Matrix product state

For periodic boundary conditions,Penrose graphical notation (tensor diagram notation) of a matrix product state of five particles.

A matrix product state (MPS) is a representation of a quantum many-body state. It is at the core of the one of the most effective algorithms for solving one dimensional strongly correlated quantum systems – the density matrix renormalization group (DMRG) algorithm.

For a system of ${\displaystyle N}$ spins of dimension ${\displaystyle d}$, the general form of the MPS for periodic boundary conditions (PBS) can be written in the following form:

For open boundary conditions, Penrose graphical notation (tensor diagram notation) of a matrix product state of five particles.

$${\displaystyle |\Psi \rangle =\sum _{\{s\}}\operatorname {Tr} \left[A_{1}^{(s_{1})}A_{2}^{(s_{2})}\cdots A_{N}^{(s_{N})}\right]|s_{1}s_{2}\ldots s_{N}\rangle .}$$For open boundary conditions (OBC), ${\displaystyle |\Psi \rangle }$ takes the form

${\displaystyle |\Psi \rangle =\sum _{\{s\}}A_{1}^{(s_{1})}A_{2}^{(s_{2})}\cdots A_{N}^{(s_{N})}|s_{1}s_{2}\ldots s_{N}\rangle .}$

Here ${\displaystyle A_{i}^{(s_{i})}}$ are the ${\displaystyle D_{i}\times D_{i+1}}$ matrices (${\displaystyle D}$ is the dimension of the virtual subsystems) and ${\displaystyle |s_{i}\rangle }$ are the single-site basis states. For periodic boundary conditions, we consider ${\displaystyle D_{N+1}=D_{1}}$, and for open boundary conditions ${\displaystyle D_{1}=1}$. The parameter ${\displaystyle D}$  is related to the entanglement between particles. In particular, if the state is a product state (i.e. not entangled at all), it can be described as a matrix product state with ${\displaystyle D=1}$. ${\displaystyle \{s_{i}\}}$ represents a ${\displaystyle d}$-dimensional local space on site ${\displaystyle i=1,2,...,N}$. For qubits, ${\displaystyle s_{i}\in \{0,1\}}$. For qudits (d-level systems), ${\displaystyle s_{i}\in \{0,1,\ldots ,d-1\}}$.

For states that are translationally symmetric, we can choose: $${\displaystyle A_{1}^{(s)}=A_{2}^{(s)}=\cdots =A_{N}^{(s)}\equiv A^{(s)}.}$$In general, every state can be written in the MPS form (with ${\displaystyle D}$ growing exponentially with the particle number N). Note that the MPS decomposition is not unique. MPS are practical when ${\displaystyle D}$ is small – for example, does not depend on the particle number. Except for a small number of specific cases (some mentioned in the section Examples), such a thing is not possible, though in many cases it serves as a good approximation.

For introductions see,[1][2][3] and.[4] In the context of finite automata see.[5] For emphasis placed on the graphical reasoning of tensor networks, see the introduction.[6]

Wave function as a Matrix Product State

For a system of ${\displaystyle N}$ lattice sites each of which has a ${\displaystyle d}$-dimensional Hilbert space, the completely general state can be written as

${\displaystyle |\Psi \rangle =\sum _{\{s\}}\psi _{s_{1}...s_{N}}|s_{1}\ldots s_{N}\rangle ,}$

where ${\displaystyle \psi _{s_{1}...s_{N}}}$ is a ${\displaystyle d^{N}}$-dimensional tensor. For example, the wave function of the system described by the Heisenberg model is defined by the ${\displaystyle 2^{N}}$ dimensional tensor, whereas for the Hubbard model the rank is ${\displaystyle 4^{N}}$.

The main idea of the MPS approach is to separate physical degrees of freedom of each site, so that the wave function can be rewritten as the product of ${\displaystyle N}$ matrices, where each matrix corresponds to one particular site. The whole procedure includes the series of reshaping and singular value decompositions (SVD).[7][8]

There are three ways to represent wave function as an MPS: left-canonical decomposition, right-canonical decomposition, and mixed-canonical decomposition.[9]

Left-Canonical Decomposition

The decomposition of the ${\displaystyle d^{N}}$-dimensional tensor starts with the separation of the very left index, i.e., the first index ${\displaystyle s_{1}}$, which describes physical degrees of freedom of the first site. It is performed by reshaping ${\displaystyle |\Psi \rangle }$ as follows

${\displaystyle |\Psi \rangle =\sum _{\{s\}}\psi _{s_{1},(s_{2}...s_{N})}|s_{1}\ldots s_{N}\rangle .}$

In this notation, ${\displaystyle s_{1}}$ is treated as a row index, ${\displaystyle (s_{2}\ldots s_{N})}$ as a column index, and the coefficient ${\displaystyle \psi _{s_{1},(s_{2}...s_{N})}}$ is of dimension ${\displaystyle (d\times d^{N-1})}$. The SVD procedure yields

${\displaystyle \psi _{s_{1},(s_{2}...s_{N})}=\sum _{\alpha _{1}}^{r_{1}}U_{s_{1},\alpha _{1}}D_{\alpha _{1},\alpha _{1}}(V^{\dagger })_{\alpha _{1},(s_{2}...s_{N})}=\sum _{\alpha _{1}}^{r_{1}}U_{s_{1},\alpha _{1}}\psi _{\alpha _{1},(s_{2}...s_{N})}=\sum _{\alpha _{1}}^{r_{1}}A_{\alpha _{1}}^{s_{1}}\psi _{\alpha _{1},(s_{2}...s_{N})}.}$

The separation of physical degrees of freedom of the first site.

In the relation above, matrices ${\displaystyle D}$ and ${\displaystyle V^{\dagger }}$ are multiplied and form the matrix ${\displaystyle \psi _{\alpha _{1},(s_{2}...s_{N})}}$ and ${\displaystyle r_{1}\leq d}$. ${\displaystyle A_{\alpha _{1}}^{s_{1}}}$ stores the information about the first lattice site. It was obtained by decomposing matrix ${\displaystyle U}$ into ${\displaystyle d}$ row vectors ${\displaystyle A^{s_{1}}}$ with entries ${\displaystyle A_{\alpha _{1}}^{s_{1}}=U_{s_{1},\alpha _{1}}}$. So, the state vector takes the form

${\displaystyle |\Psi \rangle =\sum _{\{s\}}\sum _{\alpha _{1}}A_{\alpha _{1}}^{s_{1}}\psi _{\alpha _{1},(s_{2}...s_{N})}|s_{1}\ldots s_{N}\rangle .}$

The separation of the second site is performed by grouping ${\displaystyle s_{2}}$ and ${\displaystyle \alpha _{1}}$, and representing ${\displaystyle \psi _{\alpha _{1},(s_{2}...s_{N})}}$ as a matrix ${\displaystyle \psi _{(\alpha _{1}s_{2}),(s_{3}...s_{N})}}$ of dimension ${\displaystyle (r_{1}d\times d^{N-2})}$. The subsequent SVD of ${\displaystyle \psi _{(\alpha _{1}s_{2}),(s_{3}...s_{N})}}$ can be performed as follows:

${\displaystyle \psi _{(\alpha _{1}s_{2}),(s_{3}...s_{N})}=\sum _{\alpha _{2}}^{r_{2}}U_{(\alpha _{1}s_{2}),\alpha _{2}}D_{\alpha _{2},\alpha _{2}}(V^{\dagger })_{\alpha _{2},(s_{3}...s_{N})}=\sum _{\alpha _{2}}^{r_{2}}A_{\alpha _{1},\alpha _{2}}^{s_{2}}\psi _{\alpha _{2},(s_{3}...s_{N})}}$.

The separation of physical degrees of freedom for the first two sites.

