Quantum computers are gaining importance in various applications like quantum machine learning and quantum signal processing. These applications face significant challenges in loading classical datasets into quantum memory. With numerous algorithms available and multiple quality attributes to consider, comparing data loading methods is complex. Our objective is to compare (in a structured manner) various algorithms for loading classical datasets into quantum memory (by converting statevectors to circuits). We evaluate state preparation algorithms based on five key attributes: circuit depth, qubit count, classical runtime, statevector representation (dense or sparse), and circuit alterability. We use the Pareto set as a multi-objective optimization tool to identify algorithms with the best combination of properties. To improve comprehension and speed up comparisons, we also visually compare three metrics (namely, circuit depth, qubit count, and classical runtime). We compare seven algorithms for dense statevector conversion and six for sparse statevector conversion. Our analysis reduces the initial set of algorithms to two dense and two sparse groups, highlighting inherent trade-offs. This comparison methodology offers a structured approach for selecting algorithms based on specific needs. Researchers and practitioners can use it to help select data-loading algorithms for various quantum computing tasks.
Quantum computers are gaining importance in various applications like quantum machine learning and quantum signal processing. These applications face significant challenges in loading classical datasets into quantum memory. With numerous algorithms available and multiple quality attributes to consider, comparing data loading methods is complex. Our objective is to compare (in a structured manner) various algorithms for loading classical datasets into quantum memory (by converting statevectors to circuits). We evaluate state preparation algorithms based on five key attributes: circuit depth, qubit count, classical runtime, statevector representation (dense or sparse), and circuit alterability. We use the Pareto set as a multi-objective optimization tool to identify algorithms with the best combination of properties. To improve comprehension and speed up comparisons, we also visually compare three metrics (namely, circuit depth, qubit count, and classical runtime). We compare seven algorithms for dense statevector conversion and six for sparse statevector conversion. Our analysis reduces the initial set of algorithms to two dense and two sparse groups, highlighting inherent trade-offs. This comparison methodology offers a structured approach for selecting algorithms based on specific needs. Researchers and practitioners can use it to help select data-loading algorithms for various quantum computing tasks.
In order to use a quantum device to assess a classical dataset D, we need to representthe set D in a quantum state. Applying a quantum algorithm, that is a quantum state preparation algorithm, to convert classical data into quantum data would be the common method. Loading classical data into a quantum device is required in many current applications. Efficiently preparing a quantum state to be used as the initial state of a quantum algorithm is an essential step in developing efficient quantum algorithms, since many algorithms need to reload the initial state several times during their execution. The cost to initialize a quantum state can compromise the algorithm efficiency if the process of quantum states preparation is not efficient. The topic of quantum states preparation in quantum computing has been the focus of much attention. In this scope, preparing sparse quantum states is a more specific problem that remains open since many quantum algorithms also require sparse initialization. This dissertation presents the results of an investigation on sparse quantum states preparation with the development of three algorithms, with highlight to the preparation of sparse quantum states, the main contributionof this dissertation. From a classical input dataset with M patterns formed by pairs composed of a complex number and a binary pattern with n bits, this algorithm can prepare a quantum state with n qubits and continuous amplitudes. The cost of its steps is O(nM), classical cost of o(MlogM+nM)and requires a lower CNOT number than the main quantum state preparation algorithms currently known. The preparation of a quantumstate with 2 non-zero amplitudes reveals the need of fewer CNOT gates in n>>1 relation to the main known state preparation algorithms, with even more favorable results with s higher and less 1S in the binary string.
Abstract Quantum machine learning (QML) is an emerging field that investigates the capabilities of quantum computers for learning tasks. While QML models can theoretically offer advantages such as exponential speed-ups, challenges in data loading and the ability to scale to relevant problem sizes have prevented demonstrations of such advantages on practical problems. In particular, the encoding of arbitrary classical data into quantum states usually comes at a high computational cost, either in terms of qubits or gate count. However, real-world data typically exhibits some inherent structure (such as image data) which can be leveraged to load them with a much smaller cost on a quantum computer. This work further develops an efficient algorithm for finding low-depth quantum circuits to load classical image data as quantum states. To evaluate its effectiveness, we conduct systematic studies on the MNIST, Fashion-MNIST, CIFAR-10, and Imagenette datasets. The corresponding circuits for loading the full large-scale datasets are available publicly as PennyLane datasets and can be used by the community for their own benchmarks. We further analyze the performance of various quantum classifiers, such as quantum kernel methods, parameterized quantum circuits, and tensor-network classifiers, and we compare them to convolutional neural networks. In particular, we focus on the performance of the quantum classifiers as we introduce nonlinear functions of the input state, e.g. by letting the circuit parameters depend on the input state.
Initializing an isometry in a quantum circuit is a fundamental yet challenging task, espe- cially when aiming for efficient state preparation and resource optimization. In this work, we present a comparative analysis of the Householder Decomposition for isometry implementation, leveraging three distinct state preparation methods starting from the original Pivot method and extending the study to two additional strategies: Merge and Low Rank. The study investigates how each method impacts the performance of the isometry decomposition, focusing on key metrics such as CNOT gate count and circuit depth. To provide a comprehensive evaluation, we examine variations in matrix size and sparsity levels, capturing the effects of structural complexity on resource requirements. Our results reveal that the Merge state preparation method generally outperforms the other two approaches, particularly in terms of scalability and gate efficiency. Building upon these findings, we further compare the best-performing method, Merge, with the state-of- the-art isometry decomposition implementation available in Qiskit, a widely used quantum computing framework. The analysis demonstrates that, for isometries involving up to 6 qubits, Qiskit’s implementation exhibits superior performance. However, beyond this threshold, our proposed decomposition method proves more effective, especially for highly sparse isometries or those characterized by a smaller number of columns. This work highlights the potential for optimizing isometry decompositions in scenarios where sparsity and structural constraints are critical factors. These findings contribute to advancing state preparation techniques and offer insights into improving the efficiency of quantum circuits for applications in quantum information processing.
Quantum state preparation is the process of producing a target quantum state that will be used as the input to a quantum circuit. Many quantum algorithms require an input state where amplitudes, phases, or basis probabilities encode problem data, and the cost of preparing this state can be significant in gate count and circuit depth. This review summarizes common goals, assumptions, and methods for quantum state preparation, with emphasis on preparing states from classical vectors, probability distributions, and feature data used in quantum machine learning. We organize approaches by the information they load and by the resources they require, including gate count, circuit depth, qubit overhead, and classical preprocessing. It also compares exact and approximate preparation procedures, and discusses how precision targets affect cost. The review highlights links between families of methods, typical sources of resource estimates, and criteria that help match a preparation method to a task and hardware constraints.
The increasing complexity of quantum algorithms and the limitations of current Noisy Intermediate-Scale Quantum (NISQ) hardware underscore the importance of efficient classical simulators. To support informed decision-making by users of quantum circuit simulators, we benchmark seven statevector-based quantum circuit simulators (Qiskit, Qulacs, Qibo, Qsimov, Cirq, Pennylane and the Intel Quantum Simulator (IQS)) on a multicore node of the Lusitania high-performance computing (HPC) system. We evaluate their performance in terms of execution time, memory usage and core scalability using Grover’s algorithm, the quantum Fourier transform (QFT), and quantum volume (QV) circuits, across qubit counts ranging from 3 to 30. Our results reveal that Qulacs offers the best performance for circuits below 22 qubits, while Qiskit becomes the fastest for larger and more complex circuits. Qiskit and Qulacs achieve the most efficient parallel performance across multiple cores, while others display limited scaling benefits. IQS shows the lowest memory consumption in QFT and QV benchmarks for systems under 24 qubits; however, it suffers from higher execution times, particularly for Grover’s algorithm. The experimental implementation of Qsimov consistently underperforms in both runtime and scalability; for this reason, it is employed as a baseline in our measurements, serving to highlight the importance of performance optimizations in statevector-based quantum circuit simulators. Previous findings provide a comprehensive performance landscape to guide researchers in selecting appropriate simulators for both standard and large-scale quantum workloads on HPC infrastructures.
Quantum random access memory (QRAM) is an essential ingredient for many quantum algorithms, such as quantum machine learning models. The quantum resources (qubit counts and circuit depths) required to load classical data into a circuit severely prohibit many reasonably sophisticated demonstrations in the near term. To this end, we propose a method to construct a compact QRAM circuit, leveraging classical preprocessing to design as shallow a circuit with as few auxiliary qubits as possible. The underlying idea is to map data-encoded quantum states into an orthogonal set of quantum states (characterized by a family of parametrized circuits); then one can straightforwardly entangle the data qubits with the address qubits using cnot gates to realize a data-loading QRAM circuit. The nonunitary transformation to convert the orthogonalized quantum states back to the original data-encoded quantum states can be easily achieved with the help of a small set of auxiliary qubits and postselections. Numerical simulations on handwritten digits and Iris data sets are performed to assess the effectiveness of the proposed method, and we also highlight how noise influences quantum circuit performance across diverse configurations using random data sets. The proposed method for constructing a QRAM circuit can potentially facilitate the development and testing of many quantum machine learning methods in the near term.