Quantum-classical hybrid calculation method based on tensor network structure

By constructing a quantum-classical hybrid one-dimensional matrix product state tensor network, the distributed representation and measurement of quantum states are realized. Combining classical optimization methods and variational quantum circuit optimization, the problem of unintegrated quantum-classical computing power is solved, and a deeper integration of computing resources is achieved, which is suitable for fields such as quantum chemistry and material design.

CN120671861APending Publication Date: 2025-09-19NAT UNIV OF DEFENSE TECH
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202510820137.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-19
Publication Date
2025-09-19

AI Technical Summary

Technical Problem

In current quantum-classical hybrid algorithms, the computing power of classical computing is only used for parameter optimization, while the computing power of quantum computing has not been effectively integrated, resulting in underutilization of computing resources.

Method used

A quantum-classical hybrid one-dimensional matrix product state tensor network is constructed, in which some tensor lattice points are realized as physical quantum states, and some lattice points are represented by classical tensor network algorithms. The representation, evolution and measurement of quantum states are realized through a distributed computing architecture, and the solution of global observation values ​​is achieved by combining classical optimization methods and variational quantum circuit parameter optimization.

Benefits of technology

It achieves the deep integration of quantum computing and classical computing power, improves the participation of classical computing in computing tasks, and can solve large problems more effectively. It is suitable for quantum chemistry, material design and classical combinatorial optimization problems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120671861A_ABST
    Figure CN120671861A_ABST
Patent Text Reader

Abstract

A quantum-classical hybrid calculation method based on a tensor network structure comprises the steps that a quantum-classical hybrid matrix product state tensor network is constructed, a part of tensor lattice points are achieved in a physical quantum state, the other part of tensor lattice points are achieved through a classical tensor network algorithm, and expression, evolution and measurement of a quantum state are jointly achieved through the tensor network and the classical tensor network; the method comprises the following steps: 1) initializing a quantum-classical hybrid tensor network; 2) encoding observable measurement for a target problem; 3) measuring to obtain a local observation result; 4) reducing the local observation result to obtain a global observation value; 5) optimizing the quantum-classical hybrid tensor network by using a classical optimization method, and optimizing a classical tensor by using a typical classical method; according to the method, effective support can be provided for fusion of computing power of quantum calculation and classical calculation and construction of a fusion system, and the method has high practical significance and practical value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical field:

[0001] The present invention relates to a quantum-classical hybrid computing method, in particular to a hybrid computing method of classical high-performance computing and quantum computing based on a tensor network structure, and belongs to the fields of quantum computing and high-performance computing. Background technology:

[0002] Quantum computing is a new type of computing power. Based on the superposition and entanglement properties of quantum physics, it possesses computing power far exceeding that of classical computing for specific problems. Currently, quantum computing has demonstrated quantum supremacy, entering the realm of noisy mesoscale quantum computing. It possesses a certain level of problem-solving capabilities and holds significant application potential in key fields such as cryptography, artificial intelligence, and engineering simulation. Quantum-classical fusion systems are the most likely ultimate form of quantum computing systems, but the computing power of the two has yet to be effectively integrated. Currently, hybrid quantum-classical algorithms, such as variational quantum algorithms (such as the Variational Quantum Eigensolver (VQE) and quantum approximate optimization algorithms (QAOA), have emerged as a method for solving problems using the computing power of both classical and quantum computing. However, in these algorithms, classical computing power is only used as an optimizer to perform parameter optimization, while the majority of the computational process is performed by quantum computing power. Therefore, in these computational methods, quantum and classical computing power have not been effectively integrated.

[0003] This invention is aimed at the current quantum computing physical system and high-performance computing environment, and proposes a quantum-classical hybrid computing method based on a tensor network structure. A tensor is a common data structure that can be regarded as a high-dimensional extension of an array. Its dimension is also called the order of the tensor. The zero-order tensor, the first-order tensor, and the second-order tensor correspond to scalars, vectors, and matrices, respectively. The contraction of tensors can also be regarded as an extension of matrix multiplication, as shown in Formula C. klm =∑ ij A ijk B lijm Describe a third-order tensor A ijk With the fourth-order tensor B lijm Contraction ij The dimension is the third-order tensor C klmIn the process, tensors A and B can be described by grid points in the diagram, and the dimensions i, j, k and l, i, j, m of A and B are described by three and four open edges respectively, and the contraction relationship of tensors A and B is described by the connection of edges. The array composed of the size of each dimension of the tensor is called the shape of the tensor. For example, the tensor shape of a 4-dimensional vector is [4], and the corresponding tensor shape of a 10×8 matrix is ​​[10,8]. Multiple tensor grid points can form a complex network with a contraction relationship called a tensor network. Common tensor networks include one-dimensional matrix product states, two-dimensional projected entangled pair states, etc. Tensor networks can be used as simulations of quantum computing. For example, CN202180096422.1 discloses a method for contracting tensor networks in quantum computing simulators. Unlike simulating quantum computing systems through tensor networks on classical computers, the present invention constructs a quantum-classical common data structure through tensor networks, realizing the common representation and solution of the same computing problem under two computing powers. It can more deeply integrate the computing power of classical computing and quantum computing in the current quantum-classical fusion system, and collaboratively process computing problems, which has strong practical significance and value. Summary of the invention:

[0004] The purpose of this invention is to propose a quantum-classical hybrid computing method based on a tensor network structure for current quantum computing physical systems and high-performance computing environments, so as to achieve the common representation and solution of the same computing problem under two computing powers.

[0005] The technical solution of the present invention is to construct a quantum-classical hybrid one-dimensional matrix product state tensor network, in which a portion of the tensor lattice is realized as a physical quantum state, and the other portion of the tensor lattice is realized through a classical tensor network algorithm, and the two together realize the representation, evolution and measurement of the quantum state. A quantum-classical hybrid tensor network contains at least two tensor lattices, of which at least one is a classical tensor and one is a quantum tensor. The definition of a classical tensor is the same as that of a tensor in traditional mathematics; a quantum tensor is a physical quantum state containing at least one quantum bit. The bits of the quantum tensor are grouped, and each group is equivalent to a "dimension" of the classical tensor.

[0006] Specifically, the steps of this method include:

[0007] (1) Initializing the quantum-classical hybrid tensor network

[0008] For the target problem, set the total number of quantum bits and the number of auxiliary quantum bits. Refer to the one-dimensional matrix product state (Matrix Product State) to construct all the hybrid tensor network grids (see Appendix Figure 1 , A0 and A in the figure nis a classical tensor, and the other grid points are quantum tensors). For a one-dimensional tensor network, each grid point is a third-order tensor (mathematically expressed as Among them A i represents the i-th tensor, l and r are two dimensions representing the coupling between the tensor and other adjacent tensors, and p is the physical dimension), which is equivalent to a three-dimensional array. For a quantum tensor, it itself is a quantum state |ψ i >, its quantum bits are divided into three groups, corresponding to the three dimensions l, r and p (see Appendix Figure 2 ):

[0009]

[0010] For a classic tensor, its element access is achieved by setting different subscript values ​​(e.g., given l i 、r i and p i For quantum tensors, the element values ​​correspond to the probability amplitudes of specific bases (e.g., when l i 、r i and p i When each contains 1 bit, The value of is the probability amplitude of the quantum state on the |000> basis).

[0011] If the current grid point is a classical tensor, it is generated by a classical computer and randomly initialized; if the current grid point is a quantum tensor, the quantum state is generated through a parameterized quantum circuit. The initial parameters of the parameterized quantum circuit are random, and the circuit structure used includes but is not limited to the attached Figure 3 The quantum circuit structure shown.

[0012] (2) Encoding observables for the target problem

[0013] The final result of quantum computing needs to be obtained through measurement. The measurement is performed on a global observable. Generally speaking, an observable can be decomposed into a set of Pauli matrices:

[0014]

[0015] where c i is the coefficient, P ij is the Pauli operator acting on the jth bit in the i-th summation term, which can be taken as Pauli σ x Operator, Pauli σ y Operator, Pauli σ z operator or unit operator I, that is, P ij ∈{σ x ,σ y ,σ z,I}. The Pauli matrix is ​​three 2×2 unitary Hermitian complex matrices (denoted as σ x ,σ y ,σ z ); Pauli matrices form the basis of single-qubit logic gates (such as X / Y / Z gates), which are used to manipulate quantum states in quantum computing. Similar to the tensor network representation of quantum states, this observable can also be expressed as a tensor network form of the matrix product operator:

[0016]

[0017] Among them, matrix multiplication is the direct product of matrices. Each matrix product operator lattice corresponds to a quantum-classical hybrid tensor network lattice (such as the attached Figure 1 shown),

[0018] (3) Measure and obtain local observation results

[0019] Unlike the measurement of general quantum computing processes, the evolution and measurement of each grid point of the quantum-classical hybrid tensor network can be carried out in a distributed manner, and the measurement result is a 3rd or 6th order classical tensor.

[0020] 1) If the current lattice point is at the leftmost end of the entire one-dimensional quantum-classical hybrid tensor network, its measurement result is a 3rd-order tensor Its dimensions are represented by r0, β0 and r0'. The calculation process is:

[0021] If the current grid point is a classical tensor, the result can be obtained by contracting the classical tensor:

[0022]

[0023] in is the leftmost tensor grid, with two dimensions r o and p0; for The conjugate of , also has two dimensions r' o and p0'0; is the observable MPO grid acting on the first tensor grid, with three dimensions, where p0 dimension is the same as Contraction, p0' dimension and Contraction.

[0024] If the current grid point is a quantum tensor, then its corresponding theoretical value is:

[0025]

[0026] in, The Pauli matrices when the auxiliary dimension of the observable takes the values of β0 at the first lattice point. |r0><r0'0| is the tensor The operator corresponding to the subscript values of the corresponding elements. For example, when both r0 and r0' take 0, the corresponding element is

[0027] In quantum computing, observables with physical significance are represented as Hermitian matrices, that is, complex matrices that satisfy The diagonal elements of which are real numbers, and the non-diagonal elements satisfy conjugate symmetry Its eigenvalues are real numbers, used to describe observable physical quantities (such as energy, momentum, etc.). In the calculation process of local observables, when r0≠r0', |r0><r0'| is not a Hermitian matrix. Therefore, it needs to be decomposed into a linear combination of Pauli matrices, and then the corresponding element values are solved to form the local observation result (see Example 1). <000017"">

[0028] 2) If the current lattice point is at the rightmost end of the entire one-dimensional quantum-classical hybrid tensor network, its measurement result is a third-order tensor The calculation process is as follows:

[0029] · If the current lattice point is a classical tensor, the result can be obtained by the contraction of the classical tensor:

[0030]

[0031] where is the rightmost tensor lattice point, with two dimensions l n and p n ; is the conjugate of , also with two dimensions r' n and p' n ; is the observable MPO lattice point acting on the nth tensor lattice point, with three dimensions, where the p n dimension is contracted with , and the p' n dimension is contracted with .

[0032] · If the current lattice point is a quantum tensor, its corresponding theoretical value is:

[0033]

[0034] where, is the Pauli matrix when the auxiliary dimension of the observable takes the values of β n-1 at the nth lattice point. |l <000003']><l' n | is the tensor The operator corresponding to the subscript value of the element. For example, l n and l' n When both are 0, the corresponding elements are When l n ≠l' n Time|l n > <l' n | is not a Hermitian matrix, so it needs to be decomposed into a linear combination of Pauli matrices, and then the corresponding element values ​​are obtained to form the local observation results (see Example 1).

[0035] 3) If the current lattice point is at a non-end point of the entire one-dimensional quantum-classical hybrid tensor network, the measurement result is a 6th-order tensor The calculation process is:

[0036] If the current grid point is a classical tensor, the result can be obtained by contracting the classical tensor:

[0037]

[0038] in is the i-th tensor grid point, with three dimensions l i 、r i and p i ; for The conjugate of also has three dimensions l' i 、r' i and p' i ; is the observable MPO grid acting on the i-th tensor grid point, with four dimensions, where p i Dimensions and Contraction, p' i Dimensions and Contraction.

[0039] If the current grid point is a quantum tensor, then its corresponding theoretical value is:

[0040]

[0041] in, The auxiliary dimensions of the observable quantity at the i-th grid point are β i , β i+1 The Pauli matrix when |l i > <l' i | and |r i > <r' i | is a tensor The operator corresponding to the subscript value of the element. For example, l i and l' i Both take 0, ri and r' i When both are 1, the corresponding elements are Similarly, since |l i > <l' i | and |r i > <r' i | is not necessarily a Hermitian matrix, so it needs to be decomposed into a linear combination of Pauli matrices, and then the corresponding element values ​​are obtained to form the local observation results (see Example 1).

[0042] Note that all local observations are classical tensors.

[0043] 4) Merge local observation results to obtain global observation values

[0044] Since quantum computing and classical computing may be performed using different computers, after the local observation results of all grid points are obtained, the local calculation results need to be shared. Therefore, data needs to be transmitted to other computing nodes through data broadcasting. The steps are as follows:

[0045] Convert the calculation results into basic data and remove the gradient tracking information;

[0046] Each computer in turn uses a collective communication broadcast interface such as MPI_Bcast to broadcast data to other computers;

[0047] After receiving the data, other computers will collapse all classical tensors into a globally observable value. The calculation process is:

[0048] iv. Get the result R i,j,k for

[0049] v. For the tensor grid points at non-endpoint positions, collapse their observation tensors and update R:

[0050]

[0051] vi. Collapse the tensor grid at the final position to obtain the scalar value R:

[0052]

[0053] 5) Optimizing quantum-classical hybrid tensor networks using classical optimization methods

[0054] For classical tensors, typical classical methods can be used for optimization (such as gradient descent, etc.);

[0055] For quantum tensors, optimization is achieved by optimizing the parameters of the variational quantum circuit that generates the tensor. The variational quantum circuit parameters can also be optimized via gradient descent, where the gradient is solved using the chain rule and the parameter translation rule.

[0056] Using the chain rule, find the partial derivative of the global observation result R with respect to the line output result This step can be solved by automatic differentiation method.

[0057] Using the parameter translation rule, find the line output result for a certain parameter θ i Partial derivatives of :

[0058]

[0059] in, To transform the variable into the parameter θ in the quantum circuit i Modified to Expected line output value when .

[0060] Using the chain rule, find the global observation result for a certain parameter θ i The partial derivative of

[0061]

[0062] Find the gradient of the global observation result R with respect to all parameters, and then optimize and update the parameters by gradient descent.

[0063] 6) Repeat steps 3) to 5) until the results converge.

[0064] The present invention represents the global quantum state as a distributed one-dimensional tensor network lattice. The lattice is then evolved and measured to obtain local observations at each lattice point, which are then aggregated to form the global measurement results. Calculations at each lattice point can be performed using different systems, such as classical computers, superconducting quantum computers, or optical quantum computers. This allows for the fusion of multiple computing systems to collaboratively solve the same computational problem.

[0065] Beneficial effects: The technical solution provided by the present invention is different from the idea of ​​the current quantum-classical hybrid algorithm. Through the tensor network structure, the computing load can be effectively divided into quantum computing power and classical computing power, and the fusion calculation of multiple physical systems can be realized. Compared with the traditional variational quantum algorithm, the classical computing part not only updates the parameters, but also includes the calculation of classical tensors. Therefore, it has a higher degree of participation in the computing task, and the computing power of quantum computing and classical computing has achieved a deeper fusion. In addition, through this distributed computing architecture, the decomposition of large problems can be achieved, which is convenient for execution under the current conditions of limited physical system scale, so the quantum-classical computing power fusion calculation can be realized more effectively. The calculation method proposed in the present invention can be applied to fields such as quantum chemistry, material design, and solving classical combinatorial optimization problems. The simulation results of the quantum-classical hybrid solution of a 4-bit Heisenberg one-dimensional spin chain are given in Example 1 (see Appendix). Figure 5 The present invention can provide effective support for integrating the computing power of quantum computing and classical computing and building a fusion system, and has strong practical significance and value. Description of the drawings:

[0066] Figure 1 One-dimensional quantum-classical hybrid tensor networks and their correspondence to global observables.

[0067] Figure 2 The correspondence between the lattice points of quantum tensors and their physical chip implementation. The implementation form of quantum tensors. (a) Taking a superconducting physical system as an example, a quantum tensor is implemented using multiple superconducting qubits on a superconducting quantum chip. (b) The qubits are divided into three groups, each used to represent the three dimensions of a third-order tensor. (c) The correspondence between the dimensions of the implemented third-order quantum tensor and the qubits.

[0068] Figure 3 Examples of variational quantum circuit structures employed. Figure 3 An example of a parametric quantum circuit used to prepare quantum tensors in [1]. The qubits are divided into three groups, with the attached Figure 2 Corresponding. The Ry gate is a single-bit parameter gate, and its matrix form is θ is a parameter;

[0069] Figure 4 The 4-bit mixed matrix product state tensor network and the matrix product operator corresponding to the Hamiltonian of Example 1. A 4-bit mixed tensor network and its corresponding observable. Where A1 and A2 are classical tensors, and |ψ3> and |ψ4> are quantum tensors.

[0070] Figure 5 The solution to the quantum-classical mixed ground state energy of a 4-bit Heisenberg one-dimensional spin chain. Specific implementation method:

[0071] The present invention describes the quantum-classical hybrid computing method in detail through embodiments.

[0072] Figure 1 The same quantum-classical mixed matrix product state is given in . Where A0, A n is a classical tensor, and |ψ2> and |ψ3> are quantum tensors. to The MPO grid is obtained by tensor decomposition of the observables. To keep the dimensionality corresponding to the hybrid MPS, the MPO grid is grouped;

[0073] Figure 5 Computational simulation results of quantum-classical hybrid tensors, and their correspondence with the results of the classical algorithm. The light dashed line is the result of the classical Density Matrix Renormalization Group (DMRG) algorithm with the same auxiliary dimension size.

[0074] Example 1: Hybrid computing implementation of a single quantum computing node and a single classical computing node,

[0075] This embodiment aims to solve the ground state energy of a 4-bit Heisenberg one-dimensional spin chain, as shown in the attached figure. Figure 4 As shown. The Heisenberg spin chain is the core model for describing one-dimensional quantum magnetic systems. Its Hamiltonian is usually composed of the exchange interaction between adjacent spins. For the one-dimensional antiferromagnetic Heisenberg model with spin -1 / 2, the Hamiltonian can be expressed as:

[0076]

[0077] Where J>0 represents antiferromagnetic coupling, S i is the spin operator at the i-th lattice point. This model can be extended to more complex quantum magnetic systems by introducing anisotropy (such as XXZ deformation), such as the 4-bit Heisenberg XXZ spin chain Hamiltonian shown below:

[0078]

[0079] in is the Pauli operator p acting on the i-th bit. Where J>0 represents antiferromagnetic coupling, Δ is the proportional quantity reduced to J, and h z represents the external magnetic field. The present invention is applicable to j, Δ and h z In the case of arbitrary values, in this embodiment, J and Δ are 1, h z The value is 0. Since the measurement of quantum bits in quantum computing is often only performed on a specific basis, the observable quantity (ie, Hamiltonian) is usually decomposed to meet the conditions of experimental feasibility.

[0080] The specific execution steps are:

[0081] 1. Establish a one-dimensional mixed matrix product state. The auxiliary dimension of the matrix product state is limited to 2, and it contains a total of 4 tensor grids, where the first and second tensor grids are classical tensors, and the third and fourth grids are quantum tensors.

[0082] The first tensor grid is a classical tensor, which can be expressed as The l1 dimension size is 1, and the r1 and p1 dimension sizes are both 2, so the data shape is [1, 2, 2], which contains 4 elements;

[0083] The second tensor grid is a classical tensor, which can be expressed as The dimensions of l2, r2, and p2 are all 2, and the data shape is [2,2,2], which means it contains 8 elements.

[0084] The third tensor grid is a quantum tensor, which can be expressed as

[0085]

[0086] Among them, |l3>, |p3> and |r3> each contain 1 quantum bit. The quantum state has a total of 3 quantum bits, including 8 probability amplitudes (equivalent to a 3-order classical tensor), that is, each uses 1 bit to auxiliaryly connect with the second and fourth grid points; the classical tensor is generated by a 3-qubit quantum circuit with the same quantum circuit structure. Figure 3 , depth d = 3, the line contains 9 parameters;

[0087] The fourth tensor grid is the quantum tensor, which can be expressed as:

[0088]

[0089] Among them, |l4> and |p4> each contain 1 quantum bit. The quantum state has a total of 2 quantum bits, including 4 probability amplitudes (equivalent to a 2-order classical tensor), that is, each uses 1 bit to auxiliaryly connect to the third grid point; the classical tensor is generated by a 2-qubit quantum circuit with the same quantum circuit structure. Figure 3 , depth d=2, the line contains 4 parameters.

[0090] 2. Express the Hamiltonian as a matrix product operator:

[0091]

[0092] Among them, σ x , σ y , σ z , I is the Pauli operator. This matrix product operator includes 4 grid points.

[0093] The first grid point can be expressed as For example Since each element is a 2×2 matrix, it is a 3rd-order tensor with a shape of [9, 2, 2];

[0094] The second grid point can be expressed as For example The grid point is a 4-order tensor with a shape of [9,9,2,2];

[0095] The third grid point is similar to the second grid point and can be expressed as For example The grid point is a 4-order tensor with a shape of [9,9,2,2];

[0096] The fourth grid point is similar to the first grid point and can be expressed as For example The grid is a rank 3 tensor with a shape of [9, 2, 2].

[0097] 3. Obtain the local observation results.

[0098] For the first and second tensor grid points, they can be directly obtained by classical tensor contraction:

[0099]

[0100] The third and fourth tensor grid points need to be obtained through measurement.

[0101] For the third grid point, the specific steps are:

[0102] i. Initialize the result tensor It is a 6-order tensor of all zeros, with the sizes of the 6 dimensions being 2, 9, 2, 2, 9, and 2, respectively, i.e., it contains 2×9×2×2×9×2=1296 elements;

[0103] ii. Traversal i=1,2,…,9. For a certain value of i, record it as

[0104] iii.Measure |ψ3> in The mean under P l 、P r ∈{σ x ,σ y ,σ z ,I}, get the measured mean

[0105] iv. Calculate the element values ​​of V3 based on the theoretical expectation value of the Pauli operator. For example,

[0106]

[0107] Therefore, it can be deduced and calculated:

[0108]

[0109] For the fourth grid point, the specific steps are:

[0110] i. Initialize the result sheet It is a 3rd-order tensor of all 0s, with dimensions 2, 9, and 2, i.e., it contains 2×9×2=36 elements;

[0111] ii. Traversal i=1,2,…,9. For a certain value of i, record it as

[0112] iii.Measure |ψ4> in The mean under P l ∈{σ x ,σ y ,σ z ,I}, get the measured mean M ;

[0113] iv. Calculate the element values ​​of V4 based on the theoretical expectation value of the Pauli operator. For example,

[0114]

[0115] Therefore, it can be deduced and calculated:

[0116]

[0117] 4. Obtain the global observation results.

[0118]

[0119] 5. Find the global modulus corresponding to the one-dimensional tensor network. Find the grid points for The local observation results are then shrunk and obtained. The method is the same as step 3). Only P i = I. Then, the local measurement results are reduced and combined to obtain the global modulus N.

[0120] 6. Calculate the expected value of the global observable quantity E = R / N.

[0121] 7. Using the gradient descent method, calculate the gradient of E for all variable parameters and optimize the parameters. This step can be efficiently implemented using common automatic differentiation computing tools (such as Pennylane, Torch, TensorFlow, etc.)

[0122] 8. Repeat steps 3 to 7 until convergence. Use a classical computer to simulate the results as shown in the attached Figure 5 As shown, this method effectively obtains the ground-state energy of a 4-qubit Heisenberg one-dimensional spin chain. This shows that in addition to parameter optimization and updating, the data processing of the classical tensor grid is also performed on a classical computer. Therefore, compared to algorithms such as VQE, the computing power of quantum and classical computing is more fully integrated to solve the same computational problem, thus achieving a deep fusion of the two computing powers.

Claims

1. A quantum-classical hybrid computing method based on a tensor network structure, characterized by: Construct a quantum-classical hybrid matrix product state tensor network, in which some tensor lattices are implemented as physical quantum states, and others are implemented using classical tensor network algorithms. Together, they enable the representation, evolution, and measurement of quantum states. The following steps are involved: 1) Initialize the quantum-classical hybrid tensor network: For the target problem, set the total number of quantum bits and the number of auxiliary quantum bits; refer to the one-dimensional matrix product state (Matrix Product State) to construct all the hybrid tensor network grid points; for the one-dimensional tensor network, each grid point is a third-order tensor (mathematically expressed as Among them A i represents the i-th tensor, l and r are two dimensions representing the coupling between the tensor and other adjacent tensors, and p is the physical dimension), which is equivalent to a three-dimensional array; for a quantum tensor, it itself is a quantum state |ψ i >, its qubits are divided into three groups, corresponding to the three dimensions l, r, and p (see Figure 2): For classical tensors, element access is achieved by setting different subscript values; for quantum tensors, element values ​​correspond to the probability amplitude of a specific basis: when l i 、r i and p i When each contains 1 bit, The value of is the probability amplitude of the quantum state on the |000> basis; If the current grid point is a classical tensor, it is generated by a classical computer and randomly initialized; if the current grid point is a quantum tensor, the quantum state is generated through a parameterized quantum circuit; the initial parameters of the parameterized quantum circuit are random; 2) Encoding observables for the target problem The final result of quantum computing needs to be obtained through measurement; the measurement is performed on a global observable; an observable is decomposed into a set of Pauli matrices: where c i is the coefficient, P ij is the Pauli operator acting on the jth bit in the i-th summation term, which can be taken as Pauli σ x Operator, Pauli σ y Operator, Pauli σ z operator or unit operator I, that is, P ij ∈{σ x ,σ y ,σ z ,I}; The Pauli matrix is ​​three 2×2 unitary Hermitian complex matrices (denoted as σ x ,σ y ,σ z ); Pauli matrices form the basis of single-qubit logic gates (such as X / Y / Z gates), which are used to manipulate quantum states in quantum computing; 3) Measure and obtain local observation results Unlike measurements in general quantum computing processes, the evolution and measurement of each grid point in a quantum-classical hybrid tensor network can be carried out in a distributed manner, and the measurement result is a 3rd- or 6th-order classical tensor; all local observation results are classical tensors. 4) Merge local observation results to obtain global observation values Since quantum computing and classical computing may be performed using different computers, after all local observation results of all grid points are obtained, the local calculation results need to be shared. Therefore, data broadcasting is required to transmit data to other computing nodes, and then the global observation results are obtained. 5) Optimizing quantum-classical hybrid tensor networks using classical optimization methods For classical tensors, optimization is done using typical classical methods; For quantum tensors, optimization is achieved by optimizing the parameters of the variational quantum circuit that generates the tensor; the variational quantum circuit parameters can also be optimized by gradient descent, and the gradient solution is obtained by the chain rule and the parameter translation rule; Using the chain rule, find the partial derivative of the global observation result R with respect to the line output result This step can be solved by automatic differentiation method; Using the parameter translation rule, find the line output result for a certain parameter θ i Partial derivatives of : in, To convert the variation into the parameter θ in the quantum circuit i Modified to Expected line output value when ; Using the chain rule, find the global observation result for a certain parameter θ i The partial derivative of Find the gradient of the global observation result R with respect to all parameters, and then optimize and update the parameters by gradient descent; 6) Repeat steps 3) to 5) until the results converge.

2. The quantum-classical hybrid computing method based on a tensor network structure according to claim 1, characterized in that: The global observable is written in the form of a matrix product operator. To ensure the Hermitian property of the observed quantity when the quantum system is processed, the matrix product operator is Among them, matrix multiplication is the direct product of matrices and matrices; each matrix product operator lattice corresponds to a quantum-classical hybrid tensor network lattice.

3. The quantum-classical hybrid computing method based on a tensor network structure according to claim 1, characterized in that: The local observation result acquisition has the following steps: 1) If the current lattice point is at the leftmost end of the entire one-dimensional quantum-classical hybrid tensor network, its measurement result is a 3rd-order tensor Its dimensions are denoted by r0, β0 and r0'; The calculation process is: If the current grid point is a classical tensor, the result can be obtained by contracting the classical tensor: in is the leftmost tensor grid, with two dimensions r o and p0; for The conjugate of , also has two dimensions r' o and p0'; is the observable MPO grid acting on the first tensor grid, with three dimensions, where p0 dimension is the same as Contraction, p0' dimension and contraction; If the current grid point is a quantum tensor, then its corresponding theoretical value is: Among them, is the Pauli matrix when the auxiliary dimension values of the observable at the first lattice point are β0 respectively; |r0><r0'| is the operator taken by the subscript values of the corresponding elements of the tensor For example, when both r0 and r0' take 0, the corresponding element is In quantum computing, observables with physical significance are represented as Hermitian matrices, that is, complex square matrices that satisfy , where the diagonal elements are real numbers and the off-diagonal elements satisfy conjugate symmetry Its eigenvalues are real numbers and are used to describe observable physical quantities; in the calculation process of local observables, when r0≠r0', |r0><r0'| is not a Hermitian matrix. Therefore, it needs to be decomposed into a linear combination of Pauli matrices, and then the corresponding element values are solved to form the local observation result; 2) If the current lattice point is at the rightmost end of the entire one-dimensional quantum-classical hybrid tensor network, its measurement result is a 3rd-order tensor The calculation process is: If the current grid point is a classical tensor, the result can be obtained by contracting the classical tensor: in is the rightmost tensor grid, with two dimensions l n and p n ; for The conjugate of , also has two dimensions r' n and p' n ; is the observable MPO grid acting on the nth tensor grid point, with three dimensions, where p n Dimensions and Contraction, p' n Dimensions and contraction; If the current grid point is a quantum tensor, then its corresponding theoretical value is: in, The auxiliary dimensions of the observable quantity at the nth grid point are β n-1 The Pauli matrix when |l n > <l' n | is a tensor The operator corresponding to the subscript value of the element; for example, l n and l' n When both are 0, the corresponding elements are When l n ≠l' n Time|l n > <l' n | is not a Hermitian matrix, so it needs to be decomposed into a linear combination of Pauli matrices, and then the corresponding element values ​​are solved to form the local observation results; 3) If the current lattice point is at a non-end point of the entire one-dimensional quantum-classical hybrid tensor network, the measurement result is a 6th-order tensor The calculation process is: If the current grid point is a classical tensor, the result can be obtained by contracting the classical tensor: in is the i-th tensor grid point, with three dimensions l i 、r i and p i ; for The conjugate of also has three dimensions l' i 、r' i and p' i ; is the observable MPO grid acting on the i-th tensor grid point, with four dimensions, where p i Dimensions and Contraction, p' i Dimensions and contraction; If the current grid point is a quantum tensor, then its corresponding theoretical value is: in, The auxiliary dimensions of the observable quantity at the i-th grid point are β i , β i+1 The Pauli matrix when |l i > <l' i | and |r i > <r' i | is a tensor The operator corresponding to the subscript value of the element; for example, l i and l' i Both take 0, r i and r' i When both are 1, the corresponding elements are Similarly, since |l i > <l' i | and |r i >< r ' i | is not necessarily a Hermitian matrix, so it needs to be decomposed into a linear combination of Pauli matrices, and then the corresponding element values ​​are solved to form local observations.

4. The quantum-classical hybrid computing method based on a tensor network structure according to claim 1, characterized in that: The process of sharing the local computation results and obtaining global observations has the following steps: Convert the calculation results into basic data and remove the gradient tracking information; Each computer in turn uses a collective communication broadcast interface such as MPI_Bcast to broadcast data to other computers; After receiving the data, other computers will collapse all classical tensors into a global observable value; The calculation process is: i. Get the result R i,j,k for ii. For the tensor grid points at non-endpoint positions, collapse their observation tensors and update R: iii. Collapse the tensor grid at the final position to obtain the scalar value R:

5. The quantum-classical hybrid computing method based on a tensor network structure according to claim 1, characterized in that: The parameter optimization method is implemented using the gradient descent method.

6. The quantum-classical hybrid computing method based on a tensor network structure according to claim 1, characterized in that: Aiming to solve the ground state energy of a 4-bit Heisenberg one-dimensional spin chain, a hybrid computing method using a single quantum computing node and a single classical computing node is implemented. The steps are as follows: 1) Establish a one-dimensional mixed matrix product state. The auxiliary dimension of the matrix product state is limited to 2, and it contains a total of 4 tensor grids, where the first and second tensor grids are classical tensors, and the third and fourth grids are quantum tensors; The first tensor grid is a classical tensor, expressed as The l1 dimension size is 1, and the r1 and p1 dimension sizes are both 2, so the data shape is [1, 2, 2], which contains 4 elements; The second tensor grid is a classical tensor, which can be expressed as The dimensions of l2, r2, and p2 are all 2, and the data shape is [2,2,2], which means it contains 8 elements. The third tensor grid is the quantum tensor, which can be expressed as Here, |l3>, |p3>, and |r3> each contain one qubit. This quantum state has a total of three qubits and eight probability amplitudes, equivalent to a third-order classical tensor, where each uses one bit to connect to the second and fourth lattice points. This classical tensor is generated by a three-qubit quantum circuit with a depth of d = 3 and nine parameters. The fourth tensor grid is the quantum tensor, which can be expressed as: Here, |l4> and |p4> each contain one qubit, resulting in a total of two qubits in this quantum state, containing four probability amplitudes (equivalent to a second-order classical tensor), i.e., each using one bit for auxiliary connection to the third lattice point. This classical tensor is generated by a two-qubit quantum circuit with depth d = 2 and four parameters. 2) Express the Hamiltonian as a matrix product operator: Among them, σ x , σ y , σ z , I is the Pauli operator; the matrix product operator includes 4 grid points; The first grid point, represented by when Since each element is a 2×2 matrix, it is a 3rd-order tensor with a shape of [9, 2, 2]; The second grid point is represented by when The grid point is a 4-order tensor with a shape of [9,9,2,2]; The third grid point is similar to the second grid point and can be expressed as when The grid point is a 4-order tensor with a shape of [9,9,2,2]; The fourth grid point is similar to the first grid point and is represented by when The grid point is a 3rd-order tensor with a shape of [9,2,2]; 3) Obtain local observation results; For the first and second tensor grid points, we can directly obtain them through classical tensor contraction: For the third and fourth tensor grid points, they need to be obtained through measurement; For the third grid point, the specific steps are: i. Initialize the result tensor It is a 6-order tensor of all zeros, with the sizes of the 6 dimensions being 2, 9, 2, 2, 9, and 2, respectively, i.e., it contains 2×9×2×2×9×2=1296 elements; ii. Traversal For a certain value of i, record it as iii.Measure |ψ3> in The mean under P l 、P r ∈{σ x ,σ y ,σ z ,I}, get the measured mean iv. Calculate the element values ​​of V3 based on the theoretical expectation value of the Pauli operator; Therefore, it is deduced that: For the fourth grid point, the specific steps are: i. Initialize the result sheet It is a 3rd-order tensor of all zeros, with dimensions 2, 9, and 2, i.e., it contains 2×9×2=36 elements; ii. Traversal For a certain value of i, record it as iii.Measure |ψ4> in The mean under P l ∈{σ x ,σ y ,σ z ,I}, get the measured mean ; iv. According to the theoretical expectation value of the Pauli operator, the element value of V4 is calculated; when Therefore, it is deduced that: 4) Obtain global observation results; 5) Find the global modulus corresponding to the one-dimensional tensor network; Let Find the grid points for The local observation results of , and then shrink and get, only need to measure P i =I; then, the local measurement results are reduced and merged to obtain the global modulus N; 6) Calculate the expected value of the global observable quantity E = R / N; 7) Using the gradient descent method, calculate the gradient of E with respect to all variable parameters and optimize the parameters; this step can be efficiently implemented using common automatic differentiation calculation tools; 8) Repeat steps 3 to 7 until convergence; perform simulation using a classical computer.

Citation Information

Patent Citations

  • Processor and method for tensor network contraction in quantum simulator

    CN117203647A