Pith. sign in

REVIEW 3 major objections 3 minor 48 references

Programmable k-local Ising Machines and all-optical Kolmogorov-Arnold Networks on Photonic Platforms

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a folded 4f relay with a second programmable pass through the same spatial light modulator turns each Fourier-plane window into a freely programmable polynomial of the spin-sum or projection amplitude, thereby…

desk verdict A novel per-window structural-nonlinearity architecture, clearly presented, but the central optical claim fails: a phase-only second pass makes window intensity at most quadratic in the clique sum, so the k>2 Ising and KAN-polynomial conclusions are unsupported. read the letter →

arxiv 2508.17440 v2 pith:KCNO5DY2 submitted 2025-08-24 physics.optics cs.ETcs.LG

Nikita Stroev , Natalia G. Berloff This is my paper · ORCID
classification physics.opticscs.ETcs.LG
keywords k-localIsingmachinestructuralnonlinearityfolded4frelayKolmogorov-Arnoldnetworkspatiallightmodulatorphotonictwo-framephysicalgradientsanalogopticalcomputer
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

Two open problems in photonic computing—native k-local Ising couplings and the many independent nonlinearities needed by Kolmogorov-Arnold networks—are addressed with one optical primitive: a folded 4f relay that sends the Fourier-plane field back through a second programmable pass on the same spatial light modulator. The paper's central claim is that each disjoint window carries a clique sum or a linear projection, and that a small-depth, parity-matched phase patch on that window makes the measured intensity a freely programmable polynomial of the amplitude, of degree up to k. If true, this would let spatial-photonic Ising machines, VCSEL arrays, and analog optical computers embed k-body Ising terms without nonlinear media, and train all-optical KAN ridge functions with two-frame physical gradients. The construction is developed through a worked k=4 example, closed-form seeds, a locally lower-triangular calibration map, and numerical calibration-refinement simulations for clique orders 3 to 15.

What carries the argument

Folded 4f relay with second-pass phase patches. Lens L1 Fourier-transforms the spin-encoded field so the k macropixels of a clique overlap in one window whose on-axis amplitude is proportional to S_q; lens L2 re-images that plane onto a dedicated patch of the same SLM carrying a small-depth cosine phase profile φ_q(y;θ_q)=Σ θ cos(n k_0 y). The paper's argument is that the small-phase expansion of exp(iφ) sends products of these harmonics back to the ±k_0 bank, that only DC terms survive window averaging, and that parity-matched harmonics n=2m−δ with δ=k mod 2 select even or odd powers of S_q. The map from phase depths to polynomial coefficients is locally lower-triangular, so M(k)=ceil(k/2) depths suffice, and a two-frame adjoint protocol supplies in-situ physical gradients for calibration and refinement.

What would settle it

For one k=4 clique with both second-pass harmonics enabled and no local oscillator, record the window-integrated intensity for each admissible value S_q in {-4,-2,0,2,4} while all other parameters are fixed; if the dependence is exactly constant plus quadratic in S_q, Eq. (11) with r=4 is not realized, whereas a quartic component would support the polynomial mechanism.

Watch

Extended reading notes

Core claim

The paper proposes that every k-spin product over a clique can be replaced by a parity-restricted univariate polynomial of the clique sum S_q, and that this polynomial can be manufactured optically as the measured intensity of a window after a second pass through the same SLM. Formally, Eq. (11) asserts I_q(S_q; θ_q) = Σ_{r≡k mod 2} Γ_{r,q}(θ_q) S_q^r, with M(k)=ceil(k/2) phase depths per window controlling all non-constant coefficients. The same per-window response is then reused as a KAN ridge function: holographic fan-out computes projections z_{j,m}=w^T_{j,m}x, the relay applies an independent polynomial Φ_{j,m}(z), and banked optical summation forms the output. The paper supports this with a closed-form k=4 seed from Cardano's formula, a locally lower-triangular Jacobian argument, and simulations of gradient-refinement training, and it sketches identical insertion into VCSEL and AOC feedback paths.

