Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Complexity-energy trade-off in programmable unitary interferometers

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper argues that the high programming cost of generic unitary interferometers is intrinsic, and that interferometers with cheap quadratic programming necessarily sacrifice up to a factor of n in useful output energy.

desk verdict A useful conceptual framing of programming cost vs. energy in unitary interferometers, but the "intrinsic" universal claim rests on an unproven complexity conjecture and a gap in the SVT-byproduct argument. read the letter →

arxiv 2507.22972 v1 pith:LVCI2DAR submitted 2025-07-30 physics.optics cs.ET

classification physics.opticscs.ET
keywords programmableinterferometerssingularvaluetestunitaryembeddingprogrammingcomplexitycoherentcrossbaropticalmatrixmultiplicationenergyefficiencyMach-Zehndermeshes
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

The paper argues that the well-known difficulty of programming multiport optical interferometers is not an accident of specific architectures but an intrinsic feature of unitarity. Any device that can embed an arbitrary matrix $A$ with $\sigma_{\max}(A) \leq 1$ must, as a side effect of its programming algorithm, solve the singular value test, and exact deterministic singular value tests are believed to require superquadratic time. The paper then derives the other side of the trade-off: interferometers with cheap quadratic programming, such as coherent crossbars, embed a rescaled matrix $W/\alpha(W)$ using a cruder norm $\alpha$, and the ratio $\alpha(W)/\sigma_{\max}(W)$ can be as large as $n$, or $\sqrt{n}$ for the modified crossbar. Consequently, the feature that makes an architecture easy to program also caps its useful output energy, limiting signal-to-noise ratio and accuracy. If the paper is right, designers of optical matrix multipliers must choose between fast configuration and energy-efficient, accurate operation.

What carries the argument

The carrier of the argument is the singular value test (SVT), the decision problem that returns yes exactly when $\sigma_{\max}(A) \leq 1$. Any programming algorithm for a generic unitary interferometer that embeds all $n \times n$ matrices $A$ with $\sigma_{\max}(A) \leq 1$ and must fail for $\sigma_{\max}(A) > 1$ solves SVT as a byproduct, so its worst-case complexity is at least the complexity of SVT. The energy side is carried by the ratio $\alpha(W)/\sigma_{\max}(W)$, where $\alpha(W)$ is any rescaling factor that can be verified in quadratic time. For the coherent crossbar, $\alpha(W) = n \max_{ij} |W_{ij}|$ with $1 \leq \alpha/\sigma_{\max} \leq n$; for the modified crossbar, $\alpha_{1,2}(W) = \sqrt{n}\|W\|_{1,2}$ with ratio at most $\sqrt{n}$. These inequalities translate the choice of embedding norm directly into a bound on useful output energy.

What would settle it

Find an exact deterministic $O(n^2)$ algorithm for the singular value test on arbitrary $n \times n$ matrices, or exhibit a quadratic-time unitary architecture whose embedding rescaling $\alpha(W)$ satisfies $\alpha(W)/\sigma_{\max}(W) = o(\sqrt{n})$ in the worst case; either would refute the paper's central trade-off.

Watch

Extended reading notes

Core claim

The central claim is that high programming complexity is intrinsic to any unitary interferometer that embeds arbitrary matrices with optimal energy. The paper formalizes embedding as placing $A$ in the top-left block of a larger unitary $U$, and notes that for $m \geq n$ auxiliary ports this is possible exactly when $\sigma_{\max}(A) \leq 1$. A programming algorithm that embeds every such $A$ and fails for $\sigma_{\max}(A) > 1$ therefore performs a singular value test as a byproduct, so its complexity is at least the complexity of that test. Since known exact deterministic singular value tests cost roughly $O(n^3)$, and all known quadratic-time algorithms are approximate or probabilistic, the $O(n^3)$ programming cost of SVD-based meshes is not a removable engineering defect. On the energy side, the paper proves that quadratic-time architectures must use a rescaling $A = W/\alpha(W)$ with $\alpha \geq \sigma_{\max}$, and bounds the penalty: $1 \leq \alpha(W)/\sigma_{\max}(W) \leq n$ for the coherent crossbar and $1 \leq \alpha_{1,2}(W)/\sigma_{\max}(W) \leq \sqrt{n}$ for the modified crossbar, with Eq. (10) conjecturing that $\sqrt{n}$ is the best any $O(n^2)$-computable scalar rescaling can achieve. The conclusion is an unavoidable complexity-energy trade-off.

