1. Matrix product states for the absolute beginner
Garnet Kin-Lic Chan
Princeton University
----- Meeting Notes (12/12/14 12:49) ----- b

2. Brief overview: Why tensor networks?
Graphical notation
Matrix Product States and Matrix Product Operators
Compressing Matrix Product States
Energy optimization
Time evolution
Periodic and infinite MPS
Focus on basic computations and algorithms with MPS
not covered: entanglement area laws, RG, topological aspects, symmetries etc.

3. Quantum mechanics is complex
Dirac
The fundamental laws necessary for the mathematical treatment of a large part of physics and the whole of chemistry are thus completely known …
the difficulty lies only in the fact that application of these laws leads to equations that are too complex to be solved.
So: ideas in quantum chemistry. Well let’s start at the very beginning, which is to recognize that quantum mechanics is the foundation of chemistry, which was realized by
Paul Dirac, when he wrote down his equation and concluded that the fundamental laws for all of chemistry were known.
The only problem, as recognized by Dirac also, is that the direct solution of quantum mechanics appears too complicated. Why is this?
Well, the fundamental object in quantum mechanics is the wavefunction, and this is a joint or simultaneous probability amplitude for all the electrons.
n electron positions
L spins
L particle occupancies
Exponential complexity to represent wavefunction

4. This view of QM is depressing
[The Schrodinger equation] cannot be solved accurately when
the number of particles exceeds about 10. No computer existing,
or that will ever exist, can break this barrier because it is a
catastrophe of dimension ...
Pines and Laughlin (2000)
in general the many-electron wave function Ψ … for a system of N electrons is not a legitimate scientific concept [for large N]
Kohn (Nobel lecture, 1998)

5. illusion of complexity
nature does not explore all possibilities
Nature is local: ground-states have low entanglement

6. Language for low entanglement states is tensor networks
different tensor networks reflect geometry of entanglement
Matrix Product State
1D entanglement for gapped systems
(basis of DMRG - often used in quasi-2D/3D)
MERA
1D/nD entanglement for gapless systems
Tensor Product State (PEPS)
nD entanglement for gapped systems

7. Graphical language n1 n2 n3 Algebraic form Graphical form spin 1/2
e.g.
particles
Algebraic form
Graphical form n1 n2 n3
physical
index
General
state
thick line
= single tensor

8. Graphical language, cont’d
Algebraic form
Graphical form n1 n2 n3
General
operator
n1’
n2’
n3’

9. Ex: overlap, expectation value

10. Low entanglement states
What does it mean for a state to have low entanglement?
Consider system with two parts, 1 and 2
No entanglement
local measurements on separated system 1, system 2
can be done independently. Local realism (classical)
Entangled state
Low entanglement : small number of terms in the sum

11. “bond” or “auxiliary” dimension
Matrix product states
first and last tensors have one fewer auxiliary index
1D structure
of entanglement
“bond” or “auxiliary” dimension
“M” or “D” or “χ”
amplitude is obtained as a product of matrices
General
state i2 n1 i1 n2 n3
MPS n1 n2 n3 =

12. MPS gauge
MPS are not unique: defined up to gauge on the auxiliary indices = i j
insert gauge
matrices = =

13. Ex: MPS contraction
Overlap
“d”
“M”
Efficient computation: contract in the correct order! 1 2 3
“d”
“M”
MPS overlap: total cost

14. = MPS from general state = =
recall singular value decomposition (SVD) of matrix
singular values
orthogonality conditions i = = j i =
singular values j
orthogonality conditions

15. MPS from general state, cont’di
Step 1
“n”
“m” = j
SVD = i 1 = j
Step 2 =
“m”
“n”
SVD = 2 3 = = 2 3 1

16. Common canonical forms
“Vidal” form 2 3 = 1
different canonical form: absorb singular values into the tensors
all tensors contract to unit matrix from left
left canonical 2 3 1 1 i j i = = 2
etc j 2
right canonical
all tensors contract to unit matrix from right 2 3 i i 3 = 2 =
etc 1 j j 3 2
mixed canonical
around site 2
(DMRG form) 2 3 1 i j i 3 = 1 j 3

17. Matrix product operators
each tensor has a bra and ket physical index n1 n2 n3 i2 n1 i1 n2 n3
n1’
n2’
n3’
General
operator =
MPO
n1’
n2’
n3’

18. Typical MPO’s What is bond dimension as an MPO? L R
joins pairs of operators on both sides
MPO bond dimension = 5

19. = = MPO acting on MPS M2 MPO M1 MPS MPS M1 x M2
MPO on MPS leads to new MPS with product of bond dimensions

20. MPS compression: SVD
many operations (e.g. MPOxMPS, MPS+MPS) increase bond dim.
compression: best approximate MPS with smaller bond dimension.
write MPS in Vidal gauge via SVD’s 2 3 1 1 2 3
truncate
bonds with
small singular
values
M1 singular values
Each site is compressed
independently of new
information of other sites:
“Local” update: non-optimal. 2 3 1
truncated M2 singular values

21. MPS: variational compression
solve minimization problem
original (fixed) MPS
new MPS 1 2 3 1 3 2

22. = = Gradient algorithm x 2
To minimize quantity, follow its gradient until it vanishes 1 2 3 1 3 = 2 2
linear in 1 3 2 1 3 2 = 2
x 2 2
quadratic in
Gradient step 2

23. Sweep algorithm (DMRG style)1 2 3 1 3 2
bilinear in 1 2 3 2
consider
as vector 1 2 3 1 3
where 1 3 2 1 3
where

24. Sweep algorithm cont’d
Minimization performed site by site by solving 3 2 2 3 M b 1 2 3 M 1 3 b 1 3 1 2 3
use updated tensors from previous step M 1 2 b 1 2 1 2 3

25. Sweep and mixed canonical form
in mixed canonical form M 3 2 2 3 b 1 2 3
mixed canonical around site 1
change canonical form M 1 3 b 1 3 1 2 3
mixed canonical around site 2
note: updated singular values

26. SVD vs. variational compression
variational algorithms – optimization for each site depends on
all other sites. Uses “full environment”
SVD compression: “local update”. Not as robust, but cheap!
MPS: full environment / local update same computational
scaling, only differ by number of iterations.
General tensor networks (e.g. PEPS): full environment may be expensive to compute or need further approximations.

27. Energy optimization bilinear form: similar to compression problem
commonly used algorithms
DMRG: variational sweep with full environment
imag. TEBD: local update, imag. time evolution + SVD compression

28. DMRG energy minimization2
use mixed canonical form around site 2 1 3 2 1 3 2
unit matrix
where 1 3
eigenvalue problem for
each site, in mixed canonical form
DMRG “superblock”
Hamiltonian

29. Time evolution real time evolution
imaginary time evolution: replace i by 1.
projects onto ground-state at long times.

30. General time-evolution
compress 2 3 1 2 3 1
repeat

31. Short range H: Trotter form
evolution on pairs of bonds

32. Even-odd evolution
time evolution can be broken up into even and odd bonds

33. Time-evolving block decimation
even/odd evolution easy to combine with SVD compression: TEBD
increase of bond dimension
of unconnected bonds:
SVD compression can be done
independently on each bond.
SVD

34. Periodic and infinite MPS
MPS easily extended to PBC and thermodynamic limit
Finite MPS (OBC) 2 3 1 2 3 1
Periodic MPS 2 3 1
Infinite MPS

35. Unit cell = 2 site infinite MPS
Infinite TEBD
local algorithms such as TEBD easy to extend to infinite MPS
Unit cell = 2 site infinite MPS A B A B A B

36. i-TEBD cont’d A B even bond evolution + compression updated
not updated
Step 1
Step 2
odd bond
Repeat

37. Symmetries
Given global symmetry group, local site basis can be labelled
by irreps of group – quantum numbers
U(1) – site basis labelled by integer n (particle number)
n=0, 1, 2 etc...
SU(2) symmetry – site basis labelled by j, m (spin quanta)
Total state associated with good quantum numbers

38. MPS and symmetry
bond indices can be labelled by same symmetry labels as physical sites
e.g. particle number symmetry
labelled by integer
MPS: well defined Abelian symmetry, each tensor fulfils rule
tensor with no arrows
leaving gives total
state quantum number
Choice of convention:

39. Brief overview: Why tensor networks?
Graphical notation
Matrix Product States and Matrix Product Operators
Compressing Matrix Product States
Energy optimization
Time evolution
Periodic and infinite MPS
Focus on basic computations and algorithms with MPS
not covered: entanglement area laws, RG, topological aspects, symmetries etc.

40. Language for low entanglement states is tensor networks
different tensor networks reflect geometry of entanglement
Matrix Product State
1D entanglement for gapped systems
(basis of DMRG - often used in quasi-2D/3D)
MERA
1D/nD entanglement for gapless systems
Tensor Product State (PEPS)
nD entanglement for gapped systems

41. Questions What is the dimension of MPS (M1) + MPS (M2)?
How would we graphically represent the DM of an MPS, (tracing out sites n3 to nL?)
What is the dimension of the MPO of an electronic Hamiltonian with general quartic interactions?
What happens when we use an MPS to represent a 2D system?
What happens to the bond-dimension of an MPS as we evolve it in time? Do we expect the MPS to be compressible? How about for imaginary time evolution?
How would we alter the discussion of symmetry for non-Abelian symmetry e.g. SU(2)?