Load-bearing premise

The load-bearing premise is that each window's measured intensity is a polynomial of degree up to k in the summed spin amplitude; but the optics described are linear in that amplitude, so intensity detection yields at most a quadratic dependence, and nothing in the equations produces cubic or quartic powers.

Editorial extensions

If this is right

  • Any k-body Ising term can be represented as a parity-restricted polynomial of the clique sum and embedded optically without χ^(2) or Kerr media.
  • The same relay provides per-window, independently programmable univariate polynomials, giving all-optical KAN layers whose ridge functions are trainable with two optical frames per update.
  • The hardware insertion is one lens and a fold, or an on-chip 4f loop, and the recipe carries over from free-space SPIMs to injection-locked VCSEL arrays and to the analog optical computer.
  • Quadratization is bypassed: higher-order Ising formulations keep their native variable count instead of expanding into ancilla spins, in line with reported reductions in required variables.
  • Numerical calibration-refinement on clique orders k=3 to 15 reaches relative coefficient errors around 10^-3 to 10^-4 within a few hundred gradient steps, with weak dependence on k.

Reading between the lines

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

  • A direct experimental test of Eq. (11) would measure I_q(S_q) at all admissible S_q for one k=4 window; if only S_q^2 and a constant appear without a local oscillator, the polynomial-degree claim is not realized by the linear relay as written.
  • The windowed-addressing idea is separable from the degree claim: any per-window nonlinear element, such as a saturable absorber or electro-optic modulator, could slot into the same relay and supply true higher-order terms, making the relay's durable contribution the per-window programmability rather than the polynomial order.
  • Parity-matched harmonics imply a hardware ceiling: the maximum representable degree k is bounded by the number of grating periods that fit across a patch, so scaling to very high k will demand either finer SLM pixels, more bounces, or segmentation.
  • The bank-wise local-oscillator scheme sets one odd-order coefficient per bank; independent odd coefficients for every ridge would require per-window local-oscillator phase control, which the paper leaves as an engineering target.
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 / 3 minor

Summary. The paper proposes a two-pass folded 4f relay in which a second programmable phase-only pass through the same SLM is used to convert nominally linear optical propagation into per-window polynomial nonlinearities of the clique sum S_q or projection amplitude z. The authors claim this delivers native k-local Ising couplings and all-optical KAN ridge functions on SPIM, VCSEL, and AOC platforms, with M(k)=ceil(k/2) phase depths per window and two-frame physical gradients for training. The mathematical interpolation of a k-spin product by a parity-matched univariate polynomial in S_q is correct, and the numerical simulations are clearly described, but the physical mechanism underlying Eq. (11) is the core question.

Significance. If the central claim were correct, the paper would unify two important photonic computing tasks in a simple architecture, which is a genuinely significant prospect. The paper also makes its methods reproducible: the interpolation construction, the pseudo-code for calibration (poke tests, diagonal calibration, closed-form seeds), and the gradient-descent simulations are concrete and easy to follow. However, the significance is conditional on the existence of an optical mechanism producing intensity terms of degree higher than two in the input amplitude. As detailed in the major comments, the described linear-optical relay cannot produce such terms, and the numerical experiments assume the contested map rather than validating it. The real contribution is therefore a correct algebraic reduction plus a learning algorithm for a hypothetical device, not a supported optical primitive.

