WorksPhysics and Astronomy
An Accurate Lanczos Method for the Matrix Product State Representation
Abstract
We improve the convergence of the Lanczos algorithm using the matrix product state representation. As an alternative to the density matrix renormalization group (DMRG), the Lanczos algorithm avoids local minima and can directly find multiple low-lying eigenstates. However, its performance and accuracy are affected by the truncation required to maintain the efficiency of the tensor network representation. In this work, we propose the modified thick-block Lanczos method to enhance the convergence of the Lanczos algorithm with MPS representation. We benchmark our method on one-dimensional instances of the Fermi-Hubbard model and the Heisenberg model in an external field, using numerical experiments targeting the first five lowest eigenstates. Across these tests, our approach attains the best possible accuracy permitted by the given bond dimension. This work establishes the Lanczos method as a reliable and accurate framework for finding multiple low-lying states within a tensor-network representation
Introduction
Eigenvalue problems lie at the heart of quantum many-body physics, materials science, and quantum chemistry. By finding a few low-lying eigenstates of the Hamiltonian, one can access ground-state properties [1, 2, 3, 4, 5], low-temperature thermodynamics [5, 6, 7, 8], and dynamical correlation functions [9, 10]. The Lanczos algorithm [1, 11, 12, 13, 14, 15, 16, 17] is widely used to find the low-lying eigenstates of large linear systems, and is implemented in popular packages like ARPACK [18]. Nevertheless, its capability is limited by the curse of dimensionality for many-body quantum systems, as the Hilbert space dimension grows exponentially with the system size.
Tensor network methods [19, 20, 21, 22, 23, 24, 25] representing quantum states and operators by matrix product state (MPS) and matrix product operators (MPO) significantly alleviate this issue. As a key insight, the low-lying states only stay in a small corner of the whole Hilbert space due to the area law of the entanglement entropy, allowing them to be efficiently approximated by low-rank decompositions like MPS [26]. The density-matrix renormalization-group (DMRG) algorithm exploits these properties and optimizes an MPS ansatz variationally. However, DMRG is inherently a local energy minimization method and can become trapped in local minima [27, 28, 29]. When targeting excited states, it may suffer from root flipping [30, 31, 32], where an unintended excited state is tracked instead of the desired one, especially in the presence of (near-)degeneracies. Moreover, its accuracy depends on the quality of previously obtained lower-energy states, which leads to cumulative errors [33, 34]. The Lanczos algorithm, by contrast, naturally produces a spectrum of
eigenpairs and inherently avoids local minima when initialized with the proper initial state. Moreover, through the construction of an orthogonal Krylov subspace, it can inherently resolve degenerate or nearly degenerate low-energy eigenstates when equipped with block initial vectors [35, 36]. Following the pioneering attempts [37, 38], the combination of the Lanczos algorithm and tensor network offers a promising complementary approach for exploring the low-lying eigenstates of large-scale quantum many-body systems [29, 39, 40, 41].
Unfortunately, using tensor network representations for the Lanczos algorithm introduces a new bottleneck: generating new Krylov vectors enlarges the bond dimensions, so MPS compression is required to keep the calculation feasible and efficient. The MPS truncation introduces non-negligible errors, which result in two significant issues. First, this error causes a loss of orthogonality among the Krylov vectors. A post-reorthogonalization technique [37] has been developed to address this issue. Second, because each Krylov vector depends on its predecessors, the truncation error accumulates as the subspace is generated. This accumulation causes the Krylov subspace to diverge from the ideal one and prevents it from introducing new independent information. As discussed later in § 2 C, the resulting convergence stalls at a much higher error level than expected.
In this work, we focus on improving the convergence of MPS-based plain Lanczos methods (PLM) previously proposed in Refs. [37, 38], especially in the challenging case of targeting multiple eigenstates simultaneously. We first analyze why PLM fails to converge accurately for multiple target eigenstates, and then identify why the widely used thick-restart Lanczos method (TRL) [18, 36, 42] does not resolve this issue when combined with the MPS representation: under MPS truncation, the residual directions associated with different Ritz vectors are no longer parallel. Consequently, a single restart vector is insufficient to provide the necessary correction information for all target Ritz pairs. To address this issue, we introduce a modified thick-block Lanczos (MTBL) scheme that restarts with a block formed by Ritz vectors and their associated residuals, which restores independent correction directions under truncation. This construction substantially improves the accuracy of Lanczos-type methods in the MPS representation while enabling to target several eigenstates simultaneously, thereby overcoming the main limitation that has hindered their broader use. Furthermore, by incorporating the shift-and-invert spectral transformation [43, 44, 45, 46, 47], we can directly target excited states without first computing the lower-lying eigenstates, i.e., it does not suffer cumulative errors from the lower eigenstates. It can also be employed to refine approximate eigenstates obtained from other methods.
We benchmark our algorithms on the half-filled Fermi–Hubbard chain and the spin-1/2 Heisenberg XXZ model in a longitudinal magnetic field. For both cases, MTBL significantly improves accuracy by three to seven orders of magnitude while remaining economical, compared to the previous Lanczos method within the MPS formalism. Notably, across these numerical experiments, the MTBL method reaches the accuracy limit imposed by the chosen bond dimension, achieving the best possible approximation of the target eigenstates allowed. This demonstrates that MTBL raises the accuracy of Lanczos-type methods to the level attainable by mainstream high-accuracy MPS-based approaches such as DMRG. To examine its scalability, we also apply our approach to a larger Fermi-Hubbard model of size (100 spin sites) and Heisenberg XXZ chain of size .
Theoretical Background
In this section, we review the classical Lanczos algorithm and the MPS/MPO formalisms, and then examine their integration. We also discuss the challenges posed by the MPS truncation, which motivates our subsequent improvement.
The Lanczos algorithm
The Lanczos algorithm is a widely used iterative method for computing the extremal eigenpairs of Hermitian matrices [1, 11, 12, 48], and has been widely used in diagonalizing large Hamiltonians within quantum many-body physics [49, 50]. Starting with an initial state , the Lanczos algorithm confines the evolving space to the Krylov subspace spanned by . The Lanczos procedure recursively constructs the Krylov vectors via a three-term recurrence relation:
for . The new Krylov vector is formed by applying to , orthogonalizing it against the two previous Krylov vectors with the Gram-Schmidt process [51] in Eq. (1b), and finally normalizing the result by Eq. (1c). For , one only needs to orthogonalize the vector against .
This iteration builds an orthonormal basis of the Krylov subspace . Projecting the original Hamiltonian onto the subspace yields a reduced (effective) Hamiltonian, whose elements are given by:
Diagonalizing this matrix yields the approximate eigenvalues together with the coefficient vectors . The approximation to the exact eigenvectors of the original Hamiltonian is reconstructed by back-transformation:
The quantities and are called the Ritz value and Ritz vector, respectively. After the set of Ritz pairs is obtained, each pair is tested for convergence; those that meet a desired tolerance of convergence are accepted as the numerical eigenpairs of [13, 45].
For many physically or chemically relevant Hamiltonians, only a few dozen or hundreds of basis vectors () already yield accurate low-lying eigenstates [37, 48, 52], and only these Krylov vectors need to be stored. The Lanczos algorithm turns a huge eigenvalue problem into a much smaller one inside the Krylov subspace, making it attractive for computational physics.
Matrix product states and operators
Even though the Lanczos algorithm significantly reduces the complexity of the eigenvalue problem, it still runs into the curse of dimensionality issue since the memory needed to store the vectors (wavefunctions) explicitly scales exponentially with the system size. Therefore, it is beneficial to employ the tensor network representation [19, 20, 53, 54], especially the matrix product state (MPS) ansatz, a widely used low-rank approximation of the wavefunction. Within the MPS ansatz, the wavefunction is rewritten as:
Each inner tensor () is of order three, as illustrated in Figure 1a. The superscript denotes the physical index labeling the possible states at site , and for each , the component is a matrix of size . The parameter represents the bond dimension between sites and ; in the following, we denote the maximum MPS bond dimension by . The MPS representation offers a convenient ansatz for performing the low-rank approximation of the wavefunction. Controlled truncation of the MPS bonds via singular-value decomposition (SVD) can be applied directly to an existing MPS, a process justified by the Schmidt decomposition [19, 55], as depicted in Figure 2. By employing such a truncation, the MPS ansatz can represent the quantum states more economically.

Figure 1. Diagrammatic representation of matrix product states and operators.

Figure 2. The classical SVD method to compress the MPS. First, the MPS is brought into right-canonical form via successive QR decompositions. Then, sweeping from left to right, we carry out SVDs at each site and truncate the smallest singular values. Finally, the and matrices are merged into the next site.
Alongside MPS, an analogous framework exists for operators: the matrix-product operator (MPO) formalism [19, 24, 25]:
For each site , the local tensor carries two physical indices and can be viewed as a matrix, where is the MPO bond dimension between sites and , as illustrated in Figure 1b. We denote the largest of these bond dimensions by . In Eq. (7), the operator is expressed as a summation of mappings from to , and the coefficients arise from contracting the chain of matrices. The MPO representation can fully utilize the sparsity of the physical or chemical Hamiltonian by constructing it properly [24, 25, 56, 57]. For example, the Heisenberg XXZ model can be represented exactly as an MPO of small bond dimension [55], regardless of the system size.
As with a Hamiltonian operator acting on a state vector to produce another vector, applying an MPO to an MPS yields a new MPS by contracting the physical indices at each site [19]. Likewise, adding two MPS produces a new MPS; both operations lead to boosted bond dimensions. To keep subsequent calculations feasible and efficient, especially in iterative algorithms, it is crucial to compress the resulting MPS, as illustrated in Figure 2. There are also alternative MPS compression methods, including the zip-up method [58], the density matrix method [59], and the variational method [19], each offering different trade-offs in accuracy, efficiency, and applicability. In this work, we employ the standard SVD method, which is the most controllable and accurate.
The Lanczos algorithm with MPS
Since the Lanczos algorithm is inherently “matrix free” and the MPS and MPO forms offer an economic way to represent Hamiltonians and states, it is reasonable to combine these tensor networks with the Lanczos algorithm [37].
As discussed before, the core step in the Lanczos algorithm is to compute new Krylov vectors via Eq. (1c). In the tensor network representation, the term is realized by MPO-MPS multiplication, and the orthogonalization is implemented by MPS-MPS addition, leading to an enlarged MPS bond dimension. As discussed, the MPS compression is needed to reduce the bond dimension back to the capped one, introducing extra errors.
The first consequence of this MPS truncation error is the loss of orthogonality. Because Krylov vectors are computed iteratively from preceding vectors, round-off errors accumulate and break global orthogonality. Repeatedly applying the Gram-Schmidt procedure does not solve this issue, as new errors arise during the MPS-MPS addition and subsequent compression steps. Ultimately, this is addressed by the re-orthogonalization technique proposed in [37], which constructs a suitable linear combination of the current Krylov vectors so that the resulting vectors are well-orthogonalized:
Note that we do not create new MPS for these orthogonal states ; instead, we leave the original MPS Krylov vectors unchanged, only solve the matrix (by a Gram-Schmidt-like procedure or canonical orthogonalization method [60]) and utilize in the following calculations. When calculating the effective Hamiltonian for the Lanczos algorithm, the matrix elements are calculated by:
The Ritz pairs are obtained by solving the effective Hamiltonian , e.g., finding a real diagonal matrix and a unitary matrix so that . The actual Ritz vector is obtained by:
where is the size of the current Krylov subspace. Any desired quantity, especially the expectation value for an operator , can be computed from the summation of the terms .
The second consequence of the MPS truncation error is the divergence from the desired Krylov space. Typically, only a small MPS bond dimension is needed to represent low-lying eigenstates, since these low-lying eigenstates obey an area law for entanglement. However, a larger bond dimension is required to represent if is an arbitrary state, even when itself can be approximated by an MPS with a small bond dimension [19, 51]. The reason is that the Hamiltonian is typically a sum of operators, and applying it to can generate a more complex entanglement pattern, as each term introduces additional correlations that increase the overall entanglement in the resulting state. Therefore, the MPS truncation of generates a large error, and such an error prevents the Krylov vectors from “exploring” the Hilbert space correctly; the resulting Krylov space deviates further from the intended one, and newly generated Krylov vectors no longer provide additional independent information in subsequent iterations. Consequently, as displayed later in Sec. IV, the most significant result of this error is the severely broken convergence observed for most Hamiltonians.
In this work, we focus on strategies to mitigate the impact arising from the second consequence of MPS truncation errors, i.e., improving the convergence of the Lanczos algorithms.
MPS-BASED LANCZOS METHOD WITH BLOCK-LANCZOS CONCEPTS
In this section, we first review the standard thick-restart approach [36, 42] and then illustrate that, when integrated with the block-Lanczos framework [35, 62] [61], the resulting modified Thick-Block Lanczos (MTBL) method markedly enhances the convergence of eigenstate calculations. For comparison, the original Lanczos procedure without any restart is referred to as the plain Lanczos method (PLM). We also introduce the shift-and-invert method, which reshapes the Hamiltonian spectrum, enabling the Lanczos algorithm to directly target the desired excited states. Additionally, we briefly discuss the use of energy variance as an indicator of accuracy for the approximate eigenstates.
Thick-restart Lanczos Method
Thick–restart Lanczos (TRL) method is one of the most widely used strategies to improve convergence and reduce memory usage in the Lanczos algorithm. In practice, it underpins mainstream software such as ARPACK [18] and numerous large-scale applications [36, 42]. In the TRL scheme, the iteration is interrupted after a -step Lanczos process, and a thick set of Ritz vectors, which are approximate eigenpairs obtained from the Lanczos procedure, is retained together with the next Krylov vector . In the exact-arithmetic Lanczos method, the vector is the residual vector for all the Ritz vectors, due to the three-term recurrence relation [45, 63]. Thus, such a single vector carries all the correction information for all the Ritz pairs. Concretely, the TRL exploits this property by restarting with:
and continue the Lanczos iteration from ; we present the pseudo-code for classical TRL in Appendix A. This construction preserves the solved spectral information, i.e., the kept Ritz pairs, while injecting exactly the direction that simultaneously reduces the residuals of all kept Ritz pairs. It can be shown that one cycle of TRL can simultaneously refine all targeted Ritz pairs [64].
However, with MPS representations, the ideal Lanczos iterations are perturbed by the error induced by the MPS truncation. Consequently, the three-term recurrence relation does not hold anymore, and different Ritz pairs admit different residual vectors. In such a case, the next Krylov vector cannot carry the correction information for all the Ritz vectors, and TRL thus fails, as presented in Sec. IV A.