Above we replace ${\displaystyle U}$ by a set of ${\displaystyle d}$ matrices of dimension ${\displaystyle (r_{1}\times r_{2})}$ with entries ${\displaystyle A_{\alpha _{1},\alpha _{2}}^{s_{2}}=U_{(\alpha _{1}s_{2}),\alpha _{2}}}$. The dimension of ${\displaystyle \psi _{\alpha _{2},(s_{3}...s_{N})}}$ is ${\displaystyle (r_{2}\times d^{N-2})}$ with ${\displaystyle r_{2}\leq r_{1}d\leq d^{2}}$. Hence,

${\displaystyle |\Psi \rangle =\sum _{\{s\}}\sum _{\alpha _{1}}A_{\alpha _{1}}^{s_{1}}\psi _{(\alpha _{1}s_{2}),(s_{3}...s_{N})}|s_{1}\ldots s_{N}\rangle =\sum _{\{s\}}\sum _{\alpha _{1},\alpha _{2}}A_{\alpha _{1}}^{s_{1}}A_{\alpha _{1},\alpha _{2}}^{s_{2}}\psi _{\alpha _{2},(s_{3}...s_{N})}|s_{1}\ldots s_{N}\rangle .}$

Following the steps described above, the state ${\displaystyle |\Psi \rangle }$ can be represented as a product of matrices

${\displaystyle |\Psi \rangle =\sum _{\{s\}}\sum _{\alpha _{1},\ldots ,\alpha _{N-1}}A_{\alpha _{1}}^{s_{1}}A_{\alpha _{1},\alpha _{2}}^{s_{2}}\ldots A_{\alpha _{N-2},\alpha _{N-1}}^{s_{N-1}}A_{\alpha _{N-1}}^{s_{N}}|s_{1}\ldots s_{N}\rangle .}$

The maximal dimensions of the ${\displaystyle A}$-matrices take place in the case of the exact decomposition, i.e., assuming for simplicity that ${\displaystyle N}$ is even, ${\displaystyle (1\times d),(d\times d^{2}),\ldots ,(d^{N/2-1}\times d^{N/2}),(d^{N/2}\times d^{N/2-1}),\ldots ,(d^{2}\times d),(d\times 1)}$ going from the first to the last site. However, due to the exponential growth of the matrix dimensions in most of the cases it is impossible to perform the exact decomposition.

The dual MPS is defined by replacing each matrix ${\displaystyle A}$ with ${\displaystyle A^{*}}$:

${\displaystyle \langle \Psi |=\sum \limits _{\{s\}}\sum \limits _{\alpha '_{1},...,\alpha '_{N-1}}A_{\alpha '_{1}}^{*s'_{1}}A_{\alpha '_{1},\alpha '_{2}}^{*s'_{2}}...A_{\alpha '_{N-2},\alpha '_{N-1}}^{*s'_{N-1}}A_{\alpha '_{N-1}}^{*s'_{N}}\langle s'_{1}...s'_{N}|.}$

Note that each matrix ${\displaystyle U}$ in the SVD is a semi-unitary matrix with property ${\displaystyle U^{\dagger }U=I}$. This leads to

${\displaystyle \delta _{\alpha _{i},\alpha _{j}}=\sum _{\alpha _{i-1}s_{i}}(U^{\dagger })_{\alpha _{i},(\alpha _{i-1}s_{i})}U_{(\alpha _{i-1}s_{i}),\alpha _{j}}=\sum _{\alpha _{i-1}s_{i}}(A^{s_{i}\dagger })_{\alpha _{i},\alpha _{i-1}}A_{\alpha _{i-1},\alpha _{j}}^{s_{i}}=\sum _{s_{i}}(A^{s_{i}\dagger }A^{s_{i}})_{\alpha _{i},\alpha _{j}}}$.

To be more precise, ${\displaystyle \sum _{s_{i}}A^{s_{i}\dagger }A^{s_{i}}=I}$. Since matrices are left-normalized, we call the composition left-canonical.

Right-Canonical Decomposition

Similarly, the decomposition can be started from the very right site. After the separation of the first index, the tensor ${\displaystyle \psi _{s_{1}...s_{N}}}$ transforms as follows:

${\displaystyle \psi _{s_{1}...s_{N}}=\psi _{(s_{1}...s_{N-1}),s_{N}}=\sum _{\alpha _{N-1}}U_{(s_{1}...s_{N-1}),\alpha _{N-1}}D_{\alpha _{N-1},\alpha _{N-1}}(V^{\dagger })_{\alpha _{N-1},s_{N}}=\sum _{\alpha _{N-1}}\psi _{(s_{1}...s_{N-1}),\alpha _{N-1}}B_{\alpha _{N-1}}^{s_{N}}}$.

The matrix ${\displaystyle \psi _{(s_{1}...s_{N-1}),\alpha _{N-1}}}$ was obtained by multiplying matrices ${\displaystyle U}$ and ${\displaystyle D}$, and the reshaping of ${\displaystyle (V^{\dagger })_{\alpha _{N-1},s_{N}}}$ into ${\displaystyle d}$ column vectors forms ${\displaystyle B_{\alpha _{N-1}}^{s_{N}}}$. Performing the series of reshaping and SVD, the state vector takes the form

${\displaystyle |\Psi \rangle =\sum _{\{s\}}\sum _{\alpha _{1},\ldots ,\alpha _{N-1}}B_{\alpha _{1}}^{s_{1}}B_{\alpha _{1},\alpha _{2}}^{s_{2}}\ldots B_{\alpha _{N-2},\alpha _{N-1}}^{s_{N-1}}B_{\alpha _{N-1}}^{s_{N}}|s_{1}\ldots s_{N}\rangle .}$

Since each matrix ${\displaystyle V}$ in the SVD is a semi-unitary matrix with property ${\displaystyle V^{\dagger }V=I}$, the ${\displaystyle B}$-matrices are right-normalized and obey ${\displaystyle \sum _{s_{i}}B^{s_{i}}B^{s_{i}\dagger }=I}$. Hence, the decomposition is called right-canonical.

Mixed-Canonical Decomposition

The decomposition performs from both the right and from the left. Assuming that the left-canonical decomposition was performed for the first n sites, ${\displaystyle \psi _{s_{1}...s_{N}}}$ can be rewritten as

${\displaystyle \psi _{s_{1}...s_{N}}=\sum _{\alpha _{1},\ldots ,\alpha _{n}}A_{\alpha _{1}}^{s_{1}}A_{\alpha _{1},\alpha _{2}}^{s_{2}}\ldots A_{\alpha _{n-1},\alpha _{n}}^{s_{n}}D_{\alpha _{n},\alpha _{n}}(V^{\dagger })_{\alpha _{n},(s_{n+1}...s_{N})}}$.

MPS representation obtained by the mixed-canonical decomposition.

In the next step, we reshape ${\displaystyle (V^{\dagger })_{\alpha _{n},(s_{n+1}...s_{N})}}$ as ${\displaystyle \psi _{(\alpha _{n}s_{n+1}...s_{n-1}),s_{N}}}$ and proceed with the series of reshaping and SVD from the right up to site ${\displaystyle s_{n+1}}$:

${\displaystyle {\begin{aligned}\psi _{(\alpha _{n}s_{n+1}...s_{n-1}),s_{N}}&=&\sum _{\alpha _{n+1}...\alpha _{N}}U_{(\alpha _{n}s_{n+1}),\alpha _{n+1}}D_{\alpha _{n+1},\alpha _{n+1}}B_{\alpha _{n+1},\alpha _{n+2}}^{s_{n+2}}\ldots B_{\alpha _{N-2},\alpha _{N-1}}^{s_{N-1}}B_{\alpha _{N-1}}^{s_{N}}\\&=&\sum _{\alpha _{n+1}...\alpha _{N}}B_{\alpha _{n},\alpha _{n+1}}^{s_{n+1}}B_{\alpha _{n+1},\alpha _{n+2}}^{s_{n+2}}\ldots B_{\alpha _{N-2},\alpha _{N-1}}^{s_{N-1}}B_{\alpha _{N-1}}^{s_{N}}\end{aligned}}}$.

