Numerical Methods in Many-body Physics

Transcription

1 Numerical Methods in Many-body Physics Reinhard M. Noack Philipps-Universität Marburg Exchange Lecture BME, Budapest, Spring 2007 International Research Training Group 790 Electron-Electron Interactions in Solids

2 Literature: R.M. Noack and S.R. Manmana, Diagonalization- and Numerical Renormalization-Group- Based Methods for Interacting Quantum Systems, in Lectures on the Physics of Highly Correlated Electron Systems IX, AIP Conference Proceedings 789, AIP, New York, 2005, and cond-mat/ Density Matrix Renormalization: A New Numerical Method in Physics, Lect. Notes Phys. 528, Eds. I. Peschel, X. Wang, M. Kaulke, and K. Hallberg, Springer, Berlin, E. Dagotto, Correlated electrons in high-temperature superconductors, Rev. Mod. Phys. 66, (1994). A. Booten and H. van der Vorst, Cracking large-scale eigenvalue problems, part I: Algorithms, Computers in Phys. 10, No. 3, p. 239 (May/June 1996); Cracking large-scale eigenvalue problems, part II: Implementations, Computers in Phys. 10, No. 4, p. 331 (July/August 1996). N. Laflorencie and D. Poilblanc, Simulations of pure and doped low-dimensional spin-1/2 gapped systems in Quantum Magnetism, Lect. Notes Phys. 645, Springer, Berlin, 2004, pp (2004), cond-mat/ K. G. Wilson, The renormalization group and critical phenomena, Rev. Mod. Phys. 55, (1983). U. Schollwöck, The density-matrix renormalization group, Rev. Mod. Phys. 77, 259 (2005). W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery, Numerical Recipes in C [or C++], 2nd ed., Cambridge University Press (1993).

3 Outline I. Exact Diagonalization (i) Introduction to interacting quantum systems (ii) Representation of many-body states (iii) Complete Diagonalization (iv) Iterative Diagonalization (Lanczos and Davidson) (v) Dynamics (vi) Finite temperature II. Numerical Renormalization Group (i) Anderson and Kondo problems (ii) Numerical RG for the Kondo problem (iii) Numerical RG for quantum lattice problems (iv) Numerical RG for a noninteracting particle III. From the NRG to the Density Matrix Renormalization Group (i) Better methods for the noninteracting particle (ii) Density Matrix Projection for interacting systems (iii) DMRG Algorithms (iv) DMRG-like algorithm for the noninteracting particle IV. The DMRG in Detail (i) Programming details (ii) Measurements (iii) Wavefunction transformations (iv) Extensions to higher dimension

4 Outline I. Exact Diagonalization (i) Introduction to interacting quantum systems (ii) Representation of many-body states (iii) Complete Diagonalization (iv) Iterative Diagonalization (Lanczos and Davidson) (v) Dynamics (vi) Finite temperature II. Numerical Renormalization Group (i) Anderson and Kondo problems (ii) Numerical RG for the Kondo problem (iii) Numerical RG for quantum lattice problems (iv) Numerical RG for a noninteracting particle III. From the NRG to the Density Matrix Renormalization Group (i) Better methods for the noninteracting particle (ii) Density Matrix Projection for interacting systems (iii) DMRG Algorithms (iv) DMRG-like algorithm for the noninteracting particle IV. The DMRG in Detail (i) Programming details (ii) Measurements (iii) Wavefunction transformations (iv) Extensions to higher dimension

5 Outline I. Exact Diagonalization (i) Introduction to interacting quantum systems (ii) Representation of many-body states (iii) Complete Diagonalization (iv) Iterative Diagonalization (Lanczos and Davidson) (v) Dynamics (vi) Finite temperature II. Numerical Renormalization Group (i) Anderson and Kondo problems (ii) Numerical RG for the Kondo problem (iii) Numerical RG for quantum lattice problems (iv) Numerical RG for a noninteracting particle III. From the NRG to the Density Matrix Renormalization Group (i) Better methods for the noninteracting particle (ii) Density Matrix Projection for interacting systems (iii) DMRG Algorithms (iv) DMRG-like algorithm for the noninteracting particle IV. The DMRG in Detail (i) Programming details (ii) Measurements (iii) Wavefunction transformations (iv) Extensions to higher dimension

6 Outline I. Exact Diagonalization (i) Introduction to interacting quantum systems (ii) Representation of many-body states (iii) Complete Diagonalization (iv) Iterative Diagonalization (Lanczos and Davidson) (v) Dynamics (vi) Finite temperature II. Numerical Renormalization Group (i) Anderson and Kondo problems (ii) Numerical RG for the Kondo problem (iii) Numerical RG for quantum lattice problems (iv) Numerical RG for a noninteracting particle III. From the NRG to the Density Matrix Renormalization Group (i) Better methods for the noninteracting particle (ii) Density Matrix Projection for interacting systems (iii) DMRG Algorithms (iv) DMRG-like algorithm for the noninteracting particle IV. The DMRG in Detail (i) Programming details (ii) Measurements (iii) Wavefunction transformations (iv) Extensions to higher dimension

7 Outline V. Recent Developments in the DMRG (i) Classical transfer matrices (ii) Finite temperature (iii) Dynamics (iv) Quantum chemistry (v) Time evolution (vi) Matrix product states (vii) Quantum information VI. Quantum Monte Carlo Methods (i) Review of classical Monte Carlo (ii) Variational and Green s function QMC (iii) World-line QMC (iv) Determinantal QMC (v) Loop Algorithm (vi) Stochastic Series Expansion (SSE) (vii) Diagrammatic QMC

8 Outline V. Recent Developments in the DMRG (i) Classical transfer matrices (ii) Finite temperature (iii) Dynamics (iv) Quantum chemistry (v) Time evolution (vi) Matrix product states (vii) Quantum information VI. Quantum Monte Carlo Methods (i) Review of classical Monte Carlo (ii) Variational and Green s function QMC (iii) World-line QMC (iv) Determinantal QMC (v) Loop Algorithm (vi) Stochastic Series Expansion (SSE) (vii) Diagrammatic QMC

9 I. Exact Diagonalization Direct diagonalization of Hamiltonian matrix on finite clusters ladder Goals ground state properties low-lying excitations dynamics, finite T,... y x two dimensionnal (2D) chain (1D) Advantages almost any system can be treated almost any observable can be calculated quantum-number resolved quantities numerically exact (for finite cluster) Limitation: exponential in lattice size

10 Largest sizes reached S = 1/2 spin models square lattice: N = 40 triangular lattice: N = 39, star lattice: N = 42 maximum dimension of basis: 1.5 billion t-j models checkerboard lattice with 2 holes: N = 32 square lattice with 2 holes: N = 32 maximum dimension of basis: 2.8 billion Hubbard models square lattice at half filling: N = 20 quantum dot structure: N = 20 maximum dimension of basis: 3 billion Holstein models chain with N = 14 + phonon pseudo-sites maximum dimension of basis: 30 billion

11 I (i) Interacting Quantum Systems Here: discrete, finite case system of N quantum mechanical subsystems, l = 1,..., N finite number of basis states per subsystem α l, α l = 1,..., s l more general case: s l (continuum or thermodynamic limit) Properties: N (thermodynamic limit) l x Basis direct product of component basis (continuous quantum field) α 1, α 2,..., α N α 1 α 2... α N total number of states: N l=1 s l arbitrary state in this basis ψ = {αl} ψ(α 1, α 2,..., α N ) α 1, α 2,..., α N behavior governed by Schrödinger equation H Ψ(t) = i t Ψ(t) or H ψ = E ψ (time-independent)

12 Hamiltonians In general, Hamiltonians can connect arbitrary numbers of subsystems H = H (1) l + H (2) lm H (4) lmnp +... l l,m l,m,p H (1) l usually determines α l H (2) lm, sometimes H(4) lmnp will be important here H (2) lm often short-ranged Typical terms: tight-binding term: V(x) H tb = ψ( x) l,m,σ t lm c l,σ c m,σ localized Wannier orbitals (unfilled d- or f- orbitals in transition metals) states 0,,, per orbital 4 N degrees of freedom overlap between near orbitals hopping t lm short ranged (n.n., possibly n.n.n.)

13 local (Anderson) disorder H A l = σ λ l n l,σ, n l,σ c l,σ c l,σ Coulomb interaction between electrons H C lm = screening leads to on-site (Hubbard) interaction e 2 r l r m Hl U = U n l, n l, near-neighbor Coulomb interaction Hlm V = V n l n l+ˆr, (n l σ n l,σ) etc. Spin models S i localized quantum mechanical spins (S = 1/2, 1, 3/2,...) states S S S (2S + 1) N degrees of freedom Heisenberg exchange Hlm Heis = J S l S m = J z Sl z Sm z J xy ( S + l S m + S ) l S+ m strong coupling limit of the Hubbard model at n = 1 (S = 1/2) AF exchange J = 4t2 U variations: J z J xy (Ising or XY anisotropy), Hl n H bq lm = J 2 (S l S m ) 2 (biquadratic) = D(Sz l )2 (single-ion),

14 t J model: strong-coupling limit of doped Hubbard Hlm tj = P Hlm tb P + J (S l S m 14 ) n l n m double occupancy projected out (P) - 3 states/site Anderson impurity - hybridized d (or f) orbital with on-site interaction ) Hl AI = ε d n d l + V (d l,σc l,σ + H.c. + Un d l, n d l, single impurity or lattice (PAM) possible Kondo impurity - localized d spin S Hl K = J ( ) K 2 S l c l,α σ α,β c l,β limit of symmetric Anderson impurity at strong U

15 Lattices square lattice Kagomé lattice Described by unit cell Bravais lattice: translation vectors T 1, T 2 (2D) T 2 T 1

16 finite lattices: finite multiples of T 1, T 2 and boundary conditions periodic, antiperiodic open lattice symmetries: translation multiples of Bravais lattice vector + periodic (AP) BCs rotations e.g., π/2 for a square lattice (group C 4v ) reflection about symmetry axis Tilted clusters 40-site cluster, square lattice (a = 1) T 1 = (1, 0), T 2 = (0, 1) Spanning vectors: F 1 = (6, 2), F 2 = ( 2, 6) In general, F 1 = (n, m), F 2 = ( m, n) N = n 2 + m 2 translational symmetry satisfied reflection/rotation symmetries become more complicated

17 I (ii) Representation of Many-Body States mapping to (binary) integers: spin-1/2 Heisenberg: spin flip = bit flip Hubbard with N σ = {0, 1} N 1 N N 1 1 N N l N l N l N l or N e l S z l other models (t J, S = 1 Heisenberg,...) more complicated Symmetries: given group G with generators {g p } [H, g p ] = 0 H block diagonal (Hilbert space can be divided) Continuous conservation of particle number, S z U(1) permutations of bits total spin SU(2) difficult to combine with space group spin inversion (Z 2 ) can be used Space group translation: abelian local states point group (reflections and rotations): non-abelian in general form symmetrized linear combination of local states

18 Example Reduction of Hilbert space for S = 1/2 Heisenberg on cluster full Hilbert space: constrain to S z = 0: using spin inversion: utilizing all 40 translations: using all 4 rotations: dim= 2 40 = dim= dim= dim= dim= 430, 909, 650

19 I (iii) Complete Diagonalization To solve H ψ = E ψ (H real, symmetric) Method (Numerical Recipes, Ch. 11) 1. Householder transformation - reduction to tridiagonal form T 2n 3 /3 operations (4n 3 /3 with eigenvectors) 2. Diagonalization of a tridiagonal matrix roots of secular equation: inefficient QL (QR) algorithm - factorization T = Q L, Q orthogonal, L lower triangular 30n 2 operations ( 3n 3 with eigenvectors) Useful for: Simple problems, testing Matrix H dense Many eigenstates required But H must be stored entire matrix must be diagonalized

20 I (iv) Iterative Diagonalization Idea: project H onto a cleverly chosen subspace of dimension M N good convergence of extremal eigenstates Methods Power method v n = H n v 0 conceptually simple, but converges poorly needs only two vectors, v n & v n 1 Lanczos: orthogonal vectors in Krylov subspace (spanned by { v n }) simple to implement memory efficient - only 3 vectors needed at once works well for sparse, short-range H Davidson: subspace expanded by diagonal approximation to inverse iteration higher-order convergence than Lanczos (usually) implementation more complicated works best for diagonally-dominated H Jacobi-Davidson: generalization of Davidson nontrivial problem-specific preconditioner (approximation to inverse) can be applied to generalized eigenvalue problem A x = λ B x (A, B general, complex matrices)

21 Lanczos Algorithm 0) choose u 0 (random vector, ψ 0 from last iteration,...) 1) form u n+1 = H u n a n u n b 2 n u n 1 where a n = u n H u n u n u n and b 2 n = u n u n u n 1 u n 1 2) Is u n+1 u n+1 < ε? yes: do 4) then stop no: continue 3) repeat starting with 1) until n = M (maximum dimension) 4) diagonalize u i H u j (tridiagonal) using QL algorithm diagonal elements D = (a 0, a 1,..., a n ), off-diagonal elements O = (b 1, b 2,..., b n ) eigenvalue Ẽ0, eigenvector ψ 0 5) repeat starting with 0), setting u 0 = ψ 0

22 Convergence of Lanczos Algorithm eigenvalues converge starting with extremal ones excited states can get stuck for a while at longer times: true eigenvalus converged spurious or ghost eigenvalues produced multiplicity of eigenstates increases

23 Example: 2D t-j Model Binding of 2 holes R: average hole-hole distance B : binding energy (Poiblanc, Riera, & Dagotto, 1993) holes closer than two lattice spacings pair binding for J > J c, but large finite-size effects Does binding persist for larger lattices and constant doping (more holes)?

24 I (v) Dynamics with Exact Diagonalization Time-dependent correlation functions C(t) = i ψ 0 A(t) A (0) ψ 0 Fourier transform to frequency space (retarded) C(ω + iη) = ψ 0 A (ω + iη H + E 0 ) 1 A ψ 0 (resolvent) Spectral function I(ω) = 1 π lim η 0 + Im C(ω + iη) Examples from theory and experiment name notation operators experiment single-particle spectral weight A(k, ω) A = c k,σ photoemission structure factor S zz (q, ω) A = Sq z neutron scattering optical conductivity σ xx (ω) A = j x optics 4-spin correlation R(ω) k R k S k S k Raman scattering

25 Methods Krylov space method (continued fraction) restart Lanzcos procedure with 1 u 0 = ψ0 A A ψ 0 A ψ 0 In this Lanczos basis, Interpretation: C(z = ω + iη + E 0 ) = ψ 0 A A ψ 0 z a 1 b 2 2 z a 2 b2 3 z a 3... calculation of eigenvector not needed consider Lehmann representation of spectral function I(ω) = n ψ n A ψ 0 2 δ(ω E n + E 0 ) poles and weights of C(z) determine I(ω) weight decreases with n truncate after M steps spectrum discrete finite broadening η

26 Correction vector method (Soos & Ramasesha, 1984) Calculate vectors directly, then Advantages: φ 0 = A ψ 0, φ 1 = (ω + iη H + E 0 ) 1 φ 0 I(ω) = 1 π Im φ 0 φ 1 spectral weight calculated exactly for a given range nonlinear spectral functions computed by higher order correction vectors can be run in conjunction with Davidson algorithm Disadvantage: system (H z) φ 1 = φ 0 must be solved for each ω desired

27 Example: Dynamics in 2D t-j Model Single-particle spectral weight A(k, ω) at k = k F = (π/2, π/2) for one hole 4 4 lattice (Dagotto, Joynt, Moreo, Bacci, & Gagliano, 1993) Does a single quasiparticle propagate in an antiferromagnet? strongly localized hole with string excitations at J/t = 1.0 quasiparticle peak remains until J/t = 0.4 lump with pseudogap at J/t = 0.2 pseudogap due to finite-size effects at J/t = 0 (symmetric in ω)

28 I (vi) Finite Temperature with Exact Diagonalization To calculate finite-t properties in orthonormal basis n A = 1 N N n Ae βh n, Z = n e βh n, Z n Problem: expensive to calculate for all n Idea: stochastic sampling of Krylov space (Jaklic & Prelovsek, 1994 ) A 1 N s R M e βε(r) m r Ψ (r) m Ψ (r) m A r Z R where Z s s N s R R r r M m m e βε(r) m s over symmetry sectors of dimension N s r m n r Ψ (r) m 2 average over R random starting vectors Ψ(r) 0 Lanczos propagation of starting vectors: Ψ(r) m at step m useful if convergence good when M N s and R N s

29 Properties related to high-t expansion T limit correct high to medium T properties in thermodynamic limit low-temperature limit correct (on finite lattice), up to sampling error reduction of (large) sampling error: (Aichhorn et al., 2003) start with: A = 1 Z N n e βh/2 A e βh/2 n, n twofold insertion of Lanczos basis smaller fluctuations at low T can calculate thermodynamic properties: specific heat, entropy, static susceptibility,... static correlation functions dynamics: A(k, ω), S zz (q, ω), σ xx (ω),...

30 Example: t-j Model at finite T Hole concentration c h (= x) vs. chemical potential shift µ = µ h µ 0 h 2D t-j model, 16, 18, 20 sites, t/j = 0.3, t = 0.4eV 0.3 c h 0.2 T/t= experiment (Jaklic & Prelovsek, 1998) µ (ev) experimental results for LSCO from photoemission shift (Ino et al., 1997) holes only when µ < µ 0 h 1.99t as T 0 compressibility finite no phase separation

31 Optical conductivity compared with various cuprates at intermediate doping Bi 2 Sr 2 CaCu 2 O 8 YBa 2 Cu 3 O 7 n h =3/16 J/t=0.3 σ ρ T/t=0.15 T/t=0.3 T/t=0.5 T/t=1.0 (Jaklic & Prelovsek, 1998) 0.2 La 1.8 Sr 0.2 CuO ω (ev) Cuprates measured at T < 200K, c h somewhat uncertain high-t falloff slower for materials transitions to higher excited states? experimental curves: LCSO, c h x = 0.2 (Uchida et al., 1991) BISCCO, c h 0.23 (Romero et al., 1992) YBCO, c h 0.23 (Battlogg et al., 1994)

32 Discussion: Exact Diagonalization Method conceptually straightforward, numerically exact Iterative diagonalization allows the treatment of surprisingly large matrices Efficient implementation using symmetries useful System sizes nevertheless strongly restricted Extensions to basic method can calculate dynamical correlation functions finite temperature properties Not mentioned here, but also possible: calculation of full time evolution of quantum state with Lanczos Benchmark for other methods, useful when other methods fail