betaFeedback

WorksMathematics

Tensor-Network Finite Elements for Analytic Operator Equations

Abstract

Operator equations (OEs) underpin quantitative modeling across science and engineering. Finite-element (FE) methods discretize continuous OEs into finite-dimensional algebraic systems, whereas tensor networks (TNs) provide flexible variational representations of correlated discrete systems. Here, we develop a framework that connects FE with TN for analytic OEs. The power of this method comes from its ability to convert highly non-linear partial differential equations into linear matrix equations. In particular, we show that FE discretization induces a hierarchy of multilinear interaction tensors, through which differential, integral, nonlinear, memory, and delay equations can be expressed within a common algebraic structure. The resulting systems are reformulated as weighted-residual optimization problems over TN degrees of freedom. Matrix-product-state calculations for one-dimensional linear and nonlinear diffusion reproduce conventional solutions with controlled error while preserving continuity and Neumann boundary conditions. The framework provides a common variational language for analytic OEs and establishes a direct connection between FE numerical formalism and TN variational algorithms, offering a general foundation for TN-based and quantum-inspired approaches to solving OEs.

Introduction

Operator equations (OEs) provide a general mathematical framework for quantitative modeling across science and engineering, which includes partial differential equations (PDEs), integro-differential equations (IDEs), and other nonlocal evolution equations [1, 2, 3]. Such equations arise naturally in electrodynamics, fluid dynamics, transport, quantum mechanics, statistical physics, and multi-scale materials modeling. However, only limited classes of OEs admit analytical solutions; consequently, numerical approximation methods have become indispensable for practical applications. Established discretization frameworks, including finite-difference (FD) [4, 5, 6], finite-element (FE) [7, 8, 9, 10], and spectral-element (SE) methods, can achieve high numerical accuracy, but their computational cost grows substantially as the discretization is refined and the number of degrees of freedom increases, particularly for large-scale, high-dimensional, and highly correlated systems. Among these approaches, the finite-element method (FEM) has become one of the most widely used discretization frameworks for continuous OEs [7, 8]. In conventional FEM, continuous OEs are reduced to finite-dimensional algebraic systems by expanding the solution in localized basis functions and projecting the governing equation onto the associated approximation space through a weak formulation.

Tensor networks (TNs) provide efficient variational representations of high-dimensional tensors with nontrivial correlation structure through interconnected low-rank decompositions [11]. Originally developed in quantum many-body physics, TN models such as matrix product states (MPS) [12, 13, 14], projected entangled pair states (PEPS) [15], and multiscale entanglement renormalization ansatz (MERA) [16], together with associated optimization techniques including the density matrix renormalization group (DMRG) [17, 18], have become powerful tools for representing and optimizing correlated high-dimensional systems. A central feature of TN methods is their ability to capture dominant correlation structures efficiently, often avoiding the full exponential complexity of explicit tensor representations for structured low-correlation systems. Owing to these capabilities, TN methods have found broad applications across quantum many-body physics [19, 20, 21, 22, 23, 24, 25], quantum information [26, 27, 28], machine learning [29, 30, 31, 32, 33], quantum chemistry [34, 35, 36], disordered systems [37, 38, 39], and quantum field theory [40, 41, 42]. Recent works have also explored TN approaches for PDE solving [43, 44, 45], where TN representations are primarily employed to compress discretized solution spaces and reduce computational cost [46, 47, 48, 49, 50, 51, 52].

Building on these developments, we develop a different approach to solving OEs with TNs. Rather than treating TNs as compressed surrogates of numerical solutions, we directly formulate discretized OEs in terms of TN degrees of freedom. Specifically, we generalize the conventional FEM formulation into a tensor-product representation. Through this construction, the conventional FE coefficient tensor is replaced by a structured coefficient TN. The discretized OE is then reformulated as an optimization problem over TN parameters. In this sense, TNs serve not merely as compression tools, but as variational representations of correlated FE coefficient states. Our formalism offers three key advantages: it provides a unified treatment of different classes of OEs, incorporates inter-element correlations directly into the variational space, and enables established TN algorithms to operate directly on discretized equations. More broadly, it establishes a direct connection between FEM numerical methods and TN variational algorithms, providing a general foundation for TN-based and quantum-inspired approaches to solving OEs.

The remainder of this paper is organized as follows. In Sec. II, we introduce the mathematical notations used throughout this work. In Sec. III, we formulate analytic OEs and derive their FE discretization. In Sec. IV, we develop the TN formulation from FE discretization of analytic OEs. In § 1.5, we show how the general formalism specializes to some important sub-classes of OEs. In Sec. VI, we present representative demonstrations of the method. Finally, in § 1.7, we discuss implications, limitations, and possible extensions of the proposed approach. Additional derivations and numerical results are provided in the Appendices.

Notations

Non-scalar tensors are denoted by bold Roman characters, while their components (including scalar tensors) are written in italic font with Greek tensor indices. For example, for a rank-two contravariant tensor a\mathbf{a}, its (μ,ν)(\mu,\nu) component is denoted by aμνa^{\mu\nu}. Non-tensor indices are denoted by upright Roman characters. Einstein summation convention is also assumed over repeated tensor indices.

We use bracket notation to denote sequences generated by all values of the free indices appearing inside the brackets. For any tensor component, bracket notation simply recover the tensor, e.g., [aμν]=a[a^{\mu\nu}]=\mathbf{a}. For tensors with many indices, bracket notation may also denote index strings compactly as aα1α2⋯=a[αi]a^{\alpha_{1}\alpha_{2}\cdots}=a^{[\alpha_i]} where the sequence notation does not apply to α\alpha by context. For situations in which we wish to be explicit, we will use notation of the form

∂b(a[αi])=[∂∂b[βj]a[αi]][βj].\partial_b\left(a^{[\alpha_i]}\right)=\left[\frac{\partial}{\partial b_{[\beta_j]}}a^{[\alpha_i]}\right]_{[\beta_j]}.

where a=[a[αi]]\mathbf{a}=[a^{[\alpha_i]}], b=[b[βj]]\mathbf{b}=[b^{[\beta_j]}].

We also reserve “⊗\otimes” for tensor product (TP), “⊙\odot” for tensor contraction, “⊕\oplus” for direct sum.

Finite-Element Discretization of Operator Equations

FEM reduces continuous OEs to finite-dimensional algebraic systems by approximating the solution within a finite basis and projecting the governing equation onto the approximation space through a weak formulation, which induces a hierarchy of multi-linear interaction tensors. These tensors encode the interaction geometry of the original OE, including locality, memory effects, stochastic couplings, and higher-order correlations. It therefore provides the natural starting point for our construction, converting the continuous OE into finite-dimensional residual tensors over the FE coefficients that can subsequently be represented and optimized within a TN variational space.

