pith. sign in

arxiv: 1907.05124 · v2 · pith:A556H2ZNnew · submitted 2019-07-11 · 🪐 quant-ph · cond-mat.stat-mech· stat.ML

Highly parallel algorithm for the Ising ground state searching problem

A. Yavorsky , L.A. Markovich , E.A. Polyakov , A.N. Rubtsov This is my paper

Pith reviewed 2026-05-24 23:21 UTC · model grok-4.3

classification 🪐 quant-ph cond-mat.stat-mechstat.ML
keywords Ising modelground state searchmean-field annealingparallel algorithmmaximum cutcombinatorial optimizationsimulated annealing
0
0 comments X

The pith

MARS algorithm finds near-optimal Ising ground states by parallel mean-field descents from random states and temperatures.

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

The paper introduces the MARS algorithm for locating energy minima in the Ising model, a computationally hard task linked to many combinatorial optimization problems. MARS runs mean-field annealing starting from a randomly chosen spin configuration and temperature, then repeats the process many times. Each individual run is inexpensive, so the overall method gains power from massive parallel execution rather than from any single sophisticated trajectory. The authors report strong results on large spin systems and on standard maximum-cut benchmark instances, measured by both solution quality and wall-clock time.

Core claim

The central claim is that the Mean-field Annealing from a Random State (MARS) algorithm, which is based on the mean-field descent from a randomly selected configuration and temperature, shows excellent performance both on the large Ising spin systems and on the set of exemplary maximum cut benchmark instances in terms of both solution quality and computational time.

What carries the argument

Mean-field descent performed from a randomly selected configuration and temperature, with effectiveness obtained through many independent parallel repetitions.

If this is right

  • Large Ising systems become tractable when many low-cost mean-field runs are executed in parallel.
  • The hybrid of simulated annealing ideas with mean-field annealing improves solution quality over either method alone on the tested sizes.
  • The same procedure yields competitive results on maximum-cut instances.
  • Computational time scales favorably with available parallel resources because each run is lightweight.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Lightweight parallel heuristics may serve as a practical alternative to complex serial solvers for Ising-like problems.
  • The approach could be tested on other energy-minimization tasks where mean-field approximations remain reasonable.
  • Hardware accelerators such as GPUs could amplify the speed advantage by running even more restarts simultaneously.
  • Failure on highly frustrated instances would reveal the boundary where random mean-field starts cease to suffice.

Load-bearing premise

The mean-field approximation combined with random restarts from varied temperatures is sufficient to reach near-optimal solutions with high probability on the tested problem sizes.

What would settle it

A large Ising instance or max-cut graph on which even thousands of independent MARS runs fail to match the best-known solution energy would falsify the performance claim.

Figures

Figures reproduced from arXiv: 1907.05124 by A.N. Rubtsov, A. Yavorsky, E.A. Polyakov, L.A. Markovich.

Figure 1
Figure 1. Figure 1: Energy histograms for the four Ising problem ( [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
read the original abstract

Finding an energy minimum in the Ising model is an exemplar objective, associated with many combinatorial optimization problems, that is computationally hard in general, but occurs in all areas of modern science. There are several numerical methods, providing solution for the medium size Ising spin systems. However, they are either computationally slow and badly parallelized, or do not give sufficiently good results for the large systems. In this paper, we present a highly parallel algorithm, called Mean-field Annealing from a Random State (MARS), incorporating the best features of the classical simulated annealing (SA) and Mean-Field Annealing (MFA) methods. The algorithm is based on the mean-field descent from a randomly selected configuration and temperature. Since a single run requires little computational effort, the effectiveness can be achieved by massive parallelisation. MARS shows excellent performance both on the large Ising spin systems and on the set of exemplary maximum cut benchmark instances in terms of both solution quality and computational time.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces MARS (Mean-field Annealing from a Random State), a hybrid algorithm that performs mean-field descent from random initial configurations and temperatures, combining elements of simulated annealing and mean-field annealing. It claims to be highly parallelizable with little computational effort per run and asserts excellent performance on large Ising spin systems and exemplary Max-Cut benchmark instances in both solution quality and computational time.

Significance. If the empirical claims are substantiated with data, the approach could provide a scalable, parallel method for Ising ground-state search and related combinatorial problems, building on established mean-field techniques.

major comments (2)
  1. [Abstract] Abstract: the central performance claim ('excellent performance both on the large Ising spin systems and on the set of exemplary maximum cut benchmark instances') is stated without any numerical results, error bars, baseline comparisons, tables, or figures, leaving the assertion unverifiable.
  2. [Method description] Method (mean-field descent description): the algorithm relies on repeated short trajectories from varied random starts reaching near-ground-state energies with high probability, yet no a-priori bound on the mean-field free-energy error (arising from the product-measure approximation that neglects correlations) or characterization of the fraction of failing restarts is supplied; this is load-bearing for the claim on frustrated instances.
minor comments (2)
  1. The manuscript should include pseudocode or a clear algorithmic listing of the MARS procedure, including how temperatures and initial states are sampled.
  2. Clarify the precise definition of 'randomly selected configuration and temperature' and the stopping criterion for each descent trajectory.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the careful reading and constructive feedback. We address the major comments below and indicate where revisions will be made to improve clarity and verifiability.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central performance claim ('excellent performance both on the large Ising spin systems and on the set of exemplary maximum cut benchmark instances') is stated without any numerical results, error bars, baseline comparisons, tables, or figures, leaving the assertion unverifiable.

    Authors: We agree that the abstract would be strengthened by including concrete numerical support. In the revised manuscript we will add a concise statement of key results (e.g., achieved approximation ratios or energy gaps on the largest instances and selected Max-Cut benchmarks) together with brief baseline comparisons, while remaining within the abstract length limit. revision: yes

  2. Referee: [Method description] Method (mean-field descent description): the algorithm relies on repeated short trajectories from varied random starts reaching near-ground-state energies with high probability, yet no a-priori bound on the mean-field free-energy error (arising from the product-measure approximation that neglects correlations) or characterization of the fraction of failing restarts is supplied; this is load-bearing for the claim on frustrated instances.

    Authors: The MARS approach is empirical; its justification rests on the observed high success rate of short mean-field trajectories across many random initial conditions. We do not claim or possess an a-priori theoretical bound on the product-measure error. However, we already report empirical success fractions and failure rates on the tested instances (including frustrated ones) in the results section. We will expand this characterization with additional statistics on restart success probability and will explicitly note the absence of a rigorous error bound. revision: partial

standing simulated objections not resolved
  • No a-priori theoretical bound on the mean-field free-energy error is supplied or readily derivable for the product-measure approximation on general frustrated instances.

Circularity Check

0 steps flagged

No circularity: performance claims rest on empirical benchmarks, not self-referential derivations

full rationale

The manuscript presents MARS as a heuristic combining mean-field descent with random restarts and parallel execution. All reported outcomes (solution quality on large Ising instances and Max-Cut benchmarks) are framed as numerical results from direct runs rather than quantities derived from fitted parameters or prior self-citations. No equations appear that equate a prediction to its own input by construction, and the abstract supplies no load-bearing self-citation chain. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; full manuscript would be required to audit modeling choices such as temperature schedules or convergence criteria.

pith-pipeline@v0.9.0 · 5712 in / 1164 out tokens · 18116 ms · 2026-05-24T23:21:05.000386+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages · 1 internal anchor

  1. [1]

    Edwards and P.W

    S. Edwards and P.W. Anderson. Theory of spin glasses. J. Phys. F , (5):965–974, 1975

    work page 1975

  2. [2]

    A. C. Carter, A. J. Bray, and M. A. Moore. Aspect-ratio scaling and the stiffness exponent θ for ising spin glasses. Phys. Rev. Lett. , 88:077201, Jan 2002

    work page 2002

  3. [3]

    Hartmann, Alan J

    Alexander K. Hartmann, Alan J. Bray, A. C. Carter, M. A. Moore , and A. P. Young. Stiffness exponent of two-dimensional ising spin glasses for nonperiodic boundary conditions using aspect-ratio scaling. Phys. Rev. B , 66:224401, Dec 2002

    work page 2002

  4. [4]

    Hartmann

    Alexander K. Hartmann. Scaling of stiffness energy for three-d imensional ± j ising spin glasses. Phys. Rev. E , 59:84–87, Jan 1999

    work page 1999

  5. [5]

    Optimal storage properties of neural network models

    E Gardner and B Derrida. Optimal storage properties of neural network models. Journal of Physics A: Mathematical and General , 21(1):271–284, jan 1988

    work page 1988

  6. [6]

    Spin Glass Theory and Beyond

    M Mezard, G Parisi, and M Virasoro. Spin Glass Theory and Beyond . WORLD SCIENTIFIC, 1986

    work page 1986

  7. [7]

    FRONT MATTER, pages i–xv

    Daniel L Stein. FRONT MATTER, pages i–xv

    work page

  8. [8]

    Nishimori

    H. Nishimori. Statistical Physics of Spin Glasses and Information Proces sing: An Introduction. Oxford University Press, 2001

    work page 2001

  9. [9]

    Stein and C.M

    D.L. Stein and C.M. Newman. Spin Glasses and Complexity . Online access: JSTOR Books at JSTOR. Princeton University Press, 2013

    work page 2013

  10. [10]

    Papadimitriou

    Christos H. Papadimitriou. The euclidean travelling salesman prob lem is np- complete. Theoretical Computer Science , 4(3):237 – 244, 1977

    work page 1977

  11. [11]

    Richard M. Karp. Reducibility among Combinatorial Problems , pages 85–103. Springer US, Boston, MA, 1972

    work page 1972

  12. [12]

    A hy brid parti- cle swarm optimization algorithm for the vehicle routing problem

    Yannis Marinakis, Magdalene Marinaki, and Georgios Dounias. A hy brid parti- cle swarm optimization algorithm for the vehicle routing problem. Engineering Applications of Artificial Intelligence , 23(4):463 – 472, 2010

    work page 2010

  13. [13]

    M. K. M. Ali and F. Kamoun. Neural networks for shortest pat h computation and routing in computer networks. IEEE Transactions on Neural Networks , 4(6):941–954, Nov 1993

    work page 1993

  14. [14]

    Banaszak, G

    D. Banaszak, G. A. Dale, A. N. Watkins, and J. D. Jordan. An op tical tech- nique for detecting fatigue cracks in aerospace structures. In ICIASF 99. 18th International Congress on Instrumentation in Aerospace Si mulation Facilities. Record (Cat. No.99CH37025) , pages 27/1–27/7, June 1999

    work page 1999

  15. [15]

    Sherrington, D

    Kirkpatrick S. Sherrington, D. Solvable model of a spin glass. Phys. Rev. Lett. , 35:1792–1796, 1975. 14

    work page 1975

  16. [16]

    G. Parisi. Infinite number of order parameters for spin-glasse s. Phys. Rev. Lett., 43:1754–1756, 1979

    work page 1979

  17. [17]

    G. Parisi. A sequence of approximate solutions to the s-k model for spin glasses. J. Phys. A , 13:L–115, 1980

    work page 1980

  18. [18]

    Crystal statistics

    Lars Onsager. Crystal statistics. i. a two-dimensional model with an order- disorder transition. Phys. Rev., 65:117–149, Feb 1944

    work page 1944

  19. [19]

    On the computational complexity of ising spin glass m odels

    F Barahona. On the computational complexity of ising spin glass m odels. Jour- nal of Physics A: Mathematical and General , 15(10):3241–3253, oct 1982

    work page 1982

  20. [20]

    Kempe, A

    J. Kempe, A. Kitaev, and O. Regev. The complexity of the local h amiltonian problem. SIAM Journal on Computing , 35(5):1070–1097, 2006

    work page 2006

  21. [21]

    Finke, A

    G. Finke, A. Claus, and E. Gunn. A two-commodity network flow a pproach to the traveling salesman problem. Congressus Numerantium 41 , :167–178, 1984

    work page 1984

  22. [22]

    De Simone, M

    C. De Simone, M. Diehl, M. J¨ unger, P. Mutzel, G. Reinelt, and G. R inaldi. Exact ground states of two-dimensional j ising spin glasses. Journal of Statistical Physics, 84(5):1363–1371, Sep 1996

    work page 1996

  23. [23]

    De Simone, M

    C. De Simone, M. Diehl, M. J¨ unger, P. Mutzel, G. Reinelt, and G. R inaldi. Exact ground states of ising spin glasses: New experimental result s with a branch-and-cut algorithm. Journal of Statistical Physics , 80(1):487–496, Jul 1995

    work page 1995

  24. [24]

    A hybrid genetic - partic le swarm optimization algorithm for the vehicle routing problem

    Yannis Marinakis and Magdalene Marinaki. A hybrid genetic - partic le swarm optimization algorithm for the vehicle routing problem. Expert Systems with Applications, 37(2):1446 – 1455, 2010

    work page 2010

  25. [25]

    Artificial intelligence, heuristic frameworks and ta bu search

    Fred Glover. Artificial intelligence, heuristic frameworks and ta bu search. Man- agerial and Decision Economics , 11(5):365–375

    work page

  26. [26]

    An ant colony algorith m for solving fixed destination multi-depot multiple traveling salesmen problems

    Soheil Ghafurian and Nikbakhsh Javadian. An ant colony algorith m for solving fixed destination multi-depot multiple traveling salesmen problems. Applied Soft Computing , 11(1):1256 – 1262, 2011

    work page 2011

  27. [27]

    A novel genetic algorithm based on imm unity

    Licheng Jiao and Lei Wang. A novel genetic algorithm based on imm unity. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans, 30(5):552–561, Sep. 2000

    work page 2000

  28. [28]

    An new self-organizin g maps strategy for solving the traveling salesman problem

    Yanping Bai, Wendong Zhang, and Zhen Jin. An new self-organizin g maps strategy for solving the traveling salesman problem. Chaos, Solitons and Frac- tals, 28(4):1082 – 1089, 2006. Sir Hermann Bondi 1919 - 2005

    work page 2006

  29. [29]

    An analysis of the e lastic net approach to the traveling salesman problem

    Richard Durbin, Richard Szeliski, and Alan Yuille. An analysis of the e lastic net approach to the traveling salesman problem. Neural Computation, 1(3):348– 358, 1989. 15

    work page 1989

  30. [30]

    Embedding learning capability in lagran gean relaxation: An application to the travelling salesman problem

    Reza Zamani and Sim Kim Lau. Embedding learning capability in lagran gean relaxation: An application to the travelling salesman problem. European Jour- nal of Operational Research , 201(1):82 – 88, 2010

    work page 2010

  31. [31]

    Computing with neural circuits: a model

    JJ Hopfield and DW Tank. Computing with neural circuits: a model. Science, 233(4764):625–633, 1986

    work page 1986

  32. [32]

    A memetic neur al network for the euclidean traveling salesman problem

    Jean-Charles Creput and Abderrafiaa Koukam. A memetic neur al network for the euclidean traveling salesman problem. Neurocomputing, 72(4):1250 – 1264, 2009. Brain Inspired Cognitive Systems (BICS 2006) / Inter play Between Natural and Artificial Computation (IWINAC 2007)

    work page 2009

  33. [33]

    Molecular computation of solutions to combinatorial problems

    LM Adleman. Molecular computation of solutions to combinatorial problems. Science, 266(5187):1021–1024, 1994

    work page 1994

  34. [34]

    A quantum adiabatic evolution algorithm app lied to random instances of an np-complete problem

    Edward Farhi, Jeffrey Goldstone, Sam Gutmann, Joshua Lapan , Andrew Lund- gren, and Daniel Preda. A quantum adiabatic evolution algorithm app lied to random instances of an np-complete problem. Science, 292(5516):472–475, 2001

    work page 2001

  35. [35]

    C. D. Gelett S. Kirkpatrick and M. P. Vecchi. Science, 220:671, 1983

    work page 1983

  36. [36]

    Solving the traveling salesman problem based on an adaptive simulated annealing a lgorithm with greedy search

    Xiutang Geng, Zhihua Chen, Wei Yang, Deqian Shi, and Kai Zhao. Solving the traveling salesman problem based on an adaptive simulated annealing a lgorithm with greedy search. Applied Soft Computing , 11(4):3680 – 3689, 2011

    work page 2011

  37. [37]

    Scott Kirkpatrick, C. D. Gelatt, and Mario P. Vecchi. Optimizatio n by simu- lated annealing. Science, 220 4598:671–80, 1983

    work page 1983

  38. [38]

    Brooke, D

    J. Brooke, D. Bitko, T. F., Rosenbaum, and G. Aeppli. Quantum a nnealing of a disordered magnet. Science, 284(5415):779–781, 1999

    work page 1999

  39. [39]

    Swendsen and Jian-Sheng Wang

    Robert H. Swendsen and Jian-Sheng Wang. Replica monte carlo s imulation of spin-glasses. Phys. Rev. Lett. , 57:2607–2609, Nov 1986

    work page 1986

  40. [40]

    Katzgraber

    Wenlong Wang, Jonathan Machta, and Helmut G. Katzgraber. C omparing monte carlo methods for finding ground states of ising spin glasses: Popula- tion annealing, simulated annealing, and parallel tempering. Phys. Rev. E , 92:013303, Jul 2015

    work page 2015

  41. [41]

    L. Ingber. Simulated annealing: Practice versus theory. Mathematical and Computer Modelling , 18(11):29 – 57, 1993

    work page 1993

  42. [42]

    Isakov, I.N

    S.V. Isakov, I.N. Zintchenko, T.F. Ronnow, and M. Troyer. Opt imised simulated annealing for ising spin glasses. Computer Physics Communications , 192:265 – 271, 2015

    work page 2015

  43. [43]

    Sreenivas, and K.Ganapathy Subramaniam

    D.Janaki Ram, T.H. Sreenivas, and K.Ganapathy Subramaniam. P arallel sim- ulated annealing algorithms. Journal of Parallel and Distributed Computing , 37(2):207 – 212, 1996. 16

    work page 1996

  44. [44]

    A distributed implementatio n of simulated annealing for the travelling salesman problem

    James R.A Allwright and D.B Carpenter. A distributed implementatio n of simulated annealing for the travelling salesman problem. Parallel Computing , 10(3):335 – 338, 1989

    work page 1989

  45. [45]

    Abramson

    D. Abramson. A very high speed architecture for simulated ann ealing. Com- puter, 25(5):27–36, May 1992

    work page 1992

  46. [46]

    Byer, and Yoshih isa Ya- mamoto

    Zhe Wang, Alireza Marandi, Kai Wen, Robert L. Byer, and Yoshih isa Ya- mamoto. Coherent ising machine based on degenerate optical para metric oscil- lators. Phys. Rev. A , 88:063853, Dec 2013

    work page 2013

  47. [47]

    McMahon, Takeshi Umeki, Koji Enbutsu, Osamu Tadanaga, Hirokazu Takenouchi, Kazuyuki Aihara, Ken-ichi Kawarabayashi, Kyo Inoue, Shoko Utsunomiya, and Hirok i Take- sue

    Takahiro Inagaki, Yoshitaka Haribara, Koji Igarashi, Tomohir o Sonobe, Shuhei Tamate, Toshimori Honjo, Alireza Marandi, Peter L. McMahon, Takeshi Umeki, Koji Enbutsu, Osamu Tadanaga, Hirokazu Takenouchi, Kazuyuki Aihara, Ken-ichi Kawarabayashi, Kyo Inoue, Shoko Utsunomiya, and Hirok i Take- sue. A coherent ising machine for 2000-node optimization problems...

    work page 2000

  48. [48]

    Coherent ising machines - optical neural networks operating at th e quantum limit

    Yoshihisa Yamamoto, Kazuyuki Aihara, Timothee Leleu, Kenichi Kawarabayashi, Satoshi Kako, Martin Fejer, Kyo Inoue, and Hiro ki Takesue. Coherent ising machines - optical neural networks operating at th e quantum limit. npj Quantum Information , (49):2056–6387, 2017

    work page 2056

  49. [49]

    A.D. King, W. Bernoudy, J. King, A.J. Berkley, and T. Lanting. Em ulating the coherent ising machine with a mean-field algorithm. arXiv:1806.08422, , 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  50. [50]

    Tiunov, Alexander E

    Egor S. Tiunov, Alexander E. Ulanov, and A. I. Lvovsky. Annea ling by simu- lating the coherent ising machine. arXiv:1901.08927, 2019

    work page arXiv 1901

  51. [51]

    An application of combinatorial optimization to statistical physics an d circuit layout design

    Francisco Barahona, Martin Grotschel, Michael Junger, and G erhard Reinelt. An application of combinatorial optimization to statistical physics an d circuit layout design. Operations Research, 36(3):493–513, 1988

    work page 1988

  52. [52]

    Rosenbluth, Marshall N

    Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenb luth, Au- gusta H. Teller, and Edward Teller. Equation of state calculations by fast computing machines. The Journal of Chemical Physics , 21(6):1087–1092, 1953

    work page 1953

  53. [53]

    L. Ingber. Adaptive simulated annealing (asa): Lessons learne d. Control and cybernetics, 25(1):33 – 54, 1996

    work page 1996

  54. [54]

    Miller, Wesley E

    Griff Bilbro, Reinhold Mann, Thomas K. Miller, Wesley E. Snyder, Dav id E. Van den Bout, and Mark White. Optimization by mean field annealing. In Pro- ceedings of the 1st International Conference on Neural Info rmation Processing Systems, NIPS’88, pages 91–98, Cambridge, MA, USA, 1988. MIT Press

    work page 1988

  55. [55]

    R.W. Eglese. Simulated annealing: A tool for operational resear ch. European Journal of Operational Research , 46(3):271 – 281, 1990

    work page 1990

  56. [56]

    Mahfoud and David E

    Samir W. Mahfoud and David E. Goldberg. Parallel recombinative s imulated annealing: A genetic algorithm. Parallel Computing , 21(1):1 – 28, 1995. 17

    work page 1995

  57. [57]

    Breakout local search for the max- cutproblem

    Una Benlic and Jin-Kao Hao. Breakout local search for the max- cutproblem. Engineering Applications of Artificial Intelligence , 26(3):1162 – 1173, 2013

    work page 2013

  58. [58]

    Helmberg and F

    C. Helmberg and F. Rendl. A spectral bundle method for semidefi nite program- ming. SIAM Journal on Optimization , 10(3):673–696, 2000

    work page 2000

  59. [59]

    In columns the Ising energy (absolute and mean values) and mean runtimes (in minut es) are shown

    Appendix 18 Table 3: Graph |V | fprev fbest favg t(s) P G1 800 11624 11623 11539.9 0.0857 0.11944 G2 800 11620 11620 11540.9 0.0894 8 × 10− 5 G3 800 11622 11621 11541 0.0908 0.07732 G4 800 11646 11646 11563.8 0.0938 0.01238 G5 800 11631 11631 11550.9 0.4441 8 × 10− 5 G6 800 2178 2176 2093.54 0.0883 58 × 10− 5 G7 800 2006 2006 1925.04 0.0903 14 × 10− 5 G8 ...

    work page 2093