Exact Ground States of Finite Ising Spin Glasses Obtained by
Transcription
Exact Ground States of Finite Ising Spin Glasses Obtained by
Exact Ground States of Finite Ising Spin Glasses Obtained by ”Branch-and-Bound” S. Kobe . [email protected] http://www.physik.tu-dresden.de/itp/members/kobe.html Technische Univ. Dresden, Inst. für Theor. Physik, D-01062 Dresden, Germany S. Kobe, Goettingen/Leipzig 2004 – p.1/?? Outline • • • I SING model: A short history Exact ground states of finite I SING spin glasses Results: – Lattice models: Energy landscape Relaxation – Mean field models: Energy exponents and correction to scaling Ferromagnetic – spin-glass transition S. Kobe, Goettingen/Leipzig 2004 – p.2/?? Historical remarks Letter of WOLFGANG PAULI to H. B. G. C ASIMIR: Princeton, 11. Oktober 1945 ” ... A few weeks will be sufficient for you and others to learn everything of scientific interest which happened during these ‘lost years‘. I am sending you today a package with reprints, please divide them among persons who are interested. There is a paper of Onsager included (...) of which I think that it is a masterpiece of mathematical analysis. It contains the rigorous solution of the Kramers - Wannier order-disorder problem for the two dimensional model (unfortunately the method cannot be generalized for three dimensional crystals). ... " S. Kobe, Goettingen/Leipzig 2004 – p.3/?? Ising Model: Phase transition, scaling behaviour Zero field I SING model: H=− X 1≤i<j≤N Jij Si Sj Si = 1 ∨ −1 – Jij > 0 between nearest neighbours of a lattice; J ij = 0 else – Ground state is trivial – Phase transition: (ferromagnetic) order - disorder S. Kobe, Goettingen/Leipzig 2004 – p.4/?? S. Kobe, Goettingen/Leipzig 2004 – p.5/?? Ernst Ising 1925 S. Kobe, Goettingen/Leipzig 2004 – p.6/?? Peoria 1996 S. Kobe, Goettingen/Leipzig 2004 – p.7/?? Ising Spin Glass Models Zero field I SING model: H=− X 1≤i<j≤N Jij Si Sj Si = 1 ∨ −1 – Jij arbitrary – Ground state is not known!!! S. Kobe, Goettingen/Leipzig 2004 – p.8/?? Spin Glass Models (Lattice vs. Mean-Field Models) • Lattice models: Jij = ±1 or Gij ; i.e. the couplings between spins beeing nearest neighbours in a (cubic) lattice are randomly distributed • Mean-field models: e.g S HERRINGTON -K IRKPATRICK (SK) model: Gij Jij = √ N Gij → independent identically distributed Gaussian random numbers with zero means and variance one. S. Kobe, Goettingen/Leipzig 2004 – p.9/?? A Related Mean-Field Model SK model with non-symmetric distribution of Gauss couplings: |Gij | |Gij | P (Jij ) = xδ(Jij + √ ) + (1 − x)δ(Jij − √ ) N N i.e. the sign of interactions are inversed according the probability x limiting cases: x=0 x = 0.5 x=1 → ferromagnetic system → SK model → antiferromagnetic system No shift of the Gauss distribution !!! S. Kobe, Goettingen/Leipzig 2004 – p.10/?? How much states ? N 1 2 3 4 5 ... 10 20 ... 24 ... 47 ... 58 ... 64 ... 90 ... 1010 2N 2 4 8 16 32 ... 1024 1048576 ... 16 777 216 ... ≈ 1.4 x 1014 ... ≈ 3 x 1017 ... ≈ 2 x 1019 ... ≈ 2 x 1027 ... unimaginable! notes ≈ (Maximal-)number of ancestors in 20th generation (14th century) ≈ combinations in German Lotto ≈ age of mankind in seconds (4 Mio. a) ≈ time since "big bang" in seconds (13 Mrd. a) "chessboard"-bet: fields vs. grains ≈ time since "big bang" in ns S. Kobe, Goettingen/Leipzig 2004 – p.11/?? Exact Ground States: Optimization Complexity: NP-complete Method: "branch-and-bound": exact nonlinear discrete optimization S. K., A. H ARTWIG , Comp. Phys. Commun. 