Fact-checked by Grok 2 weeks ago

Quantum machine learning

Quantum machine learning (QML) is an interdisciplinary field that integrates quantum computing with machine learning techniques to develop algorithms capable of exploiting quantum phenomena such as superposition, entanglement, and interference for enhanced data processing and pattern recognition.^[1] This approach aims to achieve computational speedups over classical machine learning for specific tasks, particularly those involving high-dimensional or exponentially complex data structures, by running machine learning models on quantum hardware or hybrid quantum-classical systems.^[1] Emerging in the intersection of quantum information science and artificial intelligence, QML has garnered significant interest for its potential to revolutionize fields requiring intensive computation.^[2] The conceptual foundations of QML trace back to early quantum computing ideas, with the modern field originating in the late 1990s through pioneering works such as quantum associative memories by Ventura and Martinez.^[3] It further developed in the 2000s with seminal algorithms that demonstrated quantum advantages in linear algebra problems central to machine learning.^[4] A pivotal milestone was the 2008 Harrow-Hassidim-Lloyd (HHL) algorithm, which solves sparse linear systems of equations in exponential speedup relative to classical methods, laying groundwork for quantum-enhanced regression and recommendation systems.^[5] The field evolved through three conceptual waves: the first focused on quantum linear algebra speedups (circa 2008), the second on hybrid variational algorithms like the Quantum Approximate Optimization Algorithm (QAOA) introduced by Farhi et al. in 2014 and the Variational Quantum Eigensolver (VQE) by Peruzzo et al. in 2013, which enabled practical implementations on noisy quantum devices; and the third on brain-inspired models such as quantum recurrent neural networks.^[4]^[6]^[7] A landmark 2017 review by Biamonte et al. in Nature synthesized these developments, highlighting QML's promise for faster-than-classical learning while noting hardware limitations.^[1] Central to QML are methods like quantum neural networks, quantum kernel estimation, and variational quantum circuits, which differ from classical machine learning by natively handling quantum data and leveraging quantum parallelism to explore vast solution spaces.^[2] Notable applications include simulating quantum materials for new energy technologies, accelerating drug discovery through molecular modeling, optimizing logistics in supply chains, and analyzing particle collision data in high-energy physics.^[2] Despite these prospects, QML faces substantial challenges, including sensitivity to noise in current Noisy Intermediate-Scale Quantum (NISQ) hardware, the "barren plateau" problem where optimization landscapes become intractable, and the requirement for fault-tolerant quantum computers to realize full potential.^[2] Ongoing research emphasizes hybrid models and error mitigation to bridge these gaps, positioning QML as a frontier for achieving quantum advantage in data-driven sciences.^[2]

Fundamentals

Definition and Core Concepts

Quantum machine learning (QML) refers to the development and implementation of algorithms that leverage quantum computers to perform machine learning tasks more efficiently than classical methods, particularly for problems involving high-dimensional data or complex pattern recognition that are intractable on classical hardware.^[1] By exploiting quantum mechanical principles such as superposition, entanglement, and interference, QML enables the processing of information in ways that can potentially offer exponential speedups over classical machine learning approaches.^[1] At its core, QML builds on foundational quantum computing elements adapted to machine learning paradigms. A qubit, the basic unit of quantum information, differs from a classical bit by existing in a superposition of states, allowing it to represent multiple values simultaneously.^[1] Quantum gates manipulate these qubits through unitary operations, while measurement collapses the quantum state into classical outcomes, providing the interface for learning tasks.^[1] In the QML context, classical concepts like supervised and unsupervised learning are reimagined; for instance, quantum states can encode feature representations of data, enabling quantum versions of clustering or classification where datasets are embedded into Hilbert spaces.^[1] One key potential advantage of QML lies in its ability to accelerate linear algebra operations central to many machine learning algorithms, such as solving systems of linear equations. The Harrow-Hassidim-Lloyd (HHL) algorithm, for example, solves Ax = b for a sparse, well-conditioned matrix A of size N \times N in time polynomial in \log N and the condition number \kappa, compared to classical methods requiring O(N \sqrt{\kappa}) time, thus providing an exponential speedup in certain cases.^[5] This capability is exemplified by quantum parallelism in transformations like the quantum Fourier transform (QFT), which on n qubits acts as

|QFT_n\rangle = \frac{1}{\sqrt{2^n}} \sum_{j,k=0}^{2^n-1} e^{2\pi i j k / 2^n} |j\rangle \langle k|,

allowing efficient frequency analysis of quantum states that underpins speedups in tasks like pattern detection.^[1] The term "quantum machine learning" was first introduced in the late 1990s, with seminal works such as Ventura and Martinez (1998) on quantum computational learning algorithms.^[3] It gained further prominence in 2016–2017, with early comprehensive reviews synthesizing prior quantum algorithms into a unified framework for ML applications.^[1] These works highlighted how variational quantum algorithms serve as a primary paradigm for implementing QML on near-term quantum devices.^[1]

Prerequisites in Quantum Computing and Machine Learning

Quantum computing relies on fundamental principles that differ markedly from classical computation. At its core is the qubit, the basic unit of quantum information, which can exist in a superposition of states described by the equation |\psi\rangle = \alpha |0\rangle + \beta |1\rangle, where \alpha and \beta are complex amplitudes satisfying |\alpha|^2 + |\beta|^2 = 1.^[8] This superposition allows a qubit to represent multiple states simultaneously, enabling quantum systems to process information in parallel unlike classical bits that are strictly 0 or 1.^[8] Another key feature is entanglement, where qubits become correlated such that the state of one cannot be described independently of the others; a canonical example is the Bell state \frac{1}{\sqrt{2}} (|00\rangle + |11\rangle), which exhibits perfect correlation upon measurement.^[8] Quantum operations are governed by unitary evolution, where the state of a quantum system transforms under a unitary operator U as U |\psi\rangle, preserving the normalization of the state vector.^[8] A significant limitation is the no-cloning theorem, which states that an arbitrary unknown quantum state cannot be perfectly copied, preventing unrestricted duplication of quantum information.^[8] These principles underpin quantum algorithms that offer speedups over classical counterparts. Grover's algorithm, for instance, provides a quadratic speedup for unstructured search problems, requiring O(\sqrt{N}) queries to find an item in a database of size N, compared to O(N) classically; its success probability after k iterations is given by \sin^2((2k+1)\theta), where \theta = \arcsin(\sqrt{M/N}) and M is the number of solutions.^[9] Shor's algorithm achieves an exponential speedup for integer factorization, solving the problem in polynomial time on a quantum computer, which has profound implications for cryptography.^[10] Quantum algorithms also demonstrate advantages in linear algebra, such as the Harrow-Hassidim-Lloyd (HHL) method for solving linear systems Ax = b in O(\log N) time under certain conditions.^[5] Classical machine learning provides the foundational framework for pattern recognition and prediction from data. Supervised learning involves training models on labeled data to perform tasks like regression, which predicts continuous outputs, or classification, which assigns discrete categories.^[11] Unsupervised learning, in contrast, operates on unlabeled data for tasks such as clustering, which groups similar instances, or dimensionality reduction, which simplifies data while preserving structure.^[11] Optimization in these models often uses gradient descent, an iterative method that minimizes a loss function, such as cross-entropy for classification, by adjusting parameters in the direction of the negative gradient.^[11] The intersection of quantum computing and machine learning requires techniques to encode classical data into quantum states. Common methods include amplitude encoding, which maps data vectors to the amplitudes of a quantum state; density matrix encoding, which represents mixed states for noisy or probabilistic data; and feature map encoding, which embeds data via quantum circuits to exploit kernel-like properties.^[12] A practical example is angle encoding, where components of a classical vector are mapped to rotation angles of qubits, such as applying R_y(\theta_i) gates with \theta_i = 2 \arcsin(x_i / \|x\|) for a normalized vector x, allowing efficient loading of data into a quantum circuit with one qubit per feature.^[12]

Quantum-Enhanced Machine Learning Algorithms

Variational Quantum Algorithms

Variational quantum algorithms (VQAs) represent a class of hybrid quantum-classical optimization methods that are pivotal in near-term quantum machine learning due to their compatibility with noisy intermediate-scale quantum (NISQ) devices. These algorithms leverage parameterized quantum circuits, known as ansatze, to approximate solutions to complex optimization problems by iteratively minimizing a cost function through classical feedback loops. The core idea draws from the variational principle in quantum mechanics, which posits that the expectation value of an observable provides an upper bound to the true ground state energy, enabling efficient approximations even with limited quantum resources.^[13] In the VQA framework, a parameterized quantum circuit U(\theta) prepares a trial state |\psi(\theta)\rangle = U(\theta) |0\rangle from an initial state, typically the all-zero state. The cost function C(\theta) is defined as the expectation value of a problem Hamiltonian H, given by

C(\theta) = \langle \psi(\theta) | H | \psi(\theta) \rangle,

which is evaluated on the quantum hardware. A classical optimizer, such as gradient descent or COBYLA, then updates the parameters \theta to minimize C(\theta), forming a closed hybrid loop where quantum circuits compute expectations and classical routines adjust parameters based on measurement outcomes. This structure ensures that VQAs can handle tasks beyond full quantum simulation, with the variational principle guaranteeing that the minimized C(\theta) upper-bounds the ground state energy of H.^[13]^[6] Prominent applications of VQAs include the variational quantum eigensolver (VQE), introduced by Peruzzo et al. in 2014, which targets molecular ground state energies in quantum chemistry by encoding the molecular Hamiltonian into H and optimizing C(\theta) to find low-energy configurations. Another key example is the quantum approximate optimization algorithm (QAOA), proposed by Farhi et al. in 2014, which applies alternating layers of problem-specific and mixing unitaries to approximate solutions for combinatorial optimization problems like MaxCut, where the cost function encodes the objective as a quadratic unconstrained binary optimization (QUBO) form. These methods have been adapted to machine learning tasks, such as classification, through variational quantum classifiers (VQCs), where the ansatz processes encoded feature vectors and the cost function measures classification error, enabling supervised learning on quantum data in the 2020s.^[13]^[6]^[14] Trainability analyses of VQAs reveal that, under certain ansatz designs like hardware-efficient circuits with local connectivity, the variance of cost function gradients scales polynomially with system size, facilitating effective optimization even for moderately large qubit counts. Quantum neural networks serve as a specific type of ansatz within VQAs, mimicking layered neural architectures to enhance expressivity for learning tasks.^[15]

Quantum Kernel Methods

Quantum kernel methods represent a class of quantum-enhanced machine learning techniques that map classical data into high-dimensional quantum Hilