A quantum reading calibration method and system based on finite element
The weighted qubit diagram and subnoise matrix are constructed in quantum reading calibration through finite element method, and combined with the sparse tensor product engine, the calculation complexity and accuracy of quantum reading calibration in the existing technology is solved, and efficient and accurate qubits are achieved calibration.
Patent Information
- Application Number
- CN202410283895.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-13
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-03-13
AI Technical Summary
Existing quantum reading calibration methods have limitations in terms of scalability and accuracy, especially the matrix-based calibration methods have too high computational complexity and memory requirements in large-scale qubit systems, which makes it difficult to effectively reduce reading errors.
Using a quantum reading calibration method based on finite element, through multiple iterative calibration processes, feature process and calibration process, weighted qubit graphs are constructed and grouped, subnoise matrix is calculated, and the calculation is accelerated by using the sparse tensor product engine to trim the intermediate value. To improve calibration accuracy.
It improves the accuracy and efficiency of quantum reading calibration, can support calibration of more than 500 qubits, reduces computing complexity and memory requirements, and achieves high scalability and high accuracy.
Smart Images

Figure CN118070914B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of quantum computing technology, and specifically relates to a quantum reading calibration method and system based on finite elements. Background Art
[0002] The quantum computing paradigm shows great potential to surpass classical computing in solving complex problems, such as physical simulation, combinatorial optimization, and many examples in the field of artificial intelligence. When solving these problems on actual quantum hardware, the calculations will inevitably encounter various noise sources, among which the readout error (the noise generated by reading the output from the quantum device and converting it into classical data) has been shown to be the most significant source of error. For most physical quantum implementations, this readout error reaches 1%-10% in each qubit operation, thus seriously hindering the application of quantum computing.
[0003] The fidelity of the readout can be improved through hardware-level optimization techniques, including frequency fine-tuning and sapphire-based processor manufacturing. These techniques make the resulting quantum states easier to distinguish by isolating the qubits from external noise. In contrast, a more cost-effective approach is to employ software-level optimizations such as matrix-based calibration, machine learning, domain-specific models, and Bayesian estimation. Among these methods, matrix-based calibration methods have been experimentally proven to be effective in superconducting, ion trap, and optical quantum hardware, and are therefore widely used in commercial quantum cloud platforms such as IBMQ and Rigetti.
[0004] Although matrix-based calibration can be easily deployed on the software stack, it faces a trade-off between scalability and calibration accuracy. In order to maximize accuracy, that is, to accurately characterize the matrix, a large number of benchmark circuits need to be executed exhaustively to cover all possible measurement outputs, and their number grows exponentially with the number of qubits. For example, obtaining such a matrix for 16 qubits requires the execution of 6.5×10^4 circuits, which takes approximately 18.2 hours on the IBMQ Guadalupe quantum device. In addition to matrix characterization, the MVM (matrix-vector multiplication) step is equally computationally overwhelming because the size of the matrix grows exponentially with the number of qubits. Taking IBMQ's recently developed M3 as an example, it utilizes a pruning strategy based on a Hamming distance threshold, however, it still requires 87.5PB of memory and more than a year to execute the MVM step.
[0005] Quantum readout errors have been shown to be the most significant source of error, greatly affecting measurement fidelity. Matrix-based calibration has been shown to be effective in various quantum platforms. However, existing methods have fundamental limitations in scalability or accuracy.
[0006] Existing methods fall into two categories: quantum-independent methods and sparsity-aware methods. The first category includes IBU and Yang et al., which compute the noise matrix by calculating the tensor product of a series of element matrices (single-bit 2×2 matrices). Meanwhile, M3 and Nation are two representatives of sparse-aware methods that exploit the underlying sparsity pattern to approximate the noise matrix.
[0007] Currently, matrix-based techniques focus on improving scalability and accuracy. To improve scalability, methods like M3 and Yang et al. exploit the sparsity of the distribution matrix, although they show the limitation that only a few matrix elements can be pruned. Other methods only simulate qubit-independent errors, resulting in a large loss of accuracy.
[0008] In terms of scalability, both the characterization time and the MVM time scale exponentially with the number of qubits being measured. The quantum-independent approach exhibits linear characterization of the noise matrix in terms of time, requiring only 2N executions. m benchmark circuits. Therefore, they support calibration of more qubits than sparse-aware methods. However, they have to introduce linear equation solvers to compute the inverse matrix (calibration matrix). The complexity of these solvers limits the upper limit to 127 qubits. On the other hand, sparse-aware methods prune the matrix at a threshold of the Hamming distance, which, however, has no reduction in complexity and leads to a slightly improved scalability. In addition, they require characterization for each calibration run, which further increases the complexity.
[0009] In terms of accuracy, quantum-independent methods achieve better scalability by sacrificing accuracy because they cannot capture noise between qubits, such as crosstalk. More importantly, the inaccuracy will scale with the increase of the tensor product length. For example, when calibrating the output of more than 81 qubits, this inaccuracy becomes intolerable. In addition, they are mainly applicable to GHZ and BV algorithms because they consist of fewer bit strings with non-zero probability, which limits their versatility. Sparse-aware methods show smaller Hilbert-Schmidt distances, however, they are still insufficient for calibration. Summary of the invention
[0010] In view of the above, the object of the present invention is to provide a finite element-based quantum reading calibration method and system, which extends the finite element method (FEM) to the quantum calibration process and comprehensively considers the characteristics of quantum bit interactions to improve the accuracy and efficiency of quantum reading calibration.