Modified Thick-Block Lanczos Method
Having identified the reason for the inaccuracy of the thick-restart scheme under the MPS representation, we are motivated to adopt the following strategy: since a single vector fails to capture the correction direction for all Ritz vectors, we compute the residual:
corresponding to each Ritz vector. However, since the standard TRL can only employ a single correction direction, it cannot efficiently exploit all the residuals . To address this limitation, we adopt the block Lanczos method [35, 62] [61], which allows multiple vectors to serve as the initial vectors.
Algorithm 1 Modified Thick-Block Lanczos (MTBL) method
Input: Hamiltonian as MPO, initial state as MPS
Output: eigenpairs as MPS
Note: MPS is always compressed back to a given bond dimension and normalized after each operation
Initialize , , initial state
Run PLM with MPS representation with initial state . Retain the lowest Ritz pairs
for , until convergence
set Ritz vectors that are not in
for : , normalize
end for
Set
Set , orthogonalize vectors in
for until a pre-set number or convergence
Orthogonalize to and
Append to
Solve and update Ritz pairs with
end for
for
if and then
set ,
end if
end for
if then
break
end if
end for
return
The original block Lanczos method begins with a block of orthonormal vectors instead of a single vector , and the Krylov basis is thus spanned by . The block Lanczos method is also used in DMRG based on the bundled-MPS framework [65], whose efficiency has been discussed in Ref. [66] by T.E Baker and N. Seif. A related block-MPS strategy is employed in the multi-target update method (MTU) [67], which likewise optimizes a block of MPSs simultaneously. The block Lanczos method is also applied in quantum computing [68]. In this work, we incorporate the block Lanczos with TRL when restarting the Lanczos procedure. Concretely, we retain two blocks of vectors, the first one contains the Ritz vectors:
and the second one contains the residual vectors for these Ritz vectors:
which are estimated by Eq. (12). The Lanczos procedure is restarted by repeating the following process:
Generating a new block by applying to , and normalize the vectors in .
Orthogonalize vectors in to all the previous vectors by the Gram-Schmidt procedure.
Finally, we solve the Ritz pairs using the re-orthogonalization discussed in § 2 C and restart the Lanczos procedure until convergence or the desired accuracy is achieved. After each restart, any Ritz vector whose residual norm falls below the selected tolerance is locked, meaning that the corresponding Ritz vector is regarded as converged and is excluded from subsequent restarts. Nevertheless, all newly generated Krylov vectors are orthogonalized against these locked vectors to ensure that no redundant components are regenerated, as detailed in Algorithm 1.
In summary, even though both TRL and block Lanczos are classical Lanczos methods, our contribution is to adapt and combine these ideas in a way that specifically addresses the unparallelism induced by the MPS truncation. In particular, the classical TRL, which can improve the convergence of PLM, relies on the fact that the next Krylov vector represents the residual direction for all Ritz vectors, whereas this property no longer holds when MPS truncation exists. MTBL solves this issue by explicitly computing residual vectors for each individual Ritz vector, making the restart compatible with the MPS truncation.
Shift-and-invert
In many applications, one is in fact more interested in excited states than in the ground state. It is therefore highly promising to directly focus on states around a given target energy, instead of successively converging to all lower-energy eigenstates. To achieve this, a method is to combine the Lanczos algorithm with a shift-and-invert spectral transformation [41], where we replace the Hamiltonian with:
where is a shift chosen near the desired value of the spectrum, so that the eigenstates whose eigenvalues lie near become extreme values of . To construct the Krylov subspace of , we need to evaluate
However, it is generally ill-conditioned and also requires a prohibitively large bond dimension to represent as an MPO. To overcome this issue, we search for the MPS which satisfies:
using the variational method [30, 69, 70, 71]. By computing the extreme eigenstates of , we directly obtain the desired excited states of . For instance, the eigenstate of whose energy is closest to from above corresponds to the largest eigenvalue of .
While the standard DMRG typically obtains excited states indirectly by solving all lower states sequentially, making its accuracy dependent on the accuracy of those preceding states, the Lanczos method with shift-and-invert can also refine the target excited state by using the DMRG result as the initial state. In this way, besides locating eigenstates near a chosen spectral position , the Lanczos method with shift-and-invert can further improve the precision of the targeted excited state without requiring convergence of all lower-energy states.
Energy variance
As the system size increases, reference eigenvalues from exact diagonalization become unavailable. In such cases, the accuracy of approximate eigenstates can be assessed using the energy variance, a widely applied diagnostic in computational physics and numerical linear algebra [45, 72, 73, 74, 75, 76]. The energy variance of an approximate eigenpair with respect to a Hamiltonian is given by:
which measures how closely the approximate eigenvalue and eigenstate approximate an exact eigenpair. Minimizing this variance is essential in iterative eigenvalue algorithms such as Lanczos and Davidson methods to achieve reliable convergence [72]. To lower the computational cost of Eq. (18), we should avoid explicitly forming . Instead, we can evaluate the energy variance directly via tensor-network contractions, analogous to that employed in the DMRG algorithm [19, 24]; in addition, a more efficient method for computing the energy variance in MPS form has also been developed [76].
Numerical experiments
To assess the performance of MTBL, we apply it to Hamiltonians of broad interest and demonstrate its accuracy. We also numerically confirm our physical intuition
underlying the method’s improvement. The total computational cost scales with the number of iterations, that is, with how many times the Hamiltonian is applied to the most recent Krylov vector (in PLM or TRL) or block (in MTBL) to generate the next one. That means, if the block size is , one MTBL iteration involves MPO-MPS multiplication-compressions. In addition, MTBL needs extra such operations to compute the residual vectors at the beginning of each restart. For convenience, we use the iteration count as a direct measure of the computational cost, while we also provide the cost described in the number of MPO-MPS multiplication-compressions. The calculation with MPS representation is implemented in PyTeNet [77].
The Fermi-Hubbard model
We first benchmark the MTBL method using the one-dimensional Fermi–Hubbard model [78, 79], one of the most extensively studied systems in condensed matter physics. It describes the interaction of fermions on a lattice system, offering crucial insights into phenomena such as Mott insulating [80, 81] and high-temperature superconductivity [82, 83, 84]. The Fermi-Hubbard Hamiltonian is given by:
where and are fermionic creation and annihilation operators at site with spin , and denotes the number operator. We set the hopping amplitude to , the chemical potential to (half-filling), and the on-site interaction to . Such a parameter setting places the system in a strongly interacting regime, where correlation effects are expected to be significant. As a result, the accurate finding of low-lying eigenstates is more challenging than in weakly interacting regimes. The number of lattice sites is (i.e., 16 spin sites), so it is convenient to achieve the reference eigenpairs from exact diagonalization (ED) [52, 85]. We cap the Krylov vectors’ bond dimension at , which means that we always compress the MPS bond dimension to after any operations, such as MPS-MPS addition or MPO-MPS multiplication. We initialize the calculation with a random state.
First, we present the energy errors for finding the first five low-lying eigenstates using the original plain Lanczos method (PLM). As illustrated in Figure 3, the convergence stalls far from the reference values, consistent with the discussion in § 2 C, where we interpreted the PLM’s failure to accurately find the eigenstates.
Figure 3. Convergence of the first five low-lying eigenstates of the Fermi–Hubbard model obtained with PLM in MPS representation. The excited states cannot converge to high accuracy. Here, each iteration refers to a single MPO-MPS multiplication-compression.
Classically, the TRL method is widely used to improve convergence. However, as discussed in § 3 A, it fails when the wavefunction is represented by an MPS ansatz, as illustrated in Figure 4. That is because the parallelism among the residual vectors of the corresponding Ritz vectors is lost due to MPS truncation, as discussed in § 3 A. Here, we provide numerical results that verify this conclusion. Specifically, we compute the residual vectors of the five lowest Ritz vectors obtained from PLM using Eq. (12) and maximum bond dimension , normalize each residual vector, and then evaluate the overlap matrix using these residuals as:
Figure 4. Convergence behavior of the first five low-lying eigenstates of the Fermi–Hubbard model using the TRL method with the MPS representation. We restart the procedure every 20 iterations, which corresponds to applying MPO-MPS multiplication-compressions for 20 times. Only the ground state converges well since TRL using MPS can capture the residual vector of the ground state correctly.
where we only show four decimal places. We see that these residual vectors are far from being mutually paral-
lel, since perfect parallelism would yield an overlap matrix with all elements 1. The parallelism of these residual vectors can be quantitatively measured by the following metric:
where is the Frobenius norm. indicates that the residual vectors are fully parallel, while corresponds to perfect mutual orthogonality. As shown in Figure 5, we present the parallelism of the residual vectors obtained from PLM under various bond dimensions, characterized by the metric .
Figure 5. Parallelism of the residual vectors associated with the five Ritz vectors for bond dimensions , measured by . The circled point corresponds to the case , which leads to the overlap matrix Eq. (19). Only when the bond dimension is set to , which corresponds to the no-truncation case, are the residual vectors parallel.
Therefore, to fully exploit the residual information
of all target eigenstates, we employ our MTBL, which brings all desired residual vectors into the restart procedure. Consequently, all target eigenstates converge in parallel and reach the optimal accuracy with the given bond dimension when employing MTBL to find the eigenstates, as shown in Figure 6. Here, the optimal accuracy refers to the best approximation attainable with a given bond dimension. It is numerically estimated by solving the exact eigenstate in vector form through ED, converting it into an MPS, and truncating this MPS via the SVD method discussed in § 2 B; the accuracy of the resulting truncated MPS defines the optimal accuracy. All five target eigenstates reach a relative energy error of within 160 block iterations, corresponding to 960 MPO-MPS multiplication-compressions. They further reach a relative energy error of and attain the optimal accuracy within 280 block iterations, corresponding to 2016 such operations in total, or 403 per target state. Compared to the PLM results shown in Figure 3, the accuracy is improved by three to four orders of magnitude. Notably, the three degenerate first excited states are all accurately captured. In addition to the high accuracy, MTBL also converges faster compared to PLM and TRL. Using the number of MPO-MPS multiplication-compression as a unified metric of computational cost, we compare the convergence speeds of the plain Lanczos method (PLM), TRL, and MTBL for the Fermi-Hubbard model using the geometric mean of the relative energy errors. Specifically, for each method, we define the averaged relative error over the target eigenstates as
Figure 6. Track of the 5 lowest eigenstates of the Fermi–Hubbard model in the MTBL method. We use a block size of 6 and restart the procedure every 5 iterations, where each iteration contains 6 MPO-MPS multiplication-compressions, i.e., 36 each restart considering extra 6 such operations required to compute the residual vectors. For each eigenstate, the dashed line of the same color marks the optimal accuracy achievable with the chosen bond dimension. The degeneracy of the first three excited states is faithfully captured.
where denotes the relative energy error of the th target eigenstate, and is the number of target eigenstates. As shown in Figure 7, the averaged error of MTBL decreases to nearly with 400 MPO-MPS multiplication-compressions, whereas those of TRL and PLM remain around . This demonstrates that MTBL not only achieves higher accuracy but also converges faster under the same cost metric. Furthermore, we also observe that during the MTBL simulation, the MPO-MPS compression error does not necessarily increase with a fixed bond dimension. We discussed this in Appendix B.
Figure 7. Convergence comparison of PLM, TRL, and MTBL using the geometric mean of relative energy errors, under the same computational cost metric by number of MPO-MPS applications.
Here, we further compare MTBL with methods based on the bundled MPS. We implemented two representative methods: the DMRG algorithm with block Lanczos (BL) [65], and the multi-target update method (MTU [67]. We used the same parameter settings for the Fermi-Hubbard Hamiltonian as above and capped the bond dimension at . We computed the lowest five eigenstates and evaluated the relative energy errors. As summarized in Table I, MTBL achieves substantially smaller relative energy errors than the two bundled-MPS-based methods for all target eigenstates.
| MTBL | DMRG w/ BL | MTU | |
| Ground state | |||
| 1st excited state 1 | |||
| 1st excited state 2 | |||
| 1st excited state 3 | |||
| 2nd excited state |
Table I. Relative energy errors for the first five target eigenstates of the Fermi-Hubbard model, obtained by MTBL, DMRG with block Lanczos (BL), and MTU. MTBL yields the smallest relative energy error for each target eigenstate.
To demonstrate the scalability and reliability of MTBL, we extend the size of the Fermi-Hubbard model to , which corresponds to 100 spin sites and Hilbert space of dimension ; we cap the maximum bond dimension at . We lock a target state once its energy variance reaches . After being locked, this state is excluded from subsequent Lanczos iterations and is no longer updated. However, the locked states are still included when solving the effective Hamiltonian, so that the higher eigenstates remain orthogonal to them. As shown in Figure 8, all five target states reach the target energy variance of , each requiring approximately 2900 MPO-MPS multiplication-compressions on average.
Figure 8. Track of the 5 lowest eigenstates of the Fermi–Hubbard model with in the MTBL method. We use a block size of 6 and restart the procedure every 3 iterations, where each iteration contains 6 MPO-MPS multiplication-compressions, i.e., 24 each restart. The degeneracy of the first three excited states is confirmed by checking the Ritz values.
The spin-1/2 Heisenberg XXZ model in a longitudinal magnetic field
The second model we chose is the one-dimensional spin-1/2 Heisenberg XXZ model [86, 87, 88, 89, 90] in a uniform longitudinal magnetic field:
We denote operators as raising/lowering operators, and () where are the Pauli matrices. We set the nearest-neighbor exchange coupling strength to , the anisotropy parameter to , and the uniform magnetic field strength to . The number of lattice sites is ; we cap the Krylov vectors’ bond dimension at . We work in the total spin sector since favors positive magnetization, and the initial state is set as a random state satisfying this total spin.
Finding low-lying states using MTBL
Similar to the previous example, the PLM again fails to accurately recover the low-lying eigenstates. As shown in Figure 9, even though the eigenstates appear to converge, the subspace size has grown to 1000, corresponding to 1000 MPO-MPS multiplication-compressions, which is already prohibitively large. Such an oversized subspace not only makes the calculation and diagonalization of the effective Hamiltonian expensive but also leads to a cascade of issues, including numerical instability, increased memory consumption, and loss of orthogonality.
Figure 9. When using the PLM to simulate the Heisenberg model, the convergence stalls far away from the reference eigenvalues, except for the ground state. Here, each iteration refers to a single MPO-MPS multiplication-compression.
In contrast, the MTBL method achieves rapid and stable convergence, as displayed in Figure 10. The relative energy errors drop below within 150 block iterations, and all five eigenstates are successfully obtained, reaching their optimal accuracies in fewer than 250 block iterations. Equivalently, reaching a relative energy error below requires 900 MPO-MPS multiplication-compressions in total, corresponding to 180 such operations per target state. To demonstrate the robustness of MTBL, we also benchmark MTBL with the parameter settings , , and . For this parameter setting, we test MTBL with two system sizes: with , as well as with . To keep the main text concise, we present the results in Appendix C, where MTBL also achieves the optimal accuracy.
Figure 10. Convergence of the five relative energy errors of the low-lying eigenstates of the Heisenberg model () obtained using the MTBL method. Each block contains 5 vectors, and we restart the procedure every 5 iterations, for simplicity. Here, each iteration contains 5 MPO-MPS multiplication-compressions, and this indicates that we restart after every 30 such operations. For each eigenstate, the dashed line of the same color indicates the optimal accuracy achievable with the chosen bond dimension. The bond dimension is capped at 2.
Directly targeting excited states
We further demonstrate that, by employing the shift-and-invert technique, we can directly target excited states. Using the same parameter settings as above, we choose an energy shift of and compute six eigenstates whose energies are larger than this shift, which corresponds to the six highest positive eigenstates of . After the convergence of these target eigenstates, we compare them with the reference spectrum and identify them as the 6th–11th excited states. As shown in Figure 11, all these eigenstates converge to an energy variance below within 120 iterations, or 840 MPO-MPS multiplication-compressions, equivalently. Also, eigenvalues closer to the shift converge more rapidly, and this behavior is consistent with the spectral transformation induced by shift-and-invert: for an eigenpair of , the corresponding eigenvalue of is . Hence, the closer is to the shift , the larger becomes, making the associated state more extreme in the spectrum of and therefore typically converging faster in the Lanczos iteration.
Figure 11. Direct targeting of excited states close to in the spectrum using the shift-and-invert technique. Each block contains 6 vectors, and we restart the procedure every 6 iterations, i.e., 42 MPO-MPS multiplication-compressions. We assess accuracy using the energy variance, as we do not know the corresponding eigenstates before comparing them with the reference after convergence. The bond dimension is capped at 4.
Benchmark on large system
We also demonstrate the scalability of MTBL with a larger Heisenberg chain. We increase the system size to , which cannot be solved by exact diagonalization because of memory limitations, while restricting the bond dimension of the Krylov vectors to . As discussed in § 3 D, we will use the energy variance as the metric of errors since the exact reference values are not available from ED for such a large system. As shown in Figure 12, all six target states converge to energy variances below . These results demonstrate that the MTBL method is both scalable and reliable for computing low-lying eigenstates of large many-body systems. As the energy variance reflects how close the Ritz vectors are to the exact eigenstates, this indicates that all target states have converged to a very high accuracy.
Figure 12. Convergence of the MTBL method for computing the six lowest eigenstates of the 120-site Heisenberg XXZ model. Each block contains 6 vectors, and we restart the procedure every 5 iterations, i.e., 36 MPO-MPS multiplication-compressions. More iterations are required because a random initial state has only a small overlap with the low-lying eigenstates. The number of iterations can be reduced by selecting a larger bond dimension.
CONCLUSION AND DISCUSSION
In this work, we propose targeted improvements to the Lanczos method within the MPS representation. The MTBL method leverages the block-Lanczos framework to capture the residual directions of all Ritz vectors, which are no longer parallel under the MPS ansatz. The block-Lanczos scheme serves as a key component that enables the efficient propagation of more information across successive iterations. This residual-augmented restart distinguishes MTBL from previous MPS-based Lanczos methods, whereas earlier approaches continue with either a single Ritz vector or a single residual direction. We demonstrated the accuracy, reliability, and scalability of our method by benchmarking on the Fermi–Hubbard model and the Heisenberg model in a longitudinal field. We expect the MTBL to provide a reliable eigenstate solver that inherently avoids local minima, decouples the accuracy of excited states from lower-lying eigenstates when combined with the shift-and-invert technique, and naturally resolves degeneracies through a block formulation. As discussed in Ref. [41, 91], the Lanczos method combined with shift-and-invert provides a powerful alternative for finding eigenstates within an energy window that has a high density of eigenstates.
To demonstrate the convergence of our MTBL method, we employ the standard singular value decomposition (SVD) for MPS compression. Nevertheless, alternative compression strategies can also be considered, such as the zip-up method [58], the density-matrix-based approach [59], the variational optimization scheme [19], and the successive randomized compression (SRC) method [92]. Each compression technique offers a distinct balance between computational cost, memory usage, and accuracy. It is therefore promising to explore these alternative schemes to further accelerate MTBL, potentially at the expense of some accuracy.
Because the MTBL only operates on MPO-MPS multiplication, MPS-MPS addition, and compression, it can be integrated into any modern tensor network library without requiring additional heavy infrastructure. Additionally, it is not confined to the MPS/MPO framework; any ansatz supporting these operations, e.g., projected entangled pair states (PEPS) [93], isometric tensor-network states (isoTNS) [94], tree tensor network states (TTNS) [95], and so on, can be incorporated into our method.
ACKNOWLEDGMENTS
We thank Chao Yang, Alec Dektor, Roel Van Beeumen, and Jan von Delft for helpful discussions. We also acknowledge funding by the Munich Quantum Valley initiative, supported by the Bavarian state government with funds from the Hightech Agenda Bayern Plus.
AUTHOR CONTRIBUTIONS
Y.W. conceived the project, proposed the main idea, developed the initial MPS-Lanczos code, performed the numerical experiments, and wrote the manuscript. Z.Y. contributed to the conception of the project, developed the MTBL code, and performed numerical experiments. X.W. implemented the shift-and-invert method and performed the benchmark comparisons among MTBL, MTU, and DMRG with the block Lanczos algorithm. C.B.M. supervised the project, participated in the discussion, and revised the manuscript.
DATA AVAILABILITY
The data that support this study are available on Zenodo [96]. The codes are available from the corresponding authors upon reasonable request.
Process of Thick-restart Lanczos Method
In this appendix, we summarize the practical steps of the Thick-Restart Lanczos (TRL) algorithm in the exact-arithmetic Lanczos method.
The TRL improves convergence and reduces memory usage by periodically restarting the Lanczos iteration with preserved spectral information from the previous iteration. This implementation serves as the foundation for our MTBL in the main text. Additionally, the Ritz vector that reaches the desired accuracy can also be locked in TRL to save computational resources.
Algorithm 2 Thick-Restart Lanczos (TRL)
Input: Hermitian matrix , initial state , subspace size , error tolerance
Output: lowest eigenpairs
Run the first cycle of the Lanczos algorithm with initial state and Krylov subspace size to obtain the Krylov subspace .
Compute the next Krylov vector
Compute Ritz pairs with Krylov subspace
for , until convergence
Set for
Set
Orthogonalize against previous vectors
Initialize Lanczos procedure with
Extend new Krylov vectors until subspace size
Define
Compute an extra Krylov vector
Update Ritz pairs with
Sort Ritz pairs in ascending order of
for
Compute
end for
if for all $\ell = 0,1,\ldots,t-1 then
Stop: convergence achieved
set for all
end if
end for
return
The MPO-MPS truncation error during MTBL simulation
Controlling the bond dimension is a central issue in MPS-based simulations. In real-time evolution, the entanglement entropy of
often grows with the evolution time . Since the bond dimension required for an accurate MPS representation typically grows exponentially with the entanglement entropy, a fixed bond dimension can lead to increasing truncation errors during long-time evolution [51].
The situation in MTBL is different. The dominant operation is not long-time evolution by , but MPO-MPS multiplication-compressions of the form
Let denote the eigenstates of , with eigenvalues . Expanding in this eigenbasis gives
and therefore
Thus, applying reweights the eigenstate components by their eigenvalues. Assuming that the target state is the ground state , i.e., for , the resulting state is even closer to the ground state since the magnitude of is largest. For many local Hamiltonians, low-lying eigenstates can be accurately represented by MPS with small bond dimensions due to the area law. Therefore, as the algorithm proceeds and becomes closer to the ground state, one can expect that smaller and smaller bond dimensions are required to represent . Equivalently, at a fixed bond dimension, the compression error in representing is not expected to grow systematically; instead, it may decrease as the iteration converges toward the ground state. To illustrate this behavior, we compute the truncation error associated with compressing at the beginning of each restart, where denotes a Ritz vector. As shown in Figure 13, the compression error decreases as the Ritz vectors approach the eigenstates. Therefore, by repeatedly restarting the Lanczos procedure, we always initialize the procedure with better initial states, and the compression errors do not necessarily become larger.
Figure 13. Relative MPO-MPS compression error for , where denotes a Ritz vector. The error is computed at each restart of the algorithm.
Benchmark on the Heisenberg Model with , , and
We also demonstrate MTBL with the nearest-neighbor exchange coupling strength , the anisotropy parameter , and the uniform magnetic field strength , and the bond dimension is capped at , and the system size is . In such a setting, the ground state corresponds to the Tomonaga–Luttinger liquid, which is gapless in the thermodynamic limit [97], and this critical regime has a dense set of low-lying excited
Figure 15. When using the MTBL to simulate the Heisenberg model with , , , , and , the relative energy error converges fast and reaches the optimal accuracy. For each eigenstate, the dashed line of the same color indicates the optimal accuracy achievable with the chosen bond dimension.
states that makes it a numerically demanding benchmark for excited-state algorithms.
Similarly, we first present the results of PLM. As shown in Figure 14, the convergence of all target eigenstates stalls around an error of . However, when applying MTBL, the errors of all five target states decrease to within 180 iterations and reach their respective optimal accuracies by less than 240 steps. These results demonstrate that MTBL is both robust and broadly applicable, delivering reliable convergence across challenging excited-state problems.
Figure 14. When using the PLM to simulate the Heisenberg model with , , and , the convergence stalls far away from the reference eigenvalues.
We also extend the system size to and use maximum bond dimension to demonstrate MTBL’s scalability and reliability. We use a block size of 4, and restart the Lanczos procedure every three iterations. A target state is locked once its energy variance falls below . As shown in Figure 16, all four target states finally reach the target energy variance.
Figure 16. Track of the 4 lowest eigenstates of the spin-1/2 Heisenberg XXZ model in a longitudinal magnetic field with , , and with , in the MTBL method. We use a block size of 4 and restart the procedure every 3 iterations.
References
- [1]C. Lanczos, An iteration method for the solution of the eigenvalue problem of linear differential and integral operators, J. Res. Nat. Bur. Stand. 45, 255 (1950).
- [2]S. R. White, Density matrix formulation for quantum renormalization groups, Phys. Rev. Lett. 69, 2863 (1992).pubmed.ncbi.nlm.nih.gov/10046608
- [3]A. W. Sandvik, Finite-size scaling of the ground-state parameters of the two-dimensional Heisenberg model, Phys. Rev. B 56, 11678 (1997).DOI
- [4]M. C. Bañuls, Tensor network algorithms: A route map, Annu. Rev. Condens. Matter Phys. 14, 173 (2023).DOI
- [5]A. Avella and F. Mancini, eds., Strongly Correlated Systems: Numerical Methods, 1st ed., Springer Series in Solid-State Sciences, Vol. 176 (Springer, Berlin, Heidelberg, 2013).DOI
- [6]S. Sugiura and A. Shimizu, Canonical thermal pure quantum state, Phys. Rev. Lett. 111, 010401 (2013).DOI
- [7]J. Jaklič and P. Prelovšek, Finite-temperature properties of doped antiferromagnets, Adv. Phys. 49, 1 (2000).DOI
- [8]V. Laguta, Y. O. Zagorodniy, R. Kuzian, I. Kondakova, V. Chlan, R. Rezníček, H. Štěpánková, D. Bohdanov, J. Hlinka, and R. Ramesh, Low-temperature ground state structure of PbTiO3, Phys. Rev. B 107, 104107 (2023).
- [9]E. R. Gagliano and C. A. Balseiro, Dynamical properties of quantum many-body systems at zero temperature, Phys. Rev. Lett. 59, 2999 (1987).pubmed.ncbi.nlm.nih.gov/10035707
- [10]S. Sun, M. Motta, R. N. Tazhigulov, A. T. K. Tan, G. K. Chan, and A. J. Minnich, Quantum computation of finite-temperature static and dynamical properties of spin systems using quantum imaginary time evolution, PRX Quantum 2, 010317 (2021).DOI
- [11]C. C. Paige, Computational variants of the Lanczos method for the eigenproblem, IMA J. Appl. Math. 10, 373 (1972).DOI
- [12]J. K. Cullum and R. A. Willoughby, Lanczos algorithms for large symmetric eigenvalue computations: Volume I: Theory, Progress in Scientific Computing, Vol. 3 (Birkhäuser Boston, Boston, MA, 1985).
- [13]Z. Bai, J. Demmel, J. Dongarra, A. Ruhe, and H. van der Vorst, eds., Templates for the solution of algebraic eigenvalue problems: a practical guide (SIAM, 2000).
- [14]A. Bhattacharya, P. Nandy, P. P. Nath, and H. Sahu, On Krylov complexity in open systems: an approach via bi-Lanczos algorithm, J. High Energy Phys. 2023 (66).DOI
- [15]P. Weinberg and M. Bukov, QuSpin: a python package for dynamics and exact diagonalisation of quantum many body systems. part ii: bosons, fermions and higher spins, SciPost Phys. 7, 020 (2019).DOI
- [16]A. M. Läuchli, J. Sudan, and R. Moessner, The s = 1/2 kagome Heisenberg antiferromagnet revisited, Phys. Rev. B 100, 155142 (2019).DOI
- [17]S. Iskakov and M. Danilov, Exact diagonalization library for quantum electron models, Comput. Phys. Commun. 225, 128 (2018).DOI
- [18]R. B. Lehoucq, D. C. Sorensen, and C. Yang, ARPACK Users’ Guide: Solution of Large-Scale Eigenvalue Problems with Implicity Restarted Arnoldi Methods, Software, Environments & Tools, Vol. 6 (Society for Industrial and Applied Mathematics, 1998).DOI
- [19]U. Schollwöck, The density-matrix renormalization group in the age of matrix product states, Ann. Phys. 326, 96 (2011).DOI
- [20]J. C. Bridgeman and C. T. Chubb, Hand-waving and interpretive dance: An introductory course on tensor networks, J. Phys. A: Math. Theor. 50, 223001 (2017).DOI
