MBL THROUGH MPS Bryan Clark and David Pekker arxiv:1410.2224 Perimeter - Urbana
Many Body Localization... It has been suggested that interactions + (strong) disorder LX H = [h i Si z X + J Ŝ i Ŝi+1] i=1 h i 2 [ W, W ] produce a many-body localized phase at all temperatures.
A many body localized phase... has the following phenomenological properties: doesn t thermalize local constants of motion has atypical eigenstates
Thermalization Large quantum system Small subsystem: small Thermalized if, in equilibrium, small = thermal at the temperature that corresponds to the energy density of the system.
What thermalizes? States which obey the eigenstate thermalization hypothesis... Energy eigenstates will decohere. Nearby eigenstates will stay coherent. = 0 0i + 1 1i +... + n ni e ith = 0 e ite 0 0i + 1 e ite 1 1i +... + n e ite n ni To thermalize, nearby eigenstates must look the same with respect to local observables. Eigenstate Thermalization Hypothesis
and what doesn t thermalize? States which don t obey the eigenstate thermalization hypothesis... Integrable systems are one such system... which don t thermalize because of the extensive number of conserved quantities. Instead the density matrix approaches the generalized Gibbs Ensemble GGE. We should anticipate then, that MBL phases shouldn t obey ETH and should be like integrable systems.
A many body localized phase... has the following phenomenological properties: doesn t thermalize local constants of motion has atypical eigenstates
Atypical Eigenstates Low Entanglement (obeys area law) Increasing disorder Notice that there is interesting quantum effects happening at non-zero temperature 0 N=12 Delta=(0.05,10.5) Average over 10 disorder realizations and 5 eigenstates Strange Spectral Statistics... Poisson vs. Gaussian orthogonal ensemble
How should we understand these strange phenomena? We need a unifying understanding of MBL? MBL is all about the eigenstates. Can we characterize them?
H = LX h i S z i Simple Structure of Eigenstates i=1 h i 2 [ W, W ] ground state finite energy density state
Simple Structure of Eigenstates LX H = [h i Si z ] i=1 h i 2 [ W, W ] Ground State: Excited State: Highly Excited State:
Is their a simple structure for the eigenstates in a many-body localized phase? LX H = [h i Si z X + J Ŝ i Ŝi+1] i=1 h i 2 [ W, W ] Ground State: Excited State: Highly Excited State: A 1, 1 A 2, 2 A 3, 3 A 4, 4 A 5, 5 A 1, 1 B 2, 2 A 3, 3 A 4, 4 A 5, 5 B 1, 1 B 2, 2 A 3, 3 B 4, 4 A 5, 5
A wave function is an object that eats a configuration of spins and generates a number: (", #, ") =0.3 A product state eats a configuration of spins and generates a number by taking the product of complex numbers. (", #, ") =c 1," c 2,# c 3," A matrix product state (MPS) eats a configuration of spins and generates a number by taking the product of matrices/vectors. (", #, ") =M 1," M 2,# i ij M 3," j 2N matrices Each of size DxD
ii 2N matrices Each of size 1x1 1 0 0 1 1 0 A product state is a very simple MPS
A operator is an object that eats two configuration of spins and generates a number: ("#, #", "") =0.9 A matrix product operator (MPO) eats two configuration of spins and generates a number by taking the product of matrices/ vectors. ("#, #", "") =O 1,"# i O 2,#" ij O 3,"" j 4N matrices Each of size DxD
X 1 (O 1 0 1 1 M 1 1 ) X 2 (O 2 0 2 2 M 2 2 ) X 3 (O 3 0 3 3 M 3 3 ) O 1 0 1 0 2 O 2 O 3 1 2 3 0 3 M1 1M 2 2M 3 3 O " 0 1 O # 0 2 O " 0 3 An operator applied to a product state picks out two of the four matrices (per site) from the operator. ii 1 0 0 1 1 0
X 1 (O 1 0 1 1 M 1 1 ) X 2 (O 2 0 2 2 M 2 2 ) X 3 (O 3 0 3 3 M 3 3 ) O 1 0 1 0 2 O 2 O 3 1 2 3 0 3 M1 1M 2 2M 3 3 O # 0 1 O # 0 2 O " 0 3 An operator applied to a product state picks out two of the four matrices (per site) from the operator. ii Different product state - different matrices. 0 0 1 1 1 0
ÛĤÛ = D Û ii = e i i All the eigenstates are encoded with 4n matrices! Ground State: G 1 G 2 G 3 G 4 G 5 Excited State: E 1 E 2 G 3 E 4 G 5 e i i O " 0 1 O # 0 2 O " 0 3 Key question: are the sizes of E and G constant with bond dimension? Û ii 1 0 0 1 1 0
Let s check this... Ĥ Û MPO What s the bond U maps product states to eigenstates. dimension? Many such mappings Which one we pick is important! 1 This is no good. Want to preserve locality in mapping.
MBL IPR Are you spread out over all of Hilbert space or just a little piece of it? IPR Px (x) 4 P x (x) 2 MBL should be localized Let map product state i to eigenstate [i] Minimize X i hi [i]i 2 over all functions This can be done with bipartite matching. Low disorder High disorder
For large disorder, the bond dimension saturates.
A more refined look at the MPO bond dimension. Large disorder Low disorder Low bond dimension System-dependent bond dimension Disorder near transition Broad bond dimension
A many body localized phase... has the following phenomenological properties: doesn t thermalize has atypical eigenstates local constants of motion
Hamiltonian H We want a basis where the Hamiltonian becomes essentially block diagonal.
Seeing the transition... # off-diagonal terms Apply rotation into blocks... Check the number of off-diagonal terms. Almost a ground state property... Disorder Strength Increasing Disorder
[Ĥ,Ô] =0 Constants of Motion Let Ô = U(I I z I I)U ÔĤ ĤÔ ÔUU Ĥ ĤUU Ô U (ÔUU Ĥ U ÔUU ĤU ĤUU Ô)U U ĤUU ÔU)
Constants of Motion Let Ô = U(I I z I I)U UẐ3U = U X i i iihi! U X! i e i ihe i i H X i i e i ihe i! X! i e i ihe i H =0 i [Ĥ,Ô] =0 These are constants of motions...are they orthogonal? local?
Let Ô 3 = U(I I z I I)U Ô 2 = U(I z I I I)U Ô 2 Ô 3 = U(I z I I I)U U(I I z I I)U = U(I z I I I)(I I z I I)U Yes they are orthogonal.
U I U z U U The constant of motion differs from the identity by a single matrix. This means it can only differ locally from the identity. It s a local constant of motion!
Thermalization If ETH then volume law Delocalized phase by exact diagonalization (16 sites) 0.5 Xs k z \, k=11 0.0-0.5-30 -20-10 0 10 20 Energy
ETH Area law A first preliminary test of ETH: 24 sites! Nearly degenerate states have very different local observables. Failure of ETH! 1.0 0.5 Xs 5 z \ 0.0-0.5-1.0 0.32 0.34 0.36 0.38 0.40 0.42 0.44 Energy
If you have n constant of motions represented by MPO s which differ by a single matrix from an identity MPO then there is a MPO representation
Doesn t Thermalize Won t overlap GGEEGEGGEGEGGEGGEEW EGEEGEGGEGEGGEGGEEE will overlap Can t get information down here. GGEEGEGGEGEGGEGGGGG GGEEGEGGEGEGGEGGGGE GGEEGEGGEGEGGEGGGEG GGEEGEGGEGEGGEGGGEE GGEEGEGGEGEGGEGGEGG GGEEGEGGEGEGGEGGEGE GGEEGEGGEGEGGEGGEEG GGEEGEGGEGEGGEGGEEE
Eigenstate Filter Need a property to filter out an eigenstate. A Entanglement: In a many-body localized phase, the eigenstates have low entanglement: area law* S = Tr[ A log A ] Key insight: Although eigenstates have low entanglement, superposition of eigenstates have larger entanglement. = 0 0 + q 1 2 0 1 Increasing disorder 0 We need a low entanglement filter! Want the lowest entangled state close to a given energy. * Bela and Chetan; Brian Swingle; Abanin, et. al N=12 Delta=(0.05,10.5) Average over 10 disorder realizations and 5 eigenstates
Eigenstates... Diagonalizing in our little block produce PGE (pretty-good eigenstates). Can we get great great eigenstates?
Getting a MPS MPS are a good representation. How do we get them? DMRG on (H E) 2 + artificially drop bonddimension during run. s= YH 2 ] - XH\ 2 Bond Dimension 30 25 20 15 10 10-4 10-5 5 10 0.1 0.01 0.001 Metric of Goodness 0 0 2000 4000 6000 8000 10 00012 00014 000 MPS step 1 0 2000 4000 6000 800010 00012 00014 000 MPS step Disorder 10; 24 Sites Level Spacing: 1.5*10-5 Energy -3.4-3.6-3.8-4.0 Typical DMRG: For a site produce an effective Hamiltonian H and solve for the ground state of H Modified DMRG: For a site produce an effective Hamiltonian H and choose the eigenstate of H closest to the current energy of your state. 0 200 400 600 800 1000 1200 10 DMRG step Disorder 20; 24 Sites Level Spacing: 3*10-5 se= YH 2 ] - XH\ 2 0.1 0.001 10-5 0 200 400 600 800 1000 1200 DMRG step * Also time evolution variants of these; easier to analyze but general experience is diagonalization approach tends to be more accurate.
Other Eigenstates New Approach: For a site produce an effective Hamiltonian H and choose the eigenstate of H closest to the current energy For a site produce an effective Hamiltonian H and choose another eigenstate. Largest overlap 1.0 0.8 0.6 0.4 0.2 0.0-2.0-1.5-1.0-0.5 0.0 0.5 1.0 Generated energy
Conclusion MBL is a phase where a local unitary transform makes you almost block diagonal with extensive # of blocks. MBL is a phase with MPS eigenstates. Blocks don t talk => No ETH => No thermalization => Any eigenstate supports only local blocks (localized) All eigenstates are (essentially) ground states => Area law => Possion level statistics We can explicitly produce the block diagonal basis and eigenstates. We can find very high quality eigenstates to test MBL on.