Pith. sign in

REVIEW 3 major objections 5 minor 41 references

iSTAR: an algebraic-collapse framework for variational reduction in quantum-inspired continuous Ising solvers

T0 review · 3 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Late-stage continuous Ising dynamics collapse to a smaller active subsystem whose frozen spins fold exactly into an induced field, cutting most dense work without changing the discrete objective.

desk verdict Clean algebraic frozen-set reduction for SB solvers with strong same-seed G-set evidence; the math is solid, the savings claim is still a dense FLOPs proxy. read the letter →

arxiv 2607.05448 v1 pith:4M5AP66B submitted 2026-07-05 math.NA cs.NAmath.OCquant-ph

Bowen Liu , Dongmei Xiao This is my paper · ORCID
classification math.NAcs.NAmath.OCquant-ph MSC 65K1090C2782B20
keywords Isingmodelsimulatedbifurcationquantum-inspiredoptimizationvariationalreductionfrozen-setmethodactive-setcombinatorial
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

Continuous Ising solvers such as simulated bifurcation evolve a high-dimensional continuous state and read out binary spins by sign, but each step costs a dense matrix-vector product that scales with the square of the number of variables. This paper argues that the cost is not fundamental: once many coordinates have chosen a stable sign branch, they can be removed exactly rather than merely ignored. Fixing those signs turns their couplings into an effective external field on the remaining variables, so the original discrete problem reduces to an equivalent lower-dimensional Ising problem on the active tail. The authors prove continuous-to-discrete recovery for the soft quartic landscape and for a hard-box confinement model, and they supply a finite-time robust-margin certificate that keeps saturated coordinates frozen under the clipped ballistic update. Their algorithm iSTAR detects certified frozen coordinates online, folds them into the induced field, and continues only on the unresolved subsystem. On the G-set benchmark with external fields the certified procedure matches the same-seed full baseline in every run while removing, on average, more than sixty percent of the dense interaction work. A sympathetic reader cares because the O(N^{2}) bottleneck has long been treated as the price of continuous Ising dynamics; if the collapse is real and the certificate is reliable, late-stage computation can be cut sharply without changing the combinatorial objective.

What carries the argument

The frozen-set identity: for a frozen set H with fixed signs v_H, the Ising energy restricted to H is equivalent to a reduced energy on the active complement Q whose induced field is µ_eff = µ_Q + S_QH v_H. A robust freezing margin ρ_i(k;H) = D_i^H(k) − B_i^Q then certifies that a saturated coordinate with zero momentum stays fixed under the next clipped ballistic step, so the reduction can be applied online.

What would settle it

Run the online certified iSTAR procedure on the same G-set instances and seeds; if any certified reduced trajectory returns a strictly worse Ising energy than its same-seed full ballistic baseline, or if the robust-margin check freezes a coordinate that later flips under the full dynamics, the central claim that certified freezing preserves the discrete objective fails.

Watch

Extended reading notes

Core claim

During late-stage simulated bifurcation the trajectory collapses onto a lower-dimensional active subspace, and saturated coordinates can be eliminated exactly by a variational frozen-set identity: once a subset of signs is fixed, their couplings become an induced external field on the unresolved subsystem, reducing the original Ising instance to an equivalent lower-dimensional problem. An online certified implementation (iSTAR) that freezes only when a robust-margin condition holds preserves the same-seed full baseline on G-set and removes on average 64.4 percent of the dense interaction work.

Load-bearing premise

Quality preservation on the tested G-set schedules rests on the local robust-margin certificate activating and on unresolved coordinates remaining inside the box under a monotone schedule, not on the global hard-box recovery theorems that the paper itself notes do not hold at the benchmark parameters.

Editorial extensions

If this is right

  • Once coordinates satisfy the robust-margin certificate they need not be iterated; subsequent steps cost only the active-set size squared rather than N squared.
  • The same induced-field reduction applies in principle to any continuous Ising solver that ends with a sign readout, not only to ballistic simulated bifurcation.
  • External fields create a more favorable reduction regime than the zero-field case: certificates trigger earlier and more uniformly when a symmetry-breaking field is present.
  • Diagnostic probe sweeps show that late-tail reducibility is broad across G-set graphs, so most instances admit a non-degrading reduced operating point well before the nominal terminal time.

Reading between the lines

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

  • If the induced-field identity is exact, hardware accelerators for continuous Ising solvers could hard-wire progressive variable elimination once saturation is detected, turning late-stage O(N²) work into sparse or low-rank updates.
  • The gap between the global PSD hard-box theorem and the local certificate used on G-set suggests a natural next theorem: a finite-time basin-stability result that guarantees entry into the certified regime without assuming the spectral condition.
  • Because the reduction preserves the discrete objective on the retained coordinates, it can be stacked with other classical Ising heuristics as a free late-stage accelerator rather than as a competing solver.
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 / 5 minor

Summary. The paper argues that the O(N²) dense interaction cost of continuous Ising solvers (simulated bifurcation and relatives) is not intrinsic in late stage: trajectories collapse onto a lower-dimensional active set, and saturated coordinates can be eliminated exactly by a frozen-set identity that folds their couplings into an induced field on the unresolved subsystem. It proves large-α recovery for the external-field aSB quartic landscape, epi-convergence and hard-box vertex recovery for bSB under a PSD condition, and a robust-margin one-step/persistent freezing certificate for the clipped bSB Euler update. The algorithmic realization, iSTAR, freezes only after that certificate is verified online; on G-set with nonzero external field it matches the same-seed full bSB baseline in all reported runs and removes on average ~56–64% of dense matvec work under a FLOPs proxy.

Significance. If the local certificate and same-seed quality preservation hold as reported, the work supplies a clean algebraic reduction principle for continuous Ising solvers rather than a heuristic early-stop. Strengths include: (i) an exact frozen-set decomposition of the Ising energy (Thm. 6(i)/A.9); (ii) an explicit one-step and persistent freezing proof for the bSB clipping rule under a robust margin (Prop. A.10, Cor. A.12); (iii) a carefully designed online experiment that does not use the baseline to choose freeze times and reports zero same-seed degradations over large seed–instance panels; and (iv) transparent discussion that the global hard-box PSD hypothesis fails on the tested G-set schedules. The contribution is a useful bridge between variational landscape theory and adaptive active-set practice for SB-type solvers, with clear scope for follow-on sparse and wall-clock evaluation.

major comments (3)
  1. Eq. (16) and §3.3 measure savings only by the dense matvec proxy FLOPs = 2n²K (stagewise 2|Qk|²). G-set instances are sparse (e.g. G1: n=800, m=19176), so a production sparse bSB already costs O(m) per step; progressive submatrix extraction, induced-field updates, and indexing can erase the proxy gain. The abstract’s “removes on average 64.4% of the dense interaction work” is literally consistent with the proxy, but the computational claim as advertised needs either a sparse-aware cost model on the same runs or a sharper scoping statement that the reported saving is dense-proxy only and not yet a wall-clock or sparse-complexity result.
  2. Table 2 and the zero-field contrast (§4.2, Appendix Table 11) show that certified freezing and quality preservation are strongly field-dependent: with ηλ=0.2 the online rule triggers in 540/540 (Table S2) and 1420/1420 (all-Gset) runs with 0 degradations, while μ=0 triggers in only 230/540 and the probe-500 diagnostic yields 1166/1420 losses. The abstract and main claim lead with the 64.4% figure without stating that the favorable regime is the nonzero-field setting used throughout the primary experiments. The central empirical claim should be conditioned on the presence of a symmetry-breaking field (or the paper should supply a field-free certificate that works at comparable rates).
  3. §2.1–2.3 and the Discussion correctly note that the global hard-box PSD hypothesis fails on G-set (e.g. G1: λmin(S)≈−48.79 would need α≳s7 while the schedule uses α≤1), so quality preservation rests entirely on the local robust-margin certificate (Eqs. 11–14 / Prop. A.10) and on unresolved coordinates remaining in the box under a monotone α schedule. The abstract and introduction still lead with large-parameter recovery and hard-box reduction as if they underwrite the G-set result. Reorder or rephrase so that the finite-time certificate is presented as the operative theory for the experiments, with the global theorems as landscape motivation rather than as covering theorems for the reported trajectories.
minor comments (5)
  1. The diagnostic active-set selector (§3.1.2: 15% margin/velocity quantiles, |xi|<0.98, cap max{32,0.15n}) is used for probe sweeps but not ablated. A short ablation of margin quantile, saturation threshold, and velocity filter would clarify how much of the probe-sweep breadth is selector-specific.
  2. Cross-variant Table 4 uses a different protocol (Goto Table S2 step counts and probe ratios) than the main all-Gset 1200-step study; the text notes this, but a single sentence in the table caption restating the protocol difference would prevent misreading the 52/54 vs 71/71 coverage gap.
  3. Theorem 2’s explicit α* is intentionally non-sharp and can be large; a brief remark on typical numerical scales of R*, M*, Δ* for G-set-sized instances would help readers judge practical relevance of the large-α aSB recovery statement.
  4. Figure 2 panel labels and the log10(1+ΔE) energy-gap plots are useful; ensure axis labels and the “mean first freeze” markers remain legible in grayscale print.
  5. Notation: S vs J for the coupling matrix is identified in the introduction but both appear in early bSB equations; a single consistent symbol after the identification would reduce friction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: frozen-set identity is algebraic, freezing certificate is proved from the bSB update, and online experiments freeze only after a trajectory certificate without using the baseline to choose freeze times.

full rationale

The load-bearing chain is self-contained and does not reduce to its inputs by construction. Theorem 6(i)/A.9 is an exact algebraic rewrite of the Ising energy under fixed signs (couplings of H fold into µ_eff on Q); it is not a fit or a prediction of the same quantity. The one-step and persistent freezing results (Prop. A.10, Cor. A.12) are derived from the explicit bSB Euler+clipping rule under the robust margin ρ_i(k;H)>0 and a monotone α schedule—they certify that already-saturated coordinates stay frozen, rather than assuming the empirical energy match. Large-α aSB recovery (Thm. 2) and hard-box reduction (Thm. 5) are static variational proofs with explicit hypotheses; the paper itself states that the PSD hypothesis fails on G-set (λ_min(S)≈−48.79 vs α≤1), so those global theorems are not smuggled in as guarantees of the benchmark. Online iSTAR freezes only after the certificate is verified along the trajectory and compares to the same-seed full bSB baseline without using the baseline to choose freeze times; probe sweeps are labeled diagnostic and retrospective. Self-citation to [15] supplies zero-field background; the external-field recovery and frozen-set reduction used by iSTAR are proved in this paper. FLOPs-proxy vs sparse/hardware cost is a correctness/deployment limitation, not circularity. No fitted-input-as-prediction, no uniqueness imported from authors, no ansatz smuggled as theorem.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The central algorithmic claim rests on standard Ising/SB models plus a few modeling and experimental choices: the discrete Ising objective, aSB/bSB continuous embeddings, the clipped Euler bSB update, the robust ℓ1 envelope certificate, and benchmark choices (G-set, ηλ, check interval, FLOPs proxy). No new physical entity is postulated; free parameters are experimental/design knobs rather than fitted universal constants.

free parameters (5)
  • external field scale ηλ
    Primary strong results use ηλ=0.2 with fixed field seed; zero-field activation and quality preservation are much weaker, so reported average savings depend on this chosen field strength.
  • certificate check interval (50 steps)
    Online freezing is only tested at this scheduled interval; earlier/later checking would change freeze timing and measured savings.
  • diagnostic active-set thresholds (15% margin/velocity quantiles, |xi|<0.98, cap max{32,0.15n})
    Used in probe-sweep diagnostics, not the certified online rule; still free design choices that affect breadth-of-reduction evidence.
  • bifurcation/schedule parameters and horizons (Table S2 per-instance steps; all-Gset 1200-step horizon)
    Savings and freeze times are schedule-dependent; α≤1 on benchmarks while global PSD recovery would need much larger α on some graphs.
  • explicit large-α threshold α* constants in Theorem 2
    Recovery threshold is intentionally non-sharp and assembled from R*, M*, Δ*, C*; it is a sufficient-condition design, not fitted to G-set, but it is a paper-chosen quantitative envelope.
assumptions (6)
  • domain assumption Target discrete problem is the external-field Ising energy E(v)=−½vᵀSv−μ·v on {−1,1}ⁿ with symmetric zero-diagonal S.
    Eq. (1); all recovery and reduction statements are relative to this objective.
  • domain assumption aSB/bSB continuous embeddings and the discrete clipped Euler bSB update with momentum reset at the walls are valid models of the solvers being accelerated.
    Sections 1 and 3 and Prop. A.10; the certificate is proved for this update rule, not arbitrary continuous Ising dynamics.
  • ad hoc to paper For hard-box global vertex recovery, (α²−1)I+S is positive semidefinite, making U□,α concave on the box.
    Theorem 5; sufficient for landscape reduction, but the paper notes it fails on raw G-set at tested α.
  • domain assumption Unresolved coordinates remain in the box |xj|≤1 and αk is nondecreasing after certification, so the robust margin persists.
    Corollary 7 / A.12–A.13; needed for persistent freezing after the one-step certificate.
  • ad hoc to paper Dense interaction work proxy FLOPs=2n²K (and stagewise 2|Qk|²) is an adequate primary compute measure for the claimed savings.
    Section 3.3; authors exclude selection overhead, memory, and wall-clock effects by design.
  • standard math Standard real analysis / variational analysis tools (critical-point and Hessian tests, epi-convergence, concavity on polytopes) apply as used in the appendix.
    Appendix A lemmas and citations to Rockafellar–Wets style epi-convergence arguments.
invented entities (2)
  • iSTAR online certified frozen-set reduction procedure
    purpose: Detect robustly saturated spins, eliminate them via induced field μeff=μQ+SQH vH, and continue only on active tail Q.
    Algorithmic object introduced by the paper; not a physical entity, but the main proposed method.
  • robust freezing margin ρi(k;H)=DH_i(k)−BQ_i independent evidence
    purpose: Sufficient finite-time certificate that clipped bSB keeps saturated coordinates frozen despite unresolved couplings.
    Defined in Section 2.2; proved sufficient for one-step/persistent freezing under stated hypotheses, validated empirically online.

how reviews work

0 comments
Cite this review

Pith. "Pith review of iSTAR: an algebraic-collapse framework for variational reduction in quantum-inspired continuous Ising solvers." pith.science (2026). https://pith.science/paper/4M5AP66B

@misc{pith2026260705448,
  author       = {Pith},
  title        = {Pith review of: iSTAR: an algebraic-collapse framework for variational reduction in quantum-inspired continuous Ising solvers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4M5AP66B}},
  note         = {Machine review of arXiv:2607.05448}
}
abstract

Continuous Ising solvers embed a discrete optimization problem into a continuous dynamical system and recover the spin configuration by sign readout, but dense interaction evaluation gives an $O(N^2)$-per-step cost. We show that this cost is not intrinsic: during late-stage simulated bifurcation the trajectory collapses onto a lower-dimensional active subspace, and saturated coordinates can be eliminated exactly by a variational frozen-set identity whose couplings fold into an induced field on the unresolved subsystem. We prove large-parameter recovery for the external-field quartic model, the hard-box limit of ballistic confinement, and a robust-margin freezing criterion. The resulting algorithm, iSTAR (Ising Stable-set Tail-Aware Reduction), exploits this collapse by detecting stabilized coordinates and continuing only on the active tail. An online certified implementation on the G-set benchmark preserves the same-seed baseline in all runs and removes on average 64.4% of the dense interaction work.

Figures

Figures reproduced from arXiv: 2607.05448 by the authors.

Figure 1
Figure 1. Illustration of iSTAR. (a) In simulated bifurcation, many coordinates select stable [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Online certified iSTAR with nonzero external fields ( [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Scale-grouped compute reduction for online certified iSTAR. Each box contains seed– [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 1 linked inside Pith

  1. [1]

    Max cut and the smallest eigenvalue.SIAM J

    Trevisan, L. Max cut and the smallest eigenvalue.SIAM J. Comput.41, 1769–1786 (2012)

    2012

  2. [2]

    & Zhang, D

    Chang, K., Shao, S. & Zhang, D. Cheeger’s cut, max cut and the spectral theory of 1-Laplacian on graphs.Sci. China Math.60, 1963–1980 (2017)

    1963

  3. [3]

    & Zhang, D

    Chang, K., Shao, S. & Zhang, D. Nodal domains of eigenvectors for 1-Laplacian on graphs.Adv. Math. 308, 529–574 (2017)

    2017

  4. [4]

    Spectrum of the 1-Laplacian and Cheeger’s constant on graphs.J

    Chang, K. Spectrum of the 1-Laplacian and Cheeger’s constant on graphs.J. Graph Theory81, 167–207 (2015)

    2015

  5. [5]

    & Zhang, W

    Chang, K., Shao, S., Zhang, D. & Zhang, W. Lov´ asz extension and graph cut.Commun. Math. Sci. 19, 761–786 (2021)

    2021

  6. [6]

    & Moore, C

    Clauset, A., Newman, M. & Moore, C. Finding community structure in very large networks.Phys. Rev. E70, 066111 (2004). 17

    2004

  7. [7]

    Modularity and community structure in networks.Proc

    Newman, M. Modularity and community structure in networks.Proc. Natl. Acad. Sci. USA103, 8577– 8582 (2006)

    2006

  8. [8]

    & Newman, M

    Leicht, E. & Newman, M. Community structure in directed networks.Phys. Rev. Lett.100, 118703 (2008)

    2008

Show all 41 references
  1. [9]

    & Mucha, P

    Porter, M., Onnela, J. & Mucha, P. Communities in networks.Notices Am. Math. Soc.56, 1082–1097 (2009)

    2009

  2. [10]

    & Mucha, P

    Rombach, P., Porter, M., Fowler, J. & Mucha, P. Core-periphery structure in networks.SIAM J. Appl. Math.74, 167–190 (2014)

    2014

  3. [11]

    & Hein, M

    Tudisco, F., Mercado, P. & Hein, M. Community detection in networks via nonlinear modularity eigen- vectors.SIAM J. Appl. Math.78, 2393–2419 (2018)

    2018

  4. [12]

    & Reinelt, G

    Barahona, F., Gr¨ otschel, M., J¨ unger, M. & Reinelt, G. An application of combinatorial optimization to statistical physics and circuit layout design.Oper. Res.36, 493–513 (1988)

    1988

  5. [13]

    Ising formulations of many NP problems.Front

    Lucas, A. Ising formulations of many NP problems.Front. Phys.2, 5 (2014)

    2014

  6. [14]

    & Mniszewski, S

    Ushijima-Mwesigwa, H., Negre, C. & Mniszewski, S. Graph partitioning using quantum annealing on the D-Wave system.Proc. 2nd Int. Workshop Post Moores Era Supercomputing(ACM, 2017)

    2017

  7. [15]

    Liu, B., Wang, K., Xiao, D. & Yu, Z. On connection among quantum-inspired algorithms of the Ising model.Commun. Math. Sci.21, 2013–2028 (2023)

    2013

  8. [16]

    & Yung, M.-H

    Li, Y., Cui, X., Xiong, Z., Liu, B., Wang, B.-Y., Shu, R., Qiao, N. & Yung, M.-H. Quantum molecular docking with a quantum-inspired algorithm.J. Chem. Theory Comput.20, 6687–6694 (2024)

    2024

  9. [17]

    & Yung, M.-H

    Li, Y., Cui, X., Xiong, Z., Zou, Z., Liu, B., Wang, B.-Y., Shu, R., Zhu, H., Qiao, N. & Yung, M.-H. Efficient molecular conformation generation with quantum-inspired algorithm.J. Mol. Model.30, 228 (2024)

    2024

  10. [18]

    & Yung, M.-H

    Zeng, Q.-G., Cui, X.-P., Liu, B., Wang, Y., Mosharev, P. & Yung, M.-H. Performance of quantum annealing inspired algorithms for combinatorial optimization problems.Commun. Phys.7, 249 (2024)

    2024

  11. [19]

    & Qiao, N

    Shu, R., Liu, B., Xiong, Z., Cui, X., Li, Y., Cui, W., Yung, M.-H. & Qiao, N. Quantum-inspired machine learning for molecular docking.arXiv:2401.12999(2024)

    2024 arXiv

  12. [20]

    & Goto, H

    Kanao, T. & Goto, H. Simulated bifurcation assisted by thermal fluctuation.Commun. Phys.5, 153 (2022)

    2022

  13. [21]

    & Dixon, A

    Goto, H., Tatsumura, K. & Dixon, A. R. Combinatorial optimization by simulating adiabatic bifurca- tions in nonlinear Hamiltonian systems.Sci. Adv.5, eaav2372 (2019)

    2019

  14. [22]

    & Tatsumura, K

    Goto, H., Endo, K., Suzuki, M., Sakai, Y., Kanao, T., Hamakawa, Y., Hidaka, R., Yamasaki, M. & Tatsumura, K. High-performance combinatorial optimization based on classical mechanics.Sci. Adv.7, eabe7953 (2021)

    2021

  15. [23]

    & Goto, H

    Tatsumura, K., Yamasaki, M. & Goto, H. High-performance combinatorial optimization based on classical mechanics.Sci. Adv.8, eabk0048 (2022)

    2022

  16. [24]

    & Yamamoto, Y

    Wang, Z., Marandi, A., Wen, K., Byer, R. & Yamamoto, Y. Coherent Ising machine based on degenerate optical parametric oscillators.Phys. Rev. A88, 063853 (2013)

    2013

  17. [25]

    L., Marandi, A., Haribara, Y., Hamerly, R., Langrock, C., Tamate, S., Inagaki, T., Takesue, H., Utsunomiya, S., Aihara, K., Byer, R

    McMahon, P. L., Marandi, A., Haribara, Y., Hamerly, R., Langrock, C., Tamate, S., Inagaki, T., Takesue, H., Utsunomiya, S., Aihara, K., Byer, R. L., Fejer, M. M., Mabuchi, H. & Yamamoto, Y. A fully programmable 100-spin coherent Ising machine with all-to-all connections.Scienc...

    2016

  18. [26]

    & Van der Sande, G

    B¨ ohm, F., Verschaffelt, G. & Van der Sande, G. A poor man’s coherent Ising machine based on opto- electronic feedback systems for solving optimization problems.Nat. Commun.10, 3538 (2019)

    2019

  19. [27]

    & Troyer, M

    Isakov, S., Zintchenko, I., Rønnow, T. & Troyer, M. Optimised simulated annealing for Ising spin glasses.Comput. Phys. Commun.192, 265–271 (2015)

    2015

  20. [28]

    & Aihara, K

    Leleu, T., Yamamoto, Y., Utsunomiya, S. & Aihara, K. Scaling advantage of chaotic amplitude control for high-performance combinatorial optimization.Commun. Phys.4, 266 (2021)

    2021

  21. [29]

    Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. USA79, 2554–2558 (1982)

    1982

  22. [30]

    G-set test problems.Stanford G-set collection.https://web.stanford.edu/ ~yyye/yyye/Gset/

    Ye, Y. G-set test problems.Stanford G-set collection.https://web.stanford.edu/ ~yyye/yyye/Gset/

  23. [31]

    Goemans, M. X. & Williamson, D. P. Improved approximation algorithms for maximum cut and sat- isfiability problems using semidefinite programming.J. ACM42, 1115–1145 (1995)

    1995

  24. [32]

    & Savar´ e, G.Gradient Flows in Metric Spaces and in the Space of Probability Measures(Birkh¨ auser, 2005)

    Ambrosio, L., Gigli, N. & Savar´ e, G.Gradient Flows in Metric Spaces and in the Space of Probability Measures(Birkh¨ auser, 2005)

    2005

  25. [33]

    Rockafellar, R. T. & Wets, R. J.-B.Variational Analysis(Springer, 1998)

    1998

  26. [34]

    & Zhang, P

    Shen, Z.-S., Pan, F., Wang, Y., Men, Y.-D., Xu, W.-B., Yung, M.-H. & Zhang, P. Free-energy machine for combinatorial optimization.Nat. Comput. Sci.5, 322–332 (2025)

    2025

  27. [35]

    Marandi, A., Wang, Z., Takata, K., Byer, R. L. & Yamamoto, Y. Network of time-multiplexed optical parametric oscillators as a coherent Ising machine.Nat. Photonics8, 937–942 (2014)

    2014

  28. [36]

    L., Umeki, T., Enbutsu, K., Tadanaga, O., Takenouchi, H., Aihara, K., Kawarabayashi, K., Inoue, K., Utsunomiya, S

    Inagaki, T., Haribara, Y., Igarashi, K., Sonobe, T., Tamate, S., Honjo, T., Marandi, A., McMahon, P. L., Umeki, T., Enbutsu, K., Tadanaga, O., Takenouchi, H., Aihara, K., Kawarabayashi, K., Inoue, K., Utsunomiya, S. & Takesue, H. A coherent Ising machine for 2000-node optimiza...

    2000

  29. [37]

    Hamerly, R., Inagaki, T., McMahon, P. L., Venturelli, D., Marandi, A., Onodera, T., Ng, E., Lan- grock, C., Inaba, K., Honjo, T., Enbutsu, K., Umeki, T., Kasahara, R., Utsunomiya, S., Kako, S., Kawarabayashi, K., Byer, R. L., Fejer, M. M., Mabuchi, H., Englund, D., Rieffel, E....

    2019

  30. [38]

    & Takesue, H

    Honjo, T., Sonobe, T., Inaba, K., Inagaki, T., Ikuta, T., Yamada, Y., Kazama, T., Enbutsu, K., Umeki, T., Kasahara, R., Kawarabayashi, K. & Takesue, H. 100,000-spin coherent Ising machine.Sci. Adv.7, eabh0952 (2021)

    2021

  31. [39]

    Afoakwa, R., Zhang, Y., Vengalam, U. K. R., Cao, Z. & Mizuno, M. BRIM: bistable resistively-coupled Ising machine.2021 IEEE Int. Symp. Circuits Syst. (ISCAS)(2021)

    2021

  32. [40]

    Y., Faria, R., Sutton, B

    Camsari, K. Y., Faria, R., Sutton, B. M., Datta, S. & Crowell, P. A. Stochasticp-bits for invertible logic.Phys. Rev. X7, 031014 (2017)

    2017

  33. [41]

    A., Pervaiz, A

    Borders, W. A., Pervaiz, A. Z., Fukami, S., Camsari, K. Y., Ohno, H. & Datta, S. Integer factorization using stochastic magnetic tunnel junctions.Nature573, 390–393 (2019). Acknowledgments The authors thank colleagues and collaborators for helpful discussions. This work was pa...

    2019

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.