major comments (3)
  1. [Section II, Eq. (11)] The load-bearing assertion is Eq. (11), which states that the window-averaged intensity admits a polynomial expansion in S_q of degree up to k. This is inconsistent with the paper's own linear-optics model. For fixed second-pass phases θ_q, the optical system described in Section II is linear, so the field at the detector is affine in the input amplitude: E_out(y) = S_q A(y;θ_q) + E_LO (at most adding a local oscillator). The detected intensity is then |E_out|^2 = |S_q|^2 ∫|A|^2 dy + 2 Re[S_q ∫ A E_LO^* dy] + const, which contains only powers 0, 1, and 2 of S_q. No mechanism in Section II generates S_q^3 or S_q^4. The structural nonlinearity of Eq. (5) is a nonlinear dependence of the output on the programmable parameters θ, not on the input amplitude; for fixed θ the output remains linear (or affine) in the input, so the intensity is at most quadratic in S_q. This invalidates Eq. (11) for k>2 and with it the central claims of k-local Ising couplings and KAN ridge nonlinearities.
  2. [Section II, Eqs. (18)-(20)] The worked k=4 example makes the error concrete. The second-pass field is E_in(y) ∝ S_q(e^{ik0 y}+e^{-ik0 y}) multiplied by exp(iφ(y)); the field amplitude returning to the ±k0 spatial order is proportional to S_q times a function of α and β through Bessel functions. The paper computes terms 2J1(α)^2, 3J1(β)^2, and 2J2(α)J1(β), which are coefficients of the spatial Fourier decomposition of the output field. The intensity at the detector is proportional to |S_q|^2 times |c(θ)|^2, so those coefficients multiply |S_q|^2, not S_q^4. Equation (20) therefore incorrectly assigns the quartic coefficient to α^2β; physically, the α^2β term appears in a function of the phase depths multiplying the quadratic intensity |S_q|^2. The same conflation of field-coefficient order with input-amplitude polynomial order appears throughout the calibration section.
  3. [Section III, Eq. (32) and Fig. 6; Section II, Eqs. (24)-(27)] The numerical experiments do not validate the optical mechanism because they presuppose the contested map Γ = Aθ (Eq. (25)) from phase depths to polynomial coefficients of S_q or z. In Fig. 2 the device map is drawn as a random lower-triangular matrix A and the loss is ||wAθ - c||^2, which optimizes within the assumed linear relationship. In the teacher-student KAN experiment, the student model is defined through the same calibrated linearization Γ_m(θ_m)=A_m θ_m (Eq. (37)), so the experiment tests whether gradient descent can fit coefficients in the assumed model; it cannot confirm the existence of an optical device whose intensity is a polynomial of degree higher than 2 in the input amplitude. The two-frame FD proxy (Eq. (41)) likewise measures gradients of the assumed model. Thus Figs. 2, 3, and 6 demonstrate a linear-algebra learning procedure, not the proposed optical primitive.
minor comments (3)
  1. [Abstract and Section II] The abstract says that 'measured intensity in each window becomes a freely programmable polynomial of the clique sum or projection amplitude'; this is precisely what is unsupported. A clarifying sentence stating that the nonlinearity is only in the programmable parameters θ, not in the input amplitude, would have made the fundamental issue evident to the reader.
  2. [Section II, Eq. (14)] The homodyne expression is correct as written for a local oscillator, but the accompanying text states E_sig ≈ α_q S_q + O(S_q^3); for a linear optical system the field is exactly linear in S_q (no higher-order terms), so the O(S_q^3) term should be removed. This is a direct consequence of the same linearity issue.
  3. [General] The paper would benefit from stating explicitly that no experimental data are presented; the claim rests entirely on simulations of the assumed model. This is not a problem per se, but it underscores that the physical validity of Eq. (11) needs independent scrutiny.

Circularity Check

3 steps flagged · score 6.0 of 10

The central k-local polynomial claim is assumed in Eq. (11), and the numerical validation uses that same assumed map, so the calibration and teacher-student demonstrations reduce by construction to the premise they purport to support.

  1. self definitional [Sec. II, Eq. (11)]
    "By symmetry and for small phase depths on the second pass, the window-averaged intensity admits an even- or odd-parity polynomial expansion in S_q, I_q(S_q;θ_q)=∑_{r=0, r≡k (mod 2)}^k Γ_{r,q}(θ_q) S_q^r."

    This equation is the load-bearing premise of the paper: it asserts that a linear two-pass relay produces window-integrated intensity with powers S_q^r up to r=k. But from the paper's own linear scattering relation, Eq. (5), the output field is affine in the input amplitude for fixed θ, so the detected intensity contains at most |S_q|^2 plus a homodyne linear term; Γ_{r,q} for r>2 is zero, not a tunable coefficient. The later conclusions that the machine realizes native k-local couplings follow only if Eq. (11) is presupposed. Thus the claimed derivation of k-local terms is equivalent to the assumed polynomial form rather than being derived from the linear-optics model.

  2. fitted input called prediction [Sec. II, Eq. (25) and Fig. 2 caption]
    "Near the small-phase operating point the device is described by a lower-triangular map from harmonic depths to polynomial coefficients, Γ(θ)=Aθ+O(|θ|^2)... The device map is drawn as a random lower-triangular A... We report the relative coefficient error ε_rel=||Aθ−c_k||_2/||c_k||_2 using the noiseless linearized map Γ=Aθ."

    The numerical calibration/refinement experiments use as their 'device map' the linearization of the very Γ assumed in Eq. (11), and the target c_k is generated by interpolating the k-spin product with the same parity-matched polynomial representation. Optimizing wAθ≈c_k therefore tests only whether a well-conditioned triangular linear-algebra model can be inverted by gradient descent. It cannot validate that optics produces this Γ; the reported convergence and coefficient errors are properties of the assumed map re-entered as a simulation result. The 'prediction' that k-local couplings are trainable is a fitted demonstration inside the contested model, not an independent test.

1 more flagged steps
  1. self definitional [Sec. II, worked k=4 example, Eqs. (18)-(20)]
    "Expanding e^{iϕ_q(y)} with e^{ia cosθ}=∑_m i^m J_m(a)e^{imθ} ... Window averaging promotes these returns to even powers of S_q in the intensity. Grouping geometry/detection factors into calibration gains yields ... (for S_q^4) a44 1/8 α^2β = −a_q."

    The phrase 'promotes these returns to even powers of S_q' is where the quartic term is inserted. In the intensity of a field whose amplitude is proportional to S_q, every diffraction return—whether of order α, α^2, β, or α^2β—multiplies |S_q|^2; no combination of phase-depth orders can change the power of S_q. Equation (20) therefore assigns the α^2β phase-depth product to an S_q^4 coefficient by fiat, exactly reproducing the polynomial structure already assumed in Eq. (11). The k=4 'derivation' is circular with respect to the polynomial ansatz: it writes in the quartic coefficient that the linear-optics model cannot produce.

full rationale

The circularity here is partial but central. The paper's headline result—freely programmable k-local couplings and KAN ridge nonlinearities from one extra linear pass—rests on Eq. (11), which asserts that the measured window intensity is an arbitrary parity-matched polynomial of the clique sum S_q of degree up to k. That assertion is not derived from the paper's own linear scattering relation, Eq. (5); for fixed control phases the output field is linear in the input amplitude, so the detected intensity is at most quadratic in S_q (plus a homodyne linear term). The subsequent k=4 algebra and the calibration/simulation workflow do not supply independent evidence for the higher powers: the simulations use Γ=Aθ, which is precisely the linearization of the assumed Eq. (11), and the targets are the same polynomial-interpolation coefficients. Thus the demonstrations reduce by construction to the model they presuppose. The self-citations in the paper are not the main problem: Ref. [39] on structural nonlinearity and Refs. [46-48] on adjoint gradients are external, and the spin-elimination citation [28] is motivational rather than load-bearing. The score is 6 rather than higher because the interpolation reduction of k-spin products to parity-matched polynomials, Eq. (9), is mathematically valid independent content, and the phase-seeding algebra would be meaningful if a physical mechanism for the asserted Γ_{r,q} existed. But the central physical claim remains unvalidated, and the numerical 'predictions' are circular with respect to Eq. (11).

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on an assumed polynomial intensity map (Eq. 11) that is not derived and is inconsistent with the linear scattering relation Eq. (5). The simulations assume the map rather than test it.

free parameters (2)
  • per-window Jacobian A_q = random lower-triangular entries in simulations
    The numerical experiments assume a linear map Gamma=A theta from phase depths to polynomial coefficients and a well-conditioned invertible A. This map is the entire content of the central claim and is never derived from the optical field equations.
  • calibration constants (e.g., a22, a42, a44) = unspecified
    The k=4 closed-form seed depends on relay geometry and detection gains that are not computed or measured; they are assumed to exist and be positive.
assumptions (6)
  • standard math A k-spin product has a unique parity-matched polynomial in the spin sum S_q (Eq. 9).
    True by interpolation on the finite set S_k; the paper cites standard references.
  • domain assumption The first-pass hologram produces an on-axis field proportional to S_q in each disjoint Fourier-plane window, with all other terms rejected.
    Needed for the window to contain a clean coherent sum; ignores diffraction cross-talk, discretization, and fabrication errors.
  • ad hoc to paper The window-averaged intensity after the second pass is a parity-matched polynomial in S_q of degree up to k (Eq. 11).
    This is the load-bearing premise. It is asserted 'by symmetry' but no derivation from the linear scattering model is given; it conflicts with Eq. (5), under which intensity is a quadratic form in the input amplitude.
  • ad hoc to paper M(k)=ceil(k/2) phase depths per window suffice because the local Jacobian A is lower-triangular and invertible.
    The counting argument depends on coefficient reachability, which is exactly what the polynomial-expansion assumption must establish.
  • domain assumption A weak local oscillator makes odd powers of S_q appear without disturbing even powers (Eq. 14).
    Homodyne detection indeed gives a linear cross-term in S_q, but cannot generate higher odd powers S_q^3, S_q^5; the assumption that it enables full odd-order coverage is not justified.
  • domain assumption Two optical frames (forward and adjoint) provide exact parameter gradients, following Ref. [39].
    Inherited from the cited linear-scattering gradient protocol; accepted for the training loop if the rest of the map existed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Programmable k-local Ising Machines and all-optical Kolmogorov-Arnold Networks on Photonic Platforms." pith.science (2026). https://pith.science/paper/KCNO5DY2

@misc{pith2026250817440,
  author       = {Pith},
  title        = {Pith review of: Programmable k-local Ising Machines and all-optical Kolmogorov-Arnold Networks on Photonic Platforms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KCNO5DY2}},
  note         = {Machine review of arXiv:2508.17440}
}
read the original abstract

Photonic computing promises energy-efficient acceleration for optimization and learning, yet discrete combinatorial search and continuous function approximation have largely required distinct devices and control stacks. Here we unify k-local Ising optimization and optical Kolmogorov-Arnold network (KAN) learning on a single photonic platform, establishing a critical convergence point in optical computing. We introduce an SLM-centric primitive that realizes, in one stroke, all-optical k-local Ising interactions and fully optical KAN layers. The key idea is to convert the structural nonlinearity of a nominally linear scatterer into a per-window computational resource by adding a single relay pass through the same spatial light modulator: a folded 4f relay re-images the first Fourier plane onto the SLM so that each selected clique or channel occupies a disjoint window with its own second pass phase patch. Propagation remains linear in the optical field, yet the measured intensity in each window becomes a freely programmable polynomial of the clique sum or projection amplitude. This yields native, per clique k-local couplings without nonlinear media and, in parallel, the many independent univariate nonlinearities required by KAN layers, all trainable with in-situ physical gradients using two frames (forward and adjoint). We outline implementations on spatial photonic Ising machines, injection-locked vertical cavity surface emitting laser (VCSEL) arrays, and Microsoft analog optical computers; in all cases the hardware change is one extra lens and a fold (or an on-chip 4f loop), enabling a minimal overhead, massively parallel route to high-order Ising optimization and trainable, all-optical KAN processing on one platform.

Figures

Figures reproduced from arXiv: 2508.17440 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 30 canonical work pages

  1. [1]

    Mohseni, P

    N. Mohseni, P. L. McMahon, and T. Byrnes, Ising ma- chines as hardware solvers of combinatorial optimization problems, Nature Reviews Physics 4, 363 (2022)

    2022

  2. [2]

    Yamamoto, K

    Y. Yamamoto, K. Aihara, T. Leleu, K.-i. Kawarabayashi, S. Kako, M. Fejer, K. Inoue, and H. Takesue, Coherent ising machines—optical neural networks operating at the quantum limit, npj Quantum Information 3, 1 (2017)

    work page 2017

  3. [3]

    Inagaki, Y

    T. Inagaki, Y. Haribara, K. Igarashi, T. Sonobe, S. Ta- mate, T. Honjo, A. Marandi, P. L. McMahon, T. Umeki, K. Enbutsu, et al. , A coherent ising machine for 2000- node optimization problems, Science 354, 603 (2016)

    work page 2016

  4. [4]

    P. L. McMahon, A. Marandi, Y. Haribara, R. Hamerly, C. Langrock, S. Tamate, T. Inagaki, H. Takesue, S. Ut- sunomiya, K. Aihara, et al. , A fully programmable 100- spin coherent ising machine with all-to-all connections, Science 354, 614 (2016)

    work page 2016

  5. [5]

    V. Pal, S. Mahler, C. Tradonsky, A. A. Friesem, and N. Davidson, Rapid fair sampling of the x y spin hamil- tonian with a laser simulator, Physical Review Research 2, 033008 (2020)

    work page 2020

  6. [6]

    Babaeian, D

    M. Babaeian, D. T. Nguyen, V. Demir, M. Akbulut, P.- A. Blanche, Y. Kaneda, S. Guha, M. A. Neifeld, and N. Peyghambarian, A single shot coherent ising machine based on a network of injection-locked multicore fiber lasers, Nature communications 10, 1 (2019)

    work page 2019

  7. [7]

    Parto, W

    M. Parto, W. Hayenga, A. Marandi, D. N. Christodoulides, and M. Khajavikhan, Realizing 15 spin hamiltonians in nanoscale active photonic lattices, Nature materials 19, 725 (2020)

    work page 2020

  8. [8]

    Pierangeli, G

    D. Pierangeli, G. Marcucci, and C. Conti, Large-scale photonic ising machine by spatial light modulation, Phys- ical review letters 122, 213902 (2019)

    work page 2019

Show all 48 references
  1. [9]

    N. G. Berloff, M. Silva, K. Kalinin, A. Askitopoulos, J. D. T¨ opfer, P. Cilibrizzi, W. Langbein, and P. G. Lagoudakis, Realizing the classical XY hamiltonian in polariton sim- ulators, Nature materials (2017)

    2017

  2. [10]

    K. P. Kalinin, A. Amo, J. Bloch, and N. G. Berloff, Po- laritonic xy-ising machine, Nanophotonics9, 4127 (2020)

    2020

  3. [11]

    Vretenar, B

    M. Vretenar, B. Kassenberg, S. Bissesar, C. Toebes, and J. Klaers, Controllable josephson junction for photon bose-einstein condensates, Physical Review Research 3, 023167 (2021)

    2021

  4. [12]

    Kalinin, G

    K. Kalinin, G. Mourgias-Alexandris, H. Ballani, N. G. Berloff, J. H. Clegg, D. Cletheroe, C. Gkantsidis, I. Haller, V. Lyutsarev, F. Parmigiani, L. Pickup, et al. , Analog iterative machine (aim): using light to solve quadratic optimization problems with mixed variables, arXiv...

    2023 arXiv

  5. [13]

    Kalinin, J

    K. Kalinin, J. Gladrow, J. Chu, J. H. Cleg, D. Cletheroe, D. J. Kelly, B. Rahmani, G. Brennan, B. Canakci, et al., Analog optical computer for ai inference and combi- natorial optimization, Nature 635, 10.1038/s41586-025- 09430-z (2025)

    2025 doi

  6. [14]

    H. Goto, K. Endo, M. Suzuki, Y. Sakai, T. Kanao, Y. Hamakawa, R. Hidaka, M. Yamasaki, and K. Tat- sumura, High-performance combinatorial optimization based on classical mechanics, Science Advances 7, eabe7953 (2021)

    2021

  7. [15]

    K. Y. Camsari, R. Faria, B. M. Sutton, and S. Datta, Stochastic p-bits for invertible logic, Physical Review X 7, 031014 (2017)

    2017

  8. [16]

    K. Y. Camsari, B. M. Sutton, and S. Datta, P-bits for probabilistic spin logic, Applied Physics Reviews 6 (2019)

    2019

  9. [17]

    Zhang, Z

    D. Zhang, Z. Yuan, T. X. Hoang, W. Fu, C. E. Png, S. T. Lim, and A. Danner, All-optical scalable and pro- grammable vcsel-based ising annealer with parallel feed- back, Optics Express 33, 22119 (2025)