Load-bearing premise

The whole conclusion rests on the belief, which the paper labels strong evidence rather than proof, that no exact fast algorithm can decide whether an arbitrary matrix's largest amplification factor is at most one.

Editorial extensions

If this is right

  • Any generic interferometer that preserves optimal output energy embeds $W/\sigma_{\max}(W)$ and therefore inherits superquadratic programming complexity, because a quadratic programming algorithm would supply an exact quadratic singular value test.
  • The coherent crossbar and its modified variant can be configured in $O(n^2)$ time, but their rescaling loses up to $n$ and $\sqrt{n}$ of the output energy of the optimal embedding, respectively; Eq. (10) states that no $O(n^2)$-computable scalar rescaling can improve the worst-case $\sqrt{n}$ bound.
  • Unitary programming of standard meshes is itself superquadratic, since any such algorithm must test whether the input matrix is unitary, and no deterministic quadratic unitarity test is known.
  • The trade-off carries over to free-space coherent optical multipliers, because the argument uses only the unitarity of the device and the block-embedding structure.
  • Approximate, randomized, or input-class-restricted programming can in principle run in quadratic time, so the bound applies to exact deterministic programming of arbitrary matrices.

Reading between the lines

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

  • If the singular-value-test hardness premise is eventually proved, the paper implies a fundamental floor for analog optical linear algebra: no unitary device can combine optimal energy efficiency with exact quadratic-time programming of arbitrary matrices.
  • The rescaling-norm viewpoint suggests a design spectrum in which interferometer architectures are classified by the matrix norm they implicitly use, and the Pareto front between programming cost and energy efficiency could be explored by interpolating between $\sigma_{\max}$ and entry-wise norms.
  • Under shot-noise-limited detection, a factor-$n$ loss of output energy costs roughly $\frac{1}{2}\log_2 n$ bits of effective amplitude resolution, making the energy penalty also a precision penalty; the paper states that accuracy is limited but does not quantify this bit-level consequence.
  • A practical escape route suggested by the paper's own limitations is randomized programming that checks unitarity on random vectors and accepts a small failure probability, shifting the trade-off from exact worst-case complexity to high-probability correctness.
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 / 4 minor

Summary. The manuscript argues that programming a generic unitary interferometer to implement an arbitrary matrix A with σmax(A) ≤ 1 requires at least the cost of a singular value test (SVT). Since exact deterministic SVT is believed to be superquadratic, the paper claims that the O(n^3) cost of SVD-based programming is intrinsic rather than an artifact of a particular architecture. Conversely, interferometers with O(n^2) programming procedures, such as coherent crossbars, embed only rescaled matrices W/α(W) for norms α that can exceed σmax(W), reducing useful output energy. The authors derive norm inequalities for αmax and α1,2 and formulate a complexity-energy trade-off.

Significance. If the central claims were established, the paper would identify a universal feature of photonic matrix-vector multipliers and would help guide architecture choice. The paper has clear strengths: the block-embedding condition and the norm inequalities in Appendices A and B are correct; the distinction between universal embeddings and norm-rescaled embeddings is sharply drawn; and the explicit rescaling analysis of crossbar architectures is a useful and falsifiable contribution. However, the two universal claims — the SVT lower bound and Eq. (10) — are unproven conjectures, and the paper itself acknowledges this in Section V. The significance therefore depends on whether the authors can either prove these conjectures or honestly reframe the conclusions as conditional.

