Matrix product operator representations

We show how to construct relevant families of matrix product operators (MPOs) in one and higher dimensions. These form the building blocks for the numerical simulation methods based on matrix product states and projected entangled pair states. In particular, we construct translationally invariant MPOs suitable for time evolution, and show how such descriptions are possible for Hamiltonians with long-range interactions. We show how these tools can be exploited for constructing new algorithms for simulating quantum spin systems.

The study of strongly correlated quantum systems is currently receiving a lot of attention.To a large extent, this is due to the formidable progress that has been made in creating such systems under controlled laboratory conditions such as in optical lattices and ion traps.From the theoretical point of view, major new insights have been obtained into characterizing the nature of the wavefunctions associated to those strongly correlated systems.The concept of matrix product states and its generalizations plays a central role in those new insights, as it provides a sound foundation and justification for the success of numerical renormalization group methods and especially of DMRG [1,2].Those insights have led to the development of new algorithms for simulating quantum spin systems; most notable are the algorithms for simulating time evolution [3,4,5,7] and the ones generalizing DMRG to higher dimensions [6].
In this work, we are concerned with the efficient construction of so-called matrix product operators (MPO), the basic building blocks for those novel algorithms.MPO were introduced in the paper [7,8] and form the operator analogue of matrix product states.We will show how to construct translational invariant MPO's in 1 and 2 dimensions that approximate real or imaginary time evolution; in contrast to the TEBD/DMRG algorithms [4,5], the translational symmetry is not broken in the Trotter step.This generalizes the constructions reported in [16].Second, we construct MPO descriptions for general Hamiltonians with decaying long-range interactions.This is very interesting in the light of simulating quantum spin systems with long-range interactions.
Similar work for constructing MPO representations of Hamiltonians has independently been reported in [9,11].Reference [11] gives a very nice presentation of matrix product operators from the point of view of DMRG, and also contains results on how to write spin chain Hamiltonians using MPO.Reference [9] explores the connection between matrix product operators and Markov processes in depth, and also contains some results on generalizations to higher dimensions.In reference [10], an algorithm is devised to simulate quantum spin chains with long-range interactions in the thermodynamic limit; it also contains similar results as reported here on the approximation of power law decaying interactions by sums of exponentials.

A. Construction
Let us first start with a simple example: suppose we want to simulate real/imaginary time evolution under the Ising Hamiltonian in transverse field where only nearest neighbour interactions are considered; both the one-and two-dimensional case will be considered.As usual, this evolution will be approximated using a Trotter expansion, but we want to do this is such a way that the translational invariance in not broken.Therefore, we split the Hamiltonian in two parts H = H z + H x where H z contains all terms with σ z operators and H x the ones with σ x .Obviously, all terms in H z commute, and therefore O z = exp(ǫH z ) can be calculated exactly.As we will show, O z has a very simple and elegant MPO description, and of course O x has a trivial MPO description as it is a product of strictly local operators.Time evolution can now be described within the formalism of matrix product states by evolving the MPS under the action of the MPO O x O z .
Let us next show how the MPO of O z can be constructed.First, observe that Here we used the notation Z 0 = I, Z 1 = σ z = Z and defined the vectors B i .Let us now consider the translational invariant 1-D case of N spins with periodic boundary conditions In the third step, we made use of the cyclic property of the trace, and in the fourth step, we made a change of variables k 1 = i 1 ⊕ j 1 where binary arithmetic is assumed.We have therefore proven that exp (ǫ i Z i Z i+1 ) has a very efficient matrix product description with the matrices C k given by A big advantage of this precise MPO formulation is that is is symmetric; the spectral properties of the associated transfer operator are hence well behaved, which is important if used in algorithms with periodic boundary conditions [17].
In two dimensions, we can repeat exactly the same argument and obtain the PEPS description of the operator exp Here B i (α) means the α component of the vector B i , x ∈ {0, 1} and the sum is taken over i, j, k, l = 0 : 1 with the condition that i + j + k + l = x in binary arithmetic.This proves that the PEPS description of the operator exp ǫ <ij> Z i Z j has bond dimension 2. Note that no approximations were made and as such this statement is valid for any value of ǫ.In particular, this gives the MPO description for the classical Ising partition function; its free energy can therefore be calculate by contracting the tensor network consisting of tensors C 0 .The previous analysis can trivially be generalized to the case of any Hamiltonian that is a sum of commuting terms: for this class of Hamiltonians, exp(ǫH) has a very simple matrix product operator description.As this holds for any ǫ, it also holds for all thermal states, and by taking ǫ → −∞ it is proven that all ground states of such Hamiltonians have exact MPO descriptions that can easily be constructed.Notable examples of this is the toric code state of Kitaev and the family of string net states [14,15].
From numerical considerations, it is useful if the matrices/tensors occurring in the MPO description are real and symmetric.There are some tricks of how to achieve this.Consider for example the Heisenberg antiferromagnetic Hamiltonian The operator exp (−βH Heis ) can be decomposed in Trotter steps consisting of H x , H y , H z , and every Trotter term involves operators of the form exp(−ǫH x ).As we saw in the previous section, the associated matrices involve terms like sinh(ǫ), which becomes complex when ǫ > 0. What we can do however is a change of basis on every second site (this obviously only works for bipartite lattices), where we rotate the spins with the unitary operator Y = σ y ; this maps On the level of the Hamiltonian, this flips the sign of the H x and H z interactions, for which the associated operators exp(+ǫH x ) have indeed real and symmetric MPO descriptions.The problem seems to remain however with the operator exp(−ǫH y ).This can however easily be cured by defining the real antisymmetric matrix Ỹ = iY for which H ỹ = −H y when we replace all operators Y by Ỹ .Next, exp(+ǫH ỹ) can again be expressed as a MPO; however, we have to be careful as Ỹ .Ỹ = −I as opposed to +I.Looking back at the derivation of the MPO for the Ising case, we can easily see that this sign can be absorbed into C, and we can express as a MPO with matrices Of course the same can be done in two dimensions.Here we get where the sum in the power √ −1 (x+i+j+k+l) is not in binary arithmetic.This clearly leads to a real translational invariant MPO parametrization.