- [21]S. Östlund and S. Rommer, Thermodynamic limit of density matrix renormalization, Phys. Rev. Lett. 75, 3537 (1995).pubmed.ncbi.nlm.nih.gov/10059611
- [22]S. Rommer and S. Östlund, Class of ansatz wave functions for one-dimensional spin systems and their relation to the density matrix renormalization group, Phys. Rev. B 55, 2164 (1997).DOI
- [23]H. Ma, U. Schollwöck, and Z. Shuai, Density Matrix Renormalization Group (DMRG)-Based Approaches in Computational Chemistry (Elsevier, 2022).DOI
- [24]G. K.-L. Chan, A. Keselman, N. Nakatani, Z. Li, and S. R. White, Matrix product operators, matrix product states, and ab initio density matrix renormalization group algorithms, J. Chem. Phys. 145, 011402 (2016).arxiv.org/abs/1605.02611
- [25]S. Keller, M. Dolfi, M. Troyer, and M. Reiher, An efficient matrix product operator representation of the quantum chemical Hamiltonian, J. Chem. Phys. 143, 244118 (2015).DOI
- [26]D. Cocchiarella, M. Yang, Y. Zhang, M. C. Bañuls, H.-H. Tu, and Y. Liu, Excited states from local effective hamiltonians of matrix product states and their entanglement spectrum transition, Phys. Rev. B 113, 205145 (2026).DOI
- [27]S. R. White, Density matrix renormalization group algorithms with a single center site, Phys. Rev. B 72, 180403 (2005).DOI
- [28]R. N. Pfeiffer, Symmetry-protected local minima in infinite DMRG, Phys. Rev. B 92, 205127 (2015).DOI
- [29]A. Dektor, P. DelMastro, E. Ye, R. Van Beeu men, and C. Yang, Inexact subspace projection methods for low-rank tensor eigenvalue problems, arXiv preprint arXiv:2502.19578 10.48550/arXiv.2502.19578 (2025).arxiv.org/abs/2502.19578
- [30]A. Baiardi, C. J. Stein, V. Barone, and M. Reiher, Optimization of highly excited matrix product states with an application to vibrational spectroscopy, J. Chem. Phys. 150, 094113 (2019).DOI
- [31]J. J. Dorando, J. Hachmann, and G. K.-L. Chan, Targeted excited state algorithms, J. Chem. Phys. 127, 084109 (2007).DOI
- [32]L. N. Tran, J. A. R. Shea, and E. Neuscamman, Tracking excited states in wave function optimization using density matrices and variational principles, J. Chem. Theory Comput. 15, 4790 (2019).arxiv.org/abs/1904.05489
- [33]K. Szenes, N. Glaser, M. Eraković, V. Barandun, M. Mörchen, R. Feldmann, S. Battaglia, A. Baiardi, and M. Reiher, QCMaquis 4.0: Multipurpose electronic, vibrational, and vibronic structure and dynamics calculations with the density matrix renormalization group, J. Phys. Chem. A 129, 7549 (2025).DOI
- [34]H. R. Larsson, Computing vibrational eigenstates with tree tensor network states (TTNS), J. Chem. Phys. 151, 204102 (2019).DOI
- [35]G. H. Golub and R. Underwood, The block Lanczos method for computing eigenvalues, in Mathematical Software, edited by J. R. Rice (Academic Press, 1977) pp. 361–377.DOI
- [36]K. Wu, A. Canning, H. D. Simon, and L.-W. Wang, Thick-restart Lanczos method for electronic structure calculations, J. Comput. Phys. 154, 156 (1999).
- [37]P. E. Dargel, A. Woel lert, A. Honecker, I. McCulloch, U. Schollwöck, and T. Pruschke, Lanczos algorithm with matrix product states for dynamical correlation functions, Phys. Rev. B 85, 205119 (2012).DOI
- [38]R.-Z. Huang, H.-J. Liao, Z.-Y. Liu, H.-D. Xie, Z.-Y. Xie, H.-H. Zhao, J. Chen, and T. Xiang, Generalized Lanczos method for systematic optimization of tensor network states, Chinese Physics B 27, 070501 (2018).DOI
- [39]P. Li, Y. Shen, M. Qin, K. Jiang, J. Hu, and T. Xiang, Bound states in doped charge transfer insulators (2024), arXiv:2408.00576.DOI
- [40]S. Paeckel, T. Köhler, S. R. Manmana, and B. Lenz, Matrix-product-state-based band-Lanczos solver for quantum cluster approaches (2023), arXiv:2310.10799.DOI
- [41]M. Rano and H. R. Larsson, Computing excited eigenstates using inexact Lanczos methods and tree tensor net-DOI
- [42]K. K. Wu and H. Simon, Thick-restart Lanczos method for large symmetric eigenvalue problems, SIAM J. Matrix Anal. Appl. 22, 602 (2000).
- [43]J. H. Wilkinson, The Algebraic Eigenvalue Problem, Monographs on Numerical Analysis (Oxford University Press, Oxford, 1965).
- [44]A. Ruhe, Rational Krylov sequence methods for eigenvalue computation, Linear Algebra Appl. 58, 391 (1984).DOI
- [45]G. H. Golub and C. F. Van Loan, Matrix computations (JHU press, 2013).
- [46]Y. Saad, Numerical Methods for Large Eigenvalue Problems, revised ed., Classics in Applied Mathematics (Society for Industrial and Applied Mathematics, Philadelphia, 2011).
- [47]F. Pietracaprina, N. Macé, D. J. Luitz, and F. Alet, Shift-invert diagonalization of large many-body localizing spin chains, SciPost Phys. 5, 045 (2018).DOI
- [48]R. Haydock, The recursive solution of the Schrödinger equation, Solid State Phys. 35, 215 (1980).
- [49]J. Laflamme Janssen, B. Rousseau, and M. Côté, Efficient dielectric matrix calculations using the Lanczos algorithm for fast many-body g𝜔0 implementations, Phys. Rev. B 91, 125120 (2015).DOI
- [50]H.-G. Weikert, H.-D. Meyer, L. Cederbaum, and F. Tarantelli, Block Lanczos and many-body theory: Application to the one-particle green’s function, J. Chem. Phys. 104, 7122 (1996).DOI
- [51]S. Paeckel, T. Köhler, A. Swoboda, S. R. Manmana, U. Schollwöck, and C. Hubig, Time-evolution methods for matrix-product states, Annals of Physics 411, 167998 (2019).DOI
- [52]E. Dagotto, Correlated electrons in high-temperature superconductors, Rev. Mod. Phys. 66, 763 (1994).DOI
- [53]S. Östlund and S. Rommer, Thermodynamic limit of density matrix renormalization, Phys. Rev. Lett. 75, 3537 (1995).pubmed.ncbi.nlm.nih.gov/10059611
- [54]S. Rommer and S. Östlund, Class of ansatz wave functions for one-dimensional spin systems and their relation to the density matrix renormalization group, Phys. Rev. B 55, 2164 (1997).DOI
- [55]J. Hauschild and F. Pollmann, Efficient numerical simulations with tensor networks: Tensor Network Python (TeNPy), SciPost Phys. Lect. Notes, 5 (2018).DOI
- [56]J. Ren, W. Li, T. Jiang, and Z. Shuai, A general automatic method for optimal construction of matrix product operators using bipartite graph theory, J. Chem. Phys. 153, 084118 (2020).
- [57]H. Çakır, R. M. Milbradt, and C. B. Mendl, Optimal symbolic construction of matrix product operators and tree tensor network operators, arXiv preprint arXiv:2502.18630 10.48550/arXiv.2502.18630 (2025).
- [58]E. M. Stoudenmire and S. R. White, Minimally entangled typical thermal state algorithms, New J. Phys. 12, 055026 (2010).DOI
- [59]L. Ma, M. Fishman, E. M. Stoudenmire, and E. Solomonik, Approximate Contraction of Arbitrary Tensor Networks with a Flexible and Efficient Density Matrix Algorithm, Quantum 8, 1580 (2024).DOI
- [60]T. Jiang, J. Ren, and Z. Shuai, Chebyshev matrix product states with canonical orthogonalization for spectral functions of many-body systems, J. Phys. Chem. Lett. 12, 9344 (2021). Y. Saad, On the rates of convergence of the Lanczos and the block-Lanczos methods, SIAM J. Numer. Anal. 17, 687 (1980).
- [62]D. C. Hackett and M. L. Wagman, Block Lanczos algorithm for lattice qcd spectroscopy and matrix elements, Phys. Rev. D 112, 014514 (2025).arxiv.org/abs/2412.04444
- [63]B. N. Parlett, The Symmetric Eigenvalue Problem, Classics in Applied Mathematics (Society for Industrial and Applied Mathematics, Philadelphia, 1998).
- [64]R. Morgan, On restarting the arnoldi method for large nonsymmetric eigenvalue problems, Mathematics of Computation 65, 1213 (1996).DOI
- [65]T. E. Baker, A. Foley, and D. Sénéchal, Direct solution of multiple excitations in a matrix product state with block lanczos, The European Physical Journal B 97, 72 (2024).arxiv.org/abs/2109.08181
- [66]T. E. Baker and N. Seif, Bundled matrix product states represent low-energy excitations faithfully, Journal of Physics A: Mathematical and Theoretical 57, 395306 (2024).DOI
- [67]X. Li, Z. Zhou, G. Xu, R. Chi, Y. Guo, T. Liu, H. Liao, and T. Xiang, Accurate determination of low-energy eigenspectra with multitarget matrix product states, Phys. Rev. B 109, 045115 (2024).arxiv.org/abs/2305.15868
- [68]T. E. Baker, Block lanczos method for excited states on a quantum computer, Phys. Rev. A 110, 012420 (2024).arxiv.org/abs/2109.14114
- [69]X. Yu, D. Pekker, and B. K. Clark, Finding matrix product state representations of highly excited eigenstates of many-body localized hamiltonians, Phys. Rev. Lett. 118, 017201 (2017).DOI
- [70]M. Rakhuba and I. Oseledets, Calculating vibrational spectra of molecules using tensor train decomposition, J. Chem. Phys. 145, 124101 (2016).DOI
- [71]N. Pomata, S. Ganeshan, and T.-C. Wei, Seeking a many-body mobility edge with matrix product states in a quasiperiodic model, Phys. Rev. B 108, 094201 (2023).DOI
- [72]J. K. Cullum and R. A. Willoughby, Lanczos Algorithms for Large Symmetric Eigenvalue Computations (SIAM, 2002).DOI
- [73]R. M. Martin, Electronic Structure: Basic Theory and Practical Methods (Cambridge University Press, 2004).
- [74]C. Hubig, J. Haegeman, and U. Schollwöck, Error estimates for extrapolations with matrix-product states, Phys. Rev. B 97, 045125 (2018).DOI
- [75]A. Gleis, J.-W. Li, and J. von Delft, Controlled bond expansion for density matrix renormalization group ground state search at single-site costs, Phys. Rev. Lett. 130, 246402 (2023).DOI
- [76]A. Gleis, J.-W. Li, and J. von Delft, Projector formalism for kept and discarded spaces of matrix product states, Phys. Rev. B 106, 195138 (2022).DOI
- [77]C. B. Mendl, PyTeNet: A concise Python implementation of quantum tensor network algorithms, J. Open Source Softw. 3, 948 (2018).
- [78]J. Hubbard, Electron correlations in narrow energy bands, Proc. R. Soc. Lond. A 276, 238 (1963).
- [79]A. Georges, G. Kotliar, W. Krauth, and M. J. Rozenberg, Dynamical mean-field theory of strongly correlated fermion systems and the limit of infinite dimensions, Rev. Mod. Phys. 68, 13 (1996).DOI
- [80]R. Jördens, N. Strohmaier, K. Günter, H. Moritz, and T. Esslinger, A mott insulator of fermionic atoms in anDOI
- [81]C. Hofrichter, L. Riegger, F. Scazza, M. Höfer, D. Rio Fernandes, I. Bloch, and S. Fölling, Direct probing of the mott crossover in the su(n) fermi-hubbard model, Phys. Rev. X 6, 021030 (2016).DOI
- [82]P. W. Anderson, The resonating valence bond state in La2CuO4 and superconductivity, Science 235, 1196 (1987).pubmed.ncbi.nlm.nih.gov/17818979
- [83]D. J. Scalapino, E. Loh Jr., and J. E. Hirsch, d-wave pairing near a spin-density-wave instability, Phys. Rev. B 34, 8190 (1986).pubmed.ncbi.nlm.nih.gov/9939525
- [84]E. Dagotto, Correlated electrons in high-temperature superconductors, Rev. Mod. Phys. 66, 763 (1994).DOI
- [85]A. W. Sandvik, Computational studies of quantum spin systems, AIP Conference Proceedings 1297, 135 (2010).DOI
- [86]W. Heisenberg, Zur theorie des ferromagnetismus, Zeitschrift für Physik 49, 619 (1928).
- [87]C. N. Yang and C. P. Yang, One-dimensional chain of anisotropic spin-spin interactions. i. proof of bethe’s hypothesis for the ground state in a finite system, Phys. Rev. 150, 321 (1966).DOI
- [88]C. N. Yang and C. P. Yang, One-dimensional chain of anisotropic spin-spin interactions. ii. properties of the ground-state energy per lattice site for an infinite system, Phys. Rev. 150, 327 (1966).DOI
- [89]F. C. Alcaraz, M. N. Barber, M. T. Batchelor, R. J. Baxter, and G. R. W. Quispel, Surface exponents of the quantum XXZ, ashkin–teller and potts models, J. Phys. A: Math. Gen. 20, 6397 (1987).DOI
- [90]B. Bertini, M. Collura, J. De Nardis, and M. Fagotti, Transport in out-of-equilibrium XXZ chains: Exact profiles of charges and currents, Phys. Rev. Lett. 117, 207201 (2016).DOI
- [91]H. R. Larsson, B. L. Dé, and G. E. Gamboni, Accurate, full-dimensional computations of thousands of complex vibrational eigenstates with tree tensor networks, arXiv preprint arXiv:2605.00998 10.48550/arXiv.2605.00998 (2026).DOI
- [92]C. Camaño, E. N. Epperly, and J. A. Tropp, Successive randomized compression: A randomized algorithm for the compressed MPO–MPS product, arXiv preprint arXiv:2504.06475, arXiv:2504.06475.arxiv.org/abs/2504.06475
- [93]F. Verstraete, M. M. Wolf, D. Pérez-García, and J. I. Cirac, Criticality, the area law, and the computational power of projected entangled pair states, Phys. Rev. Lett. 96, 220601 (2006).pubmed.ncbi.nlm.nih.gov/16803296
- [94]M. P. Zaletel and F. Pollmann, Isometric tensor network states in two dimensions, Phys. Rev. Lett. 124, 037201 (2020).DOI
- [95]Y.-Y. Shi, L.-M. Duan, and G. Vidal, Classical simulation of quantum many-body systems with a tree tensor network, Phys. Rev. A 74, 022320 (2006).DOI
- [96]Y. Wang, Z. Yang, X. Wu, and C. B. Mendl, Data repository for “A High-Accuracy Lanczos Method in the Matrix Product State Representation” (2026).DOI
- [97]Y. Kono, T. Sakakibara, C. P. Ayoyama, C. Hotta, M. M. Turnbull, C. P. Landee, and Y. Takano, Field-induced quantum criticality and universal temperature dependence of the magnetization of a spin-1/2 Heisenberg chain, Phys. Rev. Lett. 114, 037202 (2015).pubmed.ncbi.nlm.nih.gov/25659018