major comments (3)
  1. [Section III, paragraph beginning 'Consider a programming algorithm'] The reduction of a programming algorithm to an SVT is not valid as stated. A programming algorithm is naturally only a promise algorithm: it is defined on matrices with σmax(A) ≤ 1, and for invalid inputs it need not signal an error; it may simply return phases that implement a different unitary. To conclude that any such algorithm performs an SVT, the authors must require that the algorithm either rejects invalid inputs or self-certifies its output. The cost of such rejection or self-certification is never included in the argument. Without this additional property, an O(n^2) promise algorithm for the valid set would not constitute an O(n^2) SVT, so the load-bearing step in Section III is currently a gap.
  2. [Section III and Section V] The central lower-bound premise — that exact deterministic SVT cannot be done in O(n^2) time — is stated as 'strong evidence' and 'to the best of our knowledge' rather than proved. Because the abstract's claim that high programming complexity is 'intrinsic' rests entirely on this premise, the paper's main conclusion is conditional. The authors should either supply a formal lower bound or prominently mark the conclusion as a conjecture in the abstract and throughout the main text. This is not a presentation issue: existence of an O(n^2) deterministic SVT would directly invalidate the paper's central claim.
  3. [Section IV, Eq. (10)] The assertion that every O(n^2)-computable upper bound α(W) on σmax(W) has worst-case ratio max_W α(W)/σmax(W) = sqrt(n) is unproven and is not supported by any reference or derivation. This statement is the basis for the claim that the energy efficiency of the modified crossbar in Fig. 4c cannot be asymptotically improved. The authors should either prove this for a precisely defined class of upper bounds, or weaken the claim to a conjecture or a statement about the specific norms they analyze. If there exists an O(n^2)-computable α with worst-case ratio o(sqrt(n)), the energy-side conclusion of the purported trade-off changes materially.
minor comments (4)
  1. [Section III, near 'matrix multiplication complexity O(2ω)'] The expression 'O(2ω)' appears to be a typo for O(n^ω); please correct it.
  2. [References [32] and [34]] References [32] and [34] appear to refer to the same JLT article by Giamougiannis et al.; please consolidate the duplicate.
  3. [Section III, diamond mesh discussion] The claim that programming the diamond mesh 'apparently' requires superquadratic time is not backed by a quantitative analysis; consider labeling it as a heuristic observation or providing a precise operation count.
  4. [Abstract and Section V] The abstract states that high programming complexity is 'intrinsic,' while Section V says the authors 'believe' the relevant lower bounds. Please align the abstract's certainty with the conjectural status of the complexity-theoretic premises.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central reduction and energy trade-off inequalities are self-contained, with self-citations used only as examples and the main caveats explicitly disclosed as unproven assumptions.

full rationale

The derivation chain is not circular. The central claim that any exact programming algorithm for embeddings with sigma_max(A) <= 1 performs a singular value test as a byproduct is a genuine reduction: the embeddability condition is derived from unitarity in Appendix A, and the algorithm's behavior on invalid inputs (completing correctly versus signaling an error) is what constitutes the test. This does not assume the conclusion; it only assumes an exact, deterministic, error-signaling algorithm, an assumption the paper explicitly states in Section V. The energy trade-off results in Section IV are derived from norm inequalities proven in Appendix B (e.g., alpha_max(W) >= sigma_max(W) and the bounds in Eqs. (6), (7), (9)), not from fitted parameters or from the conclusions they are used to support. Equation (10) is explicitly presented as a conjecture ('to the best of our knowledge'), so it is an unproven hypothesis about O(n^2)-computable upper bounds, not a conclusion that has been smuggled in through its own assumptions. The authors' self-citations, refs. [21] and [22], appear only as examples of alternative embedding schemes and are not load-bearing for the main argument. The paper also candidly lists limitations: the assumed exactness, determinism, and totality of the primitives, and the specific amplitude-channel encoding. These are assumptions, not circular definitions. Any weakness in the paper lies in the unresolved complexity of exact deterministic singular value tests and the conjectural nature of Eq. (10), which are correctness or evidence-strength concerns, not circularity.

