Useful Packages for Measurements

---

PyVISA

PyVISA controls your instruments using raw commands.

Keithley2600

Keithley2600 controls your Keithley 2600 instruments with abstraction.

PyMeasure

PyMeasure controls a wide range of instruments.

DMRG (I): Theory

トニカクタダシイ

---

Simple DMRG

Basic Theory

Quantum renormalization in real space applies only to single-particle systems.

On each step, the Hamiltonian undergoes a truncation \[H^S(M) = V^\dagger H(M\cdot M_S) V,\] where \(V\) contains the most significant \(M\) eigenstates of \[\rho^S (M\cdot M_S),\] given by the ground state of \[H(M\cdot M_S).\]

In the algorithm, we should keep track of the creation operators on the boundary sites.

Structure of Infinite-Size DMRG

1. Set up an initial chain of length \(L\).
2. Calculate the ground state of this chain.
3. Get the reduced density matrix \(\rho^S\).
4. Diagonalize \(\rho^S{}\) and keep the most significant \(M\) eigenstates as \(V\).
5. Reduce \(H\) to a matrix of size \(M\times M\), as well as the creation operators at the boundary.
6. Add a new site.
7. Find the ground state of the new system.
8. Return to step 3.

Structure of Finite-Size DMRG

1. Carry out infinite-size DMRG until the size reaches \(L\). During this step, store all the Hamiltonians and other necessary steps in the memory.
2. Reduce the matrices (Hamiltonian and creations operators) of \(S\) to size \(M\). Before doing the projection, we should first trace out the environment part.
3. Add two sites to \(H_S\), and reduce the environment to length \(L/2 - 2\).
4. The procedure of the last two steps is repeated until \(S\) has size \(L-3\). The matrices for \(S\) should still be stored for each size.
5. Reduce the size of \(S\). For the environment, the stored matrices are used, whereas for the system, the matrices are updated at eac hstep. Carry on until the system \(S\) has reached size \(1\).
6. Repeat the up and down sweeps of the previous steps until convergence has been achieved.

Matrix Product State

Prerequisites: for SVD and QR Decomposition, see Mathematics Miscellany.

Conventions

• \(M\) for general matrices;
• \(\Lambda\) and \(S\) for diagonal matrices;
• \(A\) for left-canonical ones;
• \(B\) for right-canonical ones.
• Label \(l\) in \(\ket{a_l}\) indicates the bond.
• Superscript \([l]\) in \(O^{[l]}\) indicates that the operator lives on site \(l\).
• OBS unless otherwise stated. Site index from \(1\) to \(L\) while bond index from \(1\) to \(L-1\).

Graphical Notations

Representations and Conversions

Classical Representation

\[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} c_{\sigma_1,\cdots,\sigma_L} \ket{\sigma_1, \cdots, \sigma_L}.\]

Left-Canonical

\[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} A^{\sigma_2} \cdots A^{\sigma_{L-1}} A^{\sigma_L} \ket{\sigma_1,\cdots,\sigma_L}\] where

• for every \(l\), \[\sum_{\sigma_l} A^{\sigma_l \dagger} A^{\sigma_l} = \mathbb{1};\]
• \(A^{\sigma_1}\)'s are row matrices while \(A^{\sigma_L}\)'s are column matrices.

From Classical Representation

\[ \begin{align*} c_{\sigma_1,\cdots,\sigma_L} &\xlongequal[\mathrm{reshape}]{} \Psi_{\sigma_1,(\sigma_2,\cdots,\sigma_L)} \\ &\xlongequal{\mathrm{SVD}} \sum_{a_1}^{r_1} U_{\sigma_1,a_1} S_{a_1,a_1} (V^\dagger)_{a_1,(\sigma_2,\cdots,\sigma_L)} \\ &\xlongequal[\mathrm{reshape}]{A^{\sigma_1}_{a_1} = U_{\sigma_1,a_1}} \sum_{a_1}^{r_1} A^{\sigma_1}_{a_1} \Psi_{(a_1\sigma_2),(\sigma_3,\cdots,\sigma_L)} \\ &\xlongequal{\mathrm{SVD}} \sum_{a_1}^{r_1} \sum_{a_2}^{r_2} A^{\sigma_1}_{a_1} U_{(a_1\sigma_2),a_2} S_{a_2,a_2} (V^\dagger)_{a_2,(\sigma_3,\cdots,\sigma_L)} \\ &\xlongequal[\mathrm{reshape}]{A^{\sigma_2}_{a_1,a_2} = U_{(a_1\sigma_2),a_2}} \sum_{a_1}^{r_1} \sum_{a_2}^{r_2} A^{\sigma_1}_{a_1} A^{\sigma_2}_{a_1,a_2} \Psi_{(\sigma_2 a_3),(\sigma_4,\cdots,\sigma_L)} \\ &= \cdots \\ &= \sum_{a_1,\cdots,a_{L-1}} A^{\sigma_1}_{a_1} A^{\sigma_2}_{a_1,a_2} \cdots A^{\sigma_{L-1}}_{a_{L-2},a_{L-1}} A^{\sigma_L}_{a_{L-1}} \end{align*} \]