B. Algorithms
It is now obvious how to turn those MPO-descriptions to our advantage for constructing new algorithms for the simulation of quantum spin chains.
Let us first consider the case of imaginary time evolution, where the goal is to evolve a state in imaginary time such as to simulate a thermal (finite β) or ground state (β → ∞).Obviously, we will use the Trotterization described in the previous section.The big there is that the translational invariance is never broken, and furthermore that the matrices involved in the MPS description of the MPO are real and symmetric.In particular, that means that, if we start with a translational invariant MPS with real symmetric MPS description, then it will stay like that during the whole course of the evolution.This has a dramatic effect on the numerical conditioning and stability of the algorithm.
The algorithm for time evolution is now as follows: given a translational invariant MPS with matrices {A i } with bond dimension D and MPO with matrices {B i }, {X i } of dimension D ′ , we want to find a way of representing cutting the bond dimension of the MPS {C i } given by in an optimal way.This can easily be done as follows: calculate the leading eigenvector x of the transfer operator E = i C i ⊗ C i (note that E is symmetric and as such this is a very well conditioned problem).Rewriting x as a DD ′ × DD ′ positive semidefinite matrix, we can easily calculate its singular value decomposition x = U ΣU † .We now define the projector/isometry P as the rectangular matrix consisting of the first D rows of U , and act with this P on the matrices C i .The updates matrices A i are therefore obtained by A i ≡ P † C i P which is obviously still symmetric and real.Clearly, all those steps have to be done in such a way as to exploit the sparse nature of the problem, such as done in DMRG, which leads to a complexity that scales like D 3 .Also, if the eigenvalues that are thrown away are not small enough, we can always increase the bond dimension.
The big advantage of this procedure is that it is extremely well conditioned and very efficient to implement.This allows to go to very large bond dimensions.Notably, as compared to the T EBD algorithms, we do not have to take inverses at any time (because the gauge degrees of freedom are trivial as they consist of unitary matrices), and furthermore it works equally well if the MPO is very far from the identity operator (this is important in the context of PEPS algorithms).
The same ideas can of course be used in the case of real time evolution.In that case, the matrices involved become complex symmetric, and it might be benefitial to apply some gauge conditions to optimize the stability.This can be done as follows: given x, we want to find the complex (symplectic) matrix Q, Q.Q T = 1 1 such that the condition number (i.e.smallest divided by largest singular value) of the matrix QxQ † is as large as possible.This optimization problem can be solved recursively as follows: calculate the singular value decomposition of x = vsv † , choose the generator , and make the substitution x → exp(−iǫG)x exp(iǫG) for small enough ǫ, and repeat this until convergence.Convergence is equivalent to the derivative of the condition number being zero.The final gauge transform to be implemented is the product of all infinitesimal transformations Q = exp(iǫG k ).Note again that all of this becomes trivial in the case of real symmetric matrices (such as occurring in imaginary time evolution): in that case the Q cannot change the condition number as they are unitary.
We have tested those new algorithms on the critical Ising and Heisenberg spin chain models, and obtained results that are consistent with what we expected.In particular, for the Heisenberg antiferromagnetic spin chain, we obtain a precision of (E D=64 − E exact )/E exact = 2.83 * 10 −6 with very modest calculations.In the case of the critical Ising in transverse field, we get (E D=64 − E exact )/E exact = 1, 10 * 10 − 9.
The algorithms for the 2-D analogue will be discussed elsewhere [17].

II. MATRIX PRODUCT OPERATOR DESCRIPTIONS OF HAMILTONIANS WITH LONG-RANGE INTERACTIONS A. Construction
Let us next investigate how to represent Hamiltonians with long-range interactions of the form with f (i − j) some decaying function.The first question to ask is whether it is still possible to find an exact MPO description of Ô = exp(ǫH).It can easily be seen that this is not possible if the function f (x) does not vanish at some finite distance: otherwise, the action of Ô on a MPS could increase the Schmidt number over any cut with an arbitrary large amount, and hence no finite MPO description is possible.This is the reason why the transfer matrix approach in classical 1-D spin systems breaks down for such long-range interactions.So let's be less ambitious and try to find a MPO description of the Hamiltonian itself.This is interesting for several reasons: first, this is useful in constructing algorithms for time evolution using iterative methods like Lanczos, and second, it allows to calculate quantities like ψ|H 2 |ψ efficiently.
As a start, let us consider a general 1-D spin 1/2 Hamiltonian with nearest neighbour interactions.If the Hamiltonian is translational and reflection invariant, then there always exists a basis such that the Hamiltonian can be written as where Ô can be any one-qubit operator.Similarly to the construction of MPS descriptions of the W-state [12], a MPO can be constructed to represent this H by making use of nilpotent matrices: The simplest way of deriving this is to think about the Hamiltonian a Markov process with 5 possible symbols (remember that MPS can be constructed using Markov processes), such that a symbol X 1 , X 2 , X 3 is always followed by itself and then all zeros X 0 , and X 4 by all zeros.As such, one can easily prove that D = 5 is optimal in this case because this is the operator Schmidt number of the Hamiltonian when splitting it into two pieces.Note that if only Ising interactions would have been considered, then D = 3 would have been sufficient and we could have chosen Note that there is no need for B 2 , B 3 , B 4 in that case.
It is obvious how to generalize this description to the case of higher dimensional systems and to the case of exponentially decaying interactions.Let us first look at the case of exponentially decaying interactions.By adding diagonal terms to B 0 we can immediately check that that the corresponding Hamiltonian / MPO is given by which is a spin chain with exponentially decaying interactions.Unfortunately, it is impossible to get exact MPO descriptions when the interactions are decaying following a power law.However, inverse polynomials can pretty well be approximated by sums of exponentials (this is the reason why DMRG is able to reproduce the correlations in critical models pretty well).Hamiltonians with power law decay of correlations can therefore be well approximated by sums of MPO's, which is itself a MPO.Actually, very few exponentials are needed to get a good approximation, even at large distances.The problem of finding the optimal weights and exponents for such an approximation problem for a general function f (k), i.e.
is not completely trivial.In the appendix, we present a simple method that solves this optimization problem for general f (k) and a given number of exponentials n and a number of sites N > n (the method works for any functions, and returns complex exponents in the case of oscillating functions as should be).If we choose power law decay with cube power 3, N=1000 n=10 then the above cost function is 10 −5 (the maximal difference between the function and the approximated one is 5.10 −8 ).This maximal difference falls to 3.10 −6 for power 2 and 3.10 −4 for power 1.
In conclusion, we found the exact MPO description for Hamiltonians with exponentially decaying interactions.Hamiltonians with power law decay can be approximated very well using sums of such MPO.The matrix product operators obtained for the description of Hamiltonians are of a very different form than the ones obtained by taking the exponential.The main difference is that the corresponding transfer matrices will always contain a Jordan block structure, and one has to be careful in dealing with such situations when considering the thermodynamic limit.
Let us now turn to the 2-dimensional case.Let us again first consider the square lattice with only nearest neighbour interactions.There is a very simple way of writing down a PEPS description that achieves the task: first, consider the MPS which is the equal superposition of having one spin up and all other ones down over all sites.Note that this MPS has bond dimension 2, and can therefore trivially be represented as a PEPS with bond dimension 2. The idea is that this excitation specifies where to put an interaction.Let us next consider the tensors where we assume that the indices α, β are the left respectively top indices, and the associated operators It can readily be seen that we get the Ising Hamiltonian if we act with the |W state on the fifth index of the tensors: the |W state puts one index i equal to one, and the other terms are such that an interaction to the right and below it will be created.The total bond dimension of the corresponding PEPS (including the |W ) is therefore 4.
Decaying interactions between one spin and all other ones can however be obtained in a much more elegant way; as we will show, it is even possible to model power law decay of interactions exactly.The idea is as follows: the critical classical Ising model in 2 dimensions has power law decay of correlations.Consider the quantum state

⊗N
where |+ stands for the superposition |0 + |1 .This is obviously a PEPS, as it is obtained by acting with a MPO (see earlier) on a product state.The partition function of the Ising model at temperature β is obtained by calculating the overlap ( +|) ⊗N |ψ(β) , and correlation functions between two spins are obtained by replacing the corresponding +| at the left side of this expression by −| = 0| − 1|.Instead of the |W state in the previous example, we will use a state |W " that is the equal superposition of two excited spins as opposed to one.The MPS description of |W " has bond dimension 3 and is similar to the ones derived for the 1-D Hamiltonians with exponential decay where we put the parameter λ = 1 (i.e.B 0 = I).This gives us all the necessary ingredients to construct the MPO description of the Hamiltonian: Here the vectors |x i act on two qubits, one on the corresponding qubit of |ψ β and the other one on |W " .The |W " state enforces that exactly two operators X i will be nontrivial, and |ψ β gives the right weight to the associated As those are products of PEPS, the result is a PEPS with bond dimension 2 × 3 = 6.This is am amazing result: as opposed to the 1-D case, there is an exact PEPS description for Hamiltonians with 2-body interactions that decay as the r −ν with ν = 1 the critical exponent of the Ising model.
Obviously, this construction can be repeated for any classical spin model, and hence many different exponents can in principle be taken.It is as yet an open question how to engineer the PEPS such as to get a specific exponent, although very good approximations can again be obtained by making use of sums of exponentials.

B. Algorithms
It is obvious how to make use of all this in algorithms for simulating quantum spin chains.First of all, it is clear how to extend the variational MPS method described in [13] to the present case.For this, we have to consider finite systems with open boundary conditions and matrix product states that have site-dependent matrices in their MPS description.The optimization ψ|H|ψ can then be done using the alternating least squares method described in [13].As expected, numerical tests showed very good convergence properties.
The problem of time evolution is a little bit more challenging, as we cannot use the Trotterization tricks.However, Krylov-based methods can of course be used (see [18] for a review), and are the method of choice here.
These finite dimensional algorithms can not readily be generalized to the infinite case however.A sensible way for determining ground state energies in that limit would be to first use the finite dimensional algorithm just described, and then use a brute-force gradient-based optimization method for optimizing the infinite case.For this we need to be able to calculate expectation values ψ| Ô|ψ in the thermodynamic limit, i.e. when |ψ is an infinite MPS with matrices {A i } and Ô the MPO description of a Hamiltonian with exponentially decaying interactions.The idea is to consider a family of MPO O N whose support is limited to N sites (i.e. the Hamiltonian does only act on N sites), calculate the energy with respect to the infinite MPS (this energy will scale linearly in N ), and the take the thermodynamic limit to calculate the energy per site.It holds that The eigenstructure of E H is nontrivial because it has Jordan blocks.In the present case, the only relevant blocks are of size 2 (larger blocks would lead to an energy that scales superlinearly with the size of the support of H, which cannot be).Using the notation used before, one sees that the left/right eigenvector corresponding to the largest eigenvalue in magnitude d 0 is given by q l | = x l | 0| / |q r = |x r |0 .The generalized eigenvectors can now be found by solving the equation (E H − d 0 I) |q r = |q r , ql | (E H − d 0 I) = q l |.We next define the matrices Q l = (|q l , |q l ), Q r = (|q r , |q r ) and the 2 × 2 matrix Q = Q T l Q r −1 .In the limit of large N , it holds that in the limit of large N .The energy per site is therefore given by In a similar vein, it is possible to calculate expectation values of the operator ψ|(H − λ) 2 |ψ .This is relevant because it gives an exact bound on how far a given MPS |ψ is from an exact eigenvector of H.As we have a squared term, Jordan blocks of dimension 3 will be encountered.As before, we define The relevant right eigenvector is again of the form |q r = |x r |0 |0 and we can find the eigenstructure of the associated Jordan block as follows: start by calculating |q r as we did in the previous section using the operator E H instead of E H 2 ; next define |q ′ = sym(|q r |0 ) where the symbol 'sym' means symmetrization with respect to the part of the state acting on the Hamiltonian part of the MPO (the antisymmetrized wavefunction turns out to be an irrelevant eigenvector of E H 2 with eigenvalue d 0 ); finally solve the linear set of equations (E H 2 − d 0 I) |q" r = |q ′ r .The relevant eigenstructure is now given by the matrix Q r = (|q r , |q ′ r , |q" r ).Similarly, one can calculate The final expectation value is then given by The matrix Q therefore contains all the relevant information about the energies and their scaling when N → ∞.
The energy can now easily be optimized with a brute force gradient-base optimization routine.Concerning the 2-dimensional MPO representing Hamiltonians, it turns out that they are very valuable for speeding up actual calculations done by the PEPS method: the calculation of the expectation value of the Hamiltonian with respect to to a given PEPS can be calculated in one run using this idea, and we don't have to calculate the expectation value for every terms individually anymore.

III. CONCLUSION
In conclusion, we constructed several examples of interesting families of matrix product operators in 1 and 2 dimensions.Those descriptions turn out to be very valuable for constructing stable and scalable algorithms for simulating quantum spin systems, in one and two dimensions.

IV. APPENDIX
In this appendix, we show how to solve the problem of approximating any function f (k) as a sum of exponentials for k = 1..N : First, construct the rectangular N − n + 1 × n matrix

  
that W is a Vandermonde matrix.We observe that F and W span the same space (note that N is typically much larger than n), such that there exists a n × n matrix Q s.t.F Q ≃ W . Define F 1 as the rectangular matrix which consists of the first N − n rows of F and F 2 as the one with the last N − n rows.Due to the Vandermonde structure of W , it must be approximately true that F 1 QΛ ≃ F 2 Q with Λ the diagonal matrix containing the exponents.Therefore, denotes the pseudoinverse of U 1 ): the exponents {λ i } hence correspond to the eigenvalues of the matrix F − † 1 F 2 which can be calculated very easily.This method can be made more robust by making use of the so-called QR-decomposition.This can be done by first calculating the (economical) QR decomposition of F = U V and by defining U 1 as the rectangular matrix which consists of the first N − n rows and n columns of U and U 2 as the one with the last N − n rows: there must again exist a Q such that U Q ≃ W .The exponents Λ can therefore easily be calculated as the eigenvalues of the matrix U − † 1 U 2 .The advantage of using the QR-decomposition is that the pseudoinverse of U 1 is much better conditioned than of F 1 .
Once those exponents are found, a simple least squares algorithm can be used to find the corresponding weights {x i }.It happens that this method is very efficient and reliable, even when oscillating functions are involved.A similar method is known in the field of signal processing under the name of Hankel singular value decomposition.

ψ|
ÔN |ψ = L|E N H |R where |L = |x l |v l ,|R = |x r |v r with |x l , |x r the left/right eigenvectors of the transfer matrix E 0 = i A i ⊗ Āi ; the vectors v l , v r are the ones used in the MPO description of the Hamiltonian, and