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

No free parameters or invented entities are introduced. The derivations use standard linear algebra and unitarity. The load-bearing conjectures are the SVT complexity lower bound and the universality of the sqrt(n) ratio for quadratic-time upper bounds; both are stated honestly but not proven.

assumptions (5)
  • standard math Embedding an arbitrary n x n matrix A in an (n+m) x (n+m) unitary requires sigma_max(A) <= 1 when m >= n.
    Proved in Appendix A via SVD and positivity of C^dag C; standard linear algebra.
  • ad hoc to paper Exact deterministic singular value test cannot be done in quadratic time; likely requires at least matrix multiplication cost.
    Not proven in the paper; invoked as a conjecture to conclude programming complexity is intrinsic (Section III).
  • ad hoc to paper Every O(n^2)-computable upper bound alpha(W) on sigma_max has worst-case ratio max_W alpha/sigma_max = sqrt(n).
    Stated as 'to the best of our knowledge' after Eq. (10); no proof, and it is load-bearing for the claimed impossibility of asymptotic energy improvement.
  • domain assumption A correct programming algorithm must signal an error when sigma_max > 1.
    Needed for the SVT-byproduct argument in Section III; an algorithm that silently returns wrong phases would not perform the test.
  • domain assumption Passive fan-out and fan-in networks impose 1/sqrt(n) amplitude factors per stage.
    Used in Section IV for the crossbar rescaling factors; follows from unitarity and is standard physics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Complexity-energy trade-off in programmable unitary interferometers." pith.science (2026). https://pith.science/paper/LVCI2DAR

@misc{pith2026250722972,
  author       = {Pith},
  title        = {Pith review of: Complexity-energy trade-off in programmable unitary interferometers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LVCI2DAR}},
  note         = {Machine review of arXiv:2507.22972}
}
read the original abstract

Coherent multiport interferometers are a promising approach to realize matrix multiplication in integrated photonics. However, most known architectures - such as MZI and beamsplitter meshes, as well as more general interferometers - suffer from complicated procedures for mapping the matrix elements of the desired transformation to specific phaseshifts in the device. We point out that the high programming complexity is intrinsic, rather than accidental. At the same time, we argue that interferometers admitting efficient programming algorithms in general yield a much lower useful output energy, which ultimately limits their accuracy and energy efficiency.

Figures

Figures reproduced from arXiv: 2507.22972 by the authors.

Figure 1
Figure 1. FIG. 1: A sketch of a unitary interferometer with 3+5 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Realizing unitary and arbitrary transformations [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Diamond mesh. Blue lines denote optical paths [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Coherent optical crossbar arrays illustrating the [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

  1. [1]

    The physics of optical computing

    Peter L. McMahon. “The physics of optical computing”. Nature Reviews Physics 5, 717–734 (2023)

  2. [2]

    Photonics for artificial intelligence and neuromorphic computing

    Bhavin J. Shastri, Alexander N. Tait, Tom B. Fer- reira de Lima, Wolfram H. P. Pernice, Harish Bhaskaran, Charles D. Wright, and Paul R. Prucnal. “Photonics for artificial intelligence and neuromorphic computing”. Na- ture Photonics 15, 102–114 (2021)

  3. [3]

    Parallel programming of an in-memory photonic processor

    J. Feldmann, N. Youngblood, C. D. Wright, H. Bhaskaran, and W. H. P. Pernice. “Parallel programming of an in-memory photonic processor”. Nature 589, 52–58 (2021)

  4. [4]

    Photonic ma- trix multiplication lights up photonic accelerator and be- yond

    Hailong Zhou, Jianji Dong, Junwei Cheng, Wenchan Dong, Chaoran Huang, Yichen Shen, Qiming Zhang, Min Gu, Chao Qian, Hongsheng Chen, et al. “Photonic ma- trix multiplication lights up photonic accelerator and be- yond”. Light: Science & Applications 11, 30 (2022)