    2025

  10. [18]

    Lucas, Ising formulations of many np problems, Fron- tiers in physics 2, 5 (2014)

    A. Lucas, Ising formulations of many np problems, Fron- tiers in physics 2, 5 (2014)

    2014

  11. [19]

    Monasson and R

    R. Monasson and R. Zecchina, Statistical mechanics of the random k-satisfiability model, Physical Review E 56, 1357 (1997)

    1997

  12. [20]

    M´ ezard, G

    M. M´ ezard, G. Parisi, and R. Zecchina, Analytic and al- gorithmic solution of random satisfiability problems, Sci- ence 297, 812 (2002)

    2002

  13. [21]

    M´ ezard, F

    M. M´ ezard, F. Ricci-Tersenghi, and R. Zecchina, Two solutions to diluted p-spin models and xorsat problems, Journal of Statistical Physics 111, 505 (2003)

    2003

  14. [22]

    Kempe, A

    J. Kempe, A. Kitaev, and O. Regev, The complexity of the local hamiltonian problem, Siam journal on comput- ing 35, 1070 (2006)

    2006

  15. [23]

    Pedretti, F

    G. Pedretti, F. B¨ ohm, T. Bhattacharya, A. Heittmann, X. Zhang, M. Hizzani, G. Hutchinson, D. Kwon, J. Moon, E. Valiante, et al., Solving boolean satisfiability problems with resistive content addressable memories, npj Uncon- ventional Computing 2, 7 (2025)

    2025

  16. [24]

    Boros, Y

    E. Boros, Y. Crama, and E. Rodr´ ıguez-Heck, Compact quadratizations for pseudo-boolean functions, Journal of combinatorial optimization 39, 687 (2020)

    2020

  17. [25]

    Valiante, M

    E. Valiante, M. Hernandez, A. Barzegar, and H. G. Katz- graber, Computational overhead of locality reduction in binary optimization problems, Computer Physics Com- munications 269, 108102 (2021)

    2021

  18. [26]

    M. K. Bashar and N. Shukla, Designing ising machines with higher order spin interactions and their application in solving combinatorial optimization, Scientific Reports 13, 9558 (2023)

    2023

  19. [27]

    Bybee, D

    C. Bybee, D. Kleyko, D. E. Nikonov, A. Khosrowshahi, B. A. Olshausen, and F. T. Sommer, Efficient optimiza- tion with higher-order ising machines, Nature Communi- cations 14, 6033 (2023)

    2023

  20. [28]

    N. G. Berloff, Exact spin elimination in ising hamiltoni- ans and energy-based machine learning, arXiv preprint arXiv:2505.07163 (2025)

    2025 arXiv

  21. [29]

    A. N. Kolmogorov, On the representation of continuous functions of several variables by superpositions of contin- uous functions of a smaller number of variables (Ameri- can Mathematical Society, 1961)

    1961

  22. [30]

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halver- son, M. Soljaˇ ci´ c, T. Y. Hou, and M. Tegmark, Kan: Kolmogorov-arnold networks, arXiv preprint arXiv:2404.19756 (2024)

    2024 arXiv

  23. [31]

    Y. Peng, S. Hooten, X. Yu, T. Van Vaerenbergh, Y. Yuan, X. Xiao, B. Tossoun, S. Cheung, M. Fiorentino, and R. Beausoleil, Photonic kan: a kolmogorov-arnold network inspired efficient photonic neuromorphic archi- tecture, arXiv preprint arXiv:2408.08407 (2024)

    2024 arXiv

  24. [32]

    ai1 Technologies LLC, Ai-driven credit decisioning and risk modeling with kolmogorov–arnold networks, https://www.ai1technologies.com/ (2025), accessed: 1 September 2025

    2025

  25. [33]

    Sakellariou, A

    J. Sakellariou, A. Askitopoulos, G. Pastras, and S. I. Tsintzos, Encoding arbitrary ising hamiltonians on spa- tial photonic ising machines, Physical Review Letters 134, 203801 (2025)

    2025

  26. [34]