From General MPS

\[ \begin{align*} \sum_{\sigma_1,\cdots,\sigma_L} \sum_{a_1,\cdots,a_L} M_{(\sigma_1,1),a_1} M^{\sigma_2}_{a_1,a_2} \cdots \ket{\sigma_1\cdots \sigma_L} &= \sum_{\sigma_1,\cdots,\sigma_L} \sum_{a_1,\cdots,a_L} \sum_{s_1} A_{(\sigma_1,1),s_1} S_{s_1,s_1} V^\dagger_{s_1,a_1} M^{\sigma_2}_{a_1,a_2} \cdots \ket{\sigma_1,\cdots,\sigma_L} \\ &= \sum_{\sigma_1,\cdots,\sigma_L} \sum_{a_2,\cdots,a_L} \sum_{s_1} A^{\sigma_1}_{1,s_1} \tilde{M}^{\sigma_2}_{s_1,a_2} M^{\sigma_3}_{a_2,a_3} \cdots \ket{\sigma_1,\cdots,\sigma_L} \\ &= \cdots. \end{align*} \]

From Vidal Notation

\[\qty(\Lambda^{[0]} \Gamma^{\sigma_1})\qty(\Lambda^{[1]} \Gamma^{\sigma_2})\qty(\Lambda^{[2]} \Gamma^{\sigma_3})\cdots \rightarrow A^{\sigma_1} A^{\sigma_2} A^{\sigma_3} \cdots.\]

Right-Canonical

\[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} B^{\sigma_1} B^{\sigma_2} \cdots B^{\sigma_{L-1}} B^{\sigma_L} \ket{\sigma_1,\cdots,\sigma_L}\] where

• for every \(l\), \[\sum_{\sigma_l} B^{\sigma_l} B^{\sigma_l \dagger} = \mathbb{1};\]
• \(B^{\sigma_1}\)'s are row matrices while \(B^{\sigma_L}\)'s are column matrices.

From Classical Representation

Similar to that of left-canonical ones.

From General MPS

Similar to that of left-canonical ones.

From Vidal Notation

\[\qty(\Gamma^{\sigma_1}\Lambda^{[1]}) \qty(\Gamma^{\sigma_2}\Lambda^{[2]}) \qty(\Gamma^{\sigma_3}\Lambda^{[3]})\cdots \rightarrow B^{\sigma_1} B^{\sigma_2} B^{\sigma_3} \cdots.\]

Mixed-Canonical

\[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots A^{\sigma_l} S B^{\sigma_{l+1}} \cdots B^{\sigma_L}.\]

From Classical Representation

\[ \begin{align*} c_{\sigma_1,\cdots,\sigma_L} &= \sum_{a_l} (A^{\sigma_1} \cdots A^{\sigma_l})_{a_l} S_{a_l,a_l} (V^\dagger)_{a_l,(\sigma_{l+1},\cdots,\sigma_L)} \\ &= \sum_{a_l} (A^{\sigma_1} \cdots A^{\sigma_l})_{a_l} S_{a_l,a_l} (B^{\sigma_{l+1}} \cdots B^{\sigma_L})_{a_l}. \end{align*} \]

Vidal Notation

\[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} \Gamma^{\sigma_1} \Lambda^{[1]} \Gamma^{\sigma_2} \Lambda^{[2]} \cdots \Gamma^{\sigma_{L-1}} \Lambda^{[L-1]} \Gamma^{\sigma_L} \ket{\sigma_1,\cdots,\sigma_L},\] where

• the states on block \(A\) and \(B\) given by \[ \begin{align*} \ket{a_l}_A &= \sum_{\sigma_1,\cdots,\sigma_l} \qty(\Gamma^{\sigma_1} \Lambda^{[1]} \Gamma^{\sigma_2} \cdots \Lambda^{[l-1]}\Gamma^{\sigma_l})_{a_l} \ket{\sigma_1,\cdots,\sigma_l}, \\ \ket{a_l}_B &= \sum_{\sigma_{l+1},\cdots,\sigma_L} \qty(\Gamma^{\sigma_{l+1}} \Lambda^{[l+1]} \Gamma^{\sigma_{l+2}} \cdots \Lambda^{[L-1]} \Gamma^{\sigma_L})_{a_l} \ket{\sigma_{l+1} \cdots \sigma_L} \end{align*} \] are orthonormal respectively.
• The reduced density matrices are given by \[\rho^{[l]}_A = \rho^{[l]}_B = \qty(\Lambda^{[l]})^2.\]

From Classical Representation

\[ \begin{align*} c_{\sigma_1,\cdots,\sigma_L} &\xlongequal[\mathrm{reshape}]{} \Psi_{\sigma_1,(\sigma_2,\cdots,\sigma_L)} \\ &\xlongequal{\mathrm{SVD}} \sum_{a_1} A^{\sigma_1}_{a_1} \Lambda^{[1]}_{a_1,a_1}(V^\dagger)_{a_1,(\sigma_2,\cdots,\sigma_L)} \\ &\xlongequal[\mathrm{reshape}]{A^{\sigma_1}_{a_1} = \Gamma^{\sigma_1}_{a_1}} \sum_{a_1} \Gamma^{\sigma_1}_{a_1} \Psi_{(a_1\sigma_2),(\sigma_3,\cdots,\sigma_L)} \\ &\xlongequal{\mathrm{SVD}} \sum_{a_1,a_2} \Gamma^{\sigma_1}_{a_1} A^{\sigma_2}_{a_1,a_2} \Lambda^{[2]}_{a_2,a_2} (V^\dagger)_{a_2,(\sigma_3,\cdots,\sigma_L)} \\ &\xlongequal[\color{orange} \mathrm{problematic\ division}]{A^{\sigma_2}_{a_1,a_2} = \Lambda^{[1]}_{a_1,a_1}\Gamma^{\sigma_2}_{a_1,a_2}} \sum_{a_1,a_2} \Gamma^{\sigma_1}_{a_1} \Lambda^{[1]}_{a_1,a_1} \Gamma^{\sigma_2}_{a_1,a_2} \Psi_{(a_2,\sigma_3),(\sigma_4,\cdots,\sigma_L)} \\ &= \cdots \\ &= \sum_{a_1,\cdots,a_L} \Gamma^{\sigma_1}_{a_1} \Lambda^{[1]}_{a_1,a_1} \Gamma^{\sigma_2}_{a_1,a_2} \Lambda^{[2]}_{a_2,a_2} \cdots \Lambda^{[L-1]}_{a_{l-1},a_{l-1}} \Gamma^{\sigma_L}_{a_{l-1},a_l}. \end{align*} \]

• Substitution starting from the left: \[A^{\sigma_l}_{a_{l-1},a_l} = \Lambda^{[l-1]}_{a_{l-1},a_{l-1}} \Gamma^{\sigma_l}_{a_{l-1},a_l}.\]
• Substitution starting from the right: \[B^{\sigma_l}_{a_{l-1},a_l} = \Gamma^{\sigma_l}_{a_{l-1},a_l} \Lambda^{[l]}_{a_l,a_l}.\]
• Dummy \[\Lambda^{[0]} = \Lambda^{[L]} = 1.\]

From Left-Canonical

Similar to that of right-canonical ones.

From Right-Canonical

\[ \begin{align*} B^{\sigma_1} B^{\sigma_2} B^{\sigma_3} \cdots &\xlongequal{\mathrm{SVD}} (A^{\sigma_1} \Lambda^{[1]} V^\dagger) B^{\sigma_2} B^{\sigma_3} \cdots \\ &\xlongequal{A^{\sigma_1} = \Lambda^{[0]}\Gamma^{\sigma_1}} \Gamma^{\sigma_1} M^{\sigma_2} B^{\sigma_3} \cdots \\ &\xlongequal{\mathrm{SVD}} \Gamma^{\sigma_1} (A^{\sigma_2} \Lambda^{[2]}V^\dagger) B^{\sigma_3} \cdots \\ &\xlongequal[\color{orange}\mathrm{problematic\ division}]{A^{\sigma_2} = \Lambda^{[1]}\Gamma^{\sigma_2}} \Gamma^{\sigma_1} \Lambda^{[1]}\Gamma^{\sigma_2} M^{\sigma_3} B^{\sigma_3} \cdots. \end{align*} \]