As the result,

${\displaystyle \psi _{s_{1}...s_{N}}=\sum _{\alpha _{1},\ldots ,\alpha _{N}}A_{\alpha _{1}}^{s_{1}}A_{\alpha _{1},\alpha _{2}}^{s_{2}}\ldots A_{\alpha _{n-1},\alpha _{n}}^{s_{n}}D_{\alpha _{n},\alpha _{n}}B_{\alpha _{n},\alpha _{n+1}}^{s_{n+1}}B_{\alpha _{n+1},\alpha _{n+2}}^{s_{n+2}}\ldots B_{\alpha _{N-2},\alpha _{N-1}}^{s_{N-1}}B_{\alpha _{N-1}}^{s_{N}}}$.

Examples

Greenberger–Horne–Zeilinger state

Greenberger–Horne–Zeilinger state, which for N particles can be written as superposition of N zeros and N ones

${\displaystyle |\mathrm {GHZ} \rangle ={\frac {|0\rangle ^{\otimes N}+|1\rangle ^{\otimes N}}{\sqrt {2}}}}$

can be expressed as a Matrix Product State, up to normalization, with

${\displaystyle A^{(0)}={\begin{bmatrix}1&0\\0&0\end{bmatrix}}\quad A^{(1)}={\begin{bmatrix}0&0\\0&1\end{bmatrix}},}$

or equivalently, using notation from:[10]

${\displaystyle A={\begin{bmatrix}|0\rangle &0\\0&|1\rangle \end{bmatrix}}.}$

This notation uses matrices with entries being state vectors (instead of complex numbers), and when multiplying matrices using tensor product for its entries (instead of product of two complex numbers). Such matrix is constructed as

${\displaystyle A\equiv |0\rangle A^{(0)}+|1\rangle A^{(1)}+\ldots +|d-1\rangle A^{(d-1)}.}$

Note that tensor product is not commutative.

In this particular example, a product of two A matrices is:

${\displaystyle AA={\begin{bmatrix}|00\rangle &0\\0&|11\rangle \end{bmatrix}}.}$

W state

W state, i.e., the superposition of all the computational basis states of Hamming weight one.

${\displaystyle |\mathrm {W} \rangle ={\frac {1}{\sqrt {3}}}(|001\rangle +|010\rangle +|100\rangle )}$

Even though the state is permutation-symmetric, its simplest MPS representation is not.[1] For example:

${\displaystyle A_{1}={\begin{bmatrix}|0\rangle &0\\|0\rangle &|1\rangle \end{bmatrix}}\quad A_{2}={\begin{bmatrix}|0\rangle &|1\rangle \\0&|0\rangle \end{bmatrix}}\quad A_{3}={\begin{bmatrix}|1\rangle &0\\0&|0\rangle \end{bmatrix}}.}$

AKLT model

Main article: AKLT model

The AKLT ground state wavefunction, which is the historical example of MPS approach,[11] corresponds to the choice[9]

${\displaystyle A^{+}={\sqrt {\frac {2}{3}}}\ \sigma ^{+}={\begin{bmatrix}0&{\sqrt {2/3}}\\0&0\end{bmatrix}}}$

${\displaystyle A^{0}={\frac {-1}{\sqrt {3}}}\ \sigma ^{z}={\begin{bmatrix}-1/{\sqrt {3}}&0\\0&1/{\sqrt {3}}\end{bmatrix}}}$

${\displaystyle A^{-}=-{\sqrt {\frac {2}{3}}}\ \sigma ^{-}={\begin{bmatrix}0&0\\-{\sqrt {2/3}}&0\end{bmatrix}}}$

where the ${\displaystyle \sigma {\text{'s}}}$ are Pauli matrices, or

${\displaystyle A={\frac {1}{\sqrt {3}}}{\begin{bmatrix}-|0\rangle &{\sqrt {2}}|+\rangle \\-{\sqrt {2}}|-\rangle &|0\rangle \end{bmatrix}}.}$