    Veraldi, D

    D. Veraldi, D. Pierangeli, S. Gentilini, M. C. Stri- nati, J. Sakellariou, J. S. Cummins, A. Kamaletdinov, M. Syed, R. Z. Wang, N. G. Berloff, et al. , Fully pro- grammable spatial photonic ising machine by focal plane division, Physical Review Letters 134, 063802 (2025)

    2025

  27. [35]

    R. Z. Wang, J. S. Cummins, M. Syed, N. Stroev, G. Pas- tras, J. Sakellariou, S. Tsintzos, A. Askitopoulos, D. Ve- raldi, M. Calvanese Strinati, and N. G. Berloff, Efficient computation using spatial-photonic ising machines with low-rank and circulant matrix constraints, Commun...

    2025

  28. [36]

    B. Wu, W. Zhang, S. Zhang, H. Zhou, Z. Ruan, M. Li, D. Huang, J. Dong, and X. Zhang, A monolithically in- tegrated optical ising machine, Nature Communications 16, 4296 (2025)

    2025

  29. [37]

    Kumar, H

    S. Kumar, H. Zhang, and Y.-P. Huang, Large-scale ising emulation with four body interaction and all-to-all con- nections, Communications Physics 3, 108 (2020)

    2020

  30. [38]

    Kanao and H

    T. Kanao and H. Goto, High-accuracy ising machine using kerr-nonlinear parametric oscillators with local four-body interactions, npj Quantum Information 7, 18 (2021)

    2021

  31. [39]

    C. C. Wanjura and F. Marquardt, Fully nonlinear neuro- morphic computing with linear wave scattering, Nature Physics 20, 1434 (2024)

    2024

  32. [40]

    H. Wang, J. Hu, A. Morandi, A. Nardi, F. Xia, X. Li, R. Savo, Q. Liu, R. Grange, and S. Gigan, Large-scale photonic computing with nonlinear disordered media, 16 Nature Computational Science 4, 429 (2024)

    2024

  33. [41]

    Bandyopadhyay, A

    S. Bandyopadhyay, A. Sludds, S. Krastanov, R. Hamerly, N. Harris, D. Bunandar, M. Streshinsky, M. Hochberg, and D. Englund, Single chip photonic deep neu- ral network with accelerated training, arXiv preprint arXiv:2208.01623 (2022)

    2022 arXiv

  34. [42]

    Yildirim, N

    M. Yildirim, N. U. Dinc, I. Oguz, D. Psaltis, and C. Moser, Nonlinear processing with linear optics, Na- ture Photonics 18, 1076 (2024)

    2024

  35. [43]

    Stroev and N

    N. Stroev and N. G. Berloff, Discrete polynomial opti- mization with coherent networks of condensates and com- plex coupling switching, arXiv preprint arXiv:1910.00842 (2019)

    2019 arXiv

  36. [44]

    L. N. Trefethen, Approximation theory and approxima- tion practice, extended edition (SIAM, 2019)

    2019

  37. [45]

    Koekoek, P

    R. Koekoek, P. A. Lesky, and R. F. Swarttouw, Hyperge- ometric orthogonal polynomials, in Hypergeometric Or- thogonal Polynomials and Their q-Analogues (Springer,

  38. [46]

    S. Pai, Z. Sun, T. W. Hughes, T. Park, B. Bartlett, I. A. Williamson, M. Minkov, M. Milanizadeh, N. Abebe, F. Morichetti, et al., Experimentally realized in situ back- propagation for deep learning in photonic neural net- works, Science 380, 398 (2023)

    2023

  39. [47]

    T. W. Hughes, M. Minkov, Y. Shi, and S. Fan, Training of photonic neural networks through in situ backpropa- gation and gradient measurement, Optica 5, 864 (2018)

    2018

  40. [48]

    Momeni, B

    A. Momeni, B. Rahmani, B. Scellier, L. G. Wright, P. L. McMahon, C. C. Wanjura, Y. Li, A. Skalli, N. G. Berloff, T. Onodera, et al. , Training of physical neural networks, arXiv preprint arXiv:2406.03372 (2024)

    2024 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.