Calculation With MPS

Block States

• Left block: defined recursively by \[ \begin{align*} \ket{a_l}_A &= \sum_{a_{l-1}} \sum_{\sigma_l} A^{\sigma_l}_{a_{l-1}, a_l} \ket{a_{l-1}}_A \ket{\sigma_l} \\ &= \cdots \\ &= \sum_{\sigma_1,\cdots,\sigma_l} (A^{\sigma_1} \cdots A^{\sigma_l})_{1,a_l} \ket{\sigma_1} \cdots \ket{\sigma_l}. \end{align*} \]
• Right block: defined recursively by \[ \begin{align*} \ket{a_{l}}_B &= \sum_{a_{l+1}} \sum_{\sigma_{l+1}} B^{\sigma_{l+1}}_{a_l,a_{l+1}} \ket{\sigma_{l+1}} \ket{a_{l+1}}_B \\ &= \cdots \\ &= \sum_{\sigma_{l+1}, \cdots, \sigma_L} (B^{\sigma_{l+1}} \cdots B^{\sigma_L})_{a_{l+1},1} \ket{\sigma_{l+1}} \cdots \ket{\sigma_L}. \end{align*} \]
• The \(\qty{\ket{a_l}_A}\) (as well as \(\qty{\ket{a_l}_B}\)) constructed in this way are orthonormal.
• A general state may be written as \[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots A^{\sigma_l} \Psi B^{\sigma_{l+1}} \cdots B^{\sigma_L} \ket{\sigma_1} \cdots \ket{\sigma_L},\] where \(\Psi\) is a matrix.

Vidal Notation and DMRG

• Vidal notation to blocks: \[ \begin{gather*} \qty(\Lambda^{[0]}\Gamma)\cdots \qty(\Lambda^{[l-2]}\Gamma)\qty(\Lambda^{[l-1]}\Gamma) \Lambda^{[l]} \qty(\Gamma \Lambda^{[l+1]}) \qty(\Gamma \Lambda^{[l+2]}) \cdots \qty(\Gamma \Lambda^{L}) \\ \rightarrow \ket{\psi} = \sum_{a_l} \ket{a_l}_A s_{a_l} \ket{a_l}_B. \end{gather*} \]
• Vidal notation to one-site DMRG: \[ \begin{gather*} \qty(\Lambda^{[0]}\Gamma)\cdots \qty(\Lambda^{[l-2]}\Gamma)\qty(\Lambda^{[l-1]}\Gamma \Lambda^{[l]}) \qty(\Gamma \Lambda^{[l+1]}) \qty(\Gamma \Lambda^{[l+2]}) \cdots \qty(\Gamma \Lambda^{L}) \\ \rightarrow \ket{\psi} = \sum_{a_{l-1},a_l,\sigma_l} \ket{a_{l-1}}_A \Psi^{\sigma_l}_{a_{l-1},a_l} \ket{a_l}_B. \end{gather*} \]
• Vidal notation to two-site DMRG: \[ \begin{gather*} \qty(\Lambda^{[0]}\Gamma)\cdots \qty(\Lambda^{[l-2]}\Gamma)\qty(\Lambda^{[l-1]}\Gamma \Lambda^{[l]} \Gamma \Lambda^{[l+1]}) \qty(\Gamma \Lambda^{[l+2]}) \cdots \qty(\Gamma \Lambda^{L}) \\ \rightarrow \ket{\psi} = \sum_{a_{l-1},a_{l+1},\sigma_l,\sigma_{l+1}} \ket{a_{l-1}}_A \Psi^{\sigma_l\sigma_{l+1}}_{a_{l-1},a_{l+1}} \ket{a_{l+1}}_B. \end{gather*} \]

Matrix Elements

Notations

• \(\ket{\psi}\) as MPS: \[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} M^{\sigma_1} \cdots M^{\sigma_L} \ket{\sigma_1} \cdots \ket{\sigma_L}.\]
• \(\ket{\phi}\) as MPS: \[\ket{\phi} = \sum_{\sigma_1,\cdots,\sigma_L} N^{\sigma_1} \cdots N^{\sigma_L} \ket{\sigma_1} \cdots \ket{\sigma_L}.\]
• Operator living on site \(l\): \[\hat{O}^{[l]} = \sum_{\sigma_l,\sigma'_l} O^{\sigma_l,\sigma'_l} \ket{\sigma_l} \bra{\sigma'_l}.\]