Majumdar–Ghosh model

Main article: Majumdar–Ghosh model

Majumdar–Ghosh ground state can be written as MPS with

${\displaystyle A={\begin{bmatrix}0&\left|\uparrow \right\rangle &\left|\downarrow \right\rangle \\{\frac {-1}{\sqrt {2}}}\left|\downarrow \right\rangle &0&0\\{\frac {1}{\sqrt {2}}}\left|\uparrow \right\rangle &0&0\end{bmatrix}}.}$

See also

• Density matrix renormalization group
• Variational method (quantum mechanics)
• Renormalization
• Markov chain
• Tensor network

References

1. Perez-Garcia, D.; Verstraete, F.; Wolf, M.M. (2008). "Matrix product state representations". Quantum Inf. Comput. 7: 401. arXiv:quant-ph/0608197.
2. Orús, Román (2014). "A practical introduction to tensor networks: Matrix product states and projected entangled pair states". Annals of Physics. 349: 117-158. arXiv:1306.2164. Bibcode:2014AnPhy.349..117O. doi:10.1016/j.aop.2014.06.013.
3. Verstraete, F; Murg, V.; Cirac, J.I. (2008). "Matrix product states, projected entangled pair states, and variational renormalization group methods for quantum spin systems". Advances in Physics. 57 (2): 143–224. arXiv:0907.2796. Bibcode:2008AdPhy..57..143V. doi:10.1080/14789940801912366. S2CID 17208624.
4. Bridgeman, Jacob C; Chubb, Christopher T (2017-06-02). "Hand-waving and interpretive dance: an introductory course on tensor networks". Journal of Physics A: Mathematical and Theoretical. 50 (22): 223001. arXiv:1603.03039. Bibcode:2017JPhA...50v3001B. doi:10.1088/1751-8121/aa6dc3. ISSN 1751-8113.
5. Crosswhite, Gregory M.; Bacon, Dave (2008-07-29). "Finite automata for caching in matrix product algorithms". Physical Review A. 78 (1): 012356. arXiv:0708.1221. Bibcode:2008PhRvA..78a2356C. doi:10.1103/PhysRevA.78.012356. ISSN 1050-2947.
6. Biamonte, Jacob; Bergholm, Ville (2017). "Tensor Networks in a Nutshell". arXiv:1708.00006 [quant-ph].
7. Baker, Thomas E.; Desrosiers, Samuel; Tremblay, Maxime; Thompson, Martin P. (2021). "Méthodes de calcul avec réseaux de tenseurs en physique". Canadian Journal of Physics. 99 (4): 207–221. arXiv:1911.11566. Bibcode:2021CaJPh..99..207B. doi:10.1139/cjp-2019-0611. ISSN 0008-4204.
8. Baker, Thomas E.; Thompson, Martin P. (2021-09-07), Build your own tensor network library: DMRjulia I. Basic library for the density matrix renormalization group, arXiv:2109.03120, retrieved 2024-11-03
9. Schollwöck, Ulrich (2011). "The density-matrix renormalization group in the age of matrix product states". Annals of Physics. 326 (1): 96–192. arXiv:1008.3477. Bibcode:2011AnPhy.326...96S. doi:10.1016/j.aop.2010.09.012. S2CID 118735367.
10. Crosswhite, Gregory; Bacon, Dave (2008). "Finite automata for caching in matrix product algorithms". Physical Review A. 78 (1): 012356. arXiv:0708.1221. Bibcode:2008PhRvA..78a2356C. doi:10.1103/PhysRevA.78.012356. S2CID 4879564.
11. Affleck, Ian; Kennedy, Tom; Lieb, Elliott H.; Tasaki, Hal (1987). "Rigorous results on valence-bond ground states in antiferromagnets". Physical Review Letters. 59 (7): 799–802. Bibcode:1987PhRvL..59..799A. doi:10.1103/PhysRevLett.59.799. PMID 10035874.

External links

• Open-source review article focused on tensor network algorithms, applications, and software
• State of Matrix Product States – Physics Stack Exchange
• A Practical Introduction to Tensor Networks: Matrix Product States and Projected Entangled Pair States
• Hand-waving and Interpretive Dance: An Introductory Course on Tensor Networks
• Tensor Networks in a Nutshell: An Introduction to Tensor Networks