  5. [5]

    Hypermultiplexed integrated photon- ics–based optical tensor processor

    Shaoyuan Ou, Kaiwen Xue, Lian Zhou, Chun ho Lee, Alexander Sludds, Ryan Hamerly, Ke Zhang, Hanke Feng, Yue Yu, Reshma Kopparapu, Eric Zhong, Cheng Wang, Dirk Englund, Mengjie Yu, and Za- ijun Chen. “Hypermultiplexed integrated photon- ics–based optical tensor processor”. Science Advances 11, eadu0228 (2025)

  6. [6]

    This block can be configured based on matrix element A23. Though this is straightforward in principle, multiple optical paths now contribute, so the number of numerical operations needed to determine the parameters of MZI block 6 is non-trivial. Apparently, the numerical com- plexity of configuring this interferometer increases faster than O(n2), because ...

  7. [7]

    Integrated photonic quantum tech- nologies

    Jianwei Wang, Fabio Sciarrino, Anthony Laing, and Mark G Thompson. “Integrated photonic quantum tech- nologies”. Nature photonics 14, 273–284 (2020)

  8. [8]

    Large-scale photonic chiplet taichi empowers 160-tops/w artificial general in- telligence

    Zhihao Xu, Tiankuang Zhou, Muzhou Ma, ChenChen Deng, Qionghai Dai, and Lu Fang. “Large-scale photonic chiplet taichi empowers 160-tops/w artificial general in- telligence”. Science 384, 202–209 (2024)

Show all 38 references
  1. [9]

    Quantum machine learning with adaptive boson sampling via post-selection

    Francesco Hoch, Eugenio Caruccio, Giovanni Rodari, Tommaso Francalanci, Alessia Suprano, Taira Giordani, Gonzalo Carvacho, Nicol` o Spagnolo, Seid Koudia, Mas- similiano Proietti, et al. “Quantum machine learning with adaptive boson sampling via post-selection”. Na- ture Commu...

  2. [10]

    Reconfigurable continuously- coupled 3d photonic circuit for boson sampling experi- ments

    Francesco Hoch, Simone Piacentini, Taira Giordani, Zhen-Nan Tian, Mariagrazia Iuliano, Chiara Esposito, Anita Camillini, Gonzalo Carvacho, Francesco Ceccarelli, Nicol` o Spagnolo, et al. “Reconfigurable continuously- coupled 3d photonic circuit for boson sampling experi- ments...

  3. [11]

    A versatile single-photon-based quan- 6 tum computing platform

    Nicolas Maring, Andreas Fyrillas, Mathias Pont, Edouard Ivanov, Petr Stepanov, Nico Margaria, William Hease, Anton Pishchagin, Aristide Lema ˆ ıtre, Isabelle Sagnes, et al. “A versatile single-photon-based quan- 6 tum computing platform”. Nature Photonics 18, 603– 609 (2024)

  4. [12]

    Multidimensional quantum entanglement with large-scale integrated optics

    Jianwei Wang, Stefano Paesani, Yunhong Ding, Raf- faele Santagati, Paul Skrzypczyk, Alexandros Salavrakos, Jordi Tura, Remigiusz Augusiak, Laura Manˇ cinska, Davide Bacco, Damien Bonneau, Joshua W. Silver- stone, Mark G. Thompson, Antonio Ac ´ ın, Karsten Rot- twitt, Leif K. O...

