betaFeedback

WorksPhysics and Astronomy

Quantum State Tomography using Quantum Machine Learning

Abstract

Quantum State Tomography (QST) is a fundamental technique in Quantum Information Processing (QIP) for reconstructing unknown quantum states. However, the conventional QST methods are limited by the number of measurements required, which makes them impractical for large-scale quantum systems. To overcome this challenge, we propose the integration of Quantum Machine Learning (QML) techniques to enhance the efficiency of QST. In this paper, we conduct a comprehensive investigation into various approaches for QST, encompassing both classical and quantum methodologies; We also implement different QML approaches for QST and demonstrate their effectiveness on various simulated and experimental quantum systems, including multi-qubit networks. Our results show that our QML-based QST approach can achieve high fidelity (98%) with significantly fewer measurements than conventional methods, making it a promising tool for practical QIP applications.

Introduction

Quantum Information Processing (QIP) involves the storage, transmission, and computation of information using the principles of Quantum Mechanics as a driver for effectively carrying out these tasks. For QIP to be effective, quantum systems must be prepared, controlled, and characterised. As a result of the act of measurement on a quantum system, the system will inevitably disintegrate from its current state to one of its eigenstates via wavefunction collapse. Subsequently, the usage of a single copy of the collapsed state to access the system’s initial state is impossible. Additionally, the no-cloning theorem [1] prohibits making multiple copies of the unknown state in order to reconstruct it further. Quantum State Tomography (QST) is an experimental procedure where the ensemble of unknown, but identically prepared quantum states, is characterised by a sequence of measurements in different bases, enabling the reconstruction of its density matrix. Analogous to medical tomographic reconstructions, in QST, several measurements in the various bases are taken and combined to give a reconstruction of the complete (initial) quantum state.

Measurement on a quantum system generally gives a probabilistic result, and the measurement outcome only provides limited information about the state of the system, even when an ideal measurement device is used. QST consists of only finite measurements and the use of appropriate estimation algorithms. Hence, choosing optimal measurement sets, and designing efficient state reconstruction algorithms, are two critical issues in quantum state tomography.

Thus, we summarise the process of QST in Algorithm 1.

Algorithm 1 Quantum State Tomography

Input: Tolerance ϵ≪1\epsilon\ll1, create n∈Nn \in\mathbb{N} identical copies, {∣ϕi⟩}i=1n\{|\phi_i\rangle\}_{i=1}^{n}, of the target state to be estimated ∣ψ⟩|\psi\rangle to form an ensemble

Initialize: Parameter information M\mathbf{M}, and density matrix ρ\rho

while ∥M−ρ∥>ϵ\lVert\mathbf{M}-\rho\rVert> \epsilon do

    for each copy {∣ϕi⟩}i=1n\{|\phi_i\rangle\}_{i=1}^{n} do

    Perform measurements on each state in the ensemble, forming a complete basis

∑i=1n∣ϕi⟩⟨ϕi∣=I\sum_{i=1}^{n}|\phi_i\rangle\langle\phi_i|=I

    Apply an estimation method, E\mathcal{E}, to recover parameter information, M\mathbf{M}, from the measurement results

    Calculate density matrix ρ\rho

    end for

end while

Return: Density matrix ρ\rho

Once the density matrix is obtained for the reconstructed state, one needs to test it for physical viability, i.e. if the resulting density matrix is positive semidefinite (non-negative eigenvalues), and has unit trace. If these conditions are met, then such an estimation is correct.

Literature Review

QST is widely used in several application areas including Quantum Error Mitigation [2, 3], State Estimation [4, 5], applications to qubit measurement and experimental reconstruction [6, 7], qudit system reconstructions [8], and Photonics [9].

Quantum Machine Learning (QML) is a rapidly developing field that combines the disciplines of Classical Machine Learning (ML) with Quantum Computing (QC). Since the advent of QC as a research track, it became natural to coalesce the two fields to try and attain a computational advantage over the implementation of ML applications. These include the design of QML algorithms that have a speed-up over their classical counterparts, reduced computational complexity and resources over ML algorithms, and a reduction in the usage of physical hardware. There exists a plethora of ML applications to QST, and thus, the application of QML to QST was an organic next step. Therefore, this paper claims no novelty in the application of QML to QST, however, the novelty presented lies in the widespread coverage of the methods, the structured approaches, and the high fidelity of the results obtained.

In this paper, we discuss some of the classical approaches employed for reconstruction and thereafter build on these principles to apply quantum algorithms based on NISQ-era [10] devices that can be employed to achieve efficient QST.

This paper is segmented as follows:

In Sec. 2., we provide a comprehensive literature review of important works in the field, and describe, in sufficient detail, the methods employed, and the novelties of each research piece.

In Sec. 3., we discuss the classical approaches to state tomography by giving detailed descriptions of the theoretical underpinnings of each method.

In Sec. 4., we discuss the mathematical, computational, and circuit architecture of the QST methods used.

In Sec. 5., we discuss the results of each method employed.

In Sec. 6., we provide a conclusion to this research by discussing all that was done, all that was achieved, and future avenues of research to explore.

In [11], Torlai et al proposed a new method using Neural Networks (NNs) that can be used to effectively reconstruct quantum states for generalised highly-entangled states using simplistic, but limited in number, experimental data. For a vast array of quantum devices, namely: Adiabatic quantum simulators, higher-dimensional ion traps, highly-entangled quantum circuits, and ultra-cold atoms, the authors have shown that their method can be applied, in addition to the construction of quantifiers that could be difficult to measure directly. Lastly, and most importantly, it was shown that the proposed method was robust against noise.

In [12], Schmale et al use Convolutional Neural Networks (CNNs) to reconstruct quantum states. In particular, a novel approach to accurately estimate observables from tomographic measurement data using a variational manifold represented by a CNN is presented. It was demonstrated that their proposed method outperformed analogous classical methods, namely the Maximum Likelihood Estimate (MLE), by up to an order of magnitude, by achieving high classical fidelities, and by reducing the error of estimation. Perhaps most important was the finding that their method scaled polynomially as the size of the system grew, making the method more versatile and able to handle larger systems with more complex quantum states.

In [13], Quek et al propose a new algorithm called Neural Adaptive Quantum Tomography (NAQT) that is modular, fast, and highly flexible. Using NNs, the algorithm resiliently optimises measurements and reconstructs quantum states with high accuracy, while being agnostic to the number of qubits involved, and the type of measurements used. Further, it was demonstrated that the NAQT algorithm outperformed the Adaptive Bayesian Quantum Tomography (ABQT) and Maximum Likelihood Quantum Tomography (MLQT) methods in terms of speed and accuracy; however, NAQT achieves a comparable reconstruction accuracy when benchmarked against ABQT and MLQT. The modularity of the algorithm lies in its ability to be seamlessly retrained, in a reasonable time, in order to suit the task at hand.

In [14], Koutný et al apply Deep Learning (DL) to QST. Using identically-prepared copies of the system, NNs were used to prepare the density matrix of the system from measurements. The novelty introduced in this paper was the concept of the positivity of the density matrix baked into the NN architecture. It was demonstrated that this method achieved state-of-the-art results in terms of speed when compared to the MLE and Semidefinite Programming (SDP) approaches in reconstructing the density matrix, and hence, estimating the initial state. In particular, this novel approach was roughly four orders of magnitude faster than the MLE and three orders of magnitude faster than the SDP. Lastly, it was shown that the results naturally extend to Quantum Information tasks.

In [2], Hai and Ho introduce a new algorithm called Universal Compilation (UC) that aims to maximise the efficiency of the QST task, with a focus on quantum sensing and metrology as application domains. Several novelties were introduced, namely:

2.1. The introduction of a new cost function

C(θ)=d(Ψ,Φ(θ)),C(\theta)=d(\Psi,\Phi(\theta)),

based upon the Fubini-Study metric (distance measurement between two quantum states) [15], for the purposes of this study, it is given by

d(Ψ,Φ)=cos⁡−1(∣⟨Ψ∣Φ⟩∣2),d(\Psi,\Phi)=\cos^{-1}\left(\lvert\langle\Psi\vert\Phi\rangle\rvert^2\right),

where ∣Ψ⟩\lvert\Psi\rangle is the target (unknown) state, and ∣Φ(θ)⟩\lvert\Phi(\theta)\rangle is the parameterised variational (reconstructed) state with parameters θ=(θ1,θ2,…,θn)∈[0,1]\theta=(\theta_1,\theta_2,\ldots,\theta_n)\in[0,1]. The associated optimisation problem is stated as

θ∗=arg min⁡θ1,θ2,…,θn C(θ),\theta^*=\underset{\theta_1,\theta_2,\ldots,\theta_n}{\operatorname{arg\,min}}\ C(\theta),

using a gradient-based scheme, with updates of the form θ←θ−η∇θC(θ)\theta\leftarrow\theta-\eta\nabla_\theta C(\theta), for 0≤η≤10\leq\eta\leq1.

2.2. The versatility of UC on qubit and qudit systems.

2.3. The UC algorithm can improve the accuracy of the QST tasks in the presence of noise, and other sources of errors.

In [16], the MLE, least squares, generalised least squares, positive least squares, thresholded least squares, and projected least squares (PLS) methods have been analysed, compared, and the computational efficiencies of the methods have been presented. The measurement scenarios have been realised with respect to Pauli bases, random bases measurements, and the covariant measurement. In addition, a complete set of simulation results have been shared online via an interactive Shiny application.

In [17], a least-square inversion method, which reconstructs the density matrix from measurable time-dependent probability distributions of a physical system from oscillators (harmonic and anharmonic), has been proposed. The applicability of the method has been compared with other methods based on least-squares inversion.

The Quantum Process tomography (QPT) task, which estimates unknown quantum transformations completely from measurement data, is a powerful, resource-intensive method employed in the field of quantum technology. In [18], the PLS method has been proposed and investigated for QPT. The PLS method computes the least-squares estimator of the Choi matrix [19] of an unknown channel, and projects it onto the convex set of Choi matrices. The PLS method has been illustrated with numerical experiments involving channels on systems with up to a 7-qubit system.

In [20], a new quantum algorithm, which determines the quality of a least-squares fit over an exponentially large dataset, used to solve a system of linear system equations efficiently, has been proposed. The proposed algorithm efficiently finds a concise function that approximates the data to be fitted and gives a bound for the approximation error. For pure quantum states, the algorithm performs an efficient parametric estimation of the quantum state. A use-case of this approach is the performance of a complete QST analysis.

In [21], a QML technique for QST on an unknown quantum state is introduced. Mechanically, the learning process of the technique involved maximising the fidelity between the output of a VQC and the state. The viability of the method has been demonstrated by performing numerical simulations for the tomography of the ground state of a one-dimensional quantum spin chain using a VQC simulator.

In addition to the methods described above, there have been several other ML approaches: Using conditional Generative Adversarial Networks (GANs) – see [22], using attention mechanisms [23], and the ML literature contained there in the abovementioned research pieces.

Classical Methods for Quantum State Tomography

In this section, we present the theoretical foundations of the various classical techniques commonly used for quantum state tomography.

Linear Inversion

The first method to appear in literature was linear inversion [24], which is based on the inverse of Born’s rule. The method entails equating experimentally normalised frequencies to the probabilities predicted from Quantum Mechanical calculations. The reconstructed state, however, might not match a physical state because of experimental noise. The predicted density matrix must be confined to the space of semidefinite, positive, and unit trace matrices in order to confirm its physicality. The statistical inference problem at hand can be described as being constrained parameter estimation.

Consider the dd-dimensional Hilbert space, Hd\mathcal{H}^{d}. Let Γμ\Gamma_{\mu} represent elements of an orthonormal basis for Hermitian matrices in this space. Thus,

Tr⁡(ΓμΓν)=δμν,(1)\operatorname{Tr}(\Gamma_{\mu}\Gamma_{\nu}) = \delta_{\mu\nu}, \tag*{(1)}

where δμν\delta_{\mu\nu} is the Kronecker-delta function. The density matrix can then be written as

ρ=∑μSμΓμ,(2)\rho= \sum_{\mu} S_{\mu}\Gamma_{\mu}, \tag*{(2)}

where SμS_{\mu} are also called the Stokes parameters for this representation. Thus, measurements described by Positive Operator-Valued Measure (POVM) set elements, OμO_{\mu}, yield results given by fμ=Tr⁡(Oμρ)f_{\mu} = \operatorname{Tr}(O_{\mu}\rho). These frequencies, in terms of the basis operators, are then given by

fμ=∑νSνTr⁡(OμΓν).(3)f_{\mu} = \sum_{\nu} S_{\nu}\operatorname{Tr}(O_{\mu}\Gamma_{\nu}). \tag*{(3)}

(3) can we write vectorially as the linear equation

BsT=f,(4)B\mathbf{s}^{T} = \mathbf{f}, \tag*{(4)}

where s\mathbf{s} is the Stokes vector, BB is the matrix calculated in terms of the Basis Γν\Gamma_{\nu} and the measurement operators OμO_{\mu} as Bμν=Tr⁡(OμΓν)B_{\mu\nu} = \operatorname{Tr}(O_{\mu}\Gamma_{\nu}), and f\mathbf{f} is the frequency vector. Thus, this technique is termed the linear tomography reconstruction or the linear inversion method.

Solving the system of linear equations from Eq. (4), however, is not guaranteed to yield a physical quantum state ρ\rho through the parameters s\mathbf{s}. This is because real-world measurements have statistical noise, which leads to the solution to this equation defying the properties of unit trace and positive semi-definiteness.

The major drawback of this method is that an exponential number of measurements, with respect to the number of qubits in the system, is required to be performed. Similarly, the MLE methods suffer from this disadvantage; a discussion is left for the next section.

Maximum Likelihood Estimation

This approach was developed by Řeháček et al [25] for discrete-variable systems. It has the following basic structure: Consider a system of NN identically prepared unknown quantum states. By using the results of the measurements made on this system, the identity of this unidentified quantum system is determined as defined by the density operator. The measurements produce multinomial statistics for the number of occurrences njn_j for each result jj. The likelihood for this measurement data D\mathcal{D} is calculated.

Additionally, the Hermitian operator,

R=∑jfjpjΠj(5)R = \sum_j \frac{f_j}{p_j}\Pi_j \tag*{(5)}

was introduced as possessing the property of being positive semi-definite. This property encompasses the outcomes of the measurements and the corresponding measurement operators. In (5), fjf_j are the measured frequencies for outcome jj, Πj\Pi_j are the corresponding measurement operators, and pjp_j are the outcome expected from the quantum state ρ\rho, i.e. pj=Tr⁡(ρΠj)p_j = \operatorname{Tr}(\rho\Pi_j).

Using RR, one can find the reconstructed state, ρ\rho, according to the iterative equation

ρk+1=αRρkR,(6)\rho^{k+1} = \alpha R\rho^kR, \tag*{(6)}

where α\alpha is a normalisation constant that ensures that ρ\rho has unit trace. Note that RR here is not constant across all iterations, but itself depends on ρ\rho. The starting state ρ0\rho^0 for this procedure is taken to be the maximally-mixed state. i.e, ρ0=I\rho^0 = I.

The essence of classical MLE is the form of the loss function E(ρ)E(\rho), given by

E(ρ)=∑jfjlog⁡[Tr⁡(ρΠj)].(7)E(\rho) = \sum_j f_j \log[\operatorname{Tr}(\rho\Pi_j)]. \tag*{(7)}

Minimisation of Eq. (7), by the standard gradient descent algorithm, for a particular parameterisation of ρ\rho can be shown to be equivalent to the iterative approach that is mentioned in [26].

To ensure positivity and unit trace the parameterisation employed is of the following form.

ρ=TT†Tr⁡(TT†),(8)\rho= \frac{TT^\dagger}{\operatorname{Tr}(TT^\dagger)}, \tag*{(8)}

where TT is taken to be of upper triangular form with complex entries.

Least Squares Estimation

In this section, the details of the Least Square Estimation (LSE) for the QST are presented. The idea behind Least Square Estimation traces its roots back to Gauss in 1809 when studying the motions of celestial bodies in conic sections around the sun. Today it is a standard tool in Statistics and supervised ML for a profusion of applications, and very well known. Below, we outline the mechanics of the method in brief [27, 28].

LSE finds the best linear fit for which the summation of the squares will be minimum. To minimise the error value, we take the first derivative equation.

Suppose that you have nn samples of data (X(1),X(2),…,X(n))(X(1), X(2), \ldots, X(n)), and the corresponding output labels for this data (Yi(1),Yi(2),…,Yi(n))(Y_i(1), Y_i(2), \ldots, Y_i(n)).

X=[X1,X2,…,Xnq]T,(9)\mathbf{X} = [X_1, X_2, \ldots, X_{n_q}]^{T}, \tag*{(9)}

where X\mathbf{X} is the vector representation of the features. We assume that output is linearly related to the input, and for the ithi^{\text{th}} data point, the equation is

Yi(k)=Q1X1(k)+Q2X2(k)+…+QnqXnq(k)+E(k),(10)Y_i(k) = Q_1X_1(k)+Q_2X_2(k)+\ldots+Q_{n_q}X_{n_q}(k)+E(k), \tag*{(10)}

where E(k)E(k) denotes the error value. This is generalised in, matrix form, to

Y(k)=XT(k)Q(k)+E(k),(11)\mathbf{Y}(k) = \mathbf{X}^{T}(k)\mathbf{Q}(k) + \mathbf{E}(k), \tag*{(11)}

where

X=[X1(1)X2(1)…Xnq(1)X1(2)X2(2)…Xnq(2)⋮⋮⋱⋮X1(n)X2(n)…Xnq(n)].\mathbf{X} = \begin{bmatrix} X_1(1) & X_2(1) & \ldots& X_{n_q}(1) \\ X_1(2) & X_2(2) & \ldots& X_{n_q}(2) \\ \vdots& \vdots& \ddots& \vdots\\ X_1(n) & X_2(n) & \ldots& X_{n_q}(n) \end{bmatrix}.

The goal is to estimate the values of Q\mathbf{Q} by minimising the sum of the squares of the error values. The associated cost function, known as the Mean Squared Error (MSE) cost, denoted J(Q)J(\mathbf{Q}), is given by

J(Q)=12∑k=1nE2(k)=ETE=12(Y−XQ)T(Y−XQ)=12(YT−XTQT)(Y−XQ).(12)\begin{aligned} J(\mathbf{Q}) = \frac{1}{2}\sum_{k=1}^{n} E^2(k) = \mathbf{E}^{T}\mathbf{E} \\ &= \frac{1}{2}(\mathbf{Y} - \mathbf{XQ})^{T}(\mathbf{Y} - \mathbf{XQ}) \\ &= \frac{1}{2}(\mathbf{Y}^{T} - \mathbf{X}^{T}\mathbf{Q}^{T})(\mathbf{Y} - \mathbf{XQ}). \tag*{(12)} \end{aligned}

Taking the derivative of Eq. (12), we obtain

δJ(Q)δQ=−YTX+QTXTX.(13)\frac{\delta J(\mathbf{Q})}{\delta\mathbf{Q}} = -\mathbf{Y}^{T}\mathbf{X} + \mathbf{Q}^{T}\mathbf{X}^{T}\mathbf{X}. \tag*{(13)}

Setting Eq. (13) to zero, and solving for Q\mathbf{Q}, we obtain

Q=(XTX)−1XTY.(14)\mathbf{Q} = (\mathbf{X}^{T}\mathbf{X})^{-1}\mathbf{X}^{T}\mathbf{Y}. \tag*{(14)}

ML and Covariance Matrix Estimation

We provide a brief review of the developments in estimating large covariance matrices and precision matrices. In order to estimate these large covariance matrices, the assumption for the target matrix of interest is sparsity which is most commonly the case for precision matrices [29]. There are two general approaches.

3.4.1. Rank-based Approach: This technique is applicable when the processing of data exhibits a heavily-tailed, or non-Gaussian distribution.

3.4.2. Factor-model-based Approach: To deal with conditional sparsity, the remaining elements of the output variables become sparse. For this, a factor model is required.

Covariance estimation uses factor models by utilising the principal components method. The most common approach to the estimation of sparse precision matrices is the MLE. Let Y1,Y2,…,YT∼i.i.d.DY_1,Y_2,\ldots,Y_T \overset{\mathrm{i.i.d.}}{\sim} \mathcal{D} be random variables that are distributed independently and identically, then the negative Gaussian log-likelihood is defined as

G(Θ)=Tr⁡(SΘ)−log⁡∣Θ∣,(15)G(\Theta)=\operatorname{Tr}(S\Theta)-\log|\Theta|, \tag*{(15)}

where, for independently weak data and non-Gaussian data, G(Θ)G(\Theta) behaves as a quasi-negative log-likelihood. We consider the penalised likelihood approach

Θ^=arg⁡min⁡Θ=(θij)p×p Tr⁡(SΘ)−log⁡∣Θ∣+∑i≠jPwT(∣θij∣),(16)\widehat{\Theta}=\underset{\Theta=(\theta_{ij})_{p\times p}}{\arg\min}\ \operatorname{Tr}(S\Theta)-\log|\Theta|+\sum_{i\ne j}P_{w_T}(|\theta_{ij}|), \tag*{(16)}

where PwTP_{w_T} is the penalty function that encourages the sparsity of Θ^\widehat{\Theta}. The penalties that are commonly used are convex penalties, i.e., L1-penalty (lasso). However, in contrast, concave penalties tend to perform better. As the value of the parameter increases, it minimises the shrinkage bias of folded concave penalties.

Thereafter, we look to solve the adjacent problem of estimating the covariance matrix Σ\Sigma under the elliptical model, which poses a significant challenge, especially when sparse estimations are required. Adopting a similar approach as the previous section, we impose a sparsity assumption on Σ\Sigma. To address this issue, the EC2 (Estimation of Covariance with Eigenvalue Constraints) estimator is used, which is a regularised rank-based estimation method. This is regarded as a natural extension of the generalised thresholding operator.

An exceptional feature of the EC2 estimator is its ability to guarantee the positive definiteness of the estimated covariance matrix. This is achieved by explicitly constraining the smallest eigenvalue of the estimated covariance matrix, a feature not present in many existing methods.

The EC2 estimator is defined for the sparse covariance matrix as

R^EC2:=arg⁡min⁡diag⁡(R)=1 12∥R^−R∥F2+λ∥R∥1,off,(17)\widehat{\mathbf{R}}^{\mathrm{EC2}}:=\underset{\operatorname{diag}(\mathbf{R})=1}{\arg\min}\ \frac{1}{2}\lVert\widehat{\mathbf{R}}-\mathbf{R}\rVert_F^2+\lambda\lVert\mathbf{R}\rVert_{1,\mathrm{off}}, \tag*{(17)}

such that τ≤Λmin⁡(R)\tau\le\Lambda_{\min}(\mathbf{R}). In this equation, R^EC2\widehat{\mathbf{R}}^{\mathrm{EC2}} represents the sparse estimator for R\mathbf{R}, derived by simultaneously imposing a sparse estimation and a positive-definiteness constraint. The regularisation parameter, λ>0\lambda>0, together with the desired minimum eigenvalue lower bound, τ>0\tau>0, are predetermined. The constraint diag⁡(R)=1\operatorname{diag}(\mathbf{R})=1 ensures the resulting R^EC2\widehat{\mathbf{R}}^{\mathrm{EC2}} is a correlation matrix. To finalise the process, the covariance matrix estimator, Σ^\widehat{\Sigma}, is obtained by converting R^EC2\widehat{\mathbf{R}}^{\mathrm{EC2}}.

The EC2 estimator’s asymptotic properties are crucial in understanding its behaviour. We consider two classes:

(i) The class of sparse correlation matrices:

M(q,Mp,δ):={R:max⁡1≤j≤p∑k≠j∣Rjk∣q≤Mp and Rjj=1 for all j, Λmin⁡(R)≥δ}.(18)\mathcal{M}(q,M_p,\delta):=\left\{\mathbf{R}:\max_{1\le j\le p}\sum_{k\ne j}|\mathbf{R}_{jk}|^q\le M_p\ \text{and}\ R_{jj}=1\ \text{for all }j,\ \Lambda_{\min}(\mathbf{R})\ge\delta\right\}. \tag*{(18)}

(ii) The class of covariance matrices:

U(κ,q,Mp,δ):={Σ:max⁡jΣjj≤κ and D−1ΣD−1∈M(q,Mp,δ)}.(19)\mathcal{U}(\kappa,q,M_p,\delta):=\left\{\Sigma:\max_j\Sigma_{jj}\le\kappa\ \text{and}\ \mathbf{D}^{-1}\Sigma\mathbf{D}^{-1}\in\mathcal{M}(q,M_p,\delta)\right\}. \tag*{(19)}

Given an elliptical distribution assumption, the EC2 estimator, Σ^\widehat{\Sigma}, is bound by

sup⁡Σ∈U(κ,q,Mp,δmin⁡)E∥Σ^EC2−Σ∥2≤c1⋅Mp(log⁡pT)1−q2,(20)\sup_{\Sigma\in\mathcal{U}(\kappa,q,M_p,\delta_{\min})}\mathbb{E}\left\lVert\widehat{\Sigma}^{\mathrm{EC2}}-\Sigma\right\rVert_2\le c_1\cdot M_p\left(\frac{\log p}{T}\right)^{\frac{1-q}{2}}, \tag*{(20)}

where TT represents the sample size and c1c_1 is a constant. Notably, the EC2 estimator attains the minimax lower bound over the class U(κ,q,Md,δmin⁡)\mathcal{U}(\kappa,q,M_d,\delta_{\min}) in the Gaussian model scenario, rendering it asymptotically rate optimal.

We can use the resulting covariance matrix as our density matrix and perform Principal Components Analysis (PCA) to reduce the dimensions needed to describe our original quantum state [30].

Bayesian Methods

Bayesian Inference is a parameter estimation method that guarantees a positive semidefinite density matrix with unit trace and is robust against noise [31].

Bayesian inference hinges on the Bayes theorem from probability theory, which asserts that the probability of estimating a parameter θ\theta given some data, D\mathcal{D}, can be calculated by

P(θ∣D)=P(D∣θ)P(θ)P(D),(21)\mathbb{P}(\theta\mid\mathcal{D})=\frac{\mathbb{P}(\mathcal{D}\mid\theta)\mathbb{P}(\theta)}{\mathbb{P}(\mathcal{D})}, \tag*{(21)}

where P(θ∣D)\mathbb{P}(\theta\mid\mathcal{D}) is called the posterior probability, P(D∣θ)\mathbb{P}(\mathcal{D}\mid\theta) is the “likelihood of obtaining said data, D\mathcal{D}”, P(θ)\mathbb{P}(\theta) is the prior probability, and P(D)\mathbb{P}(\mathcal{D}) is the evidence [27].

Given a complete set of observables {Ej}j=1M\{E_j\}_{j=1}^{M}, where MM is restricted by M≥d2−1M\ge d^2-1. For NN identically prepared states, the probability distribution of getting a particular number of specific outcomes given a density matrix ρ\rho (i.e. the likelihood) is given by:

P(N=D∣ρ)=P(N1=n1,…,NM=nM∣ρ)=N!∏i=1Mni!∏j=1Mp(j∣ρ)nj,(22)\begin{aligned} \mathbb{P}(\mathbf{N}=D\mid\rho)=\mathbb{P}(N_1=n_1,\ldots,N_M=n_M\mid\rho) \\ &=\frac{N!}{\prod_{i=1}^{M}n_i!}\prod_{j=1}^{M}p(j\mid\rho)^{n_j}, \tag*{(22)} \end{aligned}

where D=(n1,n2,…,nM)D=(n_1,n_2,\ldots,n_M) is known as the data vector which contains the results of our measurement experiments, and NjN_j is the number of times we get outcome jj in our NN trials. From Born’s rule,

p(j∣ρ)=Tr⁡(ρEj).(23)p(j|\rho)=\operatorname{Tr}(\rho E_j). \tag*{(23)}

In order to estimate a density matrix ρ\rho, we first need to parameterise it. One way of doing this is by writing it in terms of the Cholesky decomposition in Eq. (8). Additionally, we note the form of T(t)\mathcal{T}(t)

T(t)=(t1td+1+itd+2⋯td2−1+itd20t2⋯⋮00⋱⋮⋮⋱⋱t3d−3+it3d−20⋯⋯td),\mathcal{T}(t)= \begin{pmatrix} t_1 & t_{d+1}+it_{d+2} & \cdots& t_{d^2-1}+it_{d^2}\\ 0 & t_2 & \cdots& \vdots\\ 0 & 0 & \ddots& \vdots\\ \vdots& \ddots& \ddots& t_{3d-3}+it_{3d-2}\\ 0 & \cdots& \cdots& t_d \end{pmatrix},

where dd is the dimension of the density matrix ρ\rho which we want to estimate. In order to ensure that the log-likelihood remains concave, the following restrictions are put on the parameters tit_i [28]:

  • Unit length: ∥ti∥22=1\lVert t_i\rVert_2^2=1.

  • Strict positivity: ti>0t_i>0.

Further, changing the parameter space from tit_i to θi\theta_i is desirable to enforce the above restrictions. This is achieved through the transformation

ti=cos⁡θi∏j=id2sin⁡θj.(24)t_i=\cos\theta_i\prod_{j=i}^{d^2}\sin\theta_j. \tag*{(24)}

Finally, we can estimate our parameters in Eq. (21)

π(θ∣D)∝N!∏i=1Mni!∏j=1MTr⁡[ρ(θ)Ej]njπ(θ).(25)\pi(\theta|D)\propto\frac{N!}{\prod_{i=1}^{M}n_i!}\prod_{j=1}^{M}\operatorname{Tr}[\rho(\theta)E_j]^{n_j}\pi(\theta). \tag*{(25)}

A careful choice of the prior probability, π(θ)\pi(\theta), must be made. As one may notice, Eq. (25) is still not tractable analytically, so numerical techniques like the MCMC Metropolis-Hastings algorithm [31] can be used for this purpose.

Quantum Machine Learning Methods for Quantum State Tomography

In this section, we present the theory of the various QML methods used in this study.

Variational Quantum Circuit Algorithm

Variational Quantum Circuits (VQCs) are widely used in QML tasks for solving problems via parameter updates which are consistent with minimising the loss function [26, 32]. The methodology of this algorithm can be summarised as follows:

  1. State Preparation: Encode the classical data into quantum states.

  2. Parameter Adjustment: Finding the optimal values of your parameters such that the loss function is a minimum.

  3. Measurement: Measuring on a classical computer.

As this circuit’s parameters increase, the number of qubits also increases polynomially. The circuit approaches highly entangled states with these polynomial parameters. In this circuit, half of the qubits are used to indicate the mixed target state, which is obtained by QST and denoted as a ρ^\hat{\rho}, and the other half of the qubits are used as auxiliary qubits. The output of the circuit is denoted as ∣Ψ⟩|\Psi\rangle as expressed in the Eq. (26)

∣Ψ⟩=V(θ)∣0⟩⊗2n.(26)|\Psi\rangle=V(\theta)|0\rangle^{\otimes2n}. \tag*{(26)}

Tomography of the first nn-qubits, which is used for the mixed target state, is obtained by Eq. (27)

ρ^T=Tr⁡(∣Ψ⟩⟨Ψ∣).(27)\hat{\rho}^{T}=\operatorname{Tr}(|\Psi\rangle\langle\Psi|). \tag*{(27)}

The fidelity and loss function are calculated using Eqs. (28) and (29), respectively. The purpose is to maximise the value of fidelity while minimising loss value.

A(θ)=Tr⁡(ρ^ρ^T)=⟨Ψ∣ρ^∣Ψ⟩,(28)A(\theta)=\operatorname{Tr}(\hat{\rho}\hat{\rho}^{T})=\langle\Psi|\hat{\rho}|\Psi\rangle, \tag*{(28)}
a(θ)=1−A(θ).(29)a(\theta)=1-\sqrt{A(\theta)}. \tag*{(29)}

A flowchart of the VQC for implementation in the QST algorithm has been illustrated in Figure 1. This algorithm uses a classical quantum framework, as shown in Figure 2.

Image not extracted

Figure 1. Flowchart of the Variational Quantum Circuit for QST Algorithm.

Figure 2. Variational Quantum Circuit diagram. The VQC initiates with nn input qubits that undergo a feature map before being processed through the circuit itself. After this transformation, all qubits are measured, and their outcomes are subsequently assessed by a classical model to evaluate the cost function.

In the first step of the VQC for the QST algorithm, the GenerateRandomPsi function is called to initialise the target state. This function gets the number of qubits, determined as 5 in this implementation, and returns the statevector object using the RandomStatevector library. Once the θ\theta-parameter vector is initialised, the algorithm calls the InitializeTheta function, which gets the CircDepth and NumQubits as parameters and returns θ\theta values.

In the second step of the algorithm, Construct VariationalCirc is called to generate a parameterised VQC as seen in Figure 3. To represent generic quantum states, we used rotational Rx, Ry, and CNOT gates. Each layer of this circuit with CNOT gates is determined as a depth ss, and it is set to 10 in this step.

Variational Quantum Circuit used for QST

Figure 3. Variational Quantum Circuit used for QST. The depicted quantum circuit is constructed using the Construct VariationalCirc algorithm, employing rotational Rx, Ry, and CNOT gates to represent generic quantum states. Each layer consists of CNOT gates, defining the circuit depth ss, which is set to 10 layers for this representation.

In the third step, the objective is to minimise the discrepancy between the predicted and target states, which is expressed by the loss function. Notably, this loss function is equivalent to the LSE from Sec. 3.3. To achieve this, the algorithm invokes the OptimizeThetaScp function, responsible for ascertaining the optimally parameterised vector denoted as θ\boldsymbol{\theta}.

To effectuate this optimisation objective, the algorithm employs the ComputeLossGradient function, which calculates the gradient of the loss function [32]. This gradient facilitates the iterative refinement of the parameter vector θ\boldsymbol{\theta} in accordance with Eq. (30):

δa(θ)δθi=δa(θ)δA(θ)δA(θ)δθ(30)\frac{\delta a(\boldsymbol{\theta})}{\delta\theta_i}=\frac{\delta a(\boldsymbol{\theta})}{\delta A(\boldsymbol{\theta})}\frac{\delta A(\boldsymbol{\theta})}{\delta\theta} \tag*{(30)}

In the fourth step, the fidelity between the predicted and target quantum states is calculated using the ComputeFidelity function from the StateFidelity library.

In the final step, the loss is calculated using the fidelity values obtained from the preceding step.

Quantum Principal Components Analysis Algorithm

In this section, we consider using Quantum Principal Component Analysis (qPCA) in lieu of classical PCA to achieve computational speed-ups and efficient density matrix estimation.

The qPCA algorithm proposed by Lloyd et al. [33] states that nn copies of a quantum system with a density matrix ρ\rho in a dd-dimensional Hilbert space can be employed to implement the unitary operator U=e−iρtU=e^{-\mathrm{i}\rho t}. This is based on the idea that the density matrix, ρ\rho, can function analogous to a Hamiltonian operator, generating transformations on other states. The idea that multiple copies of ρ\rho are used to construct a unitary transformation applied to ρ\rho demonstrates that a quantum state can play a dynamic role in its own analysis. Instead of the state being a passive entity upon which measurements are made, it actively transforms other states.

The following relation serves as the foundation for the mathematical framework of qPCA, where the unitary U=e−iρtU=e^{-\mathrm{i}\rho t} is applied to any density matrix σ\sigma up to the nthn^{\text{th}} order in tt

Tr⁡Pe−iSΔtρ⊗σeiSΔt=σ−iΔt[ρ,σ]+O(Δt2),(31)\operatorname{Tr}_{P}e^{-\mathrm{i}S\Delta t}\rho\otimes\sigma e^{\mathrm{i}S\Delta t}=\sigma-\mathrm{i}\Delta t[\rho,\sigma]+O(\Delta t^{2}), \tag*{(31)}

where Tr⁡P\operatorname{Tr}_{P} is the partial trace over the first variable, SS is the swap operator, and ρ\rho and σ\sigma are density matrices. Repeated application of the above equation, with nn copies of ρ\rho, facilitates the construction of e−iρnΔtσeiρnΔte^{-\mathrm{i}\rho n\Delta t}\sigma e^{\mathrm{i}\rho n\Delta t}. Simulating e−iρte^{-\mathrm{i}\rho t} to an accuracy of ε\varepsilon demands n=O(t2ε−1∥ρ−σ∥∞2)≤O(t2ε−1)n=O(t^{2}\varepsilon^{-1}\lVert\rho-\sigma\rVert_{\infty}^{2})\leq O(t^{2}\varepsilon^{-1}) steps, where t=nΔtt=n\Delta t and ∥…∥∞\lVert\ldots\rVert_{\infty} is the sup norm.

Having nn copies of ρ\rho allows applying the unitary transformation e−iρte^{-\mathrm{i}\rho t} to perform the quantum phase algorithm. This algorithm can transform any initial state ∣ψ⟩∣0⟩|\psi\rangle|0\rangle to

∑iψi∣χi⟩∣r~i⟩,(32)\sum_i\psi_i|\chi_i\rangle|\widetilde r_i\rangle, \tag*{(32)}

where ∣χi⟩|\chi_i\rangle are the eigenvectors of ρ\rho and r~i\widetilde r_i are estimates of the corresponding eigenvalues, and ψi=⟨χi∣ψ⟩\psi_i=\langle\chi_i|\psi\rangle. Given t=nΔtt=n\Delta t, consider the unitary can be expressed as:

∑n∣nΔt⟩⟨nΔt∣⊗∏j=1ne−iSjΔt.(33)\sum_n|n\Delta t\rangle\langle n\Delta t|\otimes\prod_{j=1}^{n}e^{-\mathrm{i}S_j\Delta t}. \tag*{(33)}

When Eq. (33) is applied to the state as expressed in

∣nΔt⟩∣nΔt⟩⊗σ⊗ρ⊗⋯⊗ρ,(34)|n\Delta t\rangle|n\Delta t\rangle\otimes\sigma\otimes\rho\otimes\cdots\otimes\rho, \tag*{(34)}

with σ=∣χ⟩⟨χ∣\sigma=|\chi\rangle\langle\chi| and SjS_j being the swap operator for the jthj^{\text{th}} copy of ρ\rho, taking the partial trace over the copies of ρ\rho provides the desired transformation

∣t⟩∣χ⟩↦∣t⟩e−iρt∣χ⟩.(35)|t\rangle|\chi\rangle\mapsto|t\rangle e^{-\mathrm{i}\rho t}|\chi\rangle. \tag*{(35)}

Integrating this into the quantum phase algorithm and using enhanced phase estimation techniques allows for the precise extraction of eigenvectors and eigenvalues of ρ\rho in time t=O(ε−1)t=\mathcal{O}(\varepsilon^{-1}), demanding n=O(1/ε3)n=\mathcal{O}(1/\varepsilon^3) copies of ρ\rho.

Applying the quantum phase algorithm with ρ\rho as the initial state yields

∑iri∣χi⟩⟨χi∣⊗∣r~i⟩⟨r~i∣.(36)\sum_i r_i|\chi_i\rangle\langle\chi_i|\otimes|\widetilde r_i\rangle\langle\widetilde r_i|. \tag*{(36)}

Sampling from this state unveils information about the eigenvectors and eigenvalues of ρ\rho.

qPCA proves invaluable when ρ\rho has a small rank, RR, or can be approximated with rank RR. In such scenarios; only the largest RR eigenvalues will be non-zero in the eigenvector/eigenvalue decomposition. If m×nm \times n copies of ρ\rho are used, mm copies of this decomposition are obtained. The eigenvalues can then be analysed by quantum measurements on their corresponding eigenvectors for a chosen Hermitian operator MM, provided MM is sparse or can be efficiently simulated. This allows the examination of the eigenvalues and eigenvectors of an unknown ρ\rho in time O(log⁡d)\mathcal{O}(\log d).

Several improved qPCA algorithms include those that are based on quantum singular-value thresholding, that reduce the number of samples of measurement required [34, 35], and others that reduce the number of ancillary qubits required for Quantum Phase Estimation (QPE) [36]. Additionally, the algorithm by Lloyd et al can be applied to classical data by representing the classical data as quantum states and constructing a covariance matrix, which is used as the density matrix [37]. However, classical algorithms that are analogous to the qPCA algorithm by Lloyd et al, using the ℓ2\ell^2-norm sampling assumptions, suggest that the speed-ups from qPCA on classical data arise from its dependence on strong input assumptions, meaning qPCA is less useful on classical data when compared to analogous classical algorithms [38]. Thus, the qPCA algorithm by Lloyd et al might be more applicable for quantum data.

We then test how the density matrix ρreconstructed\rho_{\mathrm{reconstructed}}, which is reconstructed from the eigenvectors resulting Algorithm 2 qPCA for QST

Input: Multiple copies of a quantum state ρ\rho, desired accuracy ϵ\epsilon.

Output: Eigenvectors and eigenvalues of ρ\rho in quantum form.

Step 1: Initialise the parameters

Step 2: Density Matrix Exponentiation

Use nn copies of ρ\rho to apply unitary transformations of the form e−iρte^{-\mathrm{i}\rho t}.

Step 3: SWAP Operations

Use repeated infinitesimal swap operations on ρ⊗σ\rho\otimes\sigma to construct the unitary operator e−iρte^{-\mathrm{i}\rho t}.

Step 4: Quantum Phase Algorithm

Use the ability to apply e−iρte^{-\mathrm{i}\rho t}.

Perform the quantum phase algorithm to take any initial state ∣ψ⟩∣0⟩|\psi\rangle|0\rangle to a superposition of the form ∑iψi∣χi⟩∣r~i⟩\sum_i \psi_i|\chi_i\rangle|\tilde{r}_i\rangle.

Step 5: Conditional Operation

Implement the conditional operation by replacing the SWAP operator with a conditional SWAP.

Step 6: Eigenvector and Eigenvalue Decomposition

Use the quantum phase algorithm with ρ\rho as the initial state to obtain the state ∑iri∣χi⟩⟨χi∣⊗∣r~i⟩⟨r~i∣\sum_i r_i|\chi_i\rangle\langle\chi_i|\otimes|\tilde{r}_i\rangle\langle\tilde{r}_i|.

Step 7: Sampling

Sample from this state to reveal features of the eigenvectors and eigenvalues of ρ\rho.

Step 8: Output

Return the eigenvectors and eigenvalues in quantum form.

from QPE, differs based on tt. To do so, we introduce a modified algorithm based on (2), wherein the density matrix of ρoriginal\rho_{\mathrm{original}} is known beforehand, and the Pauli strings of the state are provided. In this case, all that is required is the application of the unitary U=e−iρtU=e^{-\mathrm{i}\rho t} in QPE. Here, we use QPE using the Quantum Fourier Transform [39] with 2 auxiliary qubits for eigendecomposition. Then, the resulting eigenvectors are used to construct ρreconstructed\rho_{\mathrm{reconstructed}}. However, because the dimensionality of ρreconstructed\rho_{\mathrm{reconstructed}} is reduced when compared to that of ρoriginal\rho_{\mathrm{original}}, we need to embed both states into a larger Hilbert space such that they have the same dimension. This can be done by taking the tensor product of the smaller state with an identity matrix of the appropriate size according to the following algorithm:

  1. Given ρ\rho has dimension d1×d1d_1 \times d_1 and σ\sigma has dimension d2×d2d_2 \times d_2, determine the larger dimension, d=max⁡(d1,d2)d=\max(d_1,d_2).

  2. If d1<dd_1<d, then take the tensor product of ρ\rho with the identity matrix of size (d/d1)×(d/d1)(d/d_1)\times(d/d_1). This would look like ρ′=ρ⊗Id/d1\rho'=\rho\otimes I_{d/d_1}.

  3. If d2<dd_2<d, then take the tensor product of σ\sigma with the identity matrix of size (d/d2)×(d/d2)(d/d_2)\times(d/d_2). This would look like σ′=σ⊗Id/d2\sigma'=\sigma\otimes I_{d/d_2}.

Subsequently, the fidelity, FF, can be computed, where FF is calculated by the Uhlmann fidelity formula [40]

F(ρ,σ)=(Tr⁡ρσρ)2.(37)F(\rho,\sigma)=\left(\operatorname{Tr}\sqrt{\sqrt{\rho}\sigma\sqrt{\rho}}\right)^2. \tag*{(37)}

Bayesian Approach

In Bayesian Inference, probabilities are taken as a prior and post-prior distribution. The goal is to define and analyse these distributions. The unknown parameter θ\theta is a prior distribution π(θ)\pi(\theta).

We propose the method for Bayesian QST below as explained in Algorithm 3. The computational steps required are:

  • Perform measurements on an unknown state ρ\rho. This amounts to a total of nn individual outcomes.

  • Compute the LSE ρLS\rho_{LS}. If there are incomplete measurements, then ρLS\rho_{LS} lives in subspace determined by only observed directions which can be represented by ρLS=[PM(ρ)]LS\rho_{LS}=[P_M(\rho)]_{LS}.

  • Parameterise the N×NN \times N density matrix by NN and non-negative real numbers yN∈R\(−∞,0]y_N \in\mathbb{R}\backslash(-\infty,0], and a column vectors zNz_N of length NN. The density matrix for the parameter set θ=[y1,y2,…,yN;z1,z2,…,zN]\theta=[y_1,y_2,\ldots,y_N;z_1,z_2,\ldots,z_N] is written as,

ρ(θ)=∑i=1N(yN∑lyl)zNzN†∣zN∣2.(38)\rho(\theta)=\sum_{i=1}^{N}\left(\frac{y_N}{\sum_l y_l}\right)\frac{z_N z_N^\dagger}{\lvert z_N\rvert^2}. \tag*{(38)}
  • Take the prior distribution for θ\theta as

π0(θ)∝∏i=1NyNα−1e−yNe−12zN†zN,(39)\pi_0(\theta)\propto\prod_{i=1}^{N}y_N^{\alpha-1}e^{-y_N}e^{-\frac{1}{2}z_N^\dagger z_N}, \tag*{(39)}

where α\alpha is a non-negative vector with scalar coefficients. π0(θ)\pi_0(\theta) are random Gamma-distribution variables, and zN∼N(0,1)z_N\sim\mathcal{N}(0,1) are standard normal Gaussian distribution.

  • The Likelihood function is given by

P(D∣θ)∝exp⁡[−N2∥PM(ρ(θ))−ρLS∥F2],(40)\mathbb{P}(D|\theta)\propto\exp\left[-\frac{N}{2}\left\|P_M(\rho(\theta))-\rho_{LS}\right\|_F^2\right], \tag*{(40)}

where ∥PM(ρ(θ))−ρLS∥F2\left\|P_M(\rho(\theta))-\rho_{LS}\right\|_F^2 represents the Frobenius norm for the difference between the two matrices, PM(⋅)P_M(\cdot) is the projection operator, and ρLS\rho_{LS} is the density matrix for estimating the least squares.

  • Make samples of invariant distribution π(θ)∝P(D∣θ)π0(θ)\pi(\theta)\propto\mathbb{P}(D|\theta)\pi_0(\theta) as

π(θ∣D)=1A×P(D∣θ)π0(θ),(41)\pi(\theta|D)=\frac{1}{A}\times\mathbb{P}(D|\theta)\pi_0(\theta), \tag*{(41)}

Algorithm 3 Bayesian Inference for QST

Input: Number of qubits, NqubitsN_{\mathrm{qubits}}, circuit depth, DcircD_{\mathrm{circ}}, prior and posterior parameters (θ,θ′)(\boldsymbol{\theta},\boldsymbol{\theta}') of the same length and size.

Step 1: Initialise Parameters Define the density matrix of length dimensions NN for the parameterised quantum circuit as ρ(θ)\rho(\boldsymbol{\theta}).

Step 2: Prior Parameters Initialise the parameter vector θ\boldsymbol{\theta} for generating samples of the prior distribution P(D)\mathbb{P}(\mathcal{D}).

Step 3: Estimating Prior Sample Data Implement the log-likelihood Function P(D∣θ)\mathbb{P}(\mathcal{D}|\boldsymbol{\theta}) to estimate the sample data between the predicted and the target state based on given prior parameters.

Step 4: Posterior Parameters Optimise the parameter vector θ′\boldsymbol{\theta}'. Perform a set of operations on the predicted and the target state.

Step 5: Posterior Distribution Make samples of invariant distribution of vector θ′\boldsymbol{\theta}'. Compute the posterior probability distribution π0(θ′)\pi_0(\boldsymbol{\theta}').

Step 6: Sampling Perform measurement operations on the targeted state. Collect the samples of the prior and posterior probability distributions. Perform the sampling using the pre-conditioned Crank-Nicholson ρCN\rho_{\mathrm{CN}} Metropolis-Hastings procedure or the Markov Chain Monte Carlo (MCMC) method.

Step 7: Calculate Fidelity Compute the fidelity, FF, between the predicted quantum state ∣ψpredicted⟩|\psi_{\mathrm{predicted}}\rangle and the target quantum state ∣ψtarget⟩|\psi_{\mathrm{target}}\rangle. Compare the results of the prior and posterior distribution and analyse the results.

Step 8: Output Plot the graph of the loss function, fidelities of the quantum circuit as shown in Figure 4 (targeted and predicted state), and posterior probability distribution.

Figure 4. A schematic diagram of Bayesian Inference for Quantum State Tomography. qq defines the Quantum Registers for nn-qubits. While θ1...θn\theta_1...\theta_n are prior parameters. P(D∣θ)P(D|\theta) represents a Maximum Likelihood function. θ′\theta' are posterior parameters used for Posterior State probability distribution, and then after measurements, the sampling procedure is performed.

where AA is the normalisation constant and is defined as ∫dx π(θ)=1\int dx\,\pi(\theta)=1. Here P(D∣θ)\mathbb{P}(\mathcal{D}|\theta) is the post-prior (posterior) probability with distribution of θ\theta, π0\pi_0 is the prior-distribution and P(D∣θ)\mathbb{P}(\mathcal{D}|\theta) is the likelihood function.

Through these samples, we can estimate any function of ρ\rho using Eq. (41).

Quantum Variational Algorithm with Classical Statistics

In this subsection, we discuss a variational algorithm suitable for quantum state reconstruction when measurement statistics of multiple Pauli strings are available. We call this algorithm the Quantum Variational Algorithm with Classical Statistics (QVCS).

Consider a pure state ∣Φ⟩|\Phi\rangle defined in terms of a certain basis consisting of vectors ∣ni⟩|n_i\rangle. Without loss of generality, we can regard this as the computational basis, resulting in the following pure state representation

∣Φ⟩=∑i=0N−1Ci∣ni⟩,(42)|\Phi\rangle= \sum_{i=0}^{N-1} C^i |n_i\rangle, \tag*{(42)}

where NN is the dimension of the Hilbert space. The complex coefficients, Cλ,μiC^i_{\lambda,\mu}, can be written in the polar form

Cλ,μi=⟨ni∣Φ⟩=pλiexp⁡(iϕμi).(43)C^i_{\lambda,\mu} = \langle n_i|\Phi\rangle= \sqrt{p^i_\lambda}\exp\left(\mathrm{i}\phi^i_\mu\right). \tag*{(43)}

The parameters λ\lambda and μ\mu are introduced to indicate that the probability amplitudes and the phases are obtained in this algorithm through two different parameterised circuits. Following [26], in the classical approach, the parameters are obtained from the output nodes of two NNs. Figure 5 schematically shows how these parameters are obtained as the output nodes of two NNs.

Perceptron model architecture

Figure 5. Perceptron model architecture to represent pλp_\lambda, and ϕμ\phi_\mu for the pure state ∣ϕ⟩|\phi\rangle, according to the parameterisation defined in Eq. (43). Here, the two networks, one for the probabilities and one for the phases, have been represented in a single diagram.

The λ\lambda network outputs a probability vector {pλi}i=0N−1\{p^i_\lambda\}_{i=0}^{N-1}, while the μ\mu network outputs phase vector {ϕμi}i=0N−1\{\phi^i_\mu\}_{i=0}^{N-1}. Since each network uses hidden layers, we have that

Xλ/μ[l]=ReLU⁡(Wλ/μ[l]Xλ/μ[l−1]+bλ/μ[l])(44)X^{[l]}_{\lambda/\mu} = \operatorname{ReLU}\left(W^{[l]}_{\lambda/\mu}X^{[l-1]}_{\lambda/\mu}+b^{[l]}_{\lambda/\mu}\right) \tag*{(44)}

for l=[1,L−1]l=[1,L-1] and Xλ/μ[0]=1X^{[0]}_{\lambda/\mu}=1. We then find that the λ\lambda and μ\mu networks output

{pλi}=softmax⁡(Wλ[L]Xλ[L−1]+bλ[L]),{ϕμi}=πtanh⁡(Wμ[L]Xμ[L−1]+bμ[L]),(45)\begin{aligned} \{p^i_\lambda\} &= \operatorname{softmax}\left(W^{[L]}_\lambda X^{[L-1]}_\lambda+b^{[L]}_\lambda\right),\\ \{\phi^i_\mu\} &= \pi\tanh\left(W^{[L]}_\mu X^{[L-1]}_\mu+b^{[L]}_\mu\right), \tag*{(45)} \end{aligned}

respectively. Here, the softmax activation function has been used for the λ\lambda network, while the μ\mu network uses πtanh⁡\pi\tanh.

We propose an analogous algorithm where the hidden layers of the classical network are replaced by a Parameterised Quantum Circuit (PQC)/Quantum Neural Network (QNN). We use two kinds of circuit ansatz for the parameterised circuit. We call the circuits: circuit (A) and circuit (B) henceforth. Circuit (A) has been proposed in [21] and was found to be expressive enough to encode the entanglement structure of quantum states to good fidelities using reasonable circuit depths of 10−1710-17 in their algorithm. Circuit (B) was the TwoLocal form available in qiskit.circuit library.

These are shown in Figs 6 and 7 respectively.

The number of output nodes in the NN for an nn-qubit state is 2n2^n. In the quantum circuit, we require nn qubits, and the outputs are obtained by measuring the probability distribution of all the 2n2^n output strings. Since sampling from this distribution will give only real numbers in the range [0,1][0,1], to obtain the parameters ϕi\phi^i from the circuit, we multiply by 2π2\pi.

Circuit B for three qubits and depth four

Figure 7. Circuit (B) for three qubits and depth four.

Circuit A for three qubits and depth five

Figure 6. Circuit (A) for three qubits and depth five.

For the quantum circuit analogous to the NN in Figure 5, we assume that each quantum state is measured

in a set of NBN_B bases {n0[b],…,nNB−1[b]}\left\{\mathbf{n}^{[b]}_0,\ldots,\mathbf{n}^{[b]}_{N_B-1}\right\}, where b∈{0,…,NB−1}b\in\{0,\ldots,N_B-1\} is a basis index. We can then define the loss function EE used for this variational algorithm, which is of the maximum likelihood form

E=−∑bLb=−∑bfni[b]log⁡[P(ni[b])],(46)E=-\sum_b\mathcal{L}_b=-\sum_b f^{[b]}_{n_i}\log\left[\mathbb{P}\left(\mathbf{n}^{[b]}_i\right)\right], \tag*{(46)}

where fni[b]f^{[b]}_{n_i} is the frequency of outcomes labeled by index ii for measurement in basis bb, P(ni[b])\mathbb{P}\left(\mathbf{n}^{[b]}_i\right) refers to the probability of getting the output eigenvector ni[b]\mathbf{n}^{[b]}_i when measuring in the bb basis for a given state. Once the parameters of a state are given, this is simply calculated by the Born rule

P(ni[b])=∣⟨ni[b]|Φ⟩∣2.(47)\mathbb{P}\left(\mathbf{n}^{[b]}_i\right)=\left|\left\langle\mathbf{n}^{[b]}_i\middle|\Phi\right\rangle\right|^2. \tag*{(47)}

Given that

∣⟨ni[b]|Φ⟩∣2=∑j=0N−1⟨ni[b]|nj⟩Cλ,μj,(48)\left|\left\langle\mathbf{n}^{[b]}_i\middle|\Phi\right\rangle\right|^2=\sum_{j=0}^{N-1}\left\langle\mathbf{n}^{[b]}_i\middle|\mathbf{n}_j\right\rangle\mathcal{C}^{j}_{\lambda,\mu}, \tag*{(48)}

we find that the loss function becomes

E=∑b=0NB−1∑i=0N−1fi[b]log⁡(∑j=0N−1⟨ni[b]∣nj⟩pλjeiϕμj+c.c.)(49)E = \sum_{b=0}^{N_B-1} \sum_{i=0}^{N-1} f_i^{[b]} \log\left( \sum_{j=0}^{N-1} \langle n_i^{[b]} \mid n_j \rangle\sqrt{p_\lambda^j e^{i\phi_\mu^j}} + \mathrm{c.c.} \right) \tag*{(49)}

The frequencies fnif_{n_i} were generated for our models using a qiskit backend qasm_simulator, The measurement results for all Pauli strings were measured for a multi-qubit system, and the fidelities were benchmarked for random quantum states (see results section).

The algorithm as explained in 4 can be summarised with the following steps:

  • Firstly, a suitable dataset D\mathcal{D} containing the measurement statistics of several Pauli strings needs to be obtained. To analyse our algorithm, we generate D\mathcal{D} via simulation through a qiskit backend.

  • The parameters Θn\Theta_n of the chosen variational circuit are initialised to random values. The number of parameters is 2×Depth×Nqubits2 \times\mathrm{Depth} \times N_{\mathrm{qubits}}.

  • The circuits are sampled using SamplerQNN class in qiskit_machine_learning.neural_networks for getting 2Nqubits2^{N_{\mathrm{qubits}}} output frequencies which are interpreted as the parameters pλip_\lambda^i and ϕμi\phi_\mu^i respectively.

  • A classical optimiser, in this case, COBYLA, minimises the loss function defined according to Eq. (49) above.

  • To calculate the gradients with respect to these parameters, we use the “parameter shift rule”, as described in [41].

  • The final state parameters are obtained after the optimisation process such that the observed frequencies fni[b]f_{n_i}^{[b]} have a maximum likelihood.

Our analysis found that our proposed algorithm produced fidelities of roughly 70 − 80% based on the circuit depth. However, the variance in the fidelities was also significant, pointing to the fact that the circuit ansatz was not complex enough to capture the entanglement structure of the state and the resulting fidelity when the states were highly entangled. To test this hypothesis, only states where one qubit was disentangled from the rest were given to the algorithm, and the fidelities achieved, along with the variance, were improved.

Results and Discussion

In this section, we present the outcomes of our research experiments conducted using the AerSimulator [42]. We detail the results of utilising a pre-built quantum model made available through the qiskit framework

Algorithm 4 QVCS for QST

Input: Number of qubits, NqubitsN_{\mathrm{qubits}}, Circuit depth, DcircD_{\mathrm{circ}}.

Step 1: Initialise Parameters

Define the ansatz for the PQC as C\mathcal{C}

Generate a random target quantum state ∣ψtarget⟩\lvert\psi_{\mathrm{target}}\rangle.

Step 2: Data Preparation

Obtain measurement data D\mathcal{D} through a simulated measurement function, and normalise the measurement data D\mathcal{D} obtained from the target state.

Step 3: Parameter Initialisation

Initialise the parameter vector θ\boldsymbol{\theta}, i.e. (2×Nqubits×Dcirc)(2 \times N_{\mathrm{qubits}} \times D_{\mathrm{circ}}) with random values.

Step 4: Optimisation

Optimise the parameter vector θ\boldsymbol{\theta} to minimise the discrepancy between the predicted state and the target state through the log-likelihood function.

Step 5: Extract Optimised Parameters

Reshape the optimised parameter vector θ\boldsymbol{\theta} into a matrix Θ\boldsymbol{\Theta}.

Step 6: Predict Quantum State

Calculate the predicted quantum state amplitudes ∣ψpredicted⟩\lvert\psi_{\mathrm{predicted}}\rangle using C\mathcal{C} and the optimised parameters Θ\boldsymbol{\Theta}.

Step 7: Calculate Fidelity

Compute the fidelity FF between the predicted quantum state ∣ψpredicted⟩\lvert\psi_{\mathrm{predicted}}\rangle and the target quantum state ∣ψtarget⟩\lvert\psi_{\mathrm{target}}\rangle.

Step 8: Output

Present the optimised PQC, the predicted quantum state amplitudes ∣ψpredicted⟩\lvert\psi_{\mathrm{predicted}}\rangle, and the fidelity as the outcomes of the algorithm.

[43] Subsequently, we delve into the outcomes derived from applying four distinct quantum methods, as previously expounded: The VQC algorithm, the qPCA algorithm, the Bayesian approach, and the QVCS algorithm.

Starting with the pre-built model, we used the StateTomography Experiment from qiskit to construct a circuit to generate the Greenberger-Horne-Zeilinger (GHZ) state [44]. The circuit utilised Hadamard and controlled-X gates to create entanglement among the qubits, as shown in Figure 8.

Circuit diagram for GHZ state generation

Figure 8. Circuit diagram for GHZ state generation.

Subsequently, we performed a StateTomography experiment on the circuit. This experiment involved measuring the qubit’s states in various bases, which allowed us to reconstruct the density matrix of the GHZ state. The average state fidelity was 98.38%, indicating that the circuit could generate the GHZ state with high fidelity.

In addition, we performed a parallel tomography experiment on the circuit, measuring the state of the qubits in various bases, but it was only performed on individual circuit gates, which allowed us to investigate the discrete influence of each gate on the state of the qubits.

The results of the parallel tomography experiment showed that all of the gates in the circuit had a high fidelity, as shown in Figure 9, which indicates that the circuit could generate the GHZ state with high fidelity, even when individual gates were not perfect.

Figure 9. State fidelity comparison plot for the QST experiment - qiskit.

Secondly, the VQC algorithm was effectively applied to mixed states, as discussed in the previous section. Our investigation explored the intricate relationship between the loss value and two distinct parameters: Depth and iteration numbers, as shown in Figure 10 and Figure 11, respectively. The results of our study clearly show that an optimal fidelity value was obtained at a depth of 13. Additionally, a compelling trend emerges: Increasing the depth number correlates with a decrease in the loss value, effectively approximating the ground states of a local spin Hamiltonian.

Correlation between loss Values and circuit depth in the VQC algorithm

Figure 10. Correlation between loss Values and circuit depth in the VQC algorithm.

Correlation between loss values and iterations in the VQC algorithm

Figure 11. Correlation between loss values and iterations in the VQC algorithm.

Notably, the results highlight the algorithm’s ability to achieve high fidelity using a relatively small number of variational parameters and iterations.

This observation attests to the efficacy of the VQC approach in realising accurate quantum computations. Furthermore, the proposed algorithm’s potential extends to near-term quantum devices, presenting a promising avenue for future research exploration.

In our next method, the qPCA, we investigated how the accuracy of the reconstructed state ρreconstructed\rho_{\mathrm{reconstructed}} varies with tt by initialising the algorithm with a known random state ρoriginal\rho_{\mathrm{original}} and calculating the average fidelity from Niter=50N_{\mathrm{iter}} = 50 iterations of QPE for a tt value of t=tit = t_i.

Each iteration yields a different fidelity FjF_j for the jjth reconstructed state. The resulting values are then used to compute the average fidelity F‾i\overline{F}_i and the standard deviation σi\sigma_i for each tit_i. These average fidelity and standard deviation values are depicted in Figure 12 and Figure 13, respectively. Here, QPE was performed using 2 auxiliary qubits and calculated for ti∈T={t∣t=2+0.1n, 0≤n<280}t_i \in T = \{t \mid t = 2 + 0.1n,\ 0 \le n < 280\}.

Average fidelity plot

Figure 12. Average fidelity Fˉi\bar{F}_i between States ρoriginal\rho_{\mathrm{original}} and ρreconstructed\rho_{\mathrm{reconstructed}} for a randomised ρoriginal\rho_{\mathrm{original}} and a 2-qubit phase estimation with t=tit=t_i.

Standard deviation plot

Figure 13. Standard Deviation σi\sigma_i of the calculated Fidelities over 50 iterations between States ρoriginal\rho_{\mathrm{original}} and ρreconstructed\rho_{\mathrm{reconstructed}} for a randomised ρoriginal\rho_{\mathrm{original}} and a 2-qubit phase estimation with t=tit=t_i.

Using the results of our calculations, we can then find the optimal tt value and its associated average fidelity, which is the maximum average fidelity. Let F(t)\mathcal{F}(t) be a function that maps a value tt to its corresponding average fidelity. Then, the maximum average fidelity Fˉmax⁡≈argmax⁡t∈TF(t)\bar{F}_{\max} \approx\operatorname{argmax}_{t\in T} \mathcal{F}(t). We find that the optimal tt value is toptimal≈20.3t_{\mathrm{optimal}} \approx20.3 with an associated average fidelity of Fˉmax⁡≈89.58%\bar{F}_{\max} \approx89.58\% and a standard deviation of σtoptimal=0.10\sigma_{t_{\mathrm{optimal}}}=0.10. Using these statistics, we can establish a 95% confidence interval of (86.76%, 92.40%) for the optimal tt value’s fidelity.

Our next method used was the Bayesian approach which employs the Bernoulli model, where the posterior distribution is based on the evidence and likelihood functions. The likelihood P(D∣θ)\mathbb{P}(D|\theta) and evidence P(D)\mathbb{P}(D) are taken as input, and the behaviour of prior distribution P(D)\mathbb{P}(D) and the evidence DD will affect the posterior probability. To optimise the set of parameters θ′\theta^{\prime}, we used the COBYLA optimiser with a maxiter=150. Figure 14 shows the results of our

Posterior probability distribution histogram

Figure 14. Posterior probability distribution for Bernoulli model in Bayesian approach for N=4N=4 qubits with a circuit depth = 8. Here θ=θ′\theta=\theta^{\prime} for posterior parameters on x-axis and ρ(θ)\rho(\theta) on y-axis.

analysis for a system of N=4N=4 qubits and circuit depth range from 3 to 8. We ran 10 000 samples with 200 learning steps for each sample. The parameters are updated after the sampling process is complete, and these posterior evidence then act as prior parameters to create a new set of posterior beliefs. The Markov Chain Monte Carlo (MCMC) Metropolis-Hastings procedure was used to sample on 4-qubit data. We computed the fidelities after the sampling process.

Figure 15 shows the algorithm’s performance on a simulator with 8GB of RAM. The average fidelity was calculated for n=25n=25 randomly generated states, with circuit depth range 3−83-8 and, and n=4n=4 qubits. For each circuit depth, a total of 25 states are randomly generated, each state has its own fidelity. The plot shows the fidelity with blue dots versus the index of randomly chosen states for a circuit of depth = 8. The performance can be improved by optimising the parameters and decreasing the circuit depth.

Image not extracted

Figure 15. Fidelity Graph for Bayesian Inference with Nqubits=4N_{\mathrm{qubits}} = 4. Each Fidelity is represented by blue dots, and benchmark fidelity with the green line. The results are plotted for circuit depth of range 8 for n=25n = 25 randomly generated states.

For our last algorithm, the Quantum Variational Algorithm with Classical Statistics, we evaluated its accuracy by calculating the average fidelity across 25 randomly generated quantum states for a fixed circuit depth DcircD_{\mathrm{circ}} and the number of qubits NqubitsN_{\mathrm{qubits}}. We used the classical optimiser COBYLA with the parameters rhobeg=30 and maxiter=150. The plot shown in Figure 16 demonstrates, for instance, how the algorithm performs on average for Dcirc=3D_{\mathrm{circ}}=3 and Nqubits=3N_{\mathrm{qubits}}=3. The circuit used here is circuit (A).

Image not extracted

Figure 16. Fidelities for circuit (A) with Nqubits=3N_{\mathrm{qubits}} = 3 and Dcirc=3D_{\mathrm{circ}} = 3.

Figure 17 shows the variation of average fidelity with circuit depth for both circuit ansatz (A) and (B). The bar plot of the standard deviation of the fidelities is shown in Fig. ?? the standard deviation of the fidelities as a function of DcircuitD_{\mathrm{circuit}} for Nqubits=3N_{\mathrm{qubits}}=3.

Image not extracted

Figure 17. Comparison of average fidelities for Nqubits=3N_{\mathrm{qubits}} = 3 in circuits (A) and (B).

These results show that the average fidelity increases with circuit depth for both circuit ansatzes. However, the standard deviation of the fidelities exhibits a non-monotonic behaviour for circuit ansatz (B), decreasing at intermediate circuit depths before increasing again at deeper circuit depths. This behaviour is likely due to the interplay between the noise affecting the circuit ansatz (B) and the ability of the circuit ansatz to compensate for the noise.

The results for the circuit (A) are shown in Figure 19. The average fidelity for the circuit (A) with 4 qubits is lower than that for 3 qubits. This is likely because the noise affecting the circuit ansatz (A) is more pronounced with 4 qubits.

Image not extracted

Figure 19. Fidelities and error bars for Nqubits=4N_{\mathrm{qubits}} = 4 with parameterised circuit (A).

Conclusion

In conclusion, this paper has presented a review of the various classical and quantum approaches to QST. The goal was to explore the various QML techniques and to show the computational advantage and improved

Image not extracted

Figure 18. Comparison of standard deviations in fidelities achieved by QVCS algorithm for Nqubits=3N_{\mathrm{qubits}} = 3 in circuits (A) and (B).

accuracy over the classical approaches.

Our approach builds on the existing literature on QST and QML. In addition, we have provided a comprehensive literature review of some important works in the field. In sufficient detail, we have described the methods employed and the novelties of each approach. The classical approaches to state tomography were discussed by giving detailed descriptions of the theoretical underpinnings of each method. Furthermore, we have discussed the mathematical, computational, and circuit architecture of the QST methods and presented the results of each method employed by implementing them and verifying their fidelity.

It is important to point out the disadvantages of the QML approach. Firstly, in this current NISQ era of quantum devices, we only have access to a few qubits and thus are restricted to reconstructions of small systems. Secondly, the methods are not infallible against noise and decoherence, and thus, can lead to many potentially erroneous results. Thirdly, implementing these methods requires domain expertise which might be challenging for the non-expert to implement and validate their findings.

For our future work, we plan to explore several avenues of research to improve our QML-based QST methods further. Firstly, we will investigate the use of more advanced QML algorithms and consider quantum versions of the state-of-the-art classical methods to improve the accuracy and efficiency of our approach. Secondly, we will explore the use of hybrid classical-quantum algorithms to reduce further the number of measurements required for QST.

Overall, QML-based QST methods represent a significant advancement in the field of QIP and can potentially revolutionise how we perform state tomography. By leveraging the power of QML algorithms, we have shown that it is possible to achieve high accuracy with significantly fewer measurements than conventional methods. This has important implications for practical QIP applications, where the ability to perform QST efficiently and accurately is critical.

References

  1. [1]Wootters, W. K., & Zurek, W. H. (1982). A Single Quantum Cannot be Cloned. Nature, 299 (5886), pp. 802-803.DOI
  2. [2]Hai, V. T. & Ho, L. B. (2023). Universal Compilation for Quantum State Tomography. 13 (3750), pp. 1-12.DOI
  3. [3]Ramadhani, S., Rehman, J. U., & Shin, H. (2021). Quantum Error Mitigation for Quantum State Tomography. IEEE Access, 9, pp. 107955-10796.DOI
  4. [4]Farrand, M., & Řeháček, J. (2004). Quantum State Estimation. Springer, New York, USA.
  5. [5]Teo, Y. S. (2015). Introduction to Quantum-state Estimation. World Scientific, Singapore.DOI
  6. [6]James, D. F. V., Kwiat, P. G., Munro, W. J., & White, A. G. (2001). Measurement of Qubits. Physical Review A, 64 (5), pp. 052312 1-15.DOI
  7. [7]Rippe, L., Julsgaard, B., Walther, A., Ying, Y., & Kröll, S. (2008). Experimental Quantum-state Tomography of a Solid-state Qubit. Physical Review A, 77 (2), pp. 022307 1-5.
  8. [8]Thew, R. T. Nemoto, K., White, A. G., & Munro, W. J. (2002). Qudit Quantum-state Tomography. Physical Review A, 66 (1), pp. 012303 1-6.DOI
  9. [9]Altepeter, J. B., Jeffrey, E. R., & Kwiat, P. G. (2004). Quantum Photonic State Tomography. Advances in Atomic, Molecular, and Optical Physics, 52, pp. 105-159.
  10. [10]Preskill, J. (2018). Quantum Computing in the NISQ Era and Beyond. Quantum, 2, pp. 79-99.DOI
  11. [11]Torlai, G., Mazzola, G., Carrasquilla, J., Troyer, M., & Melko, R. G. (2018). Many-body Quantum State Tomography with Neural Networks. arXiv: https://arxiv.org/abs/1703.05334.
  12. [12]Schmale, T., Reh, M., & Gärttner, M. (2022). Efficient Quantum State Tomography with Convolutional Neural Networks. Nature: npj Quantum Information, 8 (115), pp. 1-8.DOI
  13. [13]Quek, Y., Fort, S., & Ng, H. K. (2021). Adaptive Quantum State Tomography with Neural Networks. Nature: npj Quantum Information, 7 (105), pp. 1-6.arxiv.org/abs/1812.06693
  14. [14]Koutný, D., Motka, L., Hradil, Z., Řeháček, J., & Sánchez-Soto, L. L. (2022). Neural-network Quantum State Tomography. Physical Review A, 106 (1), pp. 012409 1-8.DOI
  15. [15]Nielsen, M. A., & Chuang, I. L. (2011). Quantum Computation and Quantum Information. Cambridge University Press, Cambridge, UK.
  16. [16]Acharya, A., Kypraios, T., & Guţă, M. (2019). A Comparative Study of Estimation Methods in Quantum Tomography. Journal of Physics A: Mathematical and Theoretical, 52 (23), pp. 234001 1-36.DOI
  17. [17]Opatrny, T., Welsch, D. G., & Vogel, W. (1997). Least-Square Inversion from Density-Matrix Reconstruction. Physical Review A, 56 (3), pp. 1788—1799.
  18. [18]Surawy-Stepney, T., Kahn, J., Kueng, R., & Guţă, M. (2022). Projected Least-Squares Quantum Process Tomography. Quantum, 6, pp. 844.
  19. [19]Choi, M-D. (1975). Completely Positive Linear Maps on Complex Matrices. Linear Algebra and its Applications, 10 (3), pp. 285-290.DOI
  20. [20]Wiebe, N., Braun, D., & Lloyd, S. (2012). Quantum Algorithm for Data Fitting. Physical Review Letters, 109 (5), pp. 050505 1-5.pubmed.ncbi.nlm.nih.gov/23006156
  21. [21]Liu, Y., Wang, D., Xue, S., Huang, A., Fu, X., Qiang, X., Xu, P., Huang, H., Deng, M., Guo, C., Yang, X., & Wu, J. (2020). Variational Quantum Circuits for Quantum State Tomography. Physical Review A, 101 (5), pp. 052316 1-6.DOI
  22. [22]Ahmed, S., Muñoz, C. S., Nori, F., & Kockum, A. F. (2021). Quantum State Tomography with Conditional Generative Adversarial Networks. Physical Review Letters, 127 (14), pp. 140502 1-8.DOI
  23. [23]Cha, P., Ginsparg, P., Wu, F., Carrasquilla, J., McMahon, P. L., & Kim, E-A. (2021). Attention-based Quantum Tomography. Machine Learning: Science and Technology, 3 (1), pp. 01LT01 1-7.DOI
  24. [24]Vogel, K., & Risken, H. (1989). Determination of Quasiprobability Distributions in Terms of Probability Distributions for the Rotated Quadrature Phase. Physical Review A, 40 (5), pp. 2847-2849.pubmed.ncbi.nlm.nih.gov/9902492
  25. [25]Řeháček, J., Hradil, Z., & Ježek, M. (2001). Iterative Algorithm for Reconstruction of Quantum States. Physical Review A, 63 (4), pp. 040303 1-4.DOI
  26. [26]Kurmapu., M. K. (2020). Machine Learning Assisted Quantum State Tomography. M.S. thesis, University of Calgary, Alberta, Canada. https://prism.ucalgary.ca/items/ce19b9c3-34cb-4166-b5ba-f4582502eb38.DOI
  27. [27]Lum, D.J., and Weinstein, Y.S. (2022). Bayesian Quantum State Tomography with Python’s PyMC. arXiv: https://arxiv.org/abs/2212.10655.
  28. [28]Gonçalves, D. S., Gomes-Ruggiero, M. A., Lavor, C., Farías, O. J., and Ribeiro, P. H. (2011). Local Solutions of Maximum Likelihood Estimation in Quantum State Tomography. arXiv: https://arxiv.org/abs/1103.3682.
  29. [29]Jiang, F., Liao, Y., Liu, H. (2015). An Overview on the Estimation of Large Covariance and Precision Matrices, arXiv:https://arxiv.org/abs/1504.02995.
  30. [30]O’Donnell, R., & Wright, J. (2016). Efficient Quantum Tomography. Proceedings of the Forty-Eighth Annual ACM Symposium on Theory of Computing, PP. 899–912. https://doi.org/10.1145/2897518.2897544
  31. [31]Gonçalves, D. S., Azevedo, C. L. N., Lavor, C., and Gomes-Ruggiero, M. A. (2018). Bayesian Inference for Quantum State Tomography. Journal of Applied Statistics, 45 (10), pp. 1846–1871.
  32. [32]Sen, P., Bhatia, A. S., Bhangu, K. S., & Elbeltagi, A. (2022). Variational Quantum Classifiers through the Lens of the Hessian. Plos One, 17 (1), pp. e0262346.DOI
  33. [33]Lloyd, S., Mohseni, M., & Rebenrost, P. (2014). Quantum Principal Component Analysis. Nature Physics, 10 (9), pp. 631–633. https://doi.org/10.1038/nphys3029.
  34. [34]He, C., Li, J., Liu, W., Peng, J., & Wang, Z. J. (2022). A Low-Complexity Quantum Principal Component Analysis Algorithm. IEEE Transactions on Quantum Engineering, 3, pp. 1–13. https://doi.org/10.1109/TQE.2021.3140152.
  35. [35]Lin, J., Bao, W.-S., Zhang, S., Li, T., & Wang, X. (2019). An Improved Quantum Principal Component Analysis Algorithm Based on the Quantum Singular Threshold Method, Physical Letters A, 383 (24), pp. 2862–2868. https://doi.org/10.1016/j.physleta.2019.06.026.
  36. [36]Li, Z., Chai, Z., Guo, Y., Ji, W., Wang, M., Shi, F., Wang, Y., Lloyd, S., & Du, J. (2021). Resonant Quantum Principal Component Analysis. Science Advances, 7 (34), https://doi.org/10.1126/sciadv.abg2589.
  37. [37]Kopczyk, D. (2018). Quantum Machine Learning for Data Scientists. arXiv: https://arxiv.org/abs/1804.10068.
  38. [38]Tang, E. (2021). Quantum Principal Component Analysis Only Achieves an Exponential Speedup Because of its State Preparation Assumptions. Physical Review Letters, 127 (6), pp. 060503 1-6. https://doi.org/10.1103/PhysRevLett.127.060503
  39. [39]Lin, L. (2022). Lecture Notes on Quantum Algorithms for Scientific Computation. arXiv: https://arxiv.org/abs/2201.08309
  40. [40]Uhlmann, A. (1976). The “Transition Probability” in the State Space of a *-Algebra. Reports on Mathematical Physics, 9 (2), pp. 273-279.DOI
  41. [41]Schuld, M., Bergholm, V., Gogolin, C., Izaac,J., & Killoran, N. (2019). Evaluating Analytic Gradients on Quantum Hardware. arXiv: https://arxiv.org/abs/1811.11184.
  42. [42]IBM Quantum. https://quantum-computing.ibm.com/.
  43. [43]Qiskit Contributors. (2023). Qiskit: An Open-source Framework for Quantum Computing. doi:10.5281/zenodo.2573505.
  44. [44]Greenberger, D. M., Horne, M. A., & Zeilinger, A. (2007). Going Beyond Bell’s Theorem. arXiv: https://arxiv.org/abs/0712.0921.

Paper details

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

views · 8

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents