graphical representations and cluster algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf ·...
TRANSCRIPT
![Page 1: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/1.jpg)
Graphical Representationsand Cluster Algorithms
Jon MachtaUniversity of Massachusetts Amherst
Newton Institute, March 27, 2008
![Page 2: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/2.jpg)
Outline
• Introduction to graphical representations and cluster algorithms
• An algorithm for the FK random cluster model (q>1)
• Graphical representations and cluster algorithms for Ising
systems with external fields and/or frustration
![Page 3: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/3.jpg)
Graphical Representations
•Tool for rigorous results on spin systems
•Basis for very efficient Monte Carlo algorithms
•Source of geometrical insights into phase transitions
Fortuin & KastelynConiglio & KleinSwendsen & WangEdwards & Sokal
![Page 4: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/4.jpg)
Joint Spin-Bond Distribution
Every occupied bond is satisfied
Edwards&Sokal, PRD, 38,2009 (1988)
Bernoulli factor
![Page 5: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/5.jpg)
Marginals of ES Distribution
Fortuin-Kastelyn random cluster model for q=2
Ising model
=number of clusters
![Page 6: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/6.jpg)
Swendsen-Wang Algorithm
Occupy satisfied bonds with probability
Identify clusters of occupied bonds
Randomly flip clusters of spins with probability 1/2.!
p =1" e"2#J
![Page 7: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/7.jpg)
Connectivity and Correlation
• Broken Symmetry Giant cluster
![Page 8: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/8.jpg)
Efficiency at the critical point
• Dynamic Exponent
• Li-Sokal bound
• linear dimension
• autocorrelation time
• dynamic exponent
• specific heat exponent
• correlation length exponent
![Page 9: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/9.jpg)
Random cluster model (q>1)
For q not too large and d>1, there is a continuous phasetransition. Above a critical q, the transition is first-order. q=1is Bernoulli percolation, q=2 is Ising, q=positive integer isPotts.
![Page 10: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/10.jpg)
Edward-Sokal Joint Distribution
Every occupied bond is satisfied
Bernoulli factor
number of clusters of inactive sites
Bond marginal is q RC model
L. Chayes & JM, Physica A 254, 477 (1998)
![Page 11: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/11.jpg)
Algorithm for RC model
• Given a bond configuration identifyclusters
• Declare clusters active with prob 1/qand inactive otherwise,
• Erase active occupied bonds• Occupy active bonds with probability p
to produce new bond configuration ,
![Page 12: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/12.jpg)
Dynamic Exponent for 2D RC modelY. Deng, T. M. Garoni, JM, G. Ossola, M. Polin,and A. D. Sokal PRL 99, 055701 (2007)
![Page 13: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/13.jpg)
Dynamic Exponent for 3D RC model
![Page 14: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/14.jpg)
• z=0, 1<q<2,• z=1, q=2,• exponential slowing (first-order), q>2• obtained from an evolution equation for the
average size of the largest cluster:
Dynamic Exponent for Mean Field RC model
![Page 15: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/15.jpg)
Conclusions (Part I)
• Swendsen-Wang scheme can be extended toreal q>1 RC models.
• Li-Sokal bound not sharp for any q>1
![Page 16: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/16.jpg)
Fields and Frustration
•Ising model in a staggered field•Spin glass•Random field Ising model
![Page 17: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/17.jpg)
For H not too large and d>1, there is a continuous phasetransition in the Ising universality class to a phase with non-zero staggered magnetization
Ising model in a staggered field(bipartite lattice)
square lattice
![Page 18: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/18.jpg)
For d>2, there is a continuous phase transition to a state withnon-zero Edwards-Anderson order parameter.
i.i.d. quenched disorder
Ising spin glass
frustration%$#@!1
![Page 19: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/19.jpg)
Swendsen-Wang fails with fields or frustration
• No relation between spin correlations andconnectivity.
• At Tc one cluster occupies most of thesystem (e.g. percolation occurs in the hightemperature phase).
• External fields hi cause small acceptanceprobabilities for cluster flips.
