The density matrix renormalization group Adrian Feiguin.
-
Upload
franklin-mcdowell -
Category
Documents
-
view
233 -
download
1
Transcript of The density matrix renormalization group Adrian Feiguin.
![Page 1: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/1.jpg)
The density matrix renormalization group
Adrian Feiguin
![Page 2: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/2.jpg)
![Page 3: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/3.jpg)
![Page 4: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/4.jpg)
Some literature-S. R. White:. Density matrix formulation for quantum renormalization groups, Phys. Rev. Lett. 69, 2863 (1992).. Density-matrix algorithms for quantum renormalization groups, Phys. Rev. B 48, 10345 (1993).
-U. Schollwöck. The density-matrix renormalization group, Rev. Mod. Phys. 77, 259 (2005).
-Karen Hallberg. Density Matrix Renormalization: A Review of the Method and its Applicationsin Theoretical Methods for Strongly Correlated Electrons, CRM Series in Mathematical Physics, David Senechal, Andre-Marie Tremblay and Claude Bourbonnais (eds.), Springer, New York, 2003. New Trends in Density Matrix Renormalization, Adv. Phys. 55, 477 (2006).
-The “DMRG BOOK”: Density-Matrix Renormalization - A New Numerical Method in Physics: Lectures of a Seminar and Workshop held at the Max-Planck-Institut für Physik ... 18th, 1998 (Lecture Notes in Physics) by Ingo Peschel, Xiaoqun Wang, Matthias Kaulke and Karen Hallberg
-R. Noack and S. Manmana. Diagonalization- and Numerical Renormalization-Group-Based Methods for Interacting Quantum SystemsProceedings of the "IX. Training Course in the Physics of Correlated Electron Systems and High-Tc Superconductors", Vietri sul Mare (Salerno, Italy, October 2004)AIP Conf. Proc. 789, 93-163 (2005)
-A.E. Feiguin. The Density Matrix Renormalization Group and its time-dependent variantsVietri Lecture Notes. http://physics.uwyo.edu/~adrian/dmrg_lectures.pdf
![Page 5: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/5.jpg)
Brief history and milestones• (1992) Steve White introduces the DMRG.• (1995-…) Dynamical DMRG. (Hallberg, Ramasesha et al, Kuhner and White,
Jeckelmann)• (1995) Nishino introduces the transfer-matrix DMRG (TMRG) for classical systems. • (1996-97) Bursill, Wang and Xiang, Shibata, generalize the TMRG to quantum
problems.• (1996) Xiang adapts DMRG to momentum space.• (2001) Shibata and Yoshioka study FQH systems.• (2004) Vidal introduces the TEBD. (time-evolving block decimation)• (2005) Verstraete and Cirac introduce an alternative algorithm for MPS’s and
explain problem with DMRG and PBC.• (2006) White and AEF, and Daley, Schollwoeck et al generalize the ideas within a
DMRG framework: adaptive tDMRG.
… the DMRG has been used in a variety of fields and contexts, from classical systems to quantum chemistry, to nuclear physics…
![Page 6: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/6.jpg)
Exact diagonalization“brute force” diagonalization of the Hamiltonian matrix.
… anything you want to know… but… only small systems
H |x = E |xH : Hamiltonian operator|x : eigenstateE : eigenvalue (ENERGY)
Schrödinger's Equation:
All we need to do is to pick a basis and write the Hamiltonian matrix in that basis
![Page 7: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/7.jpg)
Exact diagonalization recipe
Ingredients●Lattice (geometry)
●Basis of states (representation)●Hamiltonian (model/interactions)
![Page 8: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/8.jpg)
Lattices
1D chain
Ladder
2D
![Page 9: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/9.jpg)
Boundary conditionsOpen Cylindrical
Periodic
![Page 10: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/10.jpg)
What degrees of freedom do we care about?
core electrons
valence band
![Page 11: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/11.jpg)
Basis of statesOccupation number representation
(1 orbital per site, spin ½)
Empty Spin up Spin down Double occupied
Dimension = number of configurations = 4N
N: number of lattice sites
|o |↑ |↓ |2States |x =|x
1|x
2|x
3|x
4....|x
N
![Page 12: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/12.jpg)
Hamiltonian H=∑i,jH
ij
Kinetic energyH
K = -t (c†
i↑cj↑+c†i↓c j↓+h.c.)
On site interaction(diagonal)H
U = U ni↑ni↓ U
HK |↑o = -t |o↑
HU |2 = U |2
U→∞
![Page 13: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/13.jpg)
Spin flipH
J = J/2 (Si
-Sj++Si
+Sj-) HJ |↑↓ = J/2 |
↓↑
Ising term (diagonal)H
Ising = Si
zSjz
HIsing |↑↓ = -J/4 |↑↓HIsing |↑↑ = J/4 |↑↑
![Page 14: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/14.jpg)
Hubbard model
HHubbard
= -t ∑<i,j>
(c†i↑cj↑+c†
i↓c j↓+h.c.)+U ∑ini↑ni↓
t-J model
Ht-J
= -t ∑<i,j>
Pt(c†i↑cj↑+c†
i↓c j↓+h.c.)P+HHeis
Heisenberg model
HHeis
= J ∑<i,j>
SizSj
z+ 1/2 (Si-Sj
++Si+Sj
-)
![Page 15: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/15.jpg)
Symmetries SH=HS
Reflections
Translations
D' = D / N
D' = D / 2
Particle number conservation => Ntotal
Spin conservations => Sztotal
Spin reversal => |↑↓ ± |↓↑
|yk = (1/M) ∑iaki
Ti |f; aki
=exp(ikxi)
![Page 16: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/16.jpg)
Block diagonalization
0
00
0
![Page 17: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/17.jpg)
ED Example: Heisenberg chain
Geometry:1D chain
Basis:
HHeis
= J ∑<i,j>
SizSj
z+ 1/2 (Si-Sj
++Si+Sj
-)
Model Hamiltonian:
|↑↓↑↓; |↓↑↓↑;|↑↑↓↓; |↓↑↑↓; |↓↓↑↑; |↑↓↓↑
![Page 18: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/18.jpg)
Applying translations:
|1=1/(22){(1+ ei2k )|↑↓↑↓+ eik(1+ei2k)|↓↑↓↑}|2=1/2{|↑↑↓↓+eik|↓↑↑↓+ei2k|↓↓↑↑+ei3k|↑↓↓↑}
Translations
With k=0,-/2, /2,
k=0) |1=1/2{|↑↓↑↓+|↓↑↓↑}|2=1/2{|↑↑↓↓+|↓↑↑↓+|↓↓↑↑+|↑↓↓↑}
k= -/2) |2=1/2{|↑↑↓↓+e-i/2|↓↑↑↓-|↓↓↑↑+ei/2 |↑↓↓↑}
k= /2) |2=1/2{|↑↑↓↓+ei/2|↓↑↑↓-|↓↓↑↑+e-i/2 |↑↓↓↑}
k= ) |1=1/2{|↑↓↑↓-|↓↑↓↑}|2=1/2{|↑↑↓↓-|↓↑↑↓+|↓↓↑↑-|↑↓↓↑}
![Page 19: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/19.jpg)
Limitations : small lattices• Hubbard model: 20 sites at half filling, 10↑ and 10↓, D=20!
(10!10!)x20!(10!10!) = 2.4e+10. After symmetries D'=1.1e+8• t-J model (only |o, |↑ and |↓ states): 32 sites with 4 holes,
14↑ and 14↓, D = 32!/(14!18!)x18!/(14!4!) = 1.4e+12; D'=5.6e9
• Heisenberg model (only |↑ and |↓ states): 36 sites, 18↑ and 18↓, D = 36!/(18!18!)=9075135300; D' =D/(36x2x2x2x2)=1.5e6 states
![Page 20: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/20.jpg)
Exact diag. is limited by system size… How can we overcome
this problem?
Po’ man’s solution: What about truncating the basis?
![Page 21: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/21.jpg)
“Classical” analogyImage compression algorithms (e.g. Jpeg)
We want to achieve “lossless compression”… or at least minimize the loss of information
![Page 22: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/22.jpg)
Idea 1: Truncated diagonalization
|gs =∑ ai|x
i
Usually, only a few important states possess most of the weight
, ∑ |ai|2 = 1
Error = 1-∑' |ai|2
Cut here
![Page 23: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/23.jpg)
Truncated diagonalization (continued)1) We choose a small set of configurations that we know (from results in
small systems) are important. E.g. |↑↓↑↓↑↓↑↓2) We apply the Hamiltonian H|x = |y, expanding the basis up to a
dimension D.
3) We diagonalize and obtain the ground state: |gs =∑ai|x
i
4) We order the weights |ai|2 in descending order
5) We truncate the basis keeping m states with larger weights
6) We go back to 2) until we reach convergence
NOTICE: We are still working in the occupation number representation
![Page 24: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/24.jpg)
Idea 2: Change of basis
Can we rotate our basis to one where the weights are more concentrated, to minimize the error?
|gs =∑ ai|x
i , ∑ |a
i|2 = 1 Error = 1-∑' |a
i|2
Cut here Cut here
![Page 25: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/25.jpg)
What does it mean “to truncate the basis”
If we truncate
This transformation is no longer unitary, does not preserve norm ->loss of information
![Page 26: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/26.jpg)
The case of spinsThe two-site basis is given by the states
|ss’ ={|↑↑;|↑↓;|↓↑; |↓↓}
We can easily diagonalize the Hamiltonian by rotating with the matrix:
That yields the eigenstates:
![Page 27: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/27.jpg)
The case of spins…
![Page 28: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/28.jpg)
Numerical Renormalization Group
![Page 29: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/29.jpg)
Let’s consider the 1d Heisenberg model
iiiii
zi
zi
iii SSSSSSSSH 1111 2
1
10
00;
01
00;
2/10
02/1000 SSS z
For a single site , the operator matrices are:
We also need to define the identity on a block of l sites
22 dimensions with ;
10
01ll
lI
![Page 30: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/30.jpg)
Building the Hamiltonian a la NRG
0000002 2
1
2
1SSSSSSH zz
1 2
ll-1
32
0000001123 2
1
2
1SSSSSSIIHH zz
221 HIIHH llll
This recursion will generate a 2lx2l Hamiltonian matrix that we can easily diagonalize
![Page 31: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/31.jpg)
Another way to put it…
ll-1
01010111
00210200211
000000211
2
1
2
12
1
2
1
2
1
2
1
SSSSSSIH
SSISSISSIIH
SSSSSSIIHH
llzz
ll
llzz
ll
zzlll
01 OIO ll with
![Page 32: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/32.jpg)
Adding a single site to the block
1ll
1ls
Before truncating we build the new basis as:
11 lll s
And the Hamiltonian for the new block as
...'0,01,1, OOHIIHH lLllLlL
01, OIO llL with
![Page 33: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/33.jpg)
|y = ∑ijy
ij|i| j
Dim=2L
Dimension of the block grows exponentially
Idea 3: Density Matrix Renormalization Group
S.R. White, Phys. Rev. Lett. 69, 2863(1992), Phys. Rev. B 48, 10345 (1993)
![Page 34: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/34.jpg)
Block decimation
Dim=2N Dim=m
constant
|y = ∑ijy
ij|i| j
![Page 35: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/35.jpg)
The density matrix projection
Universe
system
|ienvironment
| j
We need to find the transformation
that minimizes the distance
S=||' -||2
| = ∑ij
ij|i| j |' = ∑m
jaj|| j
Solution: The optimal states are the eigenvectors of the reduced density matrix
ii'
= ∑j
ij
i'jTr = 1
with the m largest eigenvalues
Solution: The optimal states are the eigenvectors of the reduced density matrix
ii'
= ∑j
ij
i'jTr = 1
with the m largest eigenvalues
![Page 36: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/36.jpg)
Understanding the density-matrix projection
ij
BAijABji
jjiijAAAiiA
ABABBA
ii *'' ')(
tr
The reduced density matrix is defined as:
Universe
system
|ienvironment
| jRegion A Region B
![Page 37: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/37.jpg)
Properties of the density matrix
ABABBA tr
• Hermitian -> eigenvalues are real• Eigenvalues are non-negative• The trace equals to unity-> Tr rA=1• Eigenvectors form an orthonormal basis.
1 and 0 with ;
AAA
![Page 38: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/38.jpg)
The singular value decomposition (SVD)Consider a matrix
ij= dimA
dimB
We can decompose it into the product of three matrices U,D,V: =UDV†
• U is a (dimAxdimB) matrix with orthonormal columns-> UU†=1; U=U†
• D is a (dimBxdimB) diagonal matrix with non-negative elements la• V is a (dimBxdimB) unitary matrix -> VV†=1
= xx U D V
(we are choosing dimB < dimA for convenience)
![Page 39: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/39.jpg)
B
BAAB
B
BA
B
jBj
iAi
ij
B
BAjiAB
B
ji
B
jt
iij
jViU
jiVU
VUVU
dim
dimdim*
dim*
dim*
dim
l!orthonorma are , bases theHereBA
This is also called the “Schmidt decomposition” of the state
![Page 40: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/40.jpg)
The SVD and the density matrix
)dim,min(dim with BA
r
BAABr
In general:
In the Schmidt basis, the reduced density matrix is
r
BBB
r
AAABABBA
2
2
and
tr
• The singular values are the eigenvalues of the reduced d.m. squared wi=li 2
• The two reduced density matrices share the spectrum• the singular vectors are the eigenvectors of the density matrix.
![Page 41: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/41.jpg)
Optimizing the wave-function
S=||' -||2We want to minimize the distance between the two states
where | is the actual ground state, and |’ is the variational approximation after rotating to a new basis and truncating
|' = ∑mjaj|| j
We reformulate the question as: Given a matrix ,what is the optimal matrix ’ with
fixed rank r that minimizes the Frobenius distance between the two matrices. It turn out, this is a well known problem, called the “low rank matrix approximation”. If we order the eigenvalues of the density matrix in descending order w1, w2,…,wm,…, wr we obtain
S=||' -||2 =
r
mi
1
Truncation error!
![Page 42: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/42.jpg)
DMRG: The AlgorithmHow do we build the reduced basis of states?
We grow our basis systematically, adding sites to our system at each step, and using the density matrix projection to truncate
![Page 43: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/43.jpg)
2) We diagonalize the system and obtain the ground state |gs=∑y
1234|a
1|s
2|s
3|b
4
3) We calculate the reduced density matrix r for blocks 1-2 and 3-4.
4) We diagonalize r obtaining the eigenvectors and eigenvalues wi
1) We start from a small superblock with 4 sites/blocks, each with a dimension m
i , small enough to be easily diagonalized
1 2 3 4m1
H1
The Algorithm
43
*34'2'1123421121 ''
s
ss
![Page 44: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/44.jpg)
5) We add a new site to blocks 1 and 4, expanding the basis for each block to m'
1 = m
m
2 and m'
4 = m
3 m
m2
m m'1=m
m
2
7) We repeat starting from 2) replacing H1 by H'
1 and H
4 by H'
4
1 2 3 41 2 3 41 2 3 41 2 3 4
6) We rotate the Hamiltonian and operators to the new basis keeping the m states with larger eigenvalues (notice that we no longer are in the occupation number representation)
![Page 45: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/45.jpg)
We add one site at a time, until we reach the desired system size
1 2 3 41 2 3 41 2 3 41 2 3 4
The finite size algorithm
![Page 46: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/46.jpg)
1 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 41 2 3 4
We sweep from left to rightWe sweep from right to left
The finite size algorithm
…Until we converge
![Page 47: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/47.jpg)
Finite-size DMRG Flow chart
![Page 48: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/48.jpg)
The discarded weight 1- ∑m=1a w a measures the
accuracy of the truncation to m states
The discarded weight 1- ∑m=1a w a measures the
accuracy of the truncation to m states
![Page 49: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/49.jpg)
Observations
• Sweeping is essential to achieve convergence• Run the finite-size DMRG and extrapolate to
the thermodynamic limit.• For each system size, extrapolate the results
with the number of states m, or fix the truncation error below certain tolerance.
![Page 50: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/50.jpg)
Density Matrix Renormalization Group
A variational method without a-priori assumptions about the physics.
•Similar capabilities as exact diagonalization.
•Can calculate properties of very large systems (1D and quasi-2D) with unprecedented accuracy.
•Results are variational, but “quasi-exact”: Accuracy is finite, but under control.
![Page 51: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/51.jpg)
Advantages of the DMRG
• DMRG is very versatile, and easy to adapt to complex geometries and Hamiltonians.
• Can be used to study models of spins, bosons, or fermions.
• General and reusable code: A single program can be used to run arbitrary models without changing a single line (e.g. ALPS DMRG)
• Symmetries are easy to implement.
![Page 52: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/52.jpg)
Limitations of the DMRG
• DMRG is the method of choice in 1d and quasi-1d systems, but it is less efficient in higher dimensions.
• Problems with (i) critical systems, (ii) long range interactions, and (iii) periodic boundary conditions.
• These limitations are due to:– The structure of the variational wave function used by the
DMRG (the MPS ansatz). – Entanglement entropy follows area law.
![Page 53: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/53.jpg)
Technicalities…Adding a single site to the block
1ll
1ls
Before truncating we build the new basis as:
11 lll s
And the Hamiltonian for the new block as
...'0,01,1, OOHIIHH lLllLlL
01, OIO llL with
![Page 54: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/54.jpg)
.. and for the right block
Before truncating we build the new basis as:
433 lll s
And the Hamiltonian for the new block as
...' 4,0)4(04,13, lRlLlRlR OOIHHIH
3l4l
3ls
)1(0, lLlR IOOwith
![Page 55: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/55.jpg)
Putting everything together to build the Hamiltonian…
...' 3,1,
3,1,3,1,
lRlL
lRlLlRlL
OO
HIIHH
1l 3l
![Page 56: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/56.jpg)
TruncationWhen we add a site to the left block we represent the new basis states as:
ll
lll
ll slls
lL
sllllll sUss
,
1,1
,1111
1
11
1
1ll
1ls
Similarly for the right block:
43
343
43 ,43,
3
,433433
ll
lll
ll slls
lR
sllllll sUss
3l4l
3ls
![Page 57: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/57.jpg)
Measuring observablesSuppose we have a chain and we want to measure a correlation between sites i and j
iOjO'
i j
We have two options:1. Measure the correlation by storing the composite operator in a block2. Measure when the two operators are on separate blocks
We shall go for option (2) for the moment: simpler and more efficient
![Page 58: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/58.jpg)
Operators on separate blocks
iOjO'ˆ
i j
We only measure when we have the following situation:
Then, it is easy to see that:
ji
ji
jiji
OO
OO
OOOO
'''
'''
''
'',
*''
'',
*''
We cannot do this if the two operators are in the same block!!!
![Page 59: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/59.jpg)
Operators on the same blockiO
jO'ˆ
i j
We need to propagate the product operator into the block, the same way as we do for the Hamiltonian
Do never do this:
ji
jiji
OO
OOOO
'''
''
',
*'
![Page 60: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/60.jpg)
Targeting states in DMRG
If we target the ground state only, we cannot expect to have a good representation of excited states (dynamics).
If the error is strictly controlled by the DMRG truncation error, we say that the algorithm is “quasiexact”.
Non quasiexact algorithms seem to be the source of almost all DMRG “mistakes”. For instance, the infinite system algorithm applied to finite systems is not quasiexact.
Our DMRG basis is only guaranteed to represent targeted states, and those only after enough sweeps!
![Page 61: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/61.jpg)
Excited states
gsgsHHH '
b) At each step of the DMRG sweep, target the ground state, and the ground state of the modified Hamiltonian:
For targeting the two states, we use the density matrix:
112
1
2
1 gsgs
a) If we use quantum numbers, we can calculate the ground states in different sectors, for instance S=0, and S=1, to obtain the spin gap
![Page 62: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/62.jpg)
2D Generalization
![Page 63: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/63.jpg)
Why does the DMRG work???wa
a
good!
bad!
In other words: what makes the density matrix eigenvalues behave so nicely?
![Page 64: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/64.jpg)
EntanglementWe say that a two quantum systems A and B are “entangled” when we cannot describe the wave function as a product state of a wave function for system A, and a wave function for a system B
For instance, let us assume we have two spins, and write a state such as:
|y =|↑↓ + |↓↑ + |↑↑ + |↓↓We can readily see that this is equivalent to:
|y =(|↑+|↓)(|↑+|↓)=|↑x |↓x
-> The two spins are not entangled! The two subsystems carry information independently
Instead, this state: |y =|↑↓ + |↓↑
is “maximally entangled”. The state of subsystem A has ALL the information about the state of subsystem B
![Page 65: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/65.jpg)
The Schmidt decomposition
Universe
system
|ienvironment
| j
ij
BAijABji
We assume the basis for the left subsystem has dimension dimA, and the right, dimB. That means that we have dimA x dimB coefficients. We go back to the original DMRG premise: Can we simplify this state by changing to a new basis? (what do we mean with “simplifying”, anyway?)
![Page 66: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/66.jpg)
The Schmidt decompositionWe have seen that through a SVD decomposition, we can rewire the state as:
r
BAAB
Where
lorthonorma are ; and 0 );dim,min(dimBABAr
Notice that if the Schmidt rank r=1, then the wave-function reduces to a product state, and we have “disentangled” the two subsystems.
After the Schmidt decomposition, the reduced density matrices for the two subsystems read:
r
BABABA
//
2/
![Page 67: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/67.jpg)
The Schmidt decomposition, entanglement and DMRG
It is clear that the efficiency of DMRG will be determined by the spectrum of the density matrices (the “entanglement spectrum”), which are related to the Schmidt coefficients:• If the coefficients decay very fast (exponentially, for
instance), then we introduce very little error by discarding the smaller ones.
• Few coefficients mean less entanglement. In the extreme case of a single non-zero coefficient, the wave function is a product state and it completely disentangled.
• NRG minimizes the energy…DMRG concentrates entanglement in a few states. The trick is to disentangle the quantum many body state!
![Page 68: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/68.jpg)
Quantifying entanglementIn general, we write the state of a bipartite system as:
ij
ij ji
We saw previously that we can pick and orthonormal basis for “left” and “right” systems such that
RL
We define the “von Neumann entanglement entropy” as:
22 log
S
Or, in terms of the reduced density matrix:
LLLLL S
logTr2
![Page 69: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/69.jpg)
Entanglement entropyLet us go back to the state:
|y =|↑↓ + |↓↑
2/10
02/1L
We obtain the reduced density matrix for the first spin, by tracing over the second spin (and after normalizing):
We say that the state is “maximally entangled” when the reduced density matrix is proportional to the identity.
2log2
1log
2
1
2
1log
2
1S
![Page 70: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/70.jpg)
Entanglement entropy
• If the state is a product state:
0,...0,0,1 SwRL • If the state maximally entangled, all the wa are equal
DSDDDw log,...1,1,1
where D is
RL HHD dim,dimmin
![Page 71: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/71.jpg)
Area law: Intuitive pictureConsider a valence bond solid in 2D
singlet
2logcut) bonds of(#2log LS
The entanglement entropy is proportional to the area of the boundary separating both regions. This is the prototypical behavior in gapped systems. Notice that this implies that the entropy in 1D is independent of the size of the partition
![Page 72: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/72.jpg)
Critical systems in 1Dc is the “central charge” of the system, a measure of the number of gapless modes
![Page 73: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/73.jpg)
Entropy and DMRGThe number of states that we need to keep is related to the entanglement entropy:
Sm exp
• Gapped system in 1D: m=const.• Critical system in 1D: m=La
• Gapped system in 2D: m=exp(L)• In 2D in general, most systems obey the area law (not free fermions, or
fermionic systems with a 1D Fermi surface, for instance)… • Periodic boundary conditions in 1D: twice the area -> m2
![Page 74: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/74.jpg)
What have we left out?
…Exploiting quantum symmetries…wave function prediction
![Page 75: The density matrix renormalization group Adrian Feiguin.](https://reader035.fdocuments.net/reader035/viewer/2022062423/56649ebb5503460f94bc2e47/html5/thumbnails/75.jpg)
The wave-function transformationBefore the transformation, the superblock state is written as:
321 ,,,
321321
llll ssllllllll ssss
l
1ls
After the transformation, we add a site to the left block, and we “spit out” one from the right block
l
ll
1
3l
2ls
4321 ,,,
43214321
llll ssllllllll ssss
After some algebra, and assuming , one readily obtains:
31 ,,
343321114321
lll sllllllllllllll ssssss