Weak Formulation

Let Ω\Omega denote the domain of the independent variable q\mathbf{q}, H(Ω)\mathcal{H}(\Omega) denote the corresponding admissible solution space, and R(Ω)\mathcal{R}(\Omega) denote the residual space, whose elements quantify the extent to which a trial solution fails to satisfy the governing equation. We write a general OE over Ω\Omega as

L[u]=0(1)\mathbf{L}[\mathbf{u}]=0 \tag*{(1)}

where L:H(Ω)→R(Ω)\mathbf{L}:\mathcal{H}(\Omega)\to\mathcal{R}(\Omega), and u∈H(Ω)\mathbf{u}\in\mathcal{H}(\Omega). Then, consider a finite collection of scalar basis functions bβ:Ω→Fb_\beta:\Omega\to\mathbb{F} where F\mathbb{F} indicates the scalar field. In FE representation, the solution can be approximated as

u≈cβbβ(2)\mathbf{u}\approx c^\beta b_\beta \tag*{(2)}

where [cβ][c^\beta] is the coefficient tensor. The Galerkin weak formulation would requires

∫Ωbβ∗L[u] dq=0(3)\int_{\Omega} b_\beta^*\mathbf{L}[\mathbf{u}]\,\mathrm{d}\mathbf{q}=0 \tag*{(3)}

for every test function bβb_\beta with decorator * indicates conjugation. This reduces the continuous equation to a finite system of algebraic equations for the coefficient tensor.

Analytic Operator Equations

We consider a general class of analytic OEs that admit a convergent Fréchet–Taylor expansion [53, 54, 55, 56, 57] of the form

L[u]=∑i=0∞1i!∫Ωik[ηi]⊙⨂j=1i∂qjηju dqj(4)\mathbf{L}[\mathbf{u}]=\sum_{i=0}^{\infty}\frac{1}{i!}\int_{\Omega_i}\mathbf{k}^{[\eta_i]}\mathbin{\odot}\bigotimes_{j=1}^{i}\partial_{\mathbf{q}_j}^{\eta_j}\mathbf{u}\,\mathrm{d}\mathbf{q}_j \tag*{(4)}

where ηj∈N0\eta_j\in\mathbb{N}_0, uj=u(qj)\mathbf{u}_j=\mathbf{u}(\mathbf{q}_j), and k[ηj]\mathbf{k}^{[\eta_j]} is a one-linear interaction kernel with proper contraction to each ∂qjηjuj\partial_{\mathbf{q}_j}^{\eta_j}\mathbf{u}_j as shown in Figure 1.

Image not extracted

Figure 1. Contraction of i-linear interaction kernel ki[ηj]\mathbf{k}_{i}^{[\eta_j]} with an i-point correlation in tensor network representation. The solid objects represent tensors, hanging lines represent free indices, lines connecting two tensors represents index contractions. The rank of the kernel is chosen such that it can contract each of the derivative term, ∂qjηjuj\partial_{q_j}^{\eta_j}u_j, completely with the left over rank (top branch) the same as L[u]\mathbf{L}[u].

If we further define

fi,[βj]=1i!∫Ωik[ηj]⊙⨂j=1i∂qjηjbβj dqj(5)\mathbf{f}_{i,[\beta_j]}=\frac{1}{i!}\int_{\Omega_i}\mathbf{k}^{[\eta_j]}\mathbin{\odot}\bigotimes_{j=1}^{i}\partial_{\mathbf{q}_j}^{\eta_j}b_{\beta_j}\,\mathrm{d}\mathbf{q}_j \tag*{(5)}

where bβj=bβj(qj)b_{\beta_j}=b_{\beta_j}(\mathbf{q}_j), the weak formulation can be written, for all β\beta, as

∑i=0∞gi,β,[βj]⊙⨂j=1icβj=0(6)\sum_{i=0}^{\infty}g_{i,\beta,[\beta_j]}\mathbin{\odot}\bigotimes_{j=1}^{i}c^{\beta_j}=0 \tag*{(6)}

where the discretized operator tensor is defined as

gi,β,[βj]=∫Ωbβ∗fi,[βj] dq.(7)g_{i,\beta,[\beta_j]}=\int_{\Omega}b_\beta^*\mathbf{f}_{i,[\beta_j]}\,\mathrm{d}\mathbf{q}. \tag*{(7)}

Finite-Element

Suppose that Ω\Omega is partitioned into NN finite elements Ωn\Omega_n, where n∈{1,…,N}n \in\{1,\ldots,N\}, each with local basis functions bn,βnb_{n,\beta_n} which are non-zero only in Ωn\Omega_n, and elements can have different numbers of basis functions. For cleanliness, we will slightly abuse our notation by acting as if all elements contain the same number of basis functions. With this assumption, we can substitute nn with a tensor index ν\nu, and we can replace Eq. (2) with

u≈cνβbνβ.(8)\mathbf{u} \approx\mathbf{c}^{\nu\beta}b_{\nu\beta}. \tag*{(8)}

We can cascade this to the case for analytic OEs, and get the replacement for Eq. (6) as

∑igi,νβ,[νjβj]⊙⨂j=1icνjβj=0(9)\sum_i \mathbf{g}_{i,\nu\beta,[\nu_j\beta_j]} \odot\bigotimes_{j=1}^{i}\mathbf{c}^{\nu_j\beta_j}=0 \tag*{(9)}

where gi,νβ,[νjβj]\mathbf{g}_{i,\nu\beta,[\nu_j\beta_j]} can be obtain from simple substitution of basis function indexings.

Tensor Network Formulation of Discretized Analytic Operator Equations

Our goal here is to convert a non-linear PDE into a linear matrix equation. To see what ingredients will be needed, begin by considering a much simpler problem, namely converting a linear PDE into a matrix equation. The solution is essentially trivial: any function can be represented as a linear combination of basis elements in a Hilbert space of square-integrable functions. (In practice, we will truncate the infinite set of basis elements and define the approximation space to be the span of the retained basis elements.) In this basis, the differential operator of the PDE can be readily converted into a matrix.

a matrix. If the PDE is nonlinear, however, we must be more careful. There is no clean Hilbert space representation. Instead, we must upgrade the Hilbert space to Fock space. In this way each state-vector in Fock space contains multiple copies of the Hilbert space vector, meaning that operations like multiplication of the function by e.g. itself or its derivative can then be represented as linear operators acting on Fock space, subject to certain constraints. The result is a set of constrained linear equations, which are substantially easier to solve than the original non-linear PDE.

The procedure for mapping non-linear PDEs to linear matrix equations on Fock space proceeds as follows. The FE discretization reduces an analytic OE to a finite-dimensional tensor system for the coefficient tensor. In conventional FEM, the solution space is represented as direct sum of all local basis functions, bn,βnb_{n,\beta_n}, which are used to construct the Hilbert space and, subsequently, the Fock space. As each of them is non-zero in only one element, the coefficients associated with different elements are independent variational degrees of freedom, and correlations between elements enter indirectly through the discretized operator tensors. In this section, we enlarge the coefficient representation so that correlations between finite elements can be encoded directly.

