Lectures on Quantum Monte Carlo Methods B.B. Beard Christian Brothers University Firenze, Italy 25.06.2001-06.07.2001 5 Progression of Lectures 1 Stochastic Integration 6 The continuum limit 2 Random Numbers 7 Observables and Estimators 3 Classical Statistical Mechanical Simulations 8 Finite-size scaling 4 Cluster algorithms for classical models 9 More about the correlation length 5 Quantum Monte Carlo 10 Survey of other applications 5. Quantum Monte Carlo Or, how quantum maps to classical with a little trickery. “Quantum Monte Carlo” is a vast subject l Variational Monte Carlo ðoptimize trial wave functions ψ (x,α) l Diffusion Monte Carlo ðindependent “walkers” search Hilbert space l Green Function Monte Carlo ðUse (part of) the explicit propagator l Path Integral Monte Carlo → cluster algorithms Classical SM ↔ Quantum SM states µ states n Hamiltonian functional Hamiltonian operator H ( p, q ) H ( p$ , q$ ) Z = ∑ exp( − βH ) µ ( = ∑ exp − βE µ µ ) ( ) Z = Tr exp( − βH ) = ∑ n exp( − βH ) n n Quantum Mech ↔ Quantum SM imaginary time τ=it inverse temperature β= 1 T time evolution operator Boltzmann factor exp( − iHt ) = exp ( − H τ) exp( − βH ) ground state low temperature Quantum Heisenberg Models generalize classical spin models r r r r H = ∑ J S ⋅S → H = J ∑ S ⋅S isotropic xy x y xy x xy [ y ] r S x is a spin operator for site x → S xi , S yj = i δxy εijk S xk r r S x ⋅S x = S ( S + 1), if S = 1 2 then S i = l Anisotropic generalizations H = ∑ J x Si1 S 1j + J y Si2 S 2j + J z Si3 S 3j ij 1 2 σi J x = J y ≠ J z ( = 0 ?) J x ≠ J y ≠ J z etc. Heisenberg model connects several important concepts l FHM⇒ ferromagnetism l AFHM⇒ precursors of high-Tc superconductors l AFHM is asymptotically free ðtoy model for QCD σ model l Quantum XY model ⇔ U(1) gauge theory l Non-linear Heisenberg interactions can be studied in any dimension d l 1d chain ðexact solution S=1/2 from Bethe ansatz l In-between 1d & 2d: Spin ladders l 2d ðAFHM on square lattice: high-Tc SC, NLσM ðtriangular, hexagonal lattices: frustration l Nd ð “N-vector model”, large-N expansions, etc. Example: 2-spin interaction (S=1/2) r r l Two-spin interaction: H = JS1 ⋅S 2 1 0 − 41 βJ n exp( − βH ) n ′= e 0 0 ↑↑ ↑↓ where n = ↓↑ ↓↓ = 1,1 = 1 2 = 1 2 ( 1,0 ( 1,0 = 1,− 1 1 2 (1 − 0 eβJ ) 0 0 βJ 1 2 (1 − e ) 0 βJ 1 e + 1 ( ) 0 2 0 1 ) " up - down basis" 0,0 ) + 0,0 − 0 βJ 1 e + 1 ( ) 2 Fundamental difficulty: H = Σ many non-commuting terms r r l Consider, e.g. HM on 1d chain: H = J ∑ S ⋅S x ðsites •ó‚ó ƒ ó „ó l Direct etc. xy evaluation of trace Z=Tr(exp(-βH)) scales exponentially with L l Explicit diagonalization only feasible for small systems y Trotter-Suzuki expansion underlies a large class of solutions l Consider 1d spin chain: ðbreak H into two sets of commuting operators H1 = J ∑ x odd ( r r S x ⋅S x + 1$ ) ( ( H2 = J = lim N→ ∞ ∑ ∏ { nk } k x even )) Z = Tr exp( − βH ) = ∑ n exp − εβ( H1 + H 2 ) n ∑ r r S x ⋅S x + 1$ N n where N ε = 1 nk exp( − εβH1 ) nk + 1 nk + 1 exp( − εβH 2 ) nk + 2 1/ε = N = “Trotter number”; error is O(ε2) Quantum trace corresponds to path integral over classical variables l d-dimensional quantum⇔ (d+1) classical l Extra ‘euclidean time’dimension corresponds to inverse temperature β ðdifferent from ‘Monte Carlo time’ ðalternating layers of Trotter-Suzuki “sandwich” give “checkerboard” of plaquettes l Transfer matrix propagates configuration from one euclidean time slice to next Integrating over path space ( ) exp( − βH ) ↔ exp − S [] φ n [] S φ1 discretized euclidean time ti Integrating over path space ( ) exp( − βH ) ↔ exp − S [] φ n [] S φ2 discretized euclidean time ti Spin-time space comprises N TS “sandwiches” N =4 N t = 2dN = 8 time spin TS sandwich { { { { Problem now focuses on transfer matrix l Like S=1/2 2-spin matrix, but with “ε” nk exp( − εβH ) nk + 1 l ONLY 1 0 1 = exp( − 4 εβJ ) 0 0 (1 + 1 2 (1 − 1 2 0 e εβJ ) (1 − e εβJ ) 21 (1 + 0 6 NON-ZERO ELEMENTS ðnon-zero=“finite action” l Total 1 2 magnetization conserved 0 e εβJ ) e εβJ ) 0 0 0 0 1k ,k + 1 Finite-action plaquettes for FHM w =1 time spin w= 1 2 (1 + e εβJ ) w= 1 2 (1 − e εβJ ) Typical discrete-time 1d FHM configuration note every row has same number of “up” spins time spin Boltzmann weight of this configuration = ( (1 − 1 2 e εβJ )) ( (1 + 6 1 2 e εβJ )) 18 Typical discrete-time 1d FHM configuration note every row has same number of “up” spins time spin Boltzmann weight of this configuration = ( (1 − 1 2 e εβJ )) ( (1 + 6 1 2 e εβJ )) 18 Typical discrete-time 1d FHM configuration note every row has same number of “up” spins time spin Boltzmann weight of this configuration = ( (1 − 1 2 e εβJ )) ( (1 + 6 1 2 e εβJ )) 18 How should we sample the Boltzmann distribution? l Local Metropolis sampling? ðsingle flips not allowed! L ðergodicity requires moves that change total magnetization L ðdynamical critical exponent z ≈2 L l Cluster algorithm ðeliminates critical slowing down J ðimproved estimators J Choose FHM transition probabilities to satisfy detailed balance w =1 1 2 εβJ 1 + e ( ) 1 2 (1 − e εβJ ) Choose FHM transition probabilities to satisfy detailed balance w =1 1 2 εβJ 1 + e ( ) 1 2 (1 − pcontinue = 1 e εβJ ) p cross = 1 Choose FHM transition probabilities to satisfy detailed balance w =1 p continue 1 + e εβJ = 2 1 2 p cross εβJ 1 + e ( ) 1 − e εβ J = 2 1 2 (1 − pcontinue = 1 e εβJ ) p cross = 1 Choose FHM transition probabilities to satisfy detailed balance w =1 p continue 1 + e εβJ = 2 1 2 p cross εβJ 1 + e ( ) 1 − e εβ J = 2 with J<0, both these are < 1 1 2 (1 − pcontinue = 1 e εβJ ) p cross = 1 Fixing the sign problem for the AFHM (on some lattices) l Note if J > 0 (antiferromagnetic), then offεβJ 1 w = 1 − e )< 0 diagonal weight 2 ( ð but Boltzmann factor must be > 0 l Solution for bipartite lattices: ðbasis change: apply U=diag(i,-i,i,-i) ðequivalent to rotating every other spin 180o ðnon-bipartite (e.g. triangular) ⇒ frustration εβJ 1 w = e − 1) > 0 l Gives off-diagonal 2 ( Finite-action plaquettes for AFHM w =1 time spin w= 1 2 (e εβJ + 1) w= 1 2 (e εβJ − 1) Typical discrete-time 1d AFHM configuration note every row has same number of “up” spins time spin Boltzmann weight of this configuration = ( (e 1 2 εβJ ) ( (e − 1) 6 1 2 εβJ ) + 1) 18 Typical discrete-time 1d AFHM configuration note every row has same number of “up” spins time spin Boltzmann weight of this configuration = ( (e 1 2 εβJ ) ( (e − 1) 6 1 2 εβJ ) + 1) 18 Typical discrete-time 1d AFHM configuration note every row has same number of “up” spins time spin Boltzmann weight of this configuration = ( (e 1 2 εβJ ) ( (e − 1) 6 1 2 εβJ ) + 1) 18 Choose AFHM transition probabilities to satisfy detailed balance w =1 1 2 εβJ e ( + 1) 1 2 (e εβJ − 1) Choose AFHM transition probabilities to satisfy detailed balance w =1 pcontinue = 1 1 2 εβJ e ( + 1) 1 2 (e εβJ − 1) p jump = 1 Choose AFHM transition probabilities to satisfy detailed balance w =1 pcontinue = 1 p continue = 1 2 εβJ e ( + 1) 2 e εβJ + 1 p jump 1 2 (e εβJ − 1) e εβJ − 1 = εβJ e + 1 p jump = 1 Example program: 1dAFHM.exe l All the ingredients for DTCA are assembled ðpick random starting site in spin-time lattice ðfollow cluster-building rules until loop closes ðflip spins specify β,L l Demo defaults to Nt = L (not required) l Can CtrlCtrl-ShiftShift-A 1dAFHM: Noteworthy l Cluster loop is self-avoiding ð“1d object” unlike Wolff clusters in FIM l Loop can wrap in periodic time ðchanges total magnetization ðergodic l Loop can wrap in space ðrelated to helicity modulus More dimensions: split H into more parts H4 H2 H1 H3 Trotter-Suzuki Sandwich for 2d Square-Lattice AFHM Higher spin S>1/2: Add layers and projection operators The bad news: Trotter error requires treatment is O(ε2) l Typical routine involves repeating simulation for various N values l Extrapolate to continuum limit N→ ∞ l Error Quantum systems are amenable to cluster algorithms quantum ⇔ (d+1) classical l Ising-like variables l “Worldline” configurations reflect conserved quantities l Cluster updates are efficient and ergodic l Trotter error is controllable l d-dimensional ðbut we can do better... Progression of Lectures 1 Stochastic Integration 6 The continuum limit 2 Random Numbers 7 Observables and Estimators 3 Classical Statistical Mechanical Simulations 8 Finite-size scaling 4 Cluster algorithms for classical models 9 More about the correlation length 5 Quantum Monte Carlo 10 Survey of other applications