![Page 20: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/20.jpg)
Two Replica Graphical RepresentationSwendsen & Wang, PRL, 57, 2607 (1986) the other SW paper!O. Redner, JM & L. Chayes, PRE 58, 2749 (1998), JSP 93, 17 (1998)JM, M. Newman & L. Chayes, PRE 62, 8782 (2000)JM, C. Newman & D. Stein, JSP 130, 113 (2008)
![Page 21: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/21.jpg)
Spin-Bond Distribution
Bernoulli factor
spin bond constraints
![Page 22: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/22.jpg)
Spin Bond Constraints•If bonds satisfied in bothreplicas then
with probability
•If bonds satisfied in only onereplica then
with probability
![Page 23: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/23.jpg)
PropertiesProperties
Prob{ i and j are connected by a path of occupied bondswith an even number of red bonds}
−Prob{ i and j are connected by a path of occupied bondswith an odd number of red bonds}
•Spin marginal is two independent Ising systems
•Correlation function of EA order parameter and connectivity
![Page 24: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/24.jpg)
Properties II
•For ferromagnetic Ising systems, including the staggeredfield and random field Ising models, percolation of bluebonds is equivalent to broken symmetry.
•For Ising systems with AF interactions, the existence of ablue cluster with a density strictly larger than all other bluecluster is equivalent to broken symmetry.
![Page 25: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/25.jpg)
For purposes of this example, assume all bonds FM
Two Replica Cluster Algorithm
•Given a two spinconfigurations•Occupy bonds•Identify clusters
![Page 26: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/26.jpg)
Re-populate spins consistent with constraints
If a field is present, the acceptance probabilityof this flip involves the Boltzmann factor of thefield energy change
Even with a field, these cluster spinconfigurations are equally probable
![Page 27: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/27.jpg)
Two Replica Cluster Algorithm• Given two spin configurations in the same environment.• Blue (red) occupy doubly (singly) satisfied bonds.• Identify blue and red clusters.• Re-populate spins with equal probability consistent with
constraints due to bonds (flip clusters).• Erase blue and red bonds.• Add-ons:
– Translations and global flips of replicas relative to each other(staggered field, bipartite lattice only)
– Metropolis sweeps– Parallel tempering
![Page 28: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/28.jpg)
How well does it work?
![Page 29: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/29.jpg)
2D Ising model in a staggered field
X. Li & JM, Int. J. Mod. Phys. C12, 257 (2001)
z<0.5
![Page 30: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/30.jpg)
2D Spin GlassJ. S. Wang & R. H. Swendsen, Prog. Theor. Phys. Suppl, vol 157, 317 (2005)
![Page 31: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/31.jpg)
Parallel Tempering
•n replicas at temperatures β1 … βn•MC (e.g. Metropolis) on each replica•Exchange replicas with energies E and E’ andtemperatures β and β’,with probability:
!
min{exp[(" #" ')(E # E ')],1}
β5β4β3β2β1
easy to equilibrate hard to equilibrate
![Page 32: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/32.jpg)
Parallel Tempering+Two Replica ClusterParallel Tempering+Two Replica ClusterMovesMoves
•n replicas at temperatures β1 … βn•Two replica cluster moves at each temperature•Exchange replicas.
β5β4β3β2β1
easy to equilibrate hard to equilibrate
![Page 33: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/33.jpg)
Diluted 3D spin glassT. Jorg, Phys. Rev. B 73, 224431 (2006)
![Page 34: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/34.jpg)
• For the d>2 spin glasses, two giant clusters(“agree” and “disagree” spins) appear in the hightemperature phase and together comprise most ofthe system.
• The signature of the spin glass transition is theonset of a density difference between the twogiant clusters.
• The algorithm is not much more efficient thatparallel tempering alone.
JM, C. Newman & D. Stein, JSP 130, 113 (2008)
3D Spin Glass
![Page 35: Graphical Representations and Cluster Algorithmspeople.umass.edu/machta/talks/newton_3-08.pdf · Graphical Representations and Cluster Algorithms Jon Machta University of Massachusetts](https://reader034.vdocument.in/reader034/viewer/2022050110/5f4791e6adf6a75090029b83/html5/thumbnails/35.jpg)
Conclusions (Part II)
• The two-replica graphical representation andassociated algorithms are promisingapproaches for spin systems with fields orfrustration.
However…• For the hardest systems, e.g. 3D Ising spin
glass,the algorithm is not much more efficientthan parallel tempering.