Augmented Element Spaces

The local approximation space on element n is the span of all local basis functions of that element. For normal FEM, the overall approximation Hilbert space Φ⊂H(Ω)\Phi\subset\mathcal{H}(\Omega) is the direct sum of the local approximation spaces:

Φ=⨁n=1NΦn(10)\Phi= \bigoplus_{n=1}^{N}\Phi_n \tag*{(10)}

where

Φn=Span⁡({bn,βn∣βn∈{1,…,Bn}})(11)\Phi_n = \operatorname{Span}\left(\{b_{n,\beta_n}\mid\beta_n\in\{1,\ldots,B_n\}\}\right) \tag*{(11)}

and BnB_n is the number of basis functions in element n. To build a TP representation, we use bra-ket notation from quantum mechanics, replacing each basis element by

bn,βn↦∣bn,βn⟩∈Φ.(12)b_{n,\beta_n}\mapsto|b_{n,\beta_n}\rangle\in\Phi. \tag*{(12)}

If we were only interested in linear PDEs, we could stop here. All such equations can (trivially) be projected down to matrix equations acting on elements of the approximation Hilbert space Φ\Phi. We are, however, primarily interested in the case of nonlinear PDEs. As such we define the Fock space by

Ψ=⨁i=0∞Φ⊗i.(13)\Psi=\bigoplus_{i=0}^{\infty}\Phi^{\otimes i}. \tag*{(13)}

This space has a choice of basis vectors of the form

∣[njβnj]i⟩=∣n1βn1,…,niβni⟩∈Ψ.(14)|[n_j\beta_{n_j}]_i\rangle=|n_1\beta_{n_1},\ldots,n_i\beta_{n_i}\rangle\in\Psi. \tag*{(14)}

By itself, the Fock space is not identified directly with approximation space over the physical domain. It serves as a space for variational coefficients. To obtain physical solution, one must define a projection. For the purpose of solving OEs, we will choose the projection from our physical interpretation of states in Ψ\Psi where each basis vector represents a sequence of basis functions involve in a correlation term, e.g., if u≈c~νβbνβu \approx\tilde{c}^{\nu\beta}b_{\nu\beta}, we will replace the coefficient for the standard 2-element correlation, c~ν1βν1⊗c~ν2βν2\tilde{c}^{\nu_1\beta_{\nu_1}} \otimes\tilde{c}^{\nu_2\beta_{\nu_2}}, with the coefficient c2,ν1βν1,ν2βν2c^{2,\nu_1\beta_{\nu_1},\nu_2\beta_{\nu_2}} of the basis vector ∣ν1βν1,ν2βν2⟩\lvert\nu_1\beta_{\nu_1},\nu_2\beta_{\nu_2}\rangle. Of course, any extra information of the system, e.g., symmetries of the kernels, can simplify the formal space.

With this interpretation, we can fully replace every term in Eq. (4) and get

L[u]=∑i=0∞Fi,[νjβνj]⊙ci,[νjβνj]=0(15)\mathbf{L}[u] = \sum_{i=0}^{\infty} \mathbf{F}_{i,[\nu_j\beta_{\nu_j}]} \odot\mathbf{c}^{i,[\nu_j\beta_{\nu_j}]} = 0 \tag*{(15)}

as our OE, where

∣u⟩=∑i=0∞ci,[νjβνj]∣νjβνj⟩i(16)\lvert u\rangle= \sum_{i=0}^{\infty} \mathbf{c}^{i,[\nu_j\beta_{\nu_j}]} \lvert\nu_j\beta_{\nu_j}\rangle_i \tag*{(16)}

is the formal solution in Ψ\Psi and

Fi,[νjβνj]=1i!∫Ωki[ηj]⊙⨂j=1i∂qjηjbνjβνj dq(17)\mathbf{F}_{i,[\nu_j\beta_{\nu_j}]} = \frac{1}{i!}\int_{\Omega} \mathbf{k}_{i}^{[\eta_j]} \odot\bigotimes_{j=1}^{i} \partial_{q_j}^{\eta_j} b_{\nu_j\beta_{\nu_j}}\,\mathrm{d}\mathbf{q} \tag*{(17)}

Similarly, the weak formulation can be written as

∑i=0∞Gi,β,[νjβνj]⊙ci,[νjβνj]=0(18)\sum_{i=0}^{\infty} \mathbf{G}_{i,\beta,[\nu_j\beta_{\nu_j}]} \odot\mathbf{c}^{i,[\nu_j\beta_{\nu_j}]} = 0 \tag*{(18)}

where

Gi,β,[νjβνj]=∫Ωbβ∗Fi,[νjβνj] dq.(19)\mathbf{G}_{i,\beta,[\nu_j\beta_{\nu_j}]} = \int_{\Omega} b_{\beta}^{*}\mathbf{F}_{i,[\nu_j\beta_{\nu_j}]}\,\mathrm{d}\mathbf{q}. \tag*{(19)}

However, this finite system of algebraic equations has many more coefficients than Eq. (9). Hence, extra conditions, which can take the form of regularization or full-blown equations, are required.

Tensor-Network Representation

Due to the ii-linearity of ki[ηj]\mathbf{k}_{i}^{[\eta_j]}, we have gauge freedom to permute the pair indices (νj,βνj)(\nu_j,\beta_{\nu_j}) of Gi,β,[νjβνj]\mathbf{G}_{i,\beta,[\nu_j\beta_{\nu_j}]} and ci,[νjβνj]\mathbf{c}^{i,[\nu_j\beta_{\nu_j}]}. We chose to follow a canonical ordering such that for any j′>jj' > j, νj′≥νj\nu_{j'} \ge\nu_j, i.e., in order of element labels. Furthermore, if νj′=νj\nu_{j'} = \nu_j, ηj′≥ηj\eta_{j'} \ge\eta_j, i.e., in order of derivative order as well.

We can also tensorized the summation over ii by introducing, for each element, empty basis function bν0b_{\nu_0} into the local approximation space basis set in Eq. (11). With this addition and let JJ be the largest correlation order contribution of the analytic OE from any element, we can map each basis vector in Ψ\Psi to a new representation such that each element is pre-allocated with JJ slots and the basis orders that appear in the basis vector are filled in the slots of their corresponding elements in the same order, e.g., if J=3J = 3 and N=4N = 4, the basis vector ∣13,14,21,22,33⟩\lvert13,14,21,22,33\rangle is mapped to

∣(3,4,0),(1,2,0),(3,0,0),(0,0,0)⟩,\lvert(3,4,0),(1,2,0),(3,0,0),(0,0,0)\rangle,

i.e., 2 contributions from element 1, 2 contributions from element 2, 1 contribution from element 3, and no contribution from element 4. Hence, Eq. (18) can be replaced with

Gβ,[βνj]⊙c[βνj]=0(20)\mathbf{G}_{\beta,[\beta_{\nu_j}]} \odot\mathbf{c}^{[\beta_{\nu_j}]} = 0 \tag*{(20)}

with appropriate substitution of Gβ,[βνj]\mathbf{G}_{\beta,[\beta_{\nu_j}]}’s definition. The equation can be neatly represented in tensor network format as shown in Figure 2.

Image not extracted

Figure 2. Tensor network representation of the left hand side of Eq. (20). Unlike in Figure 1, we omit the obvious tensor indices β\beta and [βνj][\beta_{\nu_j}] for cleanliness.

Effectively, one can directly solve Eq. (20) and obtain the desired approximate solution. However, directly storing the full coefficient tensor c\mathbf{c} classically requires O(BN)O(B^N) coefficient blocks, where B=max⁡({Bn})B = \max(\{B_n\}) which grows exponentially with the problem size. The full tensor-product space therefore cannot be introduced as a direct storage format, but as a formal correlated approximation space to be compressed by any TN ansatz.

The choice of TN ansatz is arbitrary in the perspective of our formalism, but, if one already has expectation for the solution to contain some correlation and/or entanglement behaviors, the network can be designed to support them. One crucial assumption is the locality of correlation: we expected the correlation to only appear between neighboring elements. Figure 3 illustrates representative TN ansatzes under different local correlation geometries.

Image not extracted

Figure 3. Tensor network ansatz of coefficient tensor with different locality assumptions: (a) no assumption, (b) matrix product state, (c) projected entangled pair state, and (d) higher-order tensor network assumption where shaded regions represent index hyper-contraction. β\beta’s and σ\sigma’s indicate tensor indices responsible for each bond and contraction, respectively.

Similar to Figure 2, Figure 3a is the generic tensor representation of Figure 3b shows a matrix product state (MPS) which is appropriate for one-dimensional element orderings with nearest-neighbor correlations while Figure 3c shows a projected entangled pair state (PEPS) structure adapted for two-dimensional grid-like element connectivity. Figure 3d illustrates a higher-order or hypergraph-type (HG) tensor network capable of representing multi-element correlations beyond pair-wise interactions.

The bond dimensions (ranges of bond indices) control the amount of inter-element correlation that can be represented. If the maximum bond dimension in the TN is SS, and the maximum bond degree (including hyperbond) in the TN is MM, then the storage requirement is

O(NBJSM)(21)O(NB^{J}S^{M}) \tag*{(21)}

instead of O(BNJ)O(B^{NJ}). Therefore, as long as the bond degree and correlation order are capped by a constant, the TN ansatz defines a polynomial-size variational manifold inside the exponentially large correlated coefficient space.

Least-Squares Variational Formulation

Given any formal coefficient tensor c\mathbf{c}, the residual from weak formulation can be defined as

hβ(c)=Gβ,[βνj]⊙c[βνj].(22)\mathbf{h}_{\beta}(\mathbf{c})=\mathbf{G}_{\beta,[\beta_{\nu j}]}\mathbin{\odot}\mathbf{c}^{[\beta_{\nu j}]}. \tag*{(22)}

We determine the correlated coefficients by minimizing the weighted residual norm

J(c)=Wβ′β(hβ′†(c)⊙hβ(c)),(23)J(\mathbf{c})=W^{\beta'\beta}\left(\mathbf{h}_{\beta'}^{\dagger}(\mathbf{c})\mathbin{\odot}\mathbf{h}_{\beta}(\mathbf{c})\right), \tag*{(23)}

where Wβ′βW^{\beta'\beta} is a positive-definite metric on the residual space. The TN approximation is obtained by restricting c\mathbf{c} to a chosen TN ansatz variational manifold TTN\mathcal{T}_{\mathrm{TN}}:

cTN=arg⁡min⁡c∈TTN J(c).(24)\mathbf{c}_{\mathrm{TN}}=\underset{\mathbf{c}\in\mathcal{T}_{\mathrm{TN}}}{\arg\min}\ J(\mathbf{c}). \tag*{(24)}

Although this objective is quadratic and convex in the unrestricted correlated coefficient tensor c\mathbf{c}, the TN parametrization makes the global optimization generally non-convex. However, when all tensors in TN except one are fixed, the residual becomes linear in the active tensor. Therefore, each single-site local update reduces the optimization to a convex weighted least-squares sub-problem which is part of the reason for the success of density matrix renormalization group (DMRG) algorithm for MPS ansatz.

Specialization to Important Operator-Equation Classes

The formulation above was stated for analytic operator equations in general form. In this section, we show how several common equation classes can be reduced to the same weak formulation structure instead of introducing different methods for different equations.

Partial Differential Equations

A partial differential equation (PDE) can be written as

g(q,[∂qηu])=0.(25)\mathbf{g}\left(\mathbf{q},[\partial_{\mathbf{q}}^{\eta}\mathbf{u}]\right)=0. \tag*{(25)}

If g\mathbf{g} is analytic, then it can be expanded as a convergent series

g(q,[∂qηu])=∑i1i!ai[ηj]⊙⨂j=1i∂qηju,(26)\mathbf{g}\left(\mathbf{q},[\partial_{\mathbf{q}}^{\eta}\mathbf{u}]\right)=\sum_{i}\frac{1}{i!}\mathbf{a}_{i}^{[\eta_{j}]}\mathbin{\odot}\bigotimes_{j=1}^{i}\partial_{\mathbf{q}}^{\eta_{j}}\mathbf{u}, \tag*{(26)}

which is a special case of Eq. (4) when

ki[ηj]=ai[ηj]⨂j=1iδ(qj−q).(27)\mathbf{k}_{i}^{[\eta_{j}]}=\mathbf{a}_{i}^{[\eta_{j}]}\bigotimes_{j=1}^{i}\delta(\mathbf{q}_{j}-\mathbf{q}). \tag*{(27)}

Initial-Value Problems

An initial-value problem (IVP) is simply a variation of PDE with time as one of the domain dimensions. Hence, it is of the form

g(t,q,[∂qη∂tζu])=0.(28)\mathbf{g}\left(t,\mathbf{q},[\partial_{\mathbf{q}}^{\eta}\partial_{t}^{\zeta}\mathbf{u}]\right)=0. \tag*{(28)}

If g\mathbf{g} is analytic,

g(t,q,[∂qη∂tζu])=∑i1i!ai[ζk],[ηj]⊙⨂j=1i⨂k=j+1i∂qηj∂tζku.(29)\mathbf{g}\left(t,\mathbf{q},[\partial_{\mathbf{q}}^{\eta}\partial_{t}^{\zeta}\mathbf{u}]\right) =\sum_{i}\frac{1}{i!}\mathbf{a}_{i}^{[\zeta_{k}],[\eta_{j}]}\mathbin{\odot}\bigotimes_{j=1}^{i}\bigotimes_{k=j+1}^{i}\partial_{\mathbf{q}}^{\eta_{j}}\partial_{t}^{\zeta_{k}}\mathbf{u}. \tag*{(29)}

For IVP, given u′\mathbf{u}' and some of its time derivatives at time tt, the problem is to determine u\mathbf{u} and those time derivatives at t+Δtt+\Delta t. There are many schemes one can use with varying degree of convergence criteria and computational costs. We will use a simple implicit Euler’s scheme as an example that replaces the analytic form of the IVP with

0=∑i1i!ai[ζk],[ηj]⊙⨂j=1i⨂k=j+1i∂qηj{1Δt(∂tζk−1u−∂tζk−1u′),ζk≠0,u,ζk=0.(30)0=\sum_{i}\frac{1}{i!}\mathbf{a}_{i}^{[\zeta_{k}],[\eta_{j}]}\mathbin{\odot}\bigotimes_{j=1}^{i}\bigotimes_{k=j+1}^{i}\partial_{\mathbf{q}}^{\eta_{j}} \begin{cases} \frac{1}{\Delta t}\left(\partial_{t}^{\zeta_{k}-1}\mathbf{u}-\partial_{t}^{\zeta_{k}-1}\mathbf{u}'\right), & \zeta_{k}\ne0,\\ \mathbf{u}, & \zeta_{k}=0. \end{cases} \tag*{(30)}

With some term arrangement, the equation can be written as a PDE on v=[∂tζu]\mathbf{v}=\left[\partial_t^{\zeta}\mathbf{u}\right], a special case of Eq. (4).

Integro-Differential Equations

An integro-differential equation (IDE) is a generalization of PDE to include non-local interaction mostly in the form of integration kernel as

g(q,[∂qηu],[∫Ωiai(q,[qj],[∂qjηu])∏j=1idqj])=0.(31)\mathbf{g}\left(\mathbf{q},\left[\partial_{\mathbf{q}}^{\eta}\mathbf{u}\right],\left[\int_{\Omega_i}a_i\left(\mathbf{q},[\mathbf{q}_j],\left[\partial_{\mathbf{q}_j}^{\eta}\mathbf{u}\right]\right)\prod_{j=1}^{i}d\mathbf{q}_j\right]\right)=0. \tag*{(31)}

If g\mathbf{g}, and all aia_i's are analytic, then it can be expanded as a special case of Eq. (4) with multi-linear interaction kernel containing some Dirac delta functions for the parts that are not integrations in the IDE.

Memory and Delay Equations

A memory-delay equation (MDE) is an IDE with non-local interaction in temporal dimension, i.e.,

0=g(t,q,[∂qη∂tζu],[∫τiai(t,[tj],q,[∂qη∂tjζu])∏j=1idtj]).(32)0=\mathbf{g}\left(t,\mathbf{q},\left[\partial_{\mathbf{q}}^{\eta}\partial_t^{\zeta}\mathbf{u}\right],\left[\int_{\tau_i}a_i\left(t,[t_j],\mathbf{q},\left[\partial_{\mathbf{q}}^{\eta}\partial_{t_j}^{\zeta}\mathbf{u}\right]\right)\prod_{j=1}^{i}dt_j\right]\right). \tag*{(32)}

Similar to the IDE, if g\mathbf{g}, and all aia_i's are analytic, then it can be expanded as a special case of Eq. (4).

Demonstration

The formalism developed above applies to a broad class of analytic OEs across different geometries and TN ansatzes. To illustrate the framework concretely, we specialize to one-dimensional (1D) problems which can be naturally discretized into an MPS variational problems since there is a well-behave solving algorithm (DMRG) for them. We demonstrate the approach on IVP and examine the convergence and accuracy of the resulting TN solutions.

Specialization

As a continuation from Sec. VB, the ending note of that sub-section implies that we only need to consider the IVPs with only first order time derivatives. Also, instead of u′u' and uu, we will use u(t−1)u^{(t-1)} and u(t)u^{(t)} notations where labels in parentheses indicate time step. Hence,

for 1D, we are ultimately solving for u(t)\mathbf{u}^{(t)} in

0=∑i1i!ai[ζk],[ηj]∏j=1i∏k=j+1i∂qjηj{1Δt(u(t)−u(t−1)),ζk=1u(t),ζk=0.(33)0=\sum_i\frac{1}{i!}a_i^{[\boldsymbol{\zeta}_k],[\boldsymbol{\eta}_j]}\prod_{j=1}^{i}\prod_{k=j+1}^{i}\partial_{q_j}^{\eta_j}\begin{cases}\dfrac{1}{\Delta t}\left(u^{(t)}-u^{(t-1)}\right),&\zeta_k=1\\u^{(t)},&\zeta_k=0\end{cases}. \tag*{(33)}

which is a PDE but not in the same format as Eq. (26): the existence of u(t−1)u^{(t-1)} which can be treated as part of another term with lower correlation order of u(t)u^{(t)}.

It is straight forward to simply use the weak formalism directly on the equation, but it means that we need to calculate G\mathbf{G} from scratch at every time step since u(t−1)u^{(t-1)} got updated. Instead, we can leverage the deterministic structure of the IVP to pre-compute some parts. Different classes of OEs will be different but for IVE and, to extension, PDE the spatial derivatives and multiplications can be pre-formulated and approximated into concise weak forms.

As spatial derivatives and multiplications are local operators, with well-distributed basis function, any of their actions on Φn\Phi_n can be approximated with a member of Φn\Phi_n. In other words, we can approximate the spatial derivatives and multiplications with some tensor contractions.

Consider an arbitrary spatial derivative of a basis function bn,βb_{n,\beta} which can be approximated as

∂qbn,β≈mβ′′bn,β′′,(34)\partial_q b_{n,\beta}\approx m^{\beta''}b_{n,\beta''}, \tag*{(34)}

which, by the weak formulation, gives

∫Ωbn,β∗∂qbn,β dq≈mβ′∫Ωbn,β′∗bn,β′′ dq.(35)\int_{\Omega}b_{n,\beta}^{*}\partial_qb_{n,\beta}\,d\mathbf{q}\approx m^{\beta'}\int_{\Omega}b_{n,\beta'}^{*}b_{n,\beta''}\,d\mathbf{q}. \tag*{(35)}

We then define

Pn,β′β=∫Ωbn,β′∗∂qbn,β dq,(36)P_{n,\beta'\beta}=\int_{\Omega}b_{n,\beta'}^{*}\partial_qb_{n,\beta}\,d\mathbf{q}, \tag*{(36)}
Qn,β′β′′=∫Ωbn,β′∗bn,β′′ dq,(37)Q_{n,\beta'\beta''}=\int_{\Omega}b_{n,\beta'}^{*}b_{n,\beta''}\,d\mathbf{q}, \tag*{(37)}

which can be pre-computed given the basis functions. If Qn,β′β′′−1Q_{n,\beta'\beta''}^{-1} is the matrix inverse of Qn,β′β′′Q_{n,\beta'\beta''}. Then, ∂q\partial_q on the element nn can be approximated with

Dn,β′β′′≈Qn,β′β′′−1Pn,β′β.(38)D_{n,\beta'\beta''}\approx Q_{n,\beta'\beta''}^{-1}P_{n,\beta'\beta}. \tag*{(38)}

We can also visualize it in the tensor circuit form as shown in Figure 4a. Note that the label nn is needed since each element has different sets of basis functions. Also, the index qq of the derivative tensor indicate the axis of differentiation.

Image not extracted

Figure 4. Tensor circuit representation of (a) spatial derivative and (b) multiplication act locally at element n.

Similarly, consider an arbitrary spatial multiplication of a pair of basis functions bn,β1b_{n,\beta_1} and bn,β2b_{n,\beta_2} which can be approximated as

bn,β1bn,β2≈mβ′′bn,β′′.(39)b_{n,\beta_1}b_{n,\beta_2}\approx m^{\beta''}b_{n,\beta''}. \tag*{(39)}

If we define,

Pn,β′β1β2′=∫Ωbn,β′∗bn,β1bn,β2 dq(40)P'_{n,\beta'\beta_1\beta_2}=\int_{\Omega} b^*_{n,\beta'}b_{n,\beta_1}b_{n,\beta_2}\,\mathrm{d}q \tag*{(40)}
Qn,β′β′′′=∫Ωbn,β′∗bn,β′′ dq(41)Q'_{n,\beta'\beta''}=\int_{\Omega} b^*_{n,\beta'}b_{n,\beta''}\,\mathrm{d}q \tag*{(41)}

then, any multiplication in this elemental basis can be approximated with tensor product reduction

Rn,β′′β1β2=Qn,β′′β′′−1Pn,β′β1β2′(42)R_{n,\beta''\beta_1\beta_2}=Q_{n,\beta''\beta'}^{\prime-1}P'_{n,\beta'\beta_1\beta_2} \tag*{(42)}

as depicted in Figure 4b.

With these components, we can construct other more complicated IVP operators, e.g., second derivative, dot product, cross product, divergence, curl, and Laplacian, as illustrated in Figure 5. Of course, directly applying the weak formalism for each operator would give a better approximation at the cost of additional overhead calculations per time step.

Image not extracted

Figure 5. Tensor circuit representation of more complicated operators can be constructed from the two basic building block types (spatial derivative and multiplication) together with well known constant tensors like Kronecker delta (δ\delta) and Levi-Civita (ϵ\epsilon). (a) Second spatial derivative or Hessian can be achieve with stacking two D’s with both spatial indices free. (b) Dot and (c) cross products are multiplications while (d) Divergence and (e) curl are first spatial derivatives with additional spatial index summation by δ\delta and ϵ\epsilon. Lastly, (f) Laplacian is second spatial derivatives with additional spatial index summation by δ\delta. The subscript nn make it explicitly clear that the derivatives and multiplications act locally at element n.

Example

We’ll apply the procedure of converting IVPs to TN with the following non-linear diffusion equations

∂tu=up∂x2u(43)\partial_t u=u^p\partial_x^2u \tag*{(43)}

with u=u(t,x)u=u(t,x), p∈{0,1,2}p\in\{0,1,2\}, Ω=[0,1]\Omega=[0,1], u(0,x)u(0,x) is a modified Gaussian centered at x=0.5x=0.5 with ∂xu(t,0)=∂xu(t,1)=0\partial_xu(t,0)=\partial_xu(t,1)=0, i.e., close boundaries. We can rewrite the IVP in our format as

0=u(t)p∂x2u(t)−1Δt(u(t)−u(t−1)).(44)0=u^{(t)p}\partial_x^2u^{(t)}-\frac{1}{\Delta t}\left(u^{(t)}-u^{(t-1)}\right). \tag*{(44)}

The corresponding tensor circuit of this IVP is as shown in Figure 6.

Figure 6. Tensor circuit representation of residual of non-linear diffusion IVP. Since IVPs only contain local interactions, the circuit elements (GnG_n) only connect branches at the same elements. The circuit in red box illustrates the detail of GNG_N. Since the first term in Eq. (44) is up(t)∂x2c(t)u^{(t)}_p\partial_x^2 c^{(t)}, there are 2 derivatives tensor on c(t)c^{(t)} and pp multiplications between p(t)p^{(t)}’s and output branch from the second derivative of c(t)c^{(t)} in the red box. The boxes containing −1-1 and 1/Δt1/\Delta t implies multiplication by those values and branch merging implies addition of merged branches which corresponding to the second term in Eq. (44).

Then, we solve the non-linear diffusion IVP for each exponent pp using NN uniform finite elements on Ω\Omega, each equipped with BB local Lagrange polynomial basis functions, and an MPS ansatz with bond dimension SS for the coefficient tensor. The time evolution is advanced using the implicit Euler scheme described with time step Δt\Delta t. At each time step, the DMRG sweep algorithm minimizes the weighted residual J(c)J(c) over the MPS variational manifold TMPS\mathcal{T}_{\mathrm{MPS}}.

To ensure the continuity and smoothness of the results, we also add an additional circuit to measure the residual of multiple derivative orders at each boundary between elements as well as the residual of Neumann boundary conditions at the outer edges of the domain to environment. As these continuities only involve at most 2 adjacent elements, the circuit blocks either only connect adjacent tensor index branches (element boundaries) or only connect to one branch (domain edges).

Figure 7 summarize the time-evolution results across all three cases for explicit Runge–Kutta IVP simulation (top row) and TN variational (middle row) method. For p=0p=0 (left column), the equation reduces to the standard diffusion equation ∂tu=∂x2u\partial_tu=\partial_x^2u, which admits a smooth, rapidly decaying solution. For p=1p=1 (middle column), the non-linear coupling u∂x2uu\partial_x^2u introduces amplitude-dependent diffusion, slowing the spread of uu in regions of low amplitude. Finally, for p=2p=2 (right column), the stronger non-linearity further suppresses diffusion and produces sharper features. In all cases, the TN variational method with N=10N = 10, B=4B = 4, S=1S = 1 produce a reliable results with the difference of less than two percent of the original amplitude of the modified Gaussian as shown in the bottom row of Figure 7. We also see that the Neumann boundary conditions at both ends of the domain as well as the continuity and smoothness of the results are also respected to within numerical precision confirming that the weak formulation correctly encodes the boundary constraints.

Image not extracted

Figure 7. Numerical simulation results for a non-linear diffusion IVPs (Eq. (44)) with different pp’s: 0 (left column), 1 (middle column), 2 (right column). Top row shows result from explicit Runge–Kutta method of order 5(4)[58] while middle row are the results from our TN variational method with N=11N=11, B=4B=4, S=1S=1 with their differences is shown in the bottom row..

Convergence and Accuracy

To assess the convergence properties of the TN solver, we study the behavior of the solution error as functions of Lagrange polynomial basis order BB, and the number of finite elements NN at different time step size Δt\Delta t. Since both BB and NN are, in some way, representation of spatial resolution, we should expect the behavior of the trend to be similar to standard finite difference method which exhibit the stability condition where the convergence cannot be improved purely by increasing spatial resolution while the temporal resolution is fixed. This is clearly shown in Figure 8.

Image not extracted

Figure 8. Mean absolute error trend of TN variational method for solving diffusion IVP with different (a) basis order BB and (b) number of elements NN at different time step size Δt\Delta t. For each fixed Δt\Delta t, the trends exhibit convergence up to a level after which the better temporal resolution is required for a more accurate result.

Together, these results demonstrate that the TN framework introduced in this paper can produce accurate, converged solutions to non-linear IVPs with the computational scaling controlled by the bond dimension rather than the full exponential size of the coefficient space. While the examples here are restricted to 1D MPS/DMRG for clarity, the formalism is directly extensible to higher-dimensional problems via PEPS or tree tensor network ansatzes.

Conclusion

We have introduced a tensor-network formulation for finite-element discretizations of analytic operator equations. Starting from the Galerkin weak formulation, analytic operator equations generate a hierarchy of multilinear residual tensors determined by the operator kernels, basis functions, and element geometry. By lifting the conventional finite-element coefficient space into an augmented tensor-product space over elements, the discretized problem can be expressed as a variational optimization over correlated coefficient states. Tensor networks then provide a structured parametrization of this enlarged space, allowing inter-element correlations to be represented without explicitly storing the exponentially large full coefficient tensor. The key distinction of this framework is that tensor networks are used not only as compression tools for already-discretized solutions, but as variational representations of the finite-element coefficient space itself. This perspective establishes a common formulation for a broad class of analytic operator equations, including partial differential equations, initial-value problems, integro-differential equations, and memory or delay equations. Different equation classes enter through different interaction kernels and residual tensors, while the variational tensor-network structure remains the same. For initial-value problems, we showed that implicit time discretization reduces the evolution to a sequence of stationary variational problems. Spatial derivative and multiplication operators can be precomputed at the element level and reused across time steps, reducing repeated assembly costs. The one-dimensional nonlinear diffusion examples demonstrate that an MPS-based formulation can reproduce standard numerical solutions with controlled accuracy while respecting continuity and boundary constraints. The present results should be viewed as a proof of principle rather than a guarantee of universal computational advantage. The efficiency of the method depends on whether the solution admits a compact tensor-network representation. Highly discontinuous, chaotic, or strongly correlated solutions may require large bond dimensions, and accuracy remains sensitive to both local basis quality and the chosen network geometry. Future work should investigate adaptive bond-dimension control, higher-dimensional tensor-network ansatzes such as PEPS and tree networks, rigorous convergence estimates, and comparisons with conventional finite-element solvers on large-scale benchmark problems. Overall, the proposed framework provides a direct bridge between finite-element approximation theory, multi-linear operator discretization, and tensor-network variational optimization. It offers a flexible foundation dation for classical, quantum-inspired, and potentially quantum-assisted solvers for structured operator equations.

Acknowledgments

The authors thank Alan Edelman and Nuno Loureiro for insightful discussions. We dedicate this work to Loureiro’s memory and note with sadness that he passed away during the course of this work. This work was supported by the National Science Foundation (NSF) Convergence Accelerator Award No. 2345084, the U.S. Department of Energy (DOE), Basic Energy Sciences (BES), Award No. DE-SC0020148, MIT Energy Initiative, and support from R. Wachnik.

References

References

  1. [1]M. A. Krasnosel’skii, G. M. Vainikko, P. P. Zabreiko, Ya. B. Rutitskii, and V. Ya. Stetsenko. Approximate Solution of Operator Equations. Wolters-Noordhoff Publishing, Groningen, 1972.
  2. [2]Kendall Atkinson and Weimin Han. Theoretical Numerical Analysis: A Functional Analysis Framework, volume 39 of Texts in Applied Mathematics. Springer, New York, 3 edition, 2009.
  3. [3]Haim Brezis. Functional Analysis, Sobolev Spaces and Partial Differential Equations. Universitext. Springer, New York, 2011.DOI
  4. [4]Randall J. LeVeque. Finite Difference Methods for Ordinary and Partial Differential Equations: Steady-State and Time-Dependent Problems. Classics in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2007.
  5. [5]John C. Strikwerda. Finite Difference Schemes and Partial Differential Equations. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2 edition, 2004.
  6. [6]J. W. Thomas. Numerical Partial Differential Equations: Finite Difference Methods, volume 22 of Texts in Applied Mathematics. Springer, New York, 1995.DOI
  7. [7]Philippe G. Ciarlet. The Finite Element Method for Elliptic Problems, volume 40 of Classics in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2002. Reprint of the 1978 North-Holland edition.
  8. [8]Susanne C. Brenner and L. Ridgway Scott. The Mathematical Theory of Finite Element Methods, volume 15 of Texts in Applied Mathematics. Springer, New York, 3 edition, 2008.
  9. [9]Thomas J. R. Hughes. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Dover Publications, Mineola, NY, 2000. Reprint of the 1987 Prentice-Hall edition.
  10. [10]O. C. Zienkiewicz, R. L. Taylor, and J. Z. Zhu. The Finite Element Method: Its Basis and Fundamentals. Butterworth-Heinemann (Elsevier), Oxford, 7 edition, 2013.
  11. [11]Román Orús. A practical introduction to tensor networks: Matrix product states and projected entangled pair states. Annals of Physics, 349:117–158, 2014.DOI
  12. [12]M. Fannes, B. Nachtergaele, and R. F. Werner. Finitely correlated states on quantum spin chains. Communications in Mathematical Physics, 144(3):443–490, 1992.DOI
  13. [13]A. Klumper, A. Schadschneider, and J. Zittartz. Equivalence and solution of anisotropic spin-1 models and generalized t-j fermion models in one dimension. Journal of Physics A: Mathematical and General, 24(16):L955, 1991.
  14. [14]A. Klümper, A. Schadschneider, and J. Zittartz. Matrix product ground states for one-dimensional spin-1 quantum antiferromagnets. Europhysics Letters, 24(4):293, 1993.DOI
  15. [15]F. Verstraete and J. I. Cirac. Renormalization algorithms for quantum-many body systems in two and higher dimensions. arXiv, 2004.
  16. [16]G. Vidal. Entanglement renormalization. Physical Review Letters, 99(22):220405, 2007.DOI
  17. [17]Steven R. White. Density matrix formulation for quantum renormalization groups. Phys. Rev. Lett., 69:2863–2866, Nov 1992.pubmed.ncbi.nlm.nih.gov/10046608
  18. [18]Steven R. White. Density-matrix algorithms for quantum renormalization groups. Physical Review B, 48(14):10345–10356, 1993.pubmed.ncbi.nlm.nih.gov/10007313
  19. [19]H. C. Jiang, Z. Y. Weng, and D. N. Sheng. Density matrix renormalization group numerical study of the kagome antiferromagnet. Phys. Rev. Lett., 101:117203, Sep 2008.pubmed.ncbi.nlm.nih.gov/18851324
  20. [20]Thibaut Picot, Marc Ziegler, Román Orús, and Didier Poilblanc. Spin-s kagome quantum antiferromagnets in a field with tensor networks. Phys. Rev. B, 93:060407, Feb 2016.DOI
  21. [21]Philippe Corboz. Improved energy extrapolation with infinite projected entangled-pair states applied to the two-dimensional hubbard model. Phys. Rev. B, 93:045116, Jan 2016.DOI
  22. [22]J. Dubail and N. Read. Tensor network trial states for chiral topological phases in two dimensions and a no-go theorem in any dimension. Physical Review B, 92(20):205307, 2015.DOI
  23. [23]Miguel Aguado and Guifré Vidal. Entanglement renormalization and topological order. Phys. Rev. Lett., 100:070404, Feb 2008.DOI
  24. [24]Oliver Buerschaper, Miguel Aguado, and Guifré Vidal. Explicit tensor network representation for the ground states of string-net models. Physical Review B, 79(8):085119, 2009.DOI
  25. [25]Zheng-Cheng Gu, Michael Levin, Brian Swingle, and Xiao-Gang Wen. Tensor-product representations for string-net condensed states. Physical Review B, 79(8):085118, 2009.DOI
  26. [26]Igor L. Markov and Yaoyun Shi. Simulating quantum computation by contracting tensor networks. SIAM Journal on Computing, 38(3):963–981, 2008.DOI
  27. [27]Andrew J. Ferris and David Poulin. Tensor networks and quantum error correction. Phys. Rev. Lett., 113:030501, Jul 2014.DOI
  28. [28]William Huggins, Piyush Patil, Bradley Mitchell, K Birgitta Whaley, and E Miles Stoudenmire. Towards quantum machine learning with tensor networks. Quantum Science and Technology, 4(2):024001, jan 2019.DOI
  29. [29]Edwin Stoudenmire and David J Schwab. Supervised learning with tensor networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016.arxiv.org/abs/1605.05775
  30. [30]Zhao-Yu Han, Jun Wang, Heng Fan, Lei Wang, and Pan Zhang. Unsupervised generative modeling using matrix product states. Phys. Rev. X, 8:031012, Jul 2018.DOI
  31. [31]Yoav Levine, Or Sharir, Nadav Cohen, and Amnon Shashua. Quantum entanglement in deep learning architectures. Physical Review Letters, 122(6):065301, 2019.DOI
  32. [32]Ding Liu, Shi-Ju Ran, Peter Wittek, Cheng Peng, Raul Blázquez García, Gang Su, and Maciej Lewenstein. Machine learning by unitary tensor network of hierarchical tree structure. New Journal of Physics, 21(7):073059, jul 2019.DOI
  33. [33]Ángel J. Gallego and Román Orús. Language design as information renormalization. SN Computer Science, 3(2):140, Jan 2022.DOI
  34. [34]Garnet Kin-Lic Chan, Jonathan J. Dorando, Debashree Ghosh, Johannes Hachmann, Eric Neuscamman, Haitao Wang, and Takeshi Yanai. An introduction to the density matrix renormalization group ansatz in quantum chemistry. In Frontiers in quantum systems in chemistry and physics, pages 49–65. Springer, 2008.arxiv.org/abs/0711.1398
  35. [35]Szilárd Szalay, Max Pfeffer, Valentin Murg, Gergely Barcza, Frank Verstraete, Reinhold Schneider, and Örs Legeza. Tensor product methods and entanglement optimization for ab initio quantum chemistry. International Journal of Quantum Chemistry, 115(19):1342–1391, 2015.DOI
  36. [36]C. Krumnow, L. Veis, Ö. Legeza, and J. Eisert. Fermionic orbital optimization in tensor network states. Phys. Rev. Lett., 117:210402, Nov 2016.pubmed.ncbi.nlm.nih.gov/27911544
  37. [37]B. Paredes, F. Verstraete, and J. I. Cirac. Exploiting quantum parallelism to simulate quantum random many-body systems. Phys. Rev. Lett., 95:140501, Sep 2005.DOI
  38. [38]A. Chandran, J. Carrasquilla, I. H. Kim, D. A. Abanin, and G. Vidal. Spectral tensor networks for many-body localization. Phys. Rev. B, 92:024201, Jul 2015.DOI
  39. [39]Frank Pollmann, Vedika Khemani, J. Ignacio Cirac, and S. L. Sondhi. Efficient variational diagonalization of fully many-body localized hamiltonians. Phys. Rev. B, 94:041116, Jul 2016.DOI
  40. [40]F. Verstraete and J. I. Cirac. Continuous matrix product states for quantum fields. Physical Review Letters, 104(19):190405, 2010.DOI
  41. [41]Jutho Haegeman, Tobias J. Osborne, Henri Verschelde, and Frank Verstraete. Entanglement renormalization for quantum fields in real space. Physical Review Letters, 110(10):100402, 2013.pubmed.ncbi.nlm.nih.gov/23521233
  42. [42]Brian Swingle. Entanglement renormalization and holography. Physical Review D, 86(6):065007, 2012.DOI
  43. [43]Markus Bachmayr, Reinhold Schneider, and André Uschmajew. Tensor networks and hierarchical tensors for the solution of high-dimensional partial differential equations. Foundations of Computational Mathematics, 16(6):1423–1472, Dec 2016.DOI
  44. [44]Nikita Verstraete, Michael Lubasch, Sergey Dolgov, Quincy Y. van den Berg, Hessam Babaee, Peyman Givi, Martin Kiffner, and Dieter Jaksch. A quantum-inspired approach to exploit turbulence structures. Nature Computational Science, 2(1):30–37, 1 2022.DOI
  45. [45]Erika Ye and Nuno F. G. Loureiro. Quantum-inspired method for solving the vlasov-poisson equations. Phys. Rev. E, 106:035208, Sep 2022.DOI
  46. [46]Michael Lubasch, Pierre Moinier, and Dieter Jaksch. Multigrid renormalization. Journal of Computational Physics, 372:587–602, 2018.DOI
  47. [47]Juan José García-Ripoll. Quantum-inspired algorithms for multivariate analysis: from interpolation to partial differential equations. Quantum, 5:431, 4 2021.DOI
  48. [48]Yudong Cao, Anargyros Papageorgiou, Iasonas Petras, Joseph Traub, and Sabre Kais. Quantum algorithm and circuit design solving the poisson equation. New Journal of Physics, 15(1):013021, 1 2013.DOI
  49. [49]Pedro C. S. Costa, Stephen Jordan, and Aaron Ostrander. Quantum algorithm for simulating the wave equation. Phys. Rev. A, 99:012323, 1 2019.DOI
  50. [50]Abtin Ameri, Erika Ye, Paola Cappellaro, Hari Krovi, and Nuno F. Loureiro. Quantum algorithm for the linear vlasov equation with collisions. Phys. Rev. A, 107:062412, 6 2023.DOI
  51. [51]Hai-Ling Liu, Yu-Sen Wu, Lin-Chun Wan, Shi-Jie Pan, Su-Juan Qin, Fei Gao, and Qiao-Yan Wen. Variational quantum algorithm for the poisson equation. Phys. Rev. A, 104:022418, 8 2021.DOI
  52. [52]Oleksandr Kyriienko, Annie E. Paine, and Vincent E. Elfving. Solving nonlinear differential equations with differentiable quantum circuits. Phys. Rev. A, 103:052416, 5 2021.DOI
  53. [53]Stephen Boyd, Leon O. Chua, and Charles A. Desoer. Analytical foundations of Volterra series. IMA Journal of Mathematical Control and Information, 1(3):243–282, 1984.DOI
  54. [54]Stephen Boyd and Leon O. Chua. Fading memory and the problem of approximating nonlinear operators with Volterra series. IEEE Transactions on Circuits and Systems, 32(11):1150–1161, 1985.DOI
  55. [55]Martin Schetzen. The Volterra and Wiener Theories of Nonlinear Systems. John Wiley & Sons, New York, 1980.
  56. [56]Jorge Mujica. Complex Analysis in Banach Spaces: Holomorphic Functions and Domains of Holomorphy in Finite and Infinite Dimensions, volume 120 of North-Holland Mathematical Studies. North-Holland, Amsterdam, 1986.
  57. [57]Seán Dineen. Complex Analysis on Infinite Dimensional Spaces. Springer Monographs in Mathematics. Springer, London, 1999.DOI
  58. [58]J.R. Dormand and P.J. Prince. A family of embedded runge-kutta formulae. Journal of Computational and Applied Mathematics, 6(1):19–26, 1980.DOI

Paper details

vlnc:W_kk64c8qa
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents