} M_{jk} M_{jl}^* u_k u_l^*$, which shows that intensity detection makes a linear field map $M$ quadratic in the input. The diagonal sum is a linear intensity map, while the interference cross-terms carry class-discriminative information only when $M$ couples distinct input points onto one pixel (nonlocality) and the input has a deterministic relative phase (coherence). The paper pairs this with the Bhattacharyya distance $D_B = -\ln \int \sqrt{p(y)q(y)}\,dy$, a training-free separability metric whose Bayes-error bound makes it a task-relevant predictor of accuracy. The machinery works by showing that correlation-aware joint separability, not per-pixel separability, tracks downstream accuracy, with a Spearman rank correlation of $\rho = 0.97$ across the eight 2x2 configurations.

What would settle it

Simulate or fabricate a passive volumetric linear optical frontend whose field operator approximates the unconstrained matrix operator on the MNIST task with a 2x2 sensor. If coherent readout accuracy does not exceed the 87% trained-linear ceiling, or if a local metasurface alone already matches the general operator under coherent light, the central super-linear advantage would be refuted.

Watch

Extended reading notes

Core claim

Under coherent illumination, a sufficiently general nonlocal linear operator can exceed the accuracy of any linear preprocessor because the intensity readout is quadratic in the input field. The quadratic cross-terms in Eq. (3) let the frontend encode class information in inter-pixel correlations rather than per-pixel intensities: a local metasurface raises downstream accuracy from 0.52 to 0.80 while leaving per-pixel Bhattacharyya separability essentially unchanged, with joint separability rising from 0.76 to 1.64. An unconstrained complex matrix operator under coherent light reaches 93% at a 2x2 sensor, beating the trained linear ceiling of 87%, while under incoherent light the same operator collapses to the metasurface level because the cross-terms average away. The paper therefore identifies an empirical upper bound for linear optical preprocessing and locates the physical resources needed to approach it: coherence combined with general nonlocal field mixing.

Load-bearing premise

The super-linear claim assumes that an unconstrained complex matrix operator can be realized, at least approximately, by a physical passive linear optical device; the paper identifies this as an open question rather than a demonstrated device.

Editorial extensions

If this is right

  • Optical frontends only help when the sensor is a strong bottleneck; performance gains saturate for sensor sizes $N \gtrsim 6$, since optics cannot add information, only re-encode it.
  • Under incoherent illumination the readout is a non-negative linear map of input intensity, so no amount of operator generality can beat the linear ceiling; the coherence gap and the field-mixing gap vanish together.
  • Trainable k-space amplitude filters, on either side of the mask, do not improve inference accuracy for $N \geq 2$ because amplitude attenuation removes task-relevant photons or cross-term weight rather than adding class separation.
  • A per-pixel separability metric is blind to the optical advantage; only joint, correlation-aware measures such as the Bhattacharyya distance predict accuracy.
  • The best linear preprocessor of matched dimensionality (87% at 2x2) is a digital reference that requires signed weights, while physical intensity readouts are non-negative; coherent nonlocal field mixing reaches 93%.
  • pith_inferences:
  • Extensions the paper leaves implicit: a training-free separability objective could replace end-to-end co-optimization, optimizing the frontend alone to maximize joint Bhattacharyya distance before training the backend; the paper suggests this route but does not test it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript studies hybrid optical-electronic classifiers in which a linear optical frontend (a phase mask, a metasurface with an added nonlocal k-space filter, or an idealized general operator) transforms a coherent or incoherent field before intensity detection and an MLP backend. The central thesis is that a well-designed frontend improves classification accuracy by reshaping the joint readout statistics, quantified by the Bhattacharyya distance, which rank-orders configurations consistently with measured accuracy (Spearman rho = 0.97 at 2x2). Because the sensor measures intensity, the readout is quadratic in the input field; the quadratic cross-terms require coherence and nonlocal field mixing. The authors report that an unconstrained complex matrix operator under coherent illumination reaches 93% at a 2x2 sensor, surpassing a trained linear preprocessor ceiling of 87%, while the physically modeled metasurface frontends reach 80% and incoherent systems 71%. They interpret this as evidence that coherent nonlocal systems can yield super-linear performance, with the gains concentrated in the small-sensor bottleneck regime.

Significance. If the upper-bound result holds, the paper offers a useful design principle for hybrid optical-electronic inference: optimize linear optical frontends to maximize joint Bhattacharyya separability, a correlation-aware, training-free metric. The demonstration that per-pixel separability can be unchanged while joint separability and accuracy improve (Fig. 2) is valuable and well supported. The paper also carefully distinguishes inference from imaging, and the claim that intensity detection is a computational resource for classification is physically interesting. Strengths include the nonparametric validation of the Gaussian separability estimate at 2x2 (Spearman 0.89-0.99 between Gaussian and nonparametric pairwise ranks, and 0.97 against accuracy) and the clearly specified simulation and training protocol. The principal weakness is that the headline 'super-linear' claim rests on an idealized operator class whose physical realizability is left open, which limits the strength of the conclusions as currently stated.

major comments (3)
  1. [Abstract; Sec. I.D.2; Sec. III.F] The claim that coherent nonlocal optical systems can surpass the best trained linear preprocessor is load-bearing but is supported only by the unconstrained complex matrix M in C^784x784 of Sec. III.F. This operator is trained without reciprocity, causality, or restrictions on the number of accessible spatial channels or the locality of the material response; the body explicitly leaves open which physical properties a realizing device would need (Sec. I.D.2). The abstract's statement that 'such systems can yield significant performance gains' is therefore stronger than the demonstrated 'an idealized operator reaches 93%.' Please either constrain M to a physically plausible class (e.g., complex symmetric/reciprocal scattering matrices, matrices generated by a finite-thickness local dielectric slab, or operators with explicit channel limitations) and show that the 5-6% advantage over the 87% linear ceiling survives, or reframe the abstract, title, and discussion so that the super-linear advantage is attributed to an idealized operator class as an empirical upper bound rather than to physical systems.
  2. [Sec. III.F; Sec. I.D.2] The batch-normalization argument correctly removes the overall scale of M, so rescaling to norm <= 1 addresses passivity in the operator-norm sense. It does not, however, address structural constraints: a reciprocal passive medium would require a symmetric scattering matrix in an appropriate basis, and a physical Green's function must satisfy causality and limited transverse channel count. A concrete test would be to train M restricted to complex symmetric matrices, or to matrices produced by a slab of local dielectric material with finite thickness, and to report whether the 93% accuracy at 2x2 persists. Without such a test, the super-linear performance could be an artifact of nonphysical nonreciprocal or acausal transformations.
  3. [Fig. 4; Sec. I.D.2] The manuscript's own Fig. 4 shows that the physically modeled coherent frontends, including the metasurface with and without a nonlocal k-space filter, do not surpass the trained linear ceiling; only the unconstrained operator does. This discrepancy between the modeled physical systems and the headline claim is central. The Discussion should state explicitly that no currently modeled physical frontend achieves the super-linear advantage, and should identify arbitrary nonlocal field routing as a hypothesis for future work rather than a demonstrated capability.
minor comments (5)
  1. [Sec. I.A] The term 'linear optical frontend' could mislead because the intensity readout is quadratic; please state explicitly at first use that the frontend is linear in the field, not in the measured intensity.
  2. [Sec. III.F] The statement that 'real and complex operators give statistically indistinguishable accuracy' should report the actual accuracies and a measure of spread across seeds, rather than only a qualitative claim.
  3. [Fig. 3] For the LDA projections in Fig. 3(b), please state whether the projection was fit on the readout distribution and whether the same projection is used for all panels; the caption currently leaves this unclear.
  4. [Sec. III.J] For a computational study of this type, a public repository containing the trained-model readouts and analysis scripts would substantially strengthen reproducibility; 'available from the corresponding authors on reasonable request' is generally insufficient for a modern physics-optics journal.
  5. [Sec. I.D.1] The statement that under incoherent illumination 'intensity kernels are non-negative' is correct, but please clarify that this refers to the intensity point-spread function (the autocorrelation of the field kernel), not to the field kernel itself.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the separability metric is independently defined and used post-hoc, and the super-linear claim is explicitly labeled an empirical upper bound over a stated operator class.

full rationale

The paper's derivation chain is self-contained against its own simulations and external benchmarks. The Bhattacharyya distance (Eq. 5 and Methods III.H) is defined from class-conditional readout statistics and bounds Bayes error; it is not used to fit any optical or backend parameter, and the Spearman correlation with accuracy is an empirical validation rather than a training input. The trained linear ceiling is an independently trained unconstrained linear preprocessor (Methods III.G), while the general operator is a separately trained complex matrix (Methods III.F); comparing them is the paper's result, not a construction. Equation (3) is a mathematical identity for intensity detection, and the necessity of nonlocality and coherence for the interference cross-terms follows directly from the definitions of those cross-terms rather than being imported by fiat. The paper explicitly labels the general-operator result an 'empirical (and possibly loose) upper bound' and states that physical realization remains an open question, so the acknowledged realizability gap is a physical caveat, not a circular step. Self-citations are used for background or analogy and are not load-bearing for the derivation. No self-definitional reduction, fitted-input-as-prediction, or author-imported uniqueness theorem is present.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claims rest on fitted trainable parameters (phase profile, k-space filter, general operator, trained linear frontend, backend weights) and on physical modeling assumptions about scalar diffraction, intensity detection, and especially the realizability of the unconstrained operator. No new physical entities are introduced.

free parameters (5)
  • Zernike phase coefficients w_j (j=1..210) = Trained end-to-end, values not reported in text
    Parameterize the local metasurface phase profile in Section III.C; fitted to the classification loss on MNIST.
  • k-space filter center kappa_0 and width 2 Delta kappa = Trained, with kappa_0 in [0.2,0.8] and 2 Delta kappa in [0.4,1.4]
    Trainable annular band-pass filter (Eq. (6)); fitted to the classification loss.
  • General operator entries M in C^784x784 = Trained, about 1.2e6 real parameters
    Idealized volumetric frontend (Section III.F); trained end-to-end; defines the empirical upper bound.
  • Trained linear frontend W in R^784x784 = Trained
    Digital linear ceiling benchmark with signed weights (Section III.G).
  • Backend MLP weights = Trained
    Two 256-neuron hidden layers; trained jointly with the frontend.
assumptions (6)
  • domain assumption Scalar diffraction and angular spectrum propagation accurately model the optical frontend (Eq. (1)).
    Used throughout the optical simulation; standard in meta-optics but neglects vectorial and near-field effects.
  • domain assumption The sensor measures intensity as |u_out|^2 with average pooling per pixel (Eq. (3)).
    This is the physical model of the readout; the quadratic feature structure follows from it.
  • standard math Any linear optical frontend can be represented as a finite matrix M (Hilbert-Schmidt compactness), and the unconstrained trained M bounds all linear frontends on the task.
    Used to justify the general operator as an upper bound; the bound is over the matrix class rather than over physical devices.
  • domain assumption The unconstrained complex matrix M is realizable, or approachable, by a physical linear optical device.
    Load-bearing for the super-linear claim; the paper flags this as open in Section I.D.2.
  • domain assumption MNIST and the chosen sensor sizes are representative of hybrid inference tasks.
    Conclusions are drawn from one dataset; only an initial Fashion-MNIST check is mentioned in the SI.
  • domain assumption The multivariate Gaussian model adequately approximates the joint readout statistics for separability estimation.
    Validated nonparametrically only at 2x2 (rho = 0.97); used for all joint separability numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of What should a linear optical frontend compute? Assessing the role of meta-optics, nonlocality, and coherence in hybrid inference systems." pith.science (2026). https://pith.science/paper/GO3FIXG7

@misc{pith2026260810304,
  author       = {Pith},
  title        = {Pith review of: What should a linear optical frontend compute? Assessing the role of meta-optics, nonlocality, and coherence in hybrid inference systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GO3FIXG7}},
  note         = {Machine review of arXiv:2608.10304}
}
read the original abstract

Hybrid inference systems that pair an optical frontend with a digital backend offer a route to offload computation to the physical layer. Yet what the optics should compute, and when this is beneficial, has remained unclear. Here, we show that, for classification tasks, a well-designed optical frontend reshapes the statistics of the sensor intensity readout to improve class separability, quantified by the Bhattacharyya distance. This training-free metric predicts downstream accuracy and reveals that much of the discriminative information resides in inter-pixel correlations. We then identify the roles of coherence and different forms of nonlocality. Because the sensor measures intensity, a linear frontend produces features that are quadratic in the input field; however, only nonlocal, coherent optical systems can exploit the associated information. Such systems can yield significant performance gains, surpassing the best trained linear preprocessor. These results provide physical insights and new design principles for optimal optical--electronic inference systems.

Figures

Figures reproduced from arXiv: 2608.10304 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages

  1. [21]

    Kienesberger, Z

    L. Kienesberger, Z. Kuang, Y. Liu, and O. D. Miller, End-to-end meta-imagers: Information-theoretic objectives and generalized focusing optima, arXiv (2026), arXiv:2606.16724 [physics.optics]

    arXiv 2026

  2. [1]

    optical edge detection

    Nonlocalk-space amplitude filtering does not improve performance A nonlocal optical element that is transverse-shift-invariant—such as a multilayer slab or a metasurface with subwavelength period, for which the transverse wavevectorkis conserved—acts as ak-space transfer functionH(k) [14, 15, 31]. Modulating its amplitude|H(k)|filters spatial-frequency co...

    work page

  3. [2]

    quadratic ceiling,

    Coherence and nonlocal field-mixing: toward super-linear performance As mentioned in Sec. I A, a different kind of nonlocality does help: general nonlocal input-output responses that enable field mixing from different spatial points. In the previous sections, this type of nonlocality was provided entirely by free-space propagation, whose only design param...

    work page

  4. [3]

    generalized focusing

    The contrast with imaging and communication We emphasize again that super-linear performance is a coherent phenomenon. As Eq. (4) shows, under incoherent illumination the cross-terms average away, and the readout reduces to a non-negative linear map of input intensity, 10 whose best achievable separability lies below that of the best linear map. Reaching ...

    work page

  5. [4]

    Amodei and D

    D. Amodei and D. Hernandez, AI and Compute (2018)

    work page 2018

  6. [5]

    Hoffmann, S

    J. Hoffmann, S. Borgeaud, A. Mensch, E. Buchatskaya, T. Cai, E. Rutherford, D. de Las Casas, L. A. Hendricks, J. Welbl, A. Clark, T. Hennigan, E. Noland, K. Millican, G. van den Driessche, B. Damoc, A. Guy, S. Osindero, K. Simonyan, E. Elsen, J. W. Rae, O. Vinyals, and L. Sifre, An empirical analysis of compute-optimal large language model training, in Ad...

    work page 2022

  7. [6]

    P. L. McMahon, The physics of optical computing, Nature Reviews Physics5, 717 (2023)

    work page 2023

  8. [7]

    Li and F

    Y. Li and F. Monticone, Exploring the role of metamaterials in achieving advantage in optical computing, Nature Com- putational Science4, 545 (2024)

    work page 2024

Show all 52 references
  1. [8]

    Huang, Q

    L. Huang, Q. A. A. Tanguy, J. E. Fr¨ och, S. Mukherjee, K. F. B¨ ohringer, and A. Majumdar, Photonic advantage of optical encoders, Nanophotonics13, 1191 (2024)

    2024

  2. [9]

    K. Wei, X. Li, J. Froech, P. Chakravarthula, J. Whitehead, E. Tseng, A. Majumdar, and F. Heide, Spatially varying nanophotonic neural networks, Science Advances10, eadp0391 (2024)

    2024

  3. [10]

    Q. Liu, B. T. Swartz, I. Kravchenko, J. G. Valentine, and Y. Huo, Extrememeta: High-speed lightweight image segmentation model by remodeling multi-channel metamaterial imagers, arXiv preprint arXiv:2405.17568 (2024)

    2024 arXiv

  4. [11]

    M. Choi, J. Xiang, Y. Zhang, Z. Zhou, E. Shlizerman, and A. Majumdar, Meta-optical encoder for image segmentation, Nano Letters26, 9287 (2026)

    2026

  5. [12]

    Tseng, S

    E. Tseng, S. Colburn, J. Whitehead, L. Huang, S.-H. Baek, A. Majumdar, and F. Heide, Neural nano-optics for high-quality thin lens imaging, Nature Communications12, 6493 (2021)

    2021

  6. [13]

    Z. Lin, C. Roques-Carmes, R. Pestourie, M. Soljaˇ ci´ c, A. Majumdar, and S. G. Johnson, End-to-end nanophotonic inverse design for imaging and polarimetry, Nanophotonics10, 1177 (2021)

    2021

  7. [14]

    J. Peng, M. Luo, Y. Han, S. Wu, H. Li, B. J. Shastri, C. Shu, Q. Dou, Y. Chai, and C. Huang, Optical metasurfaces for general vision processing on the edge, Nature654, 917 (2026)

    2026

  8. [15]

    Tishby, F

    N. Tishby, F. C. Pereira, and W. Bialek, The information bottleneck method (2000), arXiv:physics/0004057 [physics.data- an]

    2000 arXiv

  9. [16]

    Shwartz-Ziv and N

    R. Shwartz-Ziv and N. Tishby, Opening the black box of deep neural networks via information (2017), arXiv:1703.00810 [cs.LG]. 14

    2017 arXiv

  10. [17]

    Monticone, N

    F. Monticone, N. A. Mortensen, A. I. Fern´ andez-Dom ´ ınguez, Y. Luo, X. Zheng, C. Tserkezis, J. B. Khurgin, T. V. Shahbazyan, A. J. Chaves, N. M. Peres,et al., Nonlocality in photonic materials and metamaterials: roadmap, Optical Materials Express15, 1544 (2025)

    2025

  11. [18]

    Overvig and A

    A. Overvig and A. Al` u, Diffractive nonlocal metasurfaces, Laser & Photonics Reviews16, 2100633 (2022)

    2022

  12. [19]

    Pinkard, L

    H. Pinkard, L. Kabuli, E. Markley, T. Chien, J. Jiao, and L. Waller, Information-driven design of imaging systems, arXiv:2405.20559 (2024)

    2024

  13. [20]

    L. A. Kabuli, H. Pinkard, E. Markley, C. S. Hung, and L. Waller, Designing lensless imaging systems to maximize information capture, Optica13, 227 (2026)

    2026

  14. [22]

    Bhattacharyya, On a measure of divergence between two statistical populations defined by their probability distributions, Bulletin of the Calcutta Mathematical Society35, 99 (1943)

    A. Bhattacharyya, On a measure of divergence between two statistical populations defined by their probability distributions, Bulletin of the Calcutta Mathematical Society35, 99 (1943)

    1943

  15. [23]

    J. P. Sethna,Statistical mechanics: entropy, order parameters, and complexity, Vol. 14 (Oxford University Press Oxford, 2006)

    2006

  16. [24]

    A. Wang, J. Chen, S. Vaidya, and M. Soljaˇ ci´ c, End-to-end optimization of incoherent imaging for classification under detector-limited readout, arXiv (2026), arXiv:2606.09792 [cs.CV]

    2026 arXiv

  17. [25]

    Reshef, M

    O. Reshef, M. P. DelMastro, K. K. Bearne, A. H. Alhulaymi, L. Giner, R. W. Boyd, and J. S. Lundeen, An optic to replace space and its application towards ultra-thin imaging systems, Nature communications12, 3512 (2021)

    2021

  18. [26]

    C. Guo, H. Wang, and S. Fan, Squeeze free space with nonlocal flat optics, Optica7, 1133 (2020)

    2020

  19. [27]

    Chen and F

    A. Chen and F. Monticone, Dielectric nonlocal metasurfaces for fully solid-state ultrathin optical systems, ACS Photonics 8, 1439 (2021)

    2021

  20. [28]

    Pahlevaninezhad and F

    M. Pahlevaninezhad and F. Monticone, Multi-color spaceplates in the visible, ACS nano18, 28585 (2024)

    2024

  21. [29]

    Kailath, The divergence and Bhattacharyya distance measures in signal selection, IEEE Transactions on Communication Technology15, 52 (1967)

    T. Kailath, The divergence and Bhattacharyya distance measures in signal selection, IEEE Transactions on Communication Technology15, 52 (1967)

    1967

  22. [30]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman, Linear methods for classification, inThe Elements of Statistical Learning: Data Mining, Inference, and Prediction(Springer Science+Business Media, Llc, 2009) Chap. 4, pp. 101–37

    2009

  23. [31]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White, Multilayer feedforward networks are universal approximators, Neural Networks 2, 359 (1989)

    1989

  24. [32]

    D. Ruck, S. Rogers, M. Kabrisky, M. Oxley, and B. Suter, The multilayer perceptron as an approximation to a bayes optimal discriminant function, IEEE Transactions on Neural Networks1, 296 (1990)

    1990

  25. [33]

    Cover and J

    T. Cover and J. Thomas,Elements of Information Theory, 2nd ed. (John Wiley & Sons, Ltd, 2006)

    2006

  26. [34]

    Shastri and F

    K. Shastri and F. Monticone, Nonlocal flat optics, Nature Photonics17, 36 (2023)

    2023

  27. [35]

    H. Wang, C. Guo, Z. Zhao, and S. Fan, Compact incoherent image differentiation with nanophotonic structures, ACS Photonics7, 338 (2020)

    2020

  28. [36]

    B. T. Swartz, H. Zheng, G. T. Forcherio, and J. Valentine, Broadband and large-aperture metasurface edge encoders for incoherent infrared radiation, Science Advances10, eadk0024 (2024)

    2024

  29. [37]

    D. A. B. Miller, An introduction to functional analysis for science and engineering, arXiv preprint arXiv:1904.02539 (2019)

    2019 arXiv

  30. [38]

    O. D. Miller and F. Monticone, Fundamental limits in photonics and electromagnetics: a tutorial, arXiv preprint arXiv:2605.24738 (2026)

    2026 arXiv

  31. [39]

    F. J. Chen, A. Amaolo, P. Chao, S. Molesky, Z. Lin, and A. W. Rodriguez, Optically incoherent photonic mutual infor- mation, arXiv:2607.13153 (2026)

    2026 arXiv

  32. [40]

    Li and F

    Y. Li and F. Monticone, The spatial complexity of optical computing: toward space-efficient design, Nature Communica- tions16, 8588 (2025)

    2025

  33. [41]

    X. Yang, Q. Fu, Y. Nie, and W. Heidrich, Task-driven lens design, Opt. Express34, 8961 (2026)

    2026

  34. [42]

    Zheng, Q

    H. Zheng, Q. Liu, I. I. Kravchenko, X. Zhang, Y. Huo, and J. G. Valentine, Multichannel meta-imagers for accelerating machine vision, Nature nanotechnology19, 471 (2024)

    2024

  35. [43]

    Zhang, B

    X. Zhang, B. Bai, H.-B. Sun, G. Jin, and J. Valentine, Incoherent optoelectronic differentiation based on optimized multilayer films, Laser & Photonics Reviews16, 2200038 (2022)

    2022

  36. [44]

    Bengio, A

    Y. Bengio, A. Courville, and P. Vincent, Representation learning: A review and new perspectives, IEEE Transactions on Pattern Analysis and Machine Intelligence35, 1798 (2013)

    2013

  37. [45]

    Y. Hu, H. Chi, and H. Duan, Metaoptics merging computational optics and optical computing toward intelligent visual perception, Science Advances12, eaea8941 (2026)

    2026

  38. [46]

    K. P. Kalinin, J. Gladrow, J. Chu, J. H. Clegg, D. Cletheroe, D. J. Kelly, B. Rahmani, G. Brennan, B. Canakci, F. Falck, M. Hansen, J. Kleewein, H. Kremer, G. O’Shea, L. Pickup, S. Rajmohan, A. Rowstron, V. Ruhle, L. Braine, S. Khedekar, N. G. Berloff, C. Gkantsidis, F. Parmig...

    2025

  39. [47]

    L. G. Wright, T. Wang, T. Onodera, and P. L. McMahon, Physical foundation models: Fixed hardware implementations of large-scale neural networks, arXiv (2026), arXiv:2604.27911 [cs.LG]

    2026 arXiv

  40. [48]

    C. R. Rao, Information and the accuracy attainable in the estimation of statistical parameters, Bulletin of the Calcutta Mathematical Society37, 81 (1945)

    1945

  41. [49]

    H. K. Miyamoto, F. C. C. Meneghetti, J. Pinele, and S. I. R. Costa, On closed-form expressions for the fisher–rao distance, Information Geometry7, 311 (2024)

    2024

  42. [50]

    L. T. Skovgaard, A Riemannian geometry of the multivariate normal model, Scandinavian Journal of Statistics11, 211 (1984). 15

    1984

  43. [51]

    Calvo and J

    M. Calvo and J. M. Oller, A distance between multivariate normal distributions based in an embedding into the Siegel group, Journal of Multivariate Analysis35, 223 (1990)

    1990

  44. [52]

    Pinele, J

    J. Pinele, J. E. Strapasson, and S. I. R. Costa, The Fisher–Rao distance between multivariate normal distributions: Special cases, bounds and applications, Entropy22, 404 (2020)

    2020

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.