Overlap

• \(C^{[0]} = 1\).
• \(C^{[l]}\) defined recursively by \[C^{[l]} = \sum_{\sigma_l} N^{\sigma_l\dagger} C^{[l-1]} M^{\sigma_l}.\]
  • Transfer operator: specifically for \(N=M\), \[\hat{E}^{[l]} = \sum_{a_{l-1},a'_{l-1}} \sum_{a_l,a'_l} \qty(\sum_{\sigma_l} N^{[l] \sigma_{l*}} \otimes M^{[l]\sigma_l})_{a_{l-1} a'_{l-1}, a_l a'_l} (\ket{a'_{l-1}} \bra{a_{l-1}})(\ket{a_l}\bra{a'_l}).\]
  • Reformulated with transfer opeartor: \[C^{[l]} = E^{[l]}\qty[C^{[l-1]}].\]
• Finally we have \[C^{[L]} = \bra{\phi}\ket{\psi}.\]

Matrix Element

• \(C^{[0]} = 1\).
• For \(l\neq i, j\), \(C^{[l]}\) is defined recursively by \[C^{[l]} = \sum_{\sigma_l} N^{\sigma_l\dagger} C^{[l-1]} M^{\sigma_l}.\]
• For \(l=i\) or \(l=j\), \(C^{[l]}\) is defined by \[C^{[l]} = \sum_{\sigma_l,\sigma'_l} O^{\sigma_l,\sigma'_l} N^{\sigma_l\dagger} C^{[l-1]} M^{\sigma'_l}.\]
  • Transfer operator: specifically for \(N=M\), \[\hat{E}^{[l]}_O = \sum_{a_{l-1},a'_{l-1}} \sum_{a_l,a'_l} \qty(\sum_{\sigma_l,\sigma'_l} O^{\sigma_l,\sigma'_l} N^{[l] \sigma_{l*}} \otimes M^{[l]\sigma_l})_{a_{l-1} a'_{l-1}, a_l a'_l} (\ket{a'_{l-1}} \bra{a_{l-1}})(\ket{a_l}\bra{a'_l}).\]
  • Reformulated with transfer opeartor: \[C^{[l]} = E^{[l]}\qty[C^{[l-1]}].\]
• Finally we have \[C^{[L]} = \bra{\phi} O^{[i]} O^{[j]} \ket{\psi}.\]

Properties of Transfer Matrices

• If \(N=M\) is a left-normalized or right-normalized matrix, then all eigenvalues of \(E\) satisfies \[\abs{\lambda} \le 1.\]

Compression of MPS

SVD Compression

• Starting from a left-canonical MPS.
• \(\tilde{O}\) denoting \(O\) truncated, \[ \begin{align*} \ket{\psi} &= \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots A^{\sigma_L} \ket{\sigma_1 \cdots \sigma_L} \\ &= \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots A^{\sigma_{L-1}} USB^{\sigma_L} \ket{\sigma_1 \cdots \sigma_L} \\ &\rightarrow \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots (A^{\sigma_{L-1}} \tilde{U}\tilde{S})B^{\sigma_L} \ket{\sigma_1 \cdots \sigma_L} \\ &= \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots A^{\sigma_{L-2}} US B^{\sigma_{L-1}} B^{\sigma_L} \ket{\sigma_1 \cdots \sigma_L} \\ &\rightarrow \cdots \\ &\rightarrow \sum_{\sigma_1,\cdots,\sigma_L} (A^{\sigma_1} \cdots A^{\sigma_l} U) S (B^{\sigma_{l+1}}\cdots B^{\sigma_L}) \ket{\sigma_1\cdots \sigma_l} \ket{\sigma_{l+1}\cdots \sigma_L}. \end{align*} \]

Variational Compression

MPO

\[\hat O \ket{\psi} = \sum_{\vb*{\sigma}} N^{\sigma_1} \cdots N^{\sigma_L} \ket{\vb*{\sigma}}.\]

• Left/right normalized by taking \((\sigma_l,\sigma'_l)\) as a large index.

\[\hat{O} \ket{\psi} = \sum_{\vb*{\sigma}} N^{\sigma_1} \cdots N^{\sigma_L} \ket{\vb*{\sigma}}\] where \[N^{\sigma_i} = \sum_{\sigma'_i} W^{\sigma_i \sigma'_i} \otimes M^{\sigma'_i}.\]

Ground State Calculation

MPO Representation of Hamiltonians

1. Hamiltonian defined by \[\hat{H} = \sum_{i=1}^{L-1} \qty(\frac{J}{2} \hat{S}_i^{+} \hat{S}_{i+1}^{-} + \frac{J}{2} \hat{S}_{i}^{-} \hat{S}_{i+1}^+ + J^z \hat{S}^z_{i} \hat{S}^z_{i+1}) - h\sum_i \hat{S}^z_i.\]
2. Hamiltonian represented as \[\hat{O} = \hat{W}^{[1]} \cdots \hat{W}^{[L]}\] where \(W^{[i]}\)'s are Operator-valued matrices \[\hat{W}_{bb'} = \sum_{\sigma \sigma'} W^{\sigma \sigma'}_{bb'} \ket{\sigma}\bra{\sigma}{}.\]
3. For \(2 \le l \le L-1\), the linear space that \(W^{[l]}\) acts on may be generated by \(\qty{\vb*{e}^{[l]}_{b'}}\) defined by \[ \begin{align*} O^{[l]}_1 &= \mathbb{1} \otimes \cdots \otimes \mathbb{1}, \\ O^{[l]}_2 &= \hat{S}^+ \otimes \mathbb{1} \otimes \cdots \otimes \mathbb{1}, \\ O^{[l]}_3 &= \hat{S}^- \otimes \mathbb{1} \otimes \cdots \otimes \mathbb{1}, \\ O^{[l]}_4 &= \hat{S}^z \otimes \mathbb{1} \otimes \cdots \otimes \mathbb{1}, \\ O^{[l]}_5 &= \mathrm{sum\ of\ completed\ terms\ to\ the\ right}. \\ \end{align*} \]
4. We have the recurrence relation \[ \begin{align*} O^{[l-1]}_1 &= \mathbb{1} \otimes O^{[l]}_1, \\ O^{[l-1]}_2 &= \hat{S}^+ \otimes O^{[l]}_1, \\ O^{[l-1]}_3 &= \hat{S}^- \otimes O^{[l]}_1, \\ O^{[l-1]}_4 &= \hat{S}^{z} \otimes O^{[l]}_1, \\ O^{[l-1]}_1 &= -h \hat{S}^z \otimes O^{[l]}_1 + \frac{J}{2} \hat{S}^- \otimes O^{[l]}_2 + \frac{J}{2} \hat{S}^+ \otimes O^{[l]}_3 + J^z \hat{S}^z O^{[l]}_4 + \mathbb{1}\otimes O^{[l]}_5. \\ \end{align*} \]
5. We could therefore define \(W^{[l]}\) by \[ W^{[l]} = \begin{pmatrix} \hat{1} & & & & \\ \hat{S}^+ & & & & \\ \hat{S}^- & & & & \\ \hat{S}^z & & & & \\ -h\hat{S}^z & (J/2) \hat{S}^- & (J/2) \hat{S}^+ & J^z \hat{S}^z & \mathbb{1} \end{pmatrix}{} \] such that \[ \begin{pmatrix} O^{[l-1]}_1 \\ O^{[l-1]}_2 \\ O^{[l-1]}_3 \\ O^{[l-1]}_4 \\ O^{[l-1]}_5 \end{pmatrix} = \hat{W}^{[l]} \begin{pmatrix} O^{[l]}_1 \\ O^{[l]}_2 \\ O^{[l]}_3 \\ O^{[l]}_4 \\ O^{[l]}_5 \end{pmatrix}, \] where the multiplication in the dot product is given by kron.
6. At the border we have \[ \hat{W}^{[1]} = \begin{pmatrix} -h\hat{S}^z & (J/2) \hat{S}^- & (J/2) \hat{S}^+ & J^z \hat{S}^z & \mathbb{1} \end{pmatrix} \] and \[ \hat{W}^{[L]} = \begin{pmatrix} \mathbb{1} \\ \hat{S}^+ \\ \hat{S}^- \\ \hat{S}^z \\ -h\hat{S}^z \end{pmatrix}. \]

The lesson:

• It's not that hard to contruct \(W^{[l]}\). Instead of thinking \(W^{[l]}\) as matrices labeled by \(\sigma\) and \(\sigma'\), we take it as operator-valued matrices labeld by \(b\) and \(b'\).
• \(W^{[l]}\) is sparse.
• The MPO representation of Hamiltonian stores the exact Hamiltonian in a small space without compression.

Applying a Hamiltonian MPO to a Mixed Canonical State

\[\ket{\psi} = \sum_{\sigma_1,\cdots,\sigma_L} A^{\sigma_1} \cdots A^{\sigma_{l-1}} \Psi^{\sigma_l} B^{\sigma_{l+1}} \cdots B^{\sigma_L} \ket{\sigma_1 \cdots \sigma_L}.\]

\[\bra{a_{l-1} \sigma_l a_l} \hat{H} \ket{a'_{l-1}\sigma'_l a'_l} = \sum_{b_{l-1},b_l} L^{a_{l-1},a'_{l-1}}_{b_{l-1}} W^{\sigma_l \sigma'_l}_{b_{l-1},b_l} R^{a_l,a'_l}_{b_l}.\]

Iterative Ground State Search

1. Target: with a Lagrangian multiplier \(\lambda\), we minimize \[\bra{\psi} H \ket{\psi} - \lambda \bra{\psi} \ket{\psi}.\]

   

2. Iterative search: only \(M^{[l]}\) is allowed to vary.

   

3. Boost: we'd like
   1. \(M^{[l]}\)'s be already a good guess, and
   2. The matrices left and right are left-normalized and right-normalized, respectively, so that the above equation is a eigenvalue problem instead of a generalized one.

Algorithm

1. Start from some initial guess for \(\ket{\psi}\), right-normalized.
2. Calculate the \(R\)-expressions iteratively for all sites from \(L-1\) to \(1\).
3. Right sweep: sweeping from site \(l=1\) to \(L-1\).
   1. At each site, solve the eigenvalue problem with starting vector given by the current state, getting \(M^{\sigma_l}\).
   2. Obtain \(A^{\sigma_l}\) by SVD, with the remaining \(SV^\dagger\) multiplied to the \(M^{\sigma_{l+1}}\) to the right as the starting point of the eigensolver for the next site.
   3. Build iteratively the \(L\)-expression.
4. Left sweep: sweeping from site \(l=L\) to \(2\).
   1. Again, solve the eigenvalue problem at each site.
   2. Obtain \(B^{\sigma_l}\) by SVD, with the remaining \(US\) multiplied to the \(M^{\sigma_l}\) to the left.
   3. Build iteratively the \(R\)-expression.
5. Repeat left and right sweeps, until convergence is achieved. Convergence condition given by \[\bra{\psi} \hat{H}^2 \ket{\psi} - \qty(\bra{\psi} \hat{H} \ket{\psi})^2 \rightarrow 0.\]

Computational Tricks

Avoid Trapping

Add a correction to \(\hat{\rho}^{A\bullet}\) before doing SVD: \[\hat{\rho}^{A\bullet} = \tr_B \ket{\psi} \bra{\psi} + \alpha \sum_{b_l} \tr_B \hat{H}^{A\bullet}_{b_l}\ket{\psi} \bra{\psi} \hat{H}^{A\bullet}_{b_l},\] where \(\alpha\) takes small value \(O(10^{-4})\) and will be taken slowly to zero as sweeping goes on.

The correction term \(\hat{H}_{b_l}^{A\bullet}\ket{\psi}\) is evaluated as follows.

Miscellaneous

Entanglement

• Maximally entangled state: \(\rho \in H_A \otimes H_B\) has the form for each of the subsystems \[\begin{pmatrix} 1/n & & & \\ & 1/n & & \\ & & \ddots & \\ & & & 1/n \end{pmatrix}.\]

Mathematics Miscellany

---

Matrix Decomposition

SVD

\[M = USV^\dagger\] where

• columns of \(U\) are orthonormal;
• \(S\) is diagonal with non-negative entries;
• rows of \(V^\dagger\) are orthonormal.

QR Decomposition

Full QR

\[M = QR\] where

• \(Q\) is unitary;
• \(R\) is upper-triangular.

Thin QR

\[M = Q_1 R_1\] where

• columns of \(Q_1\) are orthonormal;
• \(R_1\) is upper-triangular.

Kronecker Product

Properties

• Associativity: \[(A\otimes B) \otimes C = A \otimes (B\otimes C).\]
• Mixed-product: \[(A\otimes B)\cdot (C\otimes D) = (A\cdot C)\otimes (B\cdot D).\] In particular, \[A \otimes B = (A \otimes \mathbb{1}) \cdot (\mathbb{1}\otimes B).\]
• Inverse, Transpose, etc: \[\begin{gather*} (A\otimes B)^{-1} = A^{-1} \otimes B^{-1}, \\ (A\otimes B)^\dagger = A^\dagger \otimes B^\dagger. \end{gather*}{}\]
• Trace, Determinant, etc: \[\begin{gather*} \det (A_{n\times n}\otimes B_{m\times m}) = (\det A)^m (\det B)^n, \\ \tr (A \otimes B) = (\tr A)(\tr B). \end{gather*}{}\]
• The kron of two matrices of orthonormal columns/rows is again a matrix of orthogonal columns/rows, respectively.

Implementations

numpy.kron(a, b)
scipy.sparse.kron(a, b)
scipy.linalg.kron(a, b)

DMRG (II)

Implementations

---

Simple DMRG

Simple DMRG Program

The Hamiltonian is given by the XXZ model \[ H = -\frac{1}{2} \sum_{j=1}^N (J \sigma_j^x \sigma_{j+1}^x + J \sigma_j^y \sigma_{j+1}^y + J_z \sigma_{j}^z \sigma_{j+1}^z). \]

Finite System

\(d\) is defined to be the number of states of a single site.

In line 40 and 41 two matrices are defined, \[ S^z_{1} = \frac{1}{2} \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}, \quad S^+_{1} = \begin{pmatrix} 0 & 1 \\ 0 & 0 \end{pmatrix} \] as well as \[ H_1 = \begin{pmatrix} 0 & 0 \\ 0 & 0 \end{pmatrix}. \] \(H_1\) will be used as the single-site Hamiltonian at initialization.

The two-site Hamiltonian is defined in line 45 to be \[ \begin{align*} H_2(S_1, S_2) &= \frac{J}{2} S^+_1 \otimes S^{-}_2 + \frac{J}{2} S^{-}_1 \otimes S^+_2 + J_z S^z_1 \otimes S^z_2 \\ &= J(S_1^x \otimes S_2^x + S_1^y \otimes S_2^y) + J_z S_1^z \otimes S_2^z. \end{align*} \]

Enlarging Block

The enlarging process renews the operaors as \[ \begin{align*} H &\rightarrow H \otimes \mathbb{1}_{d\times d} + H_2(S_{\mathrm{edge}}, S_1), \\ S_{\mathrm{edge}}^z &\rightarrow \mathbb{1}_{M\times M} \otimes S^z_1, \\ S_{\mathrm{edge}}^+ &\rightarrow \mathbb{1}_{M\times M} \otimes S^+_1. \end{align*} \]

Single DMRG Step

At each DMRG step, the following steps are carried out:

1. The system block and the environment block is enlarged once.
2. The superblock Hamiltonian is constructed as \[ H = H_{\mathrm{sys}} \otimes \mathbb{1}_{M_{\mathrm{env}}\times M_{\mathrm{env}}} + \mathbb{1}_{M_{\mathrm{sys}} \times M_{\mathrm{sys}}} \otimes H_{\mathrm{env}} + H_2(S_{\mathrm{sys,edge}}, S_{\mathrm{env,edge}}). \]
3. Calculate the lowest eigenvalue \(E\) and \(\psi\) of \(H\).
4. Calculate the projected density matrix: \[ \rho = \tr_E (\ket \psi \bra \psi). \]
5. Calculate the eigenvectors of the lowest \(m\) eigenvalues. If \(m\) is greater then the dimension of the matrix, we calculate all the eigenvectors.
6. Construct the transformation matrix \(T\), where the columns are given by the eigenvectors in the previous step. \(T\) is not necessarily a square matrix.
7. All the operators of the system, \(S_{\mathrm{sys}}\) and \(H_{\mathrm{sys}}\), should undergo a transformation \[ H \rightarrow T^\dagger H T. \]

We may take system and environment to be the same. This may reduce half of the calculations.

How is the projected density matrix calculated? Although \(\psi\) is a vector in the direct product of two Hilbert spaces, it was represented by a column vector returned from eigsh. The following snippet traces out the environment part:

psi = psi.reshape([m_sys, -1])
rho = np.dot(psi0, psi0.conjugate.transpose())

Hubbard Model on a Specific Sector

MPS

DMRG for MPS

Computational Tricks

Speed-Up

See Also

Simple DMRG

• GitHub.
• Documents.

DMRG101

• GitHub.
• Documents.

TeNPy

• Documents.

NetKet

• GitHub.

Statistical Mechanics (II)

Ensembles