16 (1978) 1 example: N = 8 0 B B B B B B J=B B B B B B @ 0 −5 0 −2 −10 0 −5 −4 0 0 −6 0 0 −3 0 −1 −2 −3 −5 −4 0 0 −1 0 −7 −5 0 0 0 0 −1 −4 −8 −1 0 0 1 C C C C C C C C C C C C A S. Kobe, Goettingen/Leipzig 2004 – p.12/?? Branch-and-Bound Tree S. Kobe, Goettingen/Leipzig 2004 – p.13/?? Branch-and-Bound Tree Energy of the branching level El (with El=N = Estates ): El = El−1 + 2 l−1 X k(kl) |Jkl | El ≥ El−1 example: N = 8; E1 = Eid = −77 S. Kobe, Goettingen/Leipzig 2004 – p.14/?? Branch-and-Bound Tree: Part S. Kobe, Goettingen/Leipzig 2004 – p.15/?? Branch-and-Bound Tree N = 8; Eid = −77; Ebound = −47 (heuristic solution: − − −) S. Kobe, Goettingen/Leipzig 2004 – p.16/?? Branch-and-Bound Tree N = 8; Eid = −77; Ebound = −47 (heuristic solution: − − −) S. Kobe, Goettingen/Leipzig 2004 – p.17/?? Branch-and-Bound Tree N = 8; Eid = −77; Ebound = −47 (heuristic solution: − − −); exact solution: E0 = −51. S. Kobe, Goettingen/Leipzig 2004 – p.18/?? Misfit parameter: A measure for frustration • • • Misfit parameter → a useful rescaling of the ground-state energy per spin ` ´ Definition: µ0 = 12 1 − e0 /eid 0 with eid 0 → reference energy of a related non-frustrated system: Jij = |Jij | (lattice model) Jij = • |Gij | √ N 1/2 (SK and related models) → eid 0 = (N − 1)/(2πN ) Properties: – µ0 is the fraction of each bond of the system, which is on average not satisfied. – Example: Antiferromagnetic triangular lattice → µ 0 = 13 , because one of three bonds of equal strength cannot be satisfied. – Maximum value: µ0 = 12 for highly frustrated systems (e.g. high-dimensional hypercubic and fcc fully frustrated ±J systems). – SK and related models belongs also to the class of systems with maximum occurring frustration. S. Kobe, Goettingen/Leipzig 2004 – p.19/?? Results I Lattice Models S. Kobe, Goettingen/Leipzig 2004 – p.20/?? Misfit parameter: Lattice models (±J spin glass) lattice type D µ0 e0 from honeycomb square triangular simple cubic hypercubic hypercubic 2 2 2 3 4 5 0.09 0.1495 0.22 0.202 0.24 0.26 W. L EBRECHT, E.E. VOGEL (1994) I.A. C AMPBELL , A.K. H ARTMANN , H.G. K ATZGRABER (2004) W. L EBRECHT, E.E. VOGEL (1994) various authors S. B ÖTTCHER , A. G. P ERCUS (2001) S. B ÖTTCHER, private communication (S. KOBE , J. K RAWCZYK in: Computational Complexity and Statistical Physics, in press) S. Kobe, Goettingen/Leipzig 2004 – p.21/?? • • • ~ >? &' ./ 67 <= FG NO VW ^_ fg no vw |} \] TU 45 $% ,- 23 :; DE LM RS Z[ de lm tu z{ rs jk bc JK BC *+ "# ! () 01 89 @A HI PQ XY `a hi pq xy Ground state clusters: Configuration vs. real space two (fixed) spin clusters in the real space: "red" and "yellow" spin domains Reversal of one spin domain ⇒ another cluster in the configuration space "free" spins between the spin domains ⇒ degeneracy of ground state clusters S. Kobe, Goettingen/Leipzig 2004 – p.22/?? Exact Landscape 420 clusters,13628 states altogether E 54 clusters, 2185 states 224 clusters, 8922 states 1503690 states E3 E2 E1 E0 System 4 • • • 4 4 the first excitation configuration space 18 states 12 states - 50 states, the second excitation - 300 states 1.635.796 states up to E3 , clusters of states form valleys (#1, #2), valleys are connected via saddle clusters S. Kobe, Goettingen/Leipzig 2004 – p.23/?? Number of configurations Landscape: Inner profile of the saddle cluster Hamming distance • • • Restriction of "transition" between clusters via saddle cluster Hamming distance of all pairs of states in the saddle cluster (right) "Bottleneck" has to be passed: "entropic barrier" S. Kobe, Goettingen/Leipzig 2004 – p.24/?? L = 12: Landscape and dynamics 1 0.98 E E1 95744 199168 54016 25344 28160 52864 20998 14080 saddel cluster >500000 states Ec0 276544 Ec2 59008 saddel cluster 55168 Ec0’ Ec2’ 14560 3840 0.96 10260 37600 5600 10848 20400 17280 10240 5600 6080 4480 q(t) 0.94 0.92 0.9 c0 E0 5632 states c2 1280 states 0.88 configuration space 0.86 0.01 0.1 1 10 100 1000 10000 100000 WTM TIME • • More complex saddle cluster structure (left) Dynamics: q vs. WTM time for 20 runs at T = 0.37 starting from one ground state (right) qpl = 0.915 ± 0.02 ⇒ hd = 73 ± 2 ⇒ "width of the valley" Ground states from A.K. H ARTMANN: Genetic Cluster-Exact Approximation S. Kobe, Goettingen/Leipzig 2004 – p.25/?? Results II Mean – Field Models S. Kobe, Goettingen/Leipzig 2004 – p.26/?? Energy exponents (zero temperature) Notation of J.-P. B OUCHAUD, F. K RZAKLA , O.C. M ARTIN , Phys. Rev. B 68 (2003) 224404 Lattice models: N = Ld EJ (L) = e0 Ld + e1 LΘs + . . . . eJ (L) = EJ (L)/Ld = e0 + e1 L−ω + . . . . Scaling of the energy fluctuations: 2 [EJ2 (L) − EJ (L) ]1/2 = σ0 LΘf + . . . . Scaling exponents: shift exponent Θs → ω = d − Θs fluctuation exponent Θf S. Kobe, Goettingen/Leipzig 2004 – p.27/?? Energy exponents ctd. L → N 1/d 0 eJ (N ) = e0 + e1 N −ω + . . . . 2 σJ (N ) = [e2J (N ) − eJ (N ) ]1/2 = σ0 N −ρ + . . . . with ω 0 = ω/d = 1 − ρ=1− Θs d Θf d S. Kobe, Goettingen/Leipzig 2004 – p.28/?? System sizes and numbers of realizations # of samples N x = 0.5 (SK) x = 1.0 19 257909 438242 25 229086 207149 32 123220 39 74827 16519 42 50797 13828 49 7933 1486 50 7724 1181 56 6274 282 59 2082 136 64 1779 70 634 75 236 81 126 85 112 90 42 S. Kobe, Goettingen/Leipzig 2004 – p.29/?? Results: Ground-state energy (SK model) 0.76 e0(N) 0.74 0.72 0.70 0.68 0.66 0.64 0.00 0.05 0.10 N 0.15 -1/2 0.20 S. Kobe, Goettingen/Leipzig 2004 – p.30/?? Results: Ground-state energy (x = 1: AFM model) No analytical solution known! 0.475 e0(N) 0.470 0.465 0.460 0.00 0.05 0.10 N 0.15 -1/2 0.20 S. Kobe, Goettingen/Leipzig 2004 – p.31/?? Results: Fluctuation exponent sigma 0.06 0.04 0.02 0.00 0 0.00 0.02 0.04 N upper curve: x = 0.5 (SK); 0.06 -0.75 lower curve: x = 1 (afm) 0.08 0.10 S. Kobe, Goettingen/Leipzig 2004 – p.32/?? Results: Misfit parameter 0.5 > <µ(N) 0 0.45 0.4 0.35 0.3 0.05 0.1 0.15 N 0.2 0.25 -1/2 upper curve: x = 1 (afm); lower curve: x = 0.5 (SK) blue stars: M. PALASSINI , cond-mat/0307713: hybrid genetic algorithm S. Kobe, Goettingen/Leipzig 2004 – p.33/?? Results of fitting procedures # x = 0.5 (SK) e0 ω0 e0 (N ) 3-p -0.7615(25) item 2-p x = 1.0 e0 ω0 0.698(23) -0.4755(5) 1.66(9) *) 0.684(2) - - µ0 (N −1/2 ) 3-p -0.7655(38) 0.652(31) -0.4756(4) 1.63(8) item 2-p *) 0.671(3) - - σJ (N ) 2-p ρ 0.710(5) ρ 0.736(9) *) The analytical value eRSB = −0.76321(3) is used for the 2-parameter fit 0 (M. A. C RISANTI , T. ROSSI , Phys. Rev. E 65, 046137 (2002)) S. Kobe, Goettingen/Leipzig 2004 – p.34/?? Phase transition 0.5 0.4 µ0 0.3 0.2 0.1 0 0.2 0.4 0.6 0.8 1 x S. Kobe, Goettingen/Leipzig 2004 – p.35/?? Summary • • • Optimization algorithms → exact results for finite spin glass models Semi-quantitative understanding of dynamics (relaxation) in lattice models For mean-field models: SK ground states for small N are consistent with RSB solution and other numerical results. • • Related models are introduced: AFM model is ”higher” frustrated for finite N . • Predictions from energy scaling: Ground-state energy e0,afm is estimated. – SK and AFM model have the same fluctuation exponent: Θ f /d ' 1/4. • – The shift exponent is different: Θs /d ' 1/3 (SK) and -2/3 (AFM). Outlook: – Phase transition between ferromagnetic and spin-glass ground state near x = 1/2? – Another related model: A fully connected ±J model S. Kobe, Goettingen/Leipzig 2004 – p.36/?? Acknowledgement K LAUS H ANDRICH A NDREAS H ARTWIG F RANK DASKE T HOMAS K LOTZ G ÜNTER M ILDE PAWEL P OLASZEK M ARIUSZ N OGALA J AROSŁAW K ŁOS J AROSŁAW K RAWCZYK E UGENIO VOGEL , J AIME C ARTES , PATRICIO VARGAS , D ORA A LTBIR S. Kobe, Goettingen/Leipzig 2004 – p.37/??