[0011] To achieve the above-mentioned object of the invention, an embodiment of the present invention provides a quantum reading calibration method based on finite elements, including multiple rounds of iterative calibration, each round of iterative calibration including a characterization process and a calibration process.
[0012] The characterization process includes: determining the weight between any two quantum bits based on the benchmark probability distribution of the benchmark circuit in the current round and constructing a weighted quantum bit graph, and dividing the quantum bits in the weighted quantum bit graph into groups based on the finite element concept to obtain a grouping scheme, wherein the benchmark probability distribution of the first round is obtained by sampling the output of the benchmark circuit, and the benchmark probability distribution of other rounds other than the first round comes from the calibration output in the iterative process;
[0013] The calibration process includes: calculating the sub-noise matrix corresponding to each quantum bit group based on the measured and collected quantum bit collection, the benchmark probability distribution and the grouping scheme to obtain a sub-noise matrix set, and calibrating the calibration output of the previous round based on the sub-noise matrix set to obtain the calibration output of the current round.
[0014] Preferably, the reference circuit is obtained by screening by setting a threshold, including: executing a candidate circuit and obtaining a measured state of each quantum bit of the candidate circuit under a given ideal ground state, and calculating the interaction between the two quantum bits and an index for distinguishing key interactions based on the ideal ground state and the measured state, and obtaining the reference circuit by screening candidate circuits whose indexes exceed the set threshold.
[0015] Preferably, each time a candidate circuit is executed, the operation and readout output of each qubit is recorded as a triple (ideal, measured, ef), where Used to record the ideal ground state prepared in the circuit. ideal=0 means the ideal ground state bit value is 0, and ideal=1 means the ideal ground state bit value is 1. Indicates that the ideal ground state bit value has not been measured, Used to record the measurement status output after executing the circuit sample. Measured = 0 means the measurement status bit value is 0, and measured = 1 means the measurement status bit value is 1. Indicates that the measurement state bit value has not been measured, ef∈{0,1} indicates an error flag, ef=0 indicates that the measurement state matches the ideal ground state, and ef=1 indicates that an error has occurred;
[0016] The interaction between two quantum bits is expressed as interact(q i .ideal=x→q j .ideal=y), calculated as:
[0017] interact(q i .ideal=x→q j .ideal=y)=p(q j .ef=1|C1,C2)-p(q j .ef
[0018] =1|C2)
[0019] C1:q i .ideal = x;
[0020] C2:
[0021] Among them, q i .ideal=x represents the i-th quantum bit q i The ideal ground state is x, q j .ideal=y represents the jth quantum bit q j The ideal ground state of is y, interact(·→·) represents the interaction from the former to the latter, q j .ef=1 indicates the jth quantum bit q j The error flag of is 1, C1 and C2 represent intermediate variables used to indicate the ideal state of the bit, and p(·|·) represents the conditional probability distribution of the actual value measured under the ideal value, which is used as the reference probability distribution;
[0022] The index θ for distinguishing key interactions based on the interaction calculation of two qubits is expressed as:
[0023]
[0024] Here, num(·→·) represents the number of interactions.
[0025] Preferably, the first-round reference probability distribution is obtained by sampling the output of the reference circuit, including: obtaining the measurement state of each qubit in the reference circuit in an ideal ground state, and recording the triplet, and calculating the conditional probability distribution of each qubit based on the triplet as the first-round reference probability distribution;
[0026] The benchmark probability distributions of other rounds other than the first round are derived from the calibration outputs in the iteration process, including: using the calibration outputs of the iteration process of the current round as element values to form the benchmark probability distribution of the next round.
[0027] Preferably, determining the weight between any two quantum bits based on the reference probability distribution of the reference circuit in the current round and constructing a weighted quantum bit graph includes:
[0028] The weight between any two quantum bits (q i ,q j ) is expressed as:
[0029]
[0030] In the constructed weighted quantum bit graph, quantum bits are nodes, and the edges between nodes represent the interactions between quantum bits, and the weights (q i ,q j ) marks the weight value of the edge;
[0031] Preferably, based on the finite element concept, the quantum bits in the weighted quantum bit graph are divided into groups to obtain a grouping scheme, including:
[0032] With the goal of maximizing the locality within the group, that is, maximizing the sum of the weights within the group, the weighted quantum bit graph is divided into quantum bit groups to obtain a grouping scheme.
[0033] Preferably, the sub-noise matrix corresponding to each qubit group is calculated based on the measured and collected qubit collection, the reference probability distribution, and the grouping scheme, including:
[0034] The collection of quantum bits collected is Q M , the grouping scheme is G l = {g l,1 ,···,g l,K}, where l represents the round, K represents the total number of groups, and g l,k represents the kth quantum bit group in the lth round;
[0035] For each g l,k Definition: g ∩ =Q M ∩g l,k , where g ∩ Indicates that g l,k and Q M Overlapping qubits, Indicates g l,k The remaining qubits in
[0036] The sub-noise matrix corresponding to each qubit group is calculated as:
[0037]
[0038] Where q represents the g ∩ The quantum bit, x q represents the position of quantum bit q in bit string x, P represents the conditional probability, based on the benchmark probability distribution BP l It is estimated that q.measure = x q Indicates that the measured state of q is x q , g ∩ .ideal=y as a condition, indicating g ∩ The ideal ground state is y, As a condition, it means The ideal ground state is M l,k [x][y] represents g l,k The corresponding sub-noise matrix M l,k The values of the x and y positions in the .
[0039] Preferably, calibrating the calibration output of the previous round based on the sub-noise matrix set to obtain the calibration output of the current round includes:
[0040]
[0041] Among them, M l,1 ,M 1,2 ,…,M l,K are the K sub-noise matrices in the sub-noise matrix set, P l+1 represents the calibration output of the lth round, P l represents the calibration output of the lth round, symbol Represents the matrix tensor product, NZ l Indicates P l The set of quantum bit strings with non-zero probability in |x l,k > represents the sub-bit string corresponding to the k-th quantum bit group in the l-th round.
[0042] Preferably, in the calibration process, a sparse tensor product engine is introduced to prune the tensor product in the calibration calculation process to accelerate the calculation, including:
[0043] First, calculate the The matrix-vector multiplication value of P is then calculated, and the tensor product of the matrix-vector multiplication values of all qubit groups is then pruned by another threshold to remove the intermediate values in the tensor product. l (x) is aggregated with the remaining intermediate values in the tensor product to obtain the calibration output.
[0044] To achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a quantum reading calibration system based on finite elements, including a characterization process module and a calibration process module, wherein the characterization process module and the calibration process module are sequentially iterated to realize quantum reading calibration;
[0045] The characterization process module is used to determine the weight between any two qubits based on the benchmark probability distribution of the benchmark circuit in the current round and construct a weighted qubit graph. The qubits in the weighted qubit graph are divided into groups based on the finite element concept to obtain a grouping scheme, where the benchmark probability distribution of the first round is obtained by sampling the output of the benchmark circuit, and the benchmark probability distribution of other rounds other than the first round comes from the calibration output in the iterative process;
[0046] The calibration process module is used to calculate the sub-noise matrix corresponding to each quantum bit group based on the measured and collected quantum bit collection, the benchmark probability distribution and the grouping scheme, and obtain the sub-noise matrix set. Based on the sub-noise matrix set, the calibration output of the previous round is calibrated to obtain the calibration output of the current round.
[0047] To achieve the above-mentioned purpose of the invention, an embodiment of the present invention further provides a quantum reading calibration system based on finite elements, including a characterization process module and a calibration process module, wherein the characterization process module and the calibration process module are sequentially iterated to realize quantum reading calibration;
[0048] The characterization process module is used to determine the weight between any two qubits based on the benchmark probability distribution of the benchmark circuit in the current round and construct a weighted qubit graph. The qubits in the weighted qubit graph are divided into groups based on the finite element concept to obtain a grouping scheme, where the benchmark probability distribution of the first round is obtained by sampling the output of the benchmark circuit, and the benchmark probability distribution of other rounds other than the first round comes from the calibration output in the iterative process;
[0049] The calibration process module is used to calculate the sub-noise matrix corresponding to each quantum bit group based on the measured and collected quantum bit collection, the benchmark probability distribution and the grouping scheme, and obtain the sub-noise matrix set. Based on the sub-noise matrix set, the calibration output of the previous round is calibrated to obtain the calibration output of the current round.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] The finite element idea is extended to the quantum calibration process, and the interaction characteristics of quantum bits are fully considered through benchmark circuits to construct a weighted quantum bit graph. The quantum bits in the weighted quantum bit graph are divided into groups to obtain a grouping scheme. During the calibration process, the sub-noise matrix is calculated for each group in the grouping scheme, and iterative calibration is performed based on the sub-noise matrix based on the finite element idea, thereby improving the accuracy and efficiency of quantum reading calibration. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0053] Figure 1 is a flow chart of a finite element-based quantum reading calibration method provided in an embodiment;
[0054] Figure 2It is a schematic diagram of the principle of extending the traditional finite element method provided by the embodiment to the quantum finite element method;
[0055] Figure 3 It is a schematic diagram of quantum reading calibration of two rounds of iterations provided in an embodiment;
[0056] Figure 4 is an example diagram of pruning intermediate values of a sparse tensor product engine provided in an embodiment;
[0057] Figure 5 It is a schematic diagram of the structure of a finite element-based quantum reading calibration system provided in an embodiment. DETAILED DESCRIPTION
[0058] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.
[0059] The technical concept of the present invention is: the existing quantum reading calibration stems from the lack of a theoretical basis for simplifying the representation of the noise matrix and the theoretical basis for comprehensively formulating the characteristics of quantum bit interactions. The dilemma is circumvented by simply avoiding quantum bit interactions or directly pruning the matrix, which will result in rough and inaccurate calibration. The present invention takes into account that the reading error is mainly attributed to the local noise between quantum bits, and finds that the finite element method (FEM) can capture and analyze this feature, and FEM has been proven to be effective in approximating many complex evolutionary processes, such as aerodynamic and thermodynamic processes, which is naturally suitable for simulating complex noise during readings. In view of this, the present invention extends the finite element method (FEM) to the calibration process and proposes a quantum reading calibration method based on finite elements.
[0060] In the method of the present invention, calibration is redefined as a series of tensor products associated with multiple sub-noise matrices, which iteratively calibrate the readings through a typical FEM workflow. Similar to FEM, these sub-noise matrices are determined by grouping qubits into multiple groups, each of which attempts to capture the interaction between qubits. In order to quickly and accurately characterize these sub-noise matrices, a novel generation method is proposed to minimize the number of reference circuits. On the other hand, in order to accelerate the calibration calculation, that is, to accelerate a series of tensor products, a sparse tensor product engine is also developed to exploit the inherent sparsity of the noise matrix. Unlike previous methods that directly prune the noise matrix, the present invention chooses to prune the intermediate values to achieve higher characterization accuracy. More importantly, the pruning technique of the present invention only shows polynomial complexity, which greatly reduces the huge computational pressure. In general, the method of the present invention has achieved a high-speed improvement in end-to-end calibration while improving the fidelity of the measurement. It can achieve high scalability and high accuracy in calibration and can be expanded to more than 500 qubits.
[0061] In the quantum reading calibration task, the quantum reading error can be expressed as the error from the ideal distribution P ideal To the measured distribution P measured Linear transformation of :
[0062] P ideal =MP measured (1)
[0063] Where M is defined as the size of The noise matrix, N q represents the number of qubits. In theory, the noise matrix M[x][y] can be characterized by running the benchmark circuit and taking each ideal ground state as input. After execution, the noise matrix is filled according to the output probability distribution P(g.measure=x|g.ideal=y):
[0064] M[x][y]=P(g.measure=x|g.ideal=y) (2)
[0065] in N q For a set of qubits q, P(g.measure=x|g.ideal=y) represents the probability that the measured state of qubit q is x (i.e., g.measure=x) under the condition that the ideal ground state is y (i.e., g.ideal=y). For example, for a 2-qubit system that executes 4 benchmark circuits, the element in (3,0) is 0.02, which means that when the ideal state is |00>, the probability of observing |11> is 0.02, expressed as M[3][0]=P(measure=11|prepare=00)=0.02.
[0066] After obtaining the noise matrix, its inverse matrix (i.e., calibration matrix) can be used. The calibration is performed by MVM (matrix-vector multiplication) operation between the measured distribution and the calibration matrix. In short, by converting the measured quantum distribution P measured With the calibration matrix M -1 Multiplying them together, we can get the probability distribution P of the calibrated output calibrated .
[0067] P calibrated =M -1 P measured (3)
[0068] Figure 2 (a) shows a typical application of the finite element method (FEM) in fluid mechanics, analyzing the force distribution of a sponge under compression. In this case, FEM adopts a divide-and-conquer strategy to simulate the evolution of the sponge, dividing the sponge into multiple lattices and dividing them into groups. Considering the locality of the force distribution, each lattice is analyzed independently to simplify the modeling process. By combining the lattices together, the state of the sponge can be obtained at a certain point in time. According to the current force distribution, the entire evolution process is iteratively updated using FEM. Inspired by this, for the above-mentioned quantum reading calibration task, in the method proposed in the present invention, the reading calibration is iteratively performed according to the FEM workflow, such as Figure 2 As shown in (b). Similar to the sponge case, the noise between qubits is also localized. Therefore, the qubits are divided into multiple groups, and the sub-noise matrix of each group is independently formulated. Unlike the traditional FEM that applies static partitioning, the method of the present invention adopts a variety of grouping schemes during iterative calibration to comprehensively cover the interactions between qubits.
[0069] like Figure 1 As shown, the finite element-based quantum reading calibration method provided in an embodiment of the present invention includes multiple rounds of iterative calibration, and each round of iterative calibration includes a characterization process and a calibration process.
[0070] The characterization process aims to generate the parameters required for calibration, namely the grouping scheme G i and the benchmark probability distribution BP used for sub-noise matrix generation in each iteration l . G l and BP l It will be used in the i-th iteration of the calibration process. Specifically, a weighted quantum graph is constructed to quantify the interactions between quantum bits and group the quantum bits, and a grouping scheme is generated in each iteration. The present invention requires applying different grouping schemes in the calibration process to cover the local interactions between quantum bits. These grouping schemes are intended to maximize the locality of the group. To this end, the interactions between quantum bits are quantified by a weighted quantum bit graph.
[0071] Specifically, in each round of iteration, the weight between any two quantum bits is determined based on the benchmark probability distribution of the benchmark circuit in the current round and a weighted quantum bit graph is constructed. The quantum bits in the weighted quantum bit graph are divided into groups based on the finite element idea to obtain a grouping scheme.
[0072] The first round of benchmark probability distribution is obtained by sampling the output of the benchmark circuit, where the benchmark circuit is screened by setting a threshold, so that a more flexible benchmark circuit can be used to accurately characterize the readout noise. The benchmark circuit involves all qubits of the device, and each qubit has three possible operations:
[0073] 1) Set the qubit to |0> and measure it;
[0074] 2) Set the qubit to |1> and measure it;
[0075] 3) Set the qubit to a random state (|0> or |1>) and do not measure it.
[0076] In order to reduce the number of reference circuits and maintain high accuracy, the present invention identifies candidate circuits that can achieve high quantum bit interactions and analyzes the execution results of the candidate circuits, while the low-interaction candidate circuits mainly involve quantum bit-independent errors, so their outputs can be easily modeled without executing the program. Obviously, some candidate circuits (4 times the number of quantum bits) are first randomly generated to obtain the initial value of the interaction between every two quantum bits. After that, a table is maintained to record the updated interaction values and the number of circuits involved in these interactions. Next, an indicator θ is defined to distinguish key interactions:
[0077]
[0078] Basically, the interaction is static and depends only on the physical device, and it can be estimated more accurately by executing more benchmark circuits. In other words, θ naturally decreases with the number of benchmark circuits. Based on this foundation, the present invention sets a threshold α to specify the accuracy, and for interactions with θ>α, iteratively executes candidate circuits involving it and updates the table.
[0079] Therefore, some candidate circuits are prepared, where q 1 .ideal=0, The measured state of each qubit in the candidate circuit under a given ideal ground state is collected, and the interaction (q i .ideal=x→q j.ideal=y) and the index θ that distinguishes the key interactions. In this way, the candidate circuits are continuously executed until the index θ of all interactions is less than the threshold α, and the benchmark circuit is obtained by screening the candidate circuits whose index exceeds the set threshold. In general, to ensure that the index θ of all interactions exceeds the threshold, more benchmark circuit executions are required to produce significant interactions, and for interactions close to 0, circuit execution is unnecessary because the output distribution of each quantum bit is the same as that of the single-qubit benchmark circuit.
[0080] Each qubit in the circuit is prepared into a state, and a subset of them are measured in the circuit. Each time a candidate or benchmark circuit is executed, the operation and readout output of each qubit are recorded as a triple (ideal, measured, ef), where Used to record the ideal ground state prepared in the circuit. ideal=0 means the ideal ground state bit value is 0, and ideal=1 means the ideal ground state bit value is 1. Indicates that the ideal ground state bit value is empty, Used to record the measurement status output after executing the circuit sample. Measured = 0 means the measurement status bit value is 0, and measured = 1 means the measurement status bit value is 1. Indicates that the measurement state bit value is empty, ef∈{0,1} indicates an error flag, ef=0 indicates that the measurement state matches the ideal ground state, ef=1 indicates that an error occurs, means that the qubit is not measured and has no output. There must be no read errors, i.e. ef = 0. For example, a reference circuit prepares the qubit q2 as |1>. When the qubit q1 is not measured and the output of q2 is |0>, the triples are assigned as follows:
[0081]
[0082] q 2 .ideal=1,q 2 .measured=0,q 2 .ef=1.
[0083] Based on the recorded triplets, it is observed that the readout error is independent of the state of the qubit. In addition, the interaction from one qubit to another changes under different operations (i.e., state preparation and readout), resulting in different readout errors. Based on this observation, the interaction between qubit operations is quantified, and the interaction between two qubits is denoted as interact(q i.ideal=x→q j .ideal=y), calculated as:
[0084] interact(q i .ideal=x→q j .ideal=y)=p(q j .ef=1|C1,C2)-
[0085] p(q j .ef=1|C2) (5)
[0086] Among them, C1:q i .ideal=x;C2: q i .ideal=x represents the i-th quantum bit q i The ideal ground state is x, q j .ideal=y represents the jth quantum bit q j The ideal ground state of is y, interact(·→·) represents the interaction from the former to the latter, q j .ef=1 indicates the jth quantum bit q j The error flag is 1, C1 and C2 represent the intermediate variables used to indicate the ideal state of the bit, p(·|·) represents the conditional probability distribution, which is used as the reference probability distribution BP1; Equation (5) represents q i .ideal = x and q j The correlation between the readout errors.
[0087] After obtaining the reference circuit, obtain the measured state of each qubit in the reference circuit in the ideal ground state, record the triples, and calculate the conditional probability distribution of each qubit based on the triples as the reference probability distribution of the first round. i ,q j ) and calculate the weight weight(q i ,q j ), expressed as:
[0088]
[0089] Based on the weight weight(q i ,q j ) in the weighted quantum bit graph, where quantum bits are nodes, the edges between nodes represent the interactions between quantum bits, and the weights weight(q i ,q j) marks the weight value of the edge, and then divides the qubits in the weighted qubit graph into groups based on the finite element idea to obtain a grouping scheme. Specifically, the weighted qubit graph is divided into qubit groups with the goal of maximizing the locality within the group, that is, maximizing the sum of the weights within the group. Such a grouping scheme ensures that in each iteration, the formulated sub-noise matrix attempts to fully capture the interaction between qubits, making them closer to the actual noise matrix. In the specific implementation, the MAXCUT solver can be used to divide the qubits in the weighted qubit graph into groups.
[0090] The grouping schemes in different iterations are designed to capture different noises. To achieve this goal, after each iteration of the qubit grouping in the characterization process, a calibration is also applied to the reference probability distribution, which eliminates the noise captured in the grouping scheme. The calibration of the reference probability distribution of other rounds other than the first round comes from the calibration output in the iteration process, including: using the calibration output of the iteration process of the current round as the element value to form the reference probability distribution of the next round. Since the measurement is multiple times, taking the current round as the first round as an example, the calibration output of the first round is P 2 , then P 2 As the element values of multiple measurements, the benchmark probability distribution BP of the next round is formed 2 .
[0091] During the calibration process, calibration of the grouping scheme will focus on the remaining noise in the output excluding the grouping scheme. The specific process includes: calculating the sub-noise matrix corresponding to each quantum bit group based on the measured and collected quantum bit collection, the benchmark probability distribution and the grouping scheme, and obtaining a sub-noise matrix set. Based on the sub-noise matrix set, calibrate the calibration output of the previous round to obtain the calibration output of the current round.
[0092] The grouping scheme generated by the characterization process is G l = {g l,1 ,···,g l,K}, where l represents the round, K represents the total number of groups, and g l,k represents the kth qubit group in round l. Due to the different interactions in the circuits measuring different qubits, a sub-noise matrix is dynamically generated for each probability distribution. The set of qubits to be measured is defined as Q M , then, for each g l,k definition:
[0093]
[0094] where g ∩ Indicates that g l,k and Q M Overlapping qubits, Indicates g l,kThe remaining qubits in
[0095] The sub-noise matrix corresponding to each qubit group is expressed as:
[0096]
[0097] The conditional probability P is based on the baseline probability distribution BP of this iteration i To make an estimate, we also introduce additional conditions To ensure that the matrix generation takes into account the unmeasured qubits. In addition, since the measured output of a qubit depends only on the operation of the qubit, equation (8) can be expressed as:
[0098]
[0099] Where q represents the g ∩ The quantum bit, x q represents the position of quantum bit q in bit string x, q.measure = x q Indicates that the measured state of q is x q , g ∩ .ideal=y as a condition, indicating g ∩ The ideal ground state is y, As a condition, it means The ideal ground state is M l,k [x][y] represents g l,k The corresponding sub-noise matrix M l,k The values of the x position and y position in the calculation of the sub-noise matrix are introduced. ∩ .ideal=φ to improve the accuracy of the characterization since these qubits are not measured.
[0100] Define the circuit output as P measured , that is, P 1 , the calibration output is denoted as P L+1 , the present invention reformulates the calibration as an iterative process involving a series of sub-noise matrices:
[0101] Iter.1:
[0102] Iter.2:
[0103] ···;
[0104] Iter.L:
[0105] Among them, M l,1 ,M l,2 ,…,M l,Kare the K sub-noise matrices in the sub-noise matrix set, M l,k is the noise matrix of the k-th qubit group in the l-th iteration, symbol represents the tensor product of these sub-noise matrices, P 2 To P L It acts as an intermediate probability distribution for iteratively approximating the ideal probability distribution. L and K represent the number of iterations and the number of groups, respectively. The lth iteration converts P l Calibrated to P l+1 .
[0106] For the first iteration, P l Calibrated to P l+1 , define p l (x) is the l The probability P of observing a bit string x in the distribution l =∑P l (x)|x>, the bit string x can be divided into sub-bit strings according to the grouping scheme of the lth iteration: |x>=|x l,1 >|x l,2 …|x l,K >, where x l,k is the quantum bit group g l,k The calibration for each iteration is rewritten according to each segment as follows:
[0107]
[0108] Among them, NZ l Indicates P l The set of quantum bit strings with non-zero probability in NZ. Equation (11) has two advantages. First, its time complexity is similar to that of the non-zero probability NZ. l The relationship is linear with the size of , which is typically lower than the number of samples in the readout. Second, the matrix-vector multiplication for each group can be performed independently in constant time, determined by the size of the sub-noise matrix.
[0109] The iterative characterization process and calibration process of the present invention are performed as follows: The calibration process follows a two-step iteration to calibrate the target probability distribution. In each iteration, the sub-noise matrix is first calculated according to equations (7) and (9). The grouping scheme g in equation (7) is l,k The conditional probability in equation (9) is based on the baseline probability distribution BP in the characterization process. l Calculated. Second, the calibration process applies MVM according to equation (11). If there are remaining iterations, it uses the MVM result for the next iteration.
[0110] The static grouping scheme and calibration dataset are based on the fact that the interactions between qubits are constant. This consistency holds in current quantum hardware because current quantum operations (including readout) rely on the hardware's ability to implement a stable (constant) system Hamiltonian. Physically, the interactions are determined after deployment and do not change unless the readout frequency of the qubits is changed. The purpose of grouping is to capture this inherent interaction. In general, the noise matrix describes the perturbation from the ideal probability distribution to the noise distribution, and the finite element method can be used to simulate this perturbation within a group of qubits, through small sub-noise matrices. Different grouping schemes are then used to capture the interactions across groups.
[0111] It should be noted that the present invention has two features: 1) Baseline probability distribution BP 1 is the probability distribution obtained by running the benchmark circuit, and the other benchmark probability distributions BP in each iteration l Updated using equation (11); 2) The sub-noise matrix is dynamically generated based on the read measured qubits in both the characterization process and the calibration process to maximize fidelity, since the interaction always changes under different combinations of measured qubits. For the target quantum device, the calibration parameter (G l , BP l ) is static.
[0112] Figure 3 2 is a schematic diagram of quantum reading calibration for two rounds of iterations provided in the embodiment. In the characterization process, the first iteration (iter=1) is based on the reference probability distribution BP 1 The grouping scheme G is obtained 1 BP 1 and G 1 is stored for use in the calibration process. Then, update BP 1 Each probability distribution of BP 2 The second iteration (iter=2) outputs the BP-based 2 The derived quantum bit partitioning scheme G 2 Finally, the characterization process outputs CP = [G 1 , G 2 ],[BP 1 , BP 2 ], which serve as calibration parameters. In the calibration process, for any measurement probability distribution P 1 and the measured quantum bit Q M , the first iteration is based on Q M , G 1 and BP to obtain the sub-noise matrix, and use equation (11) to convert P 1 Update to P 2The second iteration follows the same steps based on Q M , G 2 and BP 3 Obtain the sub-noise matrix and calibrate P 2 To P 3 .P 3 is the resulting probability of these two iterative calibrations.
[0113] In an embodiment, the iterative calibration based on equation (11) involves tensor products of multiple MVM results with exponential complexity, but it is observed that the non-zero elements of the sub-noise matrix are mainly distributed near the diagonal. This sparsity will accumulate to a higher degree in the tensor product process. Therefore, a sparse tensor product engine is proposed to speed up the calculation.
[0114] Specifically, we first calculate the Then, the tensor product of the MVM results of all qubit groups is calculated, and then the intermediate values in the tensor product are pruned by another threshold β. l (x) is aggregated with the remaining intermediate values in the tensor product to obtain the calibration output. Figure 4 An example is provided. ① For the non-zero probability (0.47) in state 000, the engine calculates and ②Then calculate the tensor product of the two MVM results; ③Then prune the intermediate values of the tensor product by the threshold β. Similarly, the engine executes ①②③ for the non-zero probability (0.53) under state 011, and aggregates the pruned intermediate values by ④. The sparse tensor product engine chooses to prune on the intermediate values instead of directly on the noise matrix, so that higher characterization accuracy can be achieved. In addition, as the chain of multiple tensor products grows, the number of pruned intermediate values also increases, which actually greatly reduces the computational complexity.
[0115] like Figure 5As shown, the embodiment also provides a quantum reading calibration system based on finite elements, including a characterization process module and a calibration process module, and the characterization process module and the calibration process module are sequentially iterated to realize quantum reading calibration; wherein, the characterization process module is used to determine the weight between any two quantum bits based on the benchmark probability distribution of the benchmark circuit in the current round and construct a weighted quantum bit map, and the quantum bits in the weighted quantum bit map are divided into groups based on the finite element idea to obtain a grouping scheme, wherein the benchmark probability distribution of the first round is obtained by sampling the output of the benchmark circuit, and the benchmark probability distribution of other rounds other than the first round comes from the calibration output in the iterative process; the calibration process module is used to calculate the sub-noise matrix corresponding to each quantum bit group based on the measured and collected quantum bit set, the benchmark probability distribution and the grouping scheme, to obtain a sub-noise matrix set, and calibrate the calibration output of the previous round based on the sub-noise matrix set to obtain the calibration output of the current round.
[0116] It should be noted that the quantum reading calibration system based on finite elements provided in the above embodiments should be illustrated by the division of the above functional modules when performing quantum reading calibration. The above functional allocation can be completed by different functional modules as needed, that is, the internal structure of the terminal or server is divided into different functional modules to complete all or part of the functions described above. In addition, the quantum reading calibration system based on finite elements provided in the above embodiments and the quantum reading calibration method embodiment based on finite elements belong to the same concept. The specific implementation process is detailed in the quantum reading calibration method embodiment based on finite elements, which will not be repeated here.
[0117] The quantum reading calibration method and system provided by the present invention uses a comprehensive framework for calibrating measurement errors using the finite element method (FEM), which provides a deeper understanding of characterization calibration; at the same time, the method for generating a reference circuit is proposed, which is comparable to the exponential time complexity (O(2 n )) is different from the characterization of the reference circuit of the present invention, the method of generating the reference circuit greatly reduces the characterization time and reduces it to the polynomial complexity (O(n 2 )); Furthermore, the proposed sparse computing engine can accelerate end-to-end calibration, and the calibration time is only proportional to the cube of the number of quantum bits, that is, O(n 3 ), and so are the memory requirements.
[0118] It is also theoretically proven that the complexity of the invention does not exceed polynomial complexity with respect to the number of qubits. Specifically, four quantum cloud platforms were used to collect a data set of 25 million real machine samples on quantum devices ranging from 7 qubits to 135 qubits. The results show that compared with the most advanced matrix-based calibration technology, the invention achieves 1.8×10 9times speedup and provided 1.2× and 1.4× fidelity improvements on real quantum devices with 18 and 36 qubits.
[0119] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A quantum reading calibration method based on finite element, characterized in that: It includes multiple rounds of iterative calibration, each round of iterative calibration includes a characterization process and a calibration process. The characterization process includes: determining the weight between any two quantum bits based on the benchmark probability distribution of the benchmark circuit in the current round and constructing a weighted quantum bit graph, and dividing the quantum bits in the weighted quantum bit graph into groups based on the finite element concept to obtain a grouping scheme, wherein the benchmark probability distribution of the first round is obtained by sampling the output of the benchmark circuit, and the benchmark probability distribution of other rounds other than the first round comes from the calibration output in the iterative process; The calibration process includes: calculating the sub-noise matrix corresponding to each qubit group based on the measured and collected qubit set, the benchmark probability distribution, and the grouping scheme to obtain a sub-noise matrix set, and calibrating the calibration output of the previous round based on the sub-noise matrix set to obtain the calibration output of the current round; The sub-noise matrix corresponding to each quantum bit group is calculated based on the measured quantum bit collection, the reference probability distribution and the grouping scheme, including: The collection of quantum bits collected is Q M , the grouping scheme is G l = {g l,1 , …, g l,K }, where l represents the round, K represents the total number of groups, and g l,k represents the kth quantum bit group in the lth round; For each g l,k Definition: g ∩ =Q M ∩g l,k , where g ∩ Indicates that g l,k and Q M Overlapping qubits, Indicates g l,k The remaining qubits in The sub-noise matrix corresponding to each qubit group is calculated as: Where q represents the g ∩ The quantum bit, x q represents the position of quantum bit q in bit string x, P represents the conditional probability, based on the benchmark probability distribution BP l It is estimated that q.measure = x q Indicates that the measured state of q is x q , g ∩ .ideal=y as a condition, indicating g ∩ The ideal ground state is y, As a condition, it means The ideal ground state is M l,k [x][y] represents g l,k The corresponding sub-noise matrix M l,k The values of the x and y positions in the .
2. The finite element based quantum reading calibration method according to claim 1, characterized in that: The reference circuit is obtained by screening by setting a threshold, including: executing a candidate circuit and obtaining a measurement state of each quantum bit of the candidate circuit under a given ideal ground state, and calculating the interaction between two quantum bits and an index for distinguishing key interactions based on the ideal ground state and the measurement state, and obtaining the reference circuit by screening candidate circuits whose indexes exceed the set threshold.
3. The finite element based quantum reading calibration method according to claim 2, characterized in that: Each time a candidate circuit is executed, the operation and readout output of each qubit is recorded as a triple (ideal, measured, ef), where Used to record the ideal ground state prepared in the circuit. ideal=0 means the ideal ground state bit value is 0, and ideal=1 means the ideal ground state bit value is 1. Indicates that the ideal ground state bit value has not been measured, Used to record the measurement status output after executing the circuit sample. Measured = 0 means the measurement status bit value is 0, and measured = 1 means the measurement status bit value is 1. Indicates that the measurement state bit value has not been measured, ef∈{0,1} indicates an error flag, ef=0 indicates that the measurement state matches the ideal ground state, and ef=1 indicates that an error has occurred; The interaction between two quantum bits is expressed as interact(q i .ideal=x→q j .ideal=y), calculated as: interact(q i .ideal=x→q j .ideal=y)=p(q j .ef=1|C1,C2)-p(q j .ef=1|C2) C1:q i .ideal=x; C2: Among them, q i .ideal=x represents the i-th quantum bit q i The ideal ground state is x, q j .ideal=y represents the jth quantum bit q j The ideal ground state of is y, interact(·→·) represents the interaction from the former to the latter, q j .ef=1 indicates the jth quantum bit q j The error flag of is 1, C1 and C2 represent intermediate variables used to indicate the ideal state of the bit, and p(·|·) represents the conditional probability distribution of the actual value measured under the ideal value, which is used as the reference probability distribution; The index θ for distinguishing key interactions based on the interaction calculation of two qubits is expressed as: Here, num(·→·) represents the number of interactions.
4. The finite element based quantum reading calibration method according to claim 3, characterized in that: The first round of benchmark probability distribution is obtained by sampling the output of the benchmark circuit, including: obtaining the measurement state of each qubit in the benchmark circuit under the ideal ground state, and recording the triplet, and calculating the conditional probability distribution of each qubit based on the triplet as the first round of benchmark probability distribution; The benchmark probability distributions of other rounds other than the first round are derived from the calibration outputs in the iteration process, including: using the calibration outputs of the iteration process of the current round as element values to form the benchmark probability distribution of the next round.
5. The finite element based quantum reading calibration method according to claim 3, characterized in that: Determine the weight between any two quantum bits based on the benchmark probability distribution of the benchmark circuit in the current round and construct a weighted quantum bit graph, including: The weight between any two quantum bits (q i ,q j ) is expressed as: In the constructed weighted quantum bit graph, quantum bits are nodes, and the edges between nodes represent the interactions between quantum bits, and the weights (q i ,q j ) marks the weight value of the edge.
6. The finite element based quantum reading calibration method according to claim 5, characterized in that: Based on the finite element concept, the quantum bits in the weighted quantum bit graph are divided into groups to obtain a grouping scheme, including: With the goal of maximizing the locality within the group, that is, maximizing the sum of the weights within the group, the weighted quantum bit graph is divided into quantum bit groups to obtain a grouping scheme.
7. The finite element based quantum reading calibration method according to claim 1, characterized in that: The calibration output of the previous round is calibrated based on the sub-noise matrix set to obtain the calibration output of the current round, including: Among them, M l,1 , M 1,2 , …, M l,K are the K sub-noise matrices in the sub-noise matrix set, M l,k is the sub-noise matrix of the k-th qubit group in the l-th iteration, P l+1 represents the calibration output of the l+1th round, P l represents the calibration output of the lth round, symbol Represents the matrix tensor product, NZ l Indicates P l The set of quantum bit strings with non-zero probability in |x l,k > represents the sub-bit string corresponding to the k-th quantum bit group in the l-th round, p l (x) is the l The probability of observing bit string x from the distribution.
8. The finite element based quantum reading calibration method according to claim 7, characterized in that: In the calibration process, a sparse tensor product engine is introduced to prune the tensor product in the calibration calculation process to accelerate the calculation, including: First, calculate the The matrix-vector multiplication value of P is then calculated, and the tensor product of the matrix-vector multiplication values of all qubit groups is then pruned by another threshold to remove the intermediate values in the tensor product. l (x) is aggregated with the remaining intermediate values in the tensor product to obtain the calibration output.
9. A quantum reading calibration system based on finite elements, characterized in that: It includes a characterization process module and a calibration process module, and the characterization process module and the calibration process module are sequentially iterated to realize quantum reading calibration; The characterization process module is used to determine the weight between any two qubits based on the benchmark probability distribution of the benchmark circuit in the current round and construct a weighted qubit graph. The qubits in the weighted qubit graph are divided into groups based on the finite element concept to obtain a grouping scheme, where the benchmark probability distribution of the first round is obtained by sampling the output of the benchmark circuit, and the benchmark probability distribution of other rounds other than the first round comes from the calibration output in the iterative process; The calibration process module is used to calculate the sub-noise matrix corresponding to each qubit group based on the measured and collected qubit collection, the benchmark probability distribution and the grouping scheme, obtain the sub-noise matrix set, and calibrate the calibration output of the previous round based on the sub-noise matrix set to obtain the calibration output of the current round; The sub-noise matrix corresponding to each quantum bit group is calculated based on the measured quantum bit collection, the reference probability distribution and the grouping scheme, including: The collection of quantum bits collected is Q M , the grouping scheme is G l = {g l,1 , …, g l,K }, where l represents the round, K represents the total number of groups, and g l,k represents the kth quantum bit group in the lth round; For each g l,k Definition: g ∩ =Q M ∩g l,k , where g ∩ Indicates that g l,k and Q M Overlapping qubits, Indicates g l,k The remaining qubits in The sub-noise matrix corresponding to each qubit group is calculated as: Where q represents the g ∩ The quantum bit, x q represents the position of quantum bit q in bit string x, P represents the conditional probability, based on the benchmark probability distribution BP l It is estimated that q.measure = x q Indicates that the measured state of q is x q , g ∩ .ideal=y as a condition, indicating g ∩ The ideal ground state is y, As a condition, it means The ideal ground state is M l,k [x][y] represents g l,k The corresponding sub-noise matrix M l,k The values of the x and y positions in the .
