approximation techniques for special eigenvalue problems ... fileapproximation techniques for...

74
Technische Universit¨ at M ¨ unchen Approximation Techniques for Special Eigenvalue Problems Konrad Waldherr June 2011 Joint work with Thomas Huckle K. Waldherr: Approximation Techniques for Special Eigenvalue Problems HDA 2011, Bonn, June 2011 1

Upload: others

Post on 16-Oct-2019

26 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Approximation Techniques for SpecialEigenvalue Problems

Konrad Waldherr

June 2011

Joint work with Thomas Huckle

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 1

Page 2: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Outline1 Problem Setting: Computation of Ground States

Physical ModelMathematical Model

2 Data-Sparse Representation Formats

3 Matrix Product StatesFormalismComputations with MPSUniqueness of MPSMinimization in terms of MPS

4 Subspace Methods in Terms of MPS

5 Numerical results

6 Conclusions

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 2

Page 3: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem: Computation of Ground StatesGiven:

• Physical system: 1D spin chain with N particles

1 2 3 4 5

• Interaction within the system (e.g. nearest-neighbor interaction)

1 2 3 4 5

• External interaction (e.g. exterior magnetic field)

1 2 3 4 5

Goal:• Find ground state, i.e. the state related to the smallest energy of

the system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 3

Page 4: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem: Computation of Ground StatesGiven:

• Physical system: 1D spin chain with N particles

1 2 3 4 5

• Interaction within the system (e.g. nearest-neighbor interaction)

1 2 3 4 5

• External interaction (e.g. exterior magnetic field)

1 2 3 4 5

Goal:• Find ground state, i.e. the state related to the smallest energy of

the system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 3

Page 5: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem: Computation of Ground StatesGiven:

• Physical system: 1D spin chain with N particles

1 2 3 4 5

• Interaction within the system (e.g. nearest-neighbor interaction)

1 2 3 4 5

• External interaction (e.g. exterior magnetic field)

1 2 3 4 5

Goal:• Find ground state, i.e. the state related to the smallest energy of

the system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 3

Page 6: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem: Computation of Ground StatesGiven:

• Physical system: 1D spin chain with N particles

1 2 3 4 5

• Interaction within the system (e.g. nearest-neighbor interaction)

1 2 3 4 5

• External interaction (e.g. exterior magnetic field)

1 2 3 4 5

Goal:• Find ground state, i.e. the state related to the smallest energy of

the system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 3

Page 7: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Mathematical Model• Any state of the system is represented by a vector x ∈ C2N

.• The physical system can be described by the HamiltonianH ∈ C2N×2N

.

• The Hamiltonian may be formulated as a weighted sum ofKronecker products of Pauli and identity matrices.

• Pauli matrices:

σx =

(0 11 0

), σy =

(0 −ii 0

), σz =

(1 00 −1

).

• Kronecker product:

A⊗B :=

a11 . . . a1n

.... . .

...am1 . . . amn

⊗B =

a11B . . . a1nB...

. . ....

am1B . . . amnB

• Then, the ground state corresponds to the eigenvector related

to the smallest eigenvalue.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 4

Page 8: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Mathematical Model• Any state of the system is represented by a vector x ∈ C2N

.• The physical system can be described by the HamiltonianH ∈ C2N×2N

.• The Hamiltonian may be formulated as a weighted sum of

Kronecker products of Pauli and identity matrices.• Pauli matrices:

σx =

(0 11 0

), σy =

(0 −ii 0

), σz =

(1 00 −1

).

• Kronecker product:

A⊗B :=

a11 . . . a1n

.... . .

...am1 . . . amn

⊗B =

a11B . . . a1nB...

. . ....

am1B . . . amnB

• Then, the ground state corresponds to the eigenvector relatedto the smallest eigenvalue.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 4

Page 9: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Mathematical Model• Any state of the system is represented by a vector x ∈ C2N

.• The physical system can be described by the HamiltonianH ∈ C2N×2N

.• The Hamiltonian may be formulated as a weighted sum of

Kronecker products of Pauli and identity matrices.• Pauli matrices:

σx =

(0 11 0

), σy =

(0 −ii 0

), σz =

(1 00 −1

).

• Kronecker product:

A⊗B :=

a11 . . . a1n

.... . .

...am1 . . . amn

⊗B =

a11B . . . a1nB...

. . ....

am1B . . . amnB

• Then, the ground state corresponds to the eigenvector related

to the smallest eigenvalue.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 4

Page 10: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I

+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 11: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I

+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 12: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I

+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 13: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz

+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 14: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I

+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 15: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I

+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 16: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I

+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 17: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I

+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 18: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Example: Ising-type Hamiltonian

1 2 3 4 5

H = σz ⊗ σz ⊗ I ⊗ I ⊗ I+ I ⊗ σz ⊗ σz ⊗ I ⊗ I+ I ⊗ I ⊗ σz ⊗ σz ⊗ I+ I ⊗ I ⊗ I ⊗ σz ⊗ σz+ σx ⊗ I ⊗ I ⊗ I ⊗ I+ I ⊗ σx ⊗ I ⊗ I ⊗ I+ I ⊗ I ⊗ σx ⊗ I ⊗ I+ I ⊗ I ⊗ I ⊗ σx ⊗ I+ I ⊗ I ⊗ I ⊗ I ⊗ σx

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 5

Page 19: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem Setting• General formulation of the Hamiltonian H:

H =

M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N , Q(k)i ∈ {id, σx, σy, σz} ⊂ C2×2

• Variational ansatz: Minimization of the Rayleigh quotient:

minx 6=0

xHHx

xHx

• Problems:

• the vector space V = C2N

grows exponentially in N ,• the problem is not generic.

• Idea: Find data-sparse representation formats that allow toovercome the curse of dimensionality.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 6

Page 20: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem Setting• General formulation of the Hamiltonian H:

H =

M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N , Q(k)i ∈ {id, σx, σy, σz} ⊂ C2×2

• Variational ansatz: Minimization of the Rayleigh quotient:

minx 6=0

xHHx

xHx

• Problems:

• the vector space V = C2N

grows exponentially in N ,• the problem is not generic.

• Idea: Find data-sparse representation formats that allow toovercome the curse of dimensionality.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 6

Page 21: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem Setting• General formulation of the Hamiltonian H:

H =

M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N , Q(k)i ∈ {id, σx, σy, σz} ⊂ C2×2

• Variational ansatz: Minimization of the Rayleigh quotient:

minx 6=0

xHHx

xHx

• Problems:• the vector space V = C2N

grows exponentially in N ,• the problem is not generic.

• Idea: Find data-sparse representation formats that allow toovercome the curse of dimensionality.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 6

Page 22: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Problem Setting• General formulation of the Hamiltonian H:

H =

M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N , Q(k)i ∈ {id, σx, σy, σz} ⊂ C2×2

• Variational ansatz: Minimization of the Rayleigh quotient:

minx 6=0

xHHx

xHx

• Problems:• the vector space V = C2N

grows exponentially in N ,• the problem is not generic.

• Idea: Find data-sparse representation formats that allow toovercome the curse of dimensionality.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 6

Page 23: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Data-sparse representation formats• The data-sparse format has to allow for

1. proper reproduction of the physical setting (e.g.nearest-neighbor interaction),

2. representations that are only polynomial in N ,3. good approximation properties,4. efficient evaluation of the Rayleigh quotient

minx∈U

xHHx

xHx(1)

=⇒ efficient computation of both Hx and yHx.

• Problem setting:• Not: Find a data-sparse format of a given data structure• But: Use the data-sparse format as an ansatz for the

minimization of the Rayleigh quotient (1).

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 7

Page 24: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Data-sparse representation formats• The data-sparse format has to allow for

1. proper reproduction of the physical setting (e.g.nearest-neighbor interaction),

2. representations that are only polynomial in N ,3. good approximation properties,4. efficient evaluation of the Rayleigh quotient

minx∈U

xHHx

xHx(1)

=⇒ efficient computation of both Hx and yHx.• Problem setting:

• Not: Find a data-sparse format of a given data structure• But: Use the data-sparse format as an ansatz for the

minimization of the Rayleigh quotient (1).

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 7

Page 25: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Tensor Decompositions• Any state x ∈ C2N

may be considered as a N th order tensor:

x = (xi1i2...iN )ij=0,1

i1 i2 iN

• Goal: Find convenient tensor decomposition as ansatz for (1)• Classical decomposition formats:

• Tucker decomposition,• Canonical decomposition.

• Physically motivated formats:• Matrix Product States (MPS) for 1D problems,• Projected Entangled Pair States (PEPS) for 2D problems.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 8

Page 26: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Tensor Decompositions• Any state x ∈ C2N

may be considered as a N th order tensor:

x = (xi1i2...iN )ij=0,1

i1 i2 iN

• Goal: Find convenient tensor decomposition as ansatz for (1)

• Classical decomposition formats:• Tucker decomposition,• Canonical decomposition.

• Physically motivated formats:• Matrix Product States (MPS) for 1D problems,• Projected Entangled Pair States (PEPS) for 2D problems.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 8

Page 27: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Tensor Decompositions• Any state x ∈ C2N

may be considered as a N th order tensor:

x = (xi1i2...iN )ij=0,1

i1 i2 iN

• Goal: Find convenient tensor decomposition as ansatz for (1)• Classical decomposition formats:

• Tucker decomposition,• Canonical decomposition.

• Physically motivated formats:• Matrix Product States (MPS) for 1D problems,• Projected Entangled Pair States (PEPS) for 2D problems.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 8

Page 28: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Tensor Decompositions• Any state x ∈ C2N

may be considered as a N th order tensor:

x = (xi1i2...iN )ij=0,1

i1 i2 iN

• Goal: Find convenient tensor decomposition as ansatz for (1)• Classical decomposition formats:

• Tucker decomposition,• Canonical decomposition.

• Physically motivated formats:• Matrix Product States (MPS) for 1D problems,• Projected Entangled Pair States (PEPS) for 2D problems.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 8

Page 29: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Decompositions of a tensor• Tucker decomposition:

Not advantageous in our case of a binary tensor.

• Canonical decomposition (CP):

Allows generalizations in several ways: (see [Huckle 2011])• blockings and mixed blockings,• permutations.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 9

Page 30: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Decompositions of a tensor• Tucker decomposition:

Not advantageous in our case of a binary tensor.• Canonical decomposition (CP):

Allows generalizations in several ways: (see [Huckle 2011])• blockings and mixed blockings,• permutations.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 9

Page 31: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product States (MPS)• Consider again the linear 1D spin chain.

b b b b b

site j

• Each site j is related to a matrix pair A(0)j , A

(1)j ∈ CDj×Dj+1

b b b b b

A(ij)j

• For i = (i1, i2, . . . , iN )2, the vector component xi is given by

xi = xi1i2...iN = tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)

b b b b b

ij−1 ij ij+1 ij+2 ij+3

mj−1 mj mj+1 mj+2

(1.X)

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 10

Page 32: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product States (MPS)• Consider again the linear 1D spin chain.

b b b b b

site j

• Each site j is related to a matrix pair A(0)j , A

(1)j ∈ CDj×Dj+1

b b b b b

A(ij)j

• For i = (i1, i2, . . . , iN )2, the vector component xi is given by

xi = xi1i2...iN = tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)

b b b b b

ij−1 ij ij+1 ij+2 ij+3

mj−1 mj mj+1 mj+2

(1.X)

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 10

Page 33: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product States (MPS)• Consider again the linear 1D spin chain.

b b b b b

site j

• Each site j is related to a matrix pair A(0)j , A

(1)j ∈ CDj×Dj+1

b b b b b

A(ij)j

• For i = (i1, i2, . . . , iN )2, the vector component xi is given by

xi = xi1i2...iN = tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)

b b b b b

ij−1 ij ij+1 ij+2 ij+3

mj−1 mj mj+1 mj+2

(1.X)

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 10

Page 34: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product States• Thus, the vector x may be written as

x =

2N∑i=1

xiei =∑

i1,...,ip

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)︸ ︷︷ ︸

=xi=xi1,...,iN

ei1i2...iN︸ ︷︷ ︸=ei

.

• Binary version of the Tensor Train concept. (see [Oseledets andTyrtyshnikov, 2009])

• MPS representation requires 2ND2 instead of 2N entries. (2.X)• Any state can be represented exactly by an MPS (but at the cost

of exponentially growing matrix dimension D),• For approximation problems, D has to scale only polynomially.

(3.X)

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 11

Page 35: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product States• Thus, the vector x may be written as

x =

2N∑i=1

xiei =∑

i1,...,ip

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)︸ ︷︷ ︸

=xi=xi1,...,iN

ei1i2...iN︸ ︷︷ ︸=ei

.

• Binary version of the Tensor Train concept. (see [Oseledets andTyrtyshnikov, 2009])

• MPS representation requires 2ND2 instead of 2N entries. (2.X)

• Any state can be represented exactly by an MPS (but at the costof exponentially growing matrix dimension D),

• For approximation problems, D has to scale only polynomially.(3.X)

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 11

Page 36: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product States• Thus, the vector x may be written as

x =

2N∑i=1

xiei =∑

i1,...,ip

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)︸ ︷︷ ︸

=xi=xi1,...,iN

ei1i2...iN︸ ︷︷ ︸=ei

.

• Binary version of the Tensor Train concept. (see [Oseledets andTyrtyshnikov, 2009])

• MPS representation requires 2ND2 instead of 2N entries. (2.X)• Any state can be represented exactly by an MPS (but at the cost

of exponentially growing matrix dimension D),• For approximation problems, D has to scale only polynomially.

(3.X)

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 11

Page 37: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product StatesExample

• Consider the case N = 2 and D = 1:

ψ =

1∑i1,i2=0

(a

(i1)1 a

(i2)2

)︸ ︷︷ ︸

=xi=xi1i2

(ei1 ⊗ ei2)︸ ︷︷ ︸=ei=ei1i2

=

a

(0)1 a

(0)2

a(0)1 a

(1)2

a(1)1 a

(0)2

a(1)1 a

(1)2

!=

x1

x2

x3

x4

• For exact representation:

x1x4 = x2x3

• e1, e2, e3, e4 can be represented,• However, e1 + e4 = (1, 0, 0, 1)T cannot be represented.

• MPS are not closed under addition!

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 12

Page 38: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product StatesExample

• Consider the case N = 2 and D = 1:

ψ =

1∑i1,i2=0

(a

(i1)1 a

(i2)2

)︸ ︷︷ ︸

=xi=xi1i2

(ei1 ⊗ ei2)︸ ︷︷ ︸=ei=ei1i2

=

a

(0)1 a

(0)2

a(0)1 a

(1)2

a(1)1 a

(0)2

a(1)1 a

(1)2

!=

x1

x2

x3

x4

• For exact representation:

x1x4 = x2x3

• e1, e2, e3, e4 can be represented,• However, e1 + e4 = (1, 0, 0, 1)T cannot be represented.

• MPS are not closed under addition!

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 12

Page 39: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product StatesExample

• Consider the case N = 2 and D = 1:

ψ =

1∑i1,i2=0

(a

(i1)1 a

(i2)2

)︸ ︷︷ ︸

=xi=xi1i2

(ei1 ⊗ ei2)︸ ︷︷ ︸=ei=ei1i2

=

a

(0)1 a

(0)2

a(0)1 a

(1)2

a(1)1 a

(0)2

a(1)1 a

(1)2

!=

x1

x2

x3

x4

• For exact representation:

x1x4 = x2x3

• e1, e2, e3, e4 can be represented,• However, e1 + e4 = (1, 0, 0, 1)T cannot be represented.

• MPS are not closed under addition!

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 12

Page 40: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Matrix Product StatesExample

• Consider the case N = 2 and D = 1:

ψ =

1∑i1,i2=0

(a

(i1)1 a

(i2)2

)︸ ︷︷ ︸

=xi=xi1i2

(ei1 ⊗ ei2)︸ ︷︷ ︸=ei=ei1i2

=

a

(0)1 a

(0)2

a(0)1 a

(1)2

a(1)1 a

(0)2

a(1)1 a

(1)2

!=

x1

x2

x3

x4

• For exact representation:

x1x4 = x2x3

• e1, e2, e3, e4 can be represented,• However, e1 + e4 = (1, 0, 0, 1)T cannot be represented.

• MPS are not closed under addition!

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 12

Page 41: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Computations with MPS• Sum of two MPS vectors

x =∑

i1,...,iN

tr(A(i1)1 A

(i2)2 . . . A

(iN )N )ei1...iN

y =∑

i1,...,iN

tr(B(i1)1 B

(i2)2 . . . B

(iN )N )ei1...iN

x+ y =∑i1,...,iN

tr[(A

(i1)1 0

0 B(i1)1

)(A

(i2)2 0

0 B(i2)2

)︸ ︷︷ ︸

D′×D′

. . .

(A

(iN )N 0

0 B(iN )N

)]ei1...iN

• Summing MPS vectors leads to an augmentation of the matrixsizes D → D′.

• Use compression techniques to keep the size of the MPSmatrices constant.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 13

Page 42: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Computations with MPS• Sum of two MPS vectors

x =∑

i1,...,iN

tr(A(i1)1 A

(i2)2 . . . A

(iN )N )ei1...iN

y =∑

i1,...,iN

tr(B(i1)1 B

(i2)2 . . . B

(iN )N )ei1...iN

x+ y =∑i1,...,iN

tr[(A

(i1)1 0

0 B(i1)1

)(A

(i2)2 0

0 B(i2)2

)︸ ︷︷ ︸

D′×D′

. . .

(A

(iN )N 0

0 B(iN )N

)]ei1...iN

• Summing MPS vectors leads to an augmentation of the matrixsizes D → D′.

• Use compression techniques to keep the size of the MPSmatrices constant.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 13

Page 43: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Computations with MPS• Sum of two MPS vectors

x =∑

i1,...,iN

tr(A(i1)1 A

(i2)2 . . . A

(iN )N )ei1...iN

y =∑

i1,...,iN

tr(B(i1)1 B

(i2)2 . . . B

(iN )N )ei1...iN

x+ y =∑i1,...,iN

tr[(A

(i1)1 0

0 B(i1)1

)(A

(i2)2 0

0 B(i2)2

)︸ ︷︷ ︸

D′×D′

. . .

(A

(iN )N 0

0 B(iN )N

)]ei1...iN

• Summing MPS vectors leads to an augmentation of the matrixsizes D → D′.

• Use compression techniques to keep the size of the MPSmatrices constant.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 13

Page 44: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Computations with MPS• Inner products

xHy =∑

i1,...,iN

xi1,...,iN︷ ︸︸ ︷(tr(A

(i1)1 · · · A(iN )

N

))·

yi1,...,iN︷ ︸︸ ︷(tr(B

(i1)1 · · ·B(iN )

N

))=∑ij

∑kj

a(i1)1;k1,k2

· · · a(iN )N ;kN ,k1

∑mj

b(i1)1;m1,m2

· · · b(iN )N ;mN ,m1

Find an efficient ordering of the summations!• Matrix × vector

Hx =

(M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N

) ∑i1,...,iN

tr(A(i1)1 . . . A

(iN )N )ei1...iN

=

M∑k=1

y(k)MPS

=⇒ 4. XK. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 14

Page 45: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Uniqueness of MPS• The MPS formalism is not unique:∑

i1,...,iN

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)ei1,...,iN

=∑

i1,...,iN

tr(

(A(i1)1 M−1

2 )(M2A(i2)2 M−1

3 ) · · · (MNA(iN )N )

)ei1,...,iN

• Replace the matrices A(0)j , A

(1)j by parts of unitary matrices using

the SVD: (A

(0)j

A(1)j

)=

(U

(0)j

U(1)j

)ΣjVj .

• Multiply the ΣjVj parts on the right neighbour.• Replace all A matrices by U matrices, up to one.• This remaining factor is considered to be optimized.• Leads to better convergence properties and numerical stability.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 15

Page 46: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Uniqueness of MPS• The MPS formalism is not unique:∑

i1,...,iN

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)ei1,...,iN

=∑

i1,...,iN

tr(

(A(i1)1 M−1

2 )(M2A(i2)2 M−1

3 ) · · · (MNA(iN )N )

)ei1,...,iN

• Replace the matrices A(0)j , A

(1)j by parts of unitary matrices using

the SVD: (A

(0)j

A(1)j

)=

(U

(0)j

U(1)j

)ΣjVj .

• Multiply the ΣjVj parts on the right neighbour.• Replace all A matrices by U matrices, up to one.• This remaining factor is considered to be optimized.• Leads to better convergence properties and numerical stability.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 15

Page 47: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Uniqueness of MPS• The MPS formalism is not unique:∑

i1,...,iN

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)ei1,...,iN

=∑

i1,...,iN

tr(

(A(i1)1 M−1

2 )(M2A(i2)2 M−1

3 ) · · · (MNA(iN )N )

)ei1,...,iN

• Replace the matrices A(0)j , A

(1)j by parts of unitary matrices using

the SVD: (A

(0)j

A(1)j

)=

(U

(0)j

U(1)j

)ΣjVj .

• Multiply the ΣjVj parts on the right neighbour.

• Replace all A matrices by U matrices, up to one.• This remaining factor is considered to be optimized.• Leads to better convergence properties and numerical stability.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 15

Page 48: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Uniqueness of MPS• The MPS formalism is not unique:∑

i1,...,iN

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)ei1,...,iN

=∑

i1,...,iN

tr(

(A(i1)1 M−1

2 )(M2A(i2)2 M−1

3 ) · · · (MNA(iN )N )

)ei1,...,iN

• Replace the matrices A(0)j , A

(1)j by parts of unitary matrices using

the SVD: (A

(0)j

A(1)j

)=

(U

(0)j

U(1)j

)ΣjVj .

• Multiply the ΣjVj parts on the right neighbour.• Replace all A matrices by U matrices, up to one.• This remaining factor is considered to be optimized.

• Leads to better convergence properties and numerical stability.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 15

Page 49: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Uniqueness of MPS• The MPS formalism is not unique:∑

i1,...,iN

tr(A

(i1)1 ·A(i2)

2 · · ·A(iN )N

)ei1,...,iN

=∑

i1,...,iN

tr(

(A(i1)1 M−1

2 )(M2A(i2)2 M−1

3 ) · · · (MNA(iN )N )

)ei1,...,iN

• Replace the matrices A(0)j , A

(1)j by parts of unitary matrices using

the SVD: (A

(0)j

A(1)j

)=

(U

(0)j

U(1)j

)ΣjVj .

• Multiply the ΣjVj parts on the right neighbour.• Replace all A matrices by U matrices, up to one.• This remaining factor is considered to be optimized.• Leads to better convergence properties and numerical stability.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 15

Page 50: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Computation of the Ground StateHow to calculate the smallest eigenvalue of the Hamiltonian H?

1. Variational ansatz:

minxMPS

xHMPSHxMPS

xHMPSxMPS

• Find optimal matrix pairs A(ij)j for all sites j

• Use Alternating Least Squares: Keep all matrix pairs up toone fixed and optimize the remaining one

2. New idea: Use subspace iteration

minx∈Um

xHHx

xHx

• Find optimal solution in the subspace Um

• Um is defined by MPS vectors

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 16

Page 51: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Computation of the Ground StateHow to calculate the smallest eigenvalue of the Hamiltonian H?

1. Variational ansatz:

minxMPS

xHMPSHxMPS

xHMPSxMPS

• Find optimal matrix pairs A(ij)j for all sites j

• Use Alternating Least Squares: Keep all matrix pairs up toone fixed and optimize the remaining one

2. New idea: Use subspace iteration

minx∈Um

xHHx

xHx

• Find optimal solution in the subspace Um

• Um is defined by MPS vectors

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 16

Page 52: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Minimization Algorithm in terms of MPS• Start with initial guesses for the A(ij)

j matrices.

• Replace A(ij)j by U (ij)

j up to A(i1)1 .

• Consider all U matrices as fixed and optimize X1 = A1:

minX1

(∑ij

tr(X

(i1)1 U

(i2)2 · · ·U(iN )

N

)ei

)H

H

(∑ij

tr(X

(i1)1 U

(i2)2 · · ·U(iN )

N

)ei

)(∑

ij

tr(X

(i1)1 U

(i2)2 · · ·U(iN )

N

)ei

)H(∑ij

tr(X

(i1)1 U

(i2)2 · · ·U(iN )

N

)ei

)

= minX1

X1HR1X1

X1HX1

with effective Hamiltonian R1

• =⇒ dense eigenvalue problem for the (2D2 × 2D2) matrix R1.• Orthogonalize the solution and continue with the right neighbor.

Modifications of this algorithm lead to the DMRG method.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 17

Page 53: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration for the smallest eigenvalue• Idea: Minimize the Rayleigh quotient xHHx

xHx over a subspace (!)Um := span{q1, . . . , qm}.

• Let Vm = (q1, . . . , qm) ∈ C2N×m

• The minimization over Um yields

minx∈Um\{0}

xHHx

xHx

x=Vmy= min

y∈Cm\{0}

(Vmy)HH(Vmy)

(Vmy)HVmy

= miny∈Cm\{0}

yH(V HmHVm

)y

yH (V HmVm) y

= miny∈Cm\{0}

yHAy

yHBy,

a generalized eigenvalue problem Ay = λminBy of size m×m.• x = Vmy or restart with q1 = Vmy

• How to choose an appropriate subspace Um?

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 18

Page 54: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration for the smallest eigenvalue• Idea: Minimize the Rayleigh quotient xHHx

xHx over a subspace (!)Um := span{q1, . . . , qm}.

• Let Vm = (q1, . . . , qm) ∈ C2N×m

• The minimization over Um yields

minx∈Um\{0}

xHHx

xHx

x=Vmy= min

y∈Cm\{0}

(Vmy)HH(Vmy)

(Vmy)HVmy

= miny∈Cm\{0}

yH(V HmHVm

)y

yH (V HmVm) y

= miny∈Cm\{0}

yHAy

yHBy,

a generalized eigenvalue problem Ay = λminBy of size m×m.

• x = Vmy or restart with q1 = Vmy

• How to choose an appropriate subspace Um?

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 18

Page 55: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration for the smallest eigenvalue• Idea: Minimize the Rayleigh quotient xHHx

xHx over a subspace (!)Um := span{q1, . . . , qm}.

• Let Vm = (q1, . . . , qm) ∈ C2N×m

• The minimization over Um yields

minx∈Um\{0}

xHHx

xHx

x=Vmy= min

y∈Cm\{0}

(Vmy)HH(Vmy)

(Vmy)HVmy

= miny∈Cm\{0}

yH(V HmHVm

)y

yH (V HmVm) y

= miny∈Cm\{0}

yHAy

yHBy,

a generalized eigenvalue problem Ay = λminBy of size m×m.• x = Vmy or restart with q1 = Vmy

• How to choose an appropriate subspace Um?

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 18

Page 56: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration for the smallest eigenvalue• Idea: Minimize the Rayleigh quotient xHHx

xHx over a subspace (!)Um := span{q1, . . . , qm}.

• Let Vm = (q1, . . . , qm) ∈ C2N×m

• The minimization over Um yields

minx∈Um\{0}

xHHx

xHx

x=Vmy= min

y∈Cm\{0}

(Vmy)HH(Vmy)

(Vmy)HVmy

= miny∈Cm\{0}

yH(V HmHVm

)y

yH (V HmVm) y

= miny∈Cm\{0}

yHAy

yHBy,

a generalized eigenvalue problem Ay = λminBy of size m×m.• x = Vmy or restart with q1 = Vmy

• How to choose an appropriate subspace Um?

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 18

Page 57: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration for the smallest eigenvalue• Choose the Krylov space Km(H, b) as subspace:

Um = Km(H, b) := span{b,Hb, . . . ,Hm−1b}

• This ansatz yields

A = V HmHVm =

bHHb bHH2b . . . bHHmbbHH2b bHH3b . . . bHHm+1b

......

. . ....

bHHmb bHHm+1b . . . bHH2m−1b

B = V HmVm =

bHb bHHb . . . bHHm−1bbHHb bHH3b . . . bHHmb

......

. . ....

bHHm−1b bHHmb . . . bHH2m−2b

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 19

Page 58: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration for the smallest eigenvalue• Choose the Krylov space Km(H, b) as subspace:

Um = Km(H, b) := span{b,Hb, . . . ,Hm−1b}

• This ansatz yields

A = V HmHVm =

bHHb bHH2b . . . bHHmbbHH2b bHH3b . . . bHHm+1b

......

. . ....

bHHmb bHHm+1b . . . bHH2m−1b

B = V HmVm =

bHb bHHb . . . bHHm−1bbHHb bHH3b . . . bHHmb

......

. . ....

bHHm−1b bHHmb . . . bHH2m−2b

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 19

Page 59: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• b = bMPS andKm(H, bMPS) = span{bMPS , HbMPS , . . . ,H

m−1bMPS}:

• In principle the same proceeding, BUT

• Computation of the matrix vector products:

HjbMPS =

(M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N

)j

bMPS

=

M∑k1=1

· · ·M∑

kj=1

b(k1,...,kj)MPS

Number of summands grows dramatically!• Application of the back-transformation:

Vmymin =m∑j=1

ymin,jHj−1bMPS

does not lie in the MPS space.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 20

Page 60: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• b = bMPS andKm(H, bMPS) = span{bMPS , HbMPS , . . . ,H

m−1bMPS}:• In principle the same proceeding, BUT

• Computation of the matrix vector products:

HjbMPS =

(M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N

)j

bMPS

=M∑

k1=1

· · ·M∑

kj=1

b(k1,...,kj)MPS

Number of summands grows dramatically!• Application of the back-transformation:

Vmymin =m∑j=1

ymin,jHj−1bMPS

does not lie in the MPS space.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 20

Page 61: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• b = bMPS andKm(H, bMPS) = span{bMPS , HbMPS , . . . ,H

m−1bMPS}:• In principle the same proceeding, BUT

• Computation of the matrix vector products:

HjbMPS =

(M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N

)j

bMPS

=

M∑k1=1

· · ·M∑

kj=1

b(k1,...,kj)MPS

Number of summands grows dramatically!

• Application of the back-transformation:

Vmymin =m∑j=1

ymin,jHj−1bMPS

does not lie in the MPS space.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 20

Page 62: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• b = bMPS andKm(H, bMPS) = span{bMPS , HbMPS , . . . ,H

m−1bMPS}:• In principle the same proceeding, BUT

• Computation of the matrix vector products:

HjbMPS =

(M∑k=1

αkQ(k)1 ⊗ · · · ⊗Q(k)

N

)j

bMPS

=

M∑k1=1

· · ·M∑

kj=1

b(k1,...,kj)MPS

Number of summands grows dramatically!• Application of the back-transformation:

Vmymin =

m∑j=1

ymin,jHj−1bMPS

does not lie in the MPS space.K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 20

Page 63: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• Solution: Application of a projection P onto the MPS space:

P :

M∑k=1

y(k)MPS 7→ argminxMPS

∥∥∥∥∥M∑k=1

y(k)MPS − xMPS

∥∥∥∥∥

• Solve the optimization problem

minA

∥∥∥∥∥∥∑

i1,...,iN

tr

(M∑k=1

(B(i1;k)1 · · ·B(iN ;k)

N )−A(i1)1 · · ·A(iN )

N

)ei1...iN

∥∥∥∥∥∥• Two possible projection methods:

• SVD-based truncation• ALS-based optimization

• Use ALS to find optimal matrix pair A(0)r , A

(1)r

• Forcing the derivative to be zero leads to a linear system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 21

Page 64: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• Solution: Application of a projection P onto the MPS space:

P :

M∑k=1

y(k)MPS 7→ argminxMPS

∥∥∥∥∥M∑k=1

y(k)MPS − xMPS

∥∥∥∥∥• Solve the optimization problem

minA

∥∥∥∥∥∥∑

i1,...,iN

tr

(M∑k=1

(B(i1;k)1 · · ·B(iN ;k)

N )−A(i1)1 · · ·A(iN )

N

)ei1...iN

∥∥∥∥∥∥

• Two possible projection methods:• SVD-based truncation• ALS-based optimization

• Use ALS to find optimal matrix pair A(0)r , A

(1)r

• Forcing the derivative to be zero leads to a linear system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 21

Page 65: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• Solution: Application of a projection P onto the MPS space:

P :

M∑k=1

y(k)MPS 7→ argminxMPS

∥∥∥∥∥M∑k=1

y(k)MPS − xMPS

∥∥∥∥∥• Solve the optimization problem

minA

∥∥∥∥∥∥∑

i1,...,iN

tr

(M∑k=1

(B(i1;k)1 · · ·B(iN ;k)

N )−A(i1)1 · · ·A(iN )

N

)ei1...iN

∥∥∥∥∥∥• Two possible projection methods:

• SVD-based truncation• ALS-based optimization

• Use ALS to find optimal matrix pair A(0)r , A

(1)r

• Forcing the derivative to be zero leads to a linear system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 21

Page 66: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Subspace iteration in terms of MPS• Solution: Application of a projection P onto the MPS space:

P :

M∑k=1

y(k)MPS 7→ argminxMPS

∥∥∥∥∥M∑k=1

y(k)MPS − xMPS

∥∥∥∥∥• Solve the optimization problem

minA

∥∥∥∥∥∥∑

i1,...,iN

tr

(M∑k=1

(B(i1;k)1 · · ·B(iN ;k)

N )−A(i1)1 · · ·A(iN )

N

)ei1...iN

∥∥∥∥∥∥• Two possible projection methods:

• SVD-based truncation• ALS-based optimization

• Use ALS to find optimal matrix pair A(0)r , A

(1)r

• Forcing the derivative to be zero leads to a linear system.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 21

Page 67: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Numerical resultsIsing-type Hamiltonian, N = 10, dim = 2

1 5 10 15 20 25 30 35 40 45 5010

−6

10−4

10−2

100

102

Number of iterations

App

roxi

mat

ion

erro

r

Full vectorMPS, D=1MPS, D=2MPS, D=3MPS, D=4

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 22

Page 68: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Numerical resultsIsing-type Hamiltonian, N = 10, dim = 3

1 5 10 15 20 25 30 35 40 45 5010

−6

10−4

10−2

100

102

Number of iterations

App

roxi

mat

ion

erro

r

Full vectorMPS, D=1MPS, D=2MPS, D=3MPS, D=4

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 23

Page 69: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Numerical resultsIsing-type Hamiltonian, N = 10, dim = 4

5 10 15 20 25 30 35 40 45 50110

−6

10−4

10−2

100

102

Number of iterations

App

roxi

mat

ion

erro

r

Full vectorMPS, D=1MPS, D=2MPS, D=3MPS, D=4

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 24

Page 70: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Numerical resultsIsing-type Hamiltonian, N = 25, dim = 5

1 5 10 1510

−4

10−3

10−2

10−1

100

101

102

Number of iterations

App

roxi

mat

ion

erro

r

MPS, D=2MPS, D=3MPS, D=4MPS, D=5

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 25

Page 71: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Conclusions• Matrix Product States

• data-sparse representation format• represent ground states faithfully• DMRG in terms of MPS

• Subspace methods in terms of MPS• more flexibility (choice of subspace dimension)• approximation not only of the lowest eigenvalue• projection required to keep the number of summands limited• projection can even improve the approximation• convergence to DMRG solution• convergence properties of subspace iteration and

approximation properties of MPS

Thank you for your attention.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 26

Page 72: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Conclusions• Matrix Product States

• data-sparse representation format• represent ground states faithfully• DMRG in terms of MPS

• Subspace methods in terms of MPS• more flexibility (choice of subspace dimension)• approximation not only of the lowest eigenvalue• projection required to keep the number of summands limited• projection can even improve the approximation• convergence to DMRG solution• convergence properties of subspace iteration and

approximation properties of MPS

Thank you for your attention.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 26

Page 73: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Conclusions• Matrix Product States

• data-sparse representation format• represent ground states faithfully• DMRG in terms of MPS

• Subspace methods in terms of MPS• more flexibility (choice of subspace dimension)• approximation not only of the lowest eigenvalue• projection required to keep the number of summands limited• projection can even improve the approximation• convergence to DMRG solution• convergence properties of subspace iteration and

approximation properties of MPS

Thank you for your attention.

K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 26

Page 74: Approximation Techniques for Special Eigenvalue Problems ... fileApproximation Techniques for Special Eigenvalue Problems

Technische Universitat Munchen

Conclusions• Matrix Product States

• data-sparse representation format• represent ground states faithfully• DMRG in terms of MPS

• Subspace methods in terms of MPS• more flexibility (choice of subspace dimension)• approximation not only of the lowest eigenvalue• projection required to keep the number of summands limited• projection can even improve the approximation• convergence to DMRG solution• convergence properties of subspace iteration and

approximation properties of MPS

Thank you for your attention.K. Waldherr: Approximation Techniques for Special Eigenvalue Problems

HDA 2011, Bonn, June 2011 26