  5. [13]

    Programmable integrated photonic coherent matrix: Principle, configuring, and applications

    Bo Wu, Hailong Zhou, Jianji Dong, and Xinliang Zhang. “Programmable integrated photonic coherent matrix: Principle, configuring, and applications”. Ap- plied Physics Reviews 11, 011309 (2024)

  6. [14]

    A manufacturable platform for pho- tonic quantum computing

    PsiQuantum team. “A manufacturable platform for pho- tonic quantum computing”. Nature 641, 876–883 (2025)

  7. [15]

    Accurate self-configuration of rectangular multiport interferometers

    Ryan Hamerly, Saumil Bandyopadhyay, and Dirk En- glund. “Accurate self-configuration of rectangular multiport interferometers”. Phys. Rev. Appl. 18, 024019 (2022)

  8. [16]

    Programmable pho- tonic unitary circuits for light computing

    Kyuho Kim, Kunwoo Park, Hyungchul Park, Sunkyu Yu, Namkyoo Park, and Xianji Piao. “Programmable pho- tonic unitary circuits for light computing”. Nanophoton- ics 14, 1429–1449 (2025)

  9. [17]

    Op- timal design of error-tolerant reprogrammable multiport interferometers

    S. A. Fldzhyan, M. Yu. Saygin, and S. P. Kulik. “Op- timal design of error-tolerant reprogrammable multiport interferometers”. Opt. Lett. 45, 2632–2635 (2020)

  10. [18]

    Hardware error correction for programmable photonics

    Saumil Bandyopadhyay, Ryan Hamerly, and Dirk En- glund. “Hardware error correction for programmable photonics”. Optica 8, 1247–1255 (2021)

  11. [19]

    Ma- trix computations

    Gene H. Golub and Charles F. van Loan. “Ma- trix computations”. JHU Press. (2013). Fourth edition. url: http://www.cs.cornell.edu/cv/GVL4/ golubandvanloan.htm

  12. [20]

    Optimal design for universal multiport interferometers

    William R. Clements, Peter C. Humphreys, Benjamin J. Metcalf, W. Steven Kolthammer, and Ian A. Walmsley. “Optimal design for universal multiport interferometers”. Optica 3, 1460–1465 (2016)

  13. [21]

    Low-depth, compact, and error-tolerant pho- tonic matrix-vector multiplication beyond the unitary group

    Suren A. Fldzhyan, Mikhail Yu. Saygin, and Stanislav S. Straupe. “Low-depth, compact, and error-tolerant pho- tonic matrix-vector multiplication beyond the unitary group”. Opt. Express 32, 46239–46250 (2024)

  14. [22]

    Lower-depth programmable linear op- tical processors

    Rui Tang, Ryota Tanomura, Takuo Tanemura, and Yoshiaki Nakano. “Lower-depth programmable linear op- tical processors”. Phys. Rev. Appl. 21, 014054 (2024)

  15. [23]

    Towards the Information-Theoretic Limit of Programmable Pho- tonics

    Ryan Hamerly, Jasvith Raj Basani, Alexander Sludds, Sri Krishna Vadlamani, and Dirk Englund. “Towards the Information-Theoretic Limit of Programmable Pho- tonics” (2024). arXiv:2408.09673

  16. [24]

    For example, the Clements mesh (see Fig

    and Clements [18] meshes of Mach-Zehnder interfer- ometers (MZIs). For example, the Clements mesh (see Fig. 2a) can be configured to implement an arbitrary uni- tary matrix U . To obtain a general non-unitary matrix A one uses the singular value decomposition (SVD) A = U1ΣU2 ,...

  17. [25]

    Uni- versal low-depth two-unitary design of programmable photonic circuits

    S. A. Fldzhyan, M. Yu. Saygin, and S. S. Straupe. “Uni- versal low-depth two-unitary design of programmable photonic circuits” (2025). arXiv:2504.19358

  18. [26]

    Experimental realization of any dis- crete unitary operator

    Michael Reck, Anton Zeilinger, Herbert J. Bernstein, and Philip Bertani. “Experimental realization of any dis- crete unitary operator”. Physical Review Letters 73, 58–61 (1994)

  19. [27]

    Probabilistic Upper Bounds for the Matrix Two-Norm

    Michiel E. Hochstenbach. “Probabilistic Upper Bounds for the Matrix Two-Norm”. Journal of Scientific Com- puting 57, 464–476 (2013)

  20. [28]

    Computing Probabilistic Bounds for Extreme Eigenvalues of Symmetric Matrices with the Lanczos Method

    Jos L. M. van Dorsselaer, Michiel E. Hochstenbach, and Henk A. van der Vorst. “Computing Probabilistic Bounds for Extreme Eigenvalues of Symmetric Matrices with the Lanczos Method”. SIAM Journal on Matrix Analysis and Applications 22, 837–852 (2001)

  21. [29]

    Numerical methods for large eigenvalue problems

    Yousef Saad. “Numerical methods for large eigenvalue problems”. Society for Indus- trial and Applied Mathematics. (2011). arXiv:https://epubs.siam.org/doi/pdf/10.1137/1.9781611970739

  22. [30]

    Gaussian elimination is not optimal

    Volker Strassen. “Gaussian elimination is not optimal”. Numerische Mathematik 13, 354–356 (1969)

  23. [31]

    Parallel optical coher- ent dot-product architecture for large-scale matrix multi- plication with compatibility for diverse phase shifters

    Shaofu Xu, Jing Wang, Sicheng Yi, Xinrui Zhao, Binshuo Liu, Jiayi Shao, and Weiwen Zou. “Parallel optical coher- ent dot-product architecture for large-scale matrix multi- plication with compatibility for diverse phase shifters.”. Optics express 30, 42057–42068 (2022)

  24. [32]

    Paral- lel convolution processing using an integrated photonic tensor core

    Johannes Feldmann, Nathan Youngblood, Maxim Kar- pov, Helge Gehring, Xuan Li, Maik Stappers, Manuel Le Gallo, Xin Fu, Anton Lukashchuk, Arslan Raja, Jun- qiu Liu, David Wright, Abu Sebastian, Tobias Kippen- berg, Wolfram Pernice, and Harish Bhaskaran. “Paral- lel convolution p...

  25. [33]

    Optical coherent dot-product chip for so- phisticated deep learning regression

    Shaofu Xu, Jing Wang, Haowen Shu, Zhike Zhang, Sicheng Yi, Bowen Bai, Xingjun Wang, Jianguo Liu, and Weiwen Zou. “Optical coherent dot-product chip for so- phisticated deep learning regression”. Light: Science & Applications 10, 221 (2021)

  26. [35]

    Scalable Coherent Optical Crossbar Architecture using PCM for AI Acceler- ation

    Daniel Sturm and Sajjad Moazeni. “Scalable Coherent Optical Crossbar Architecture using PCM for AI Acceler- ation”. 2023 Design, Automation & Test in Europe Con- ference & Exhibition (DATE) 2023-April, 1–6 (2022). arXiv:2210.10851

  27. [36]

    Coherent photonic crossbar as a universal linear op- erator

    George Giamougiannis, Apostolos Tsakyridis, Yangjin Ma, Angelina Totovic, David Lazovsky, and Nikos Pleros. “Coherent photonic crossbar as a universal linear op- erator”. Journal of Lightwave Technology 41, 2425– 2442 (2022). arXiv:2208.12033

  28. [37]

    Fully reconfigurable coherent optical vector- matrix multiplication

    James Spall, Xianxin Guo, Thomas D. Barrett, and A. I. Lvovsky. “Fully reconfigurable coherent optical vector- matrix multiplication”. Optics Letters 45, 5752 (2020). arXiv:2009.12095v1

  29. [38]

    Commentary on the decomposition of universal multiport interferometers: how it works in practice

    Dario Cilluffo. “Commentary on the decomposition of universal multiport interferometers: how it works in practice” (2024). arXiv:2412.11955

  30. [39]

    Fast rectangular matrix multiplica- tion and qr decomposition

    Philip A. Knight. “Fast rectangular matrix multiplica- tion and qr decomposition”. Linear Algebra and its Ap- plications 221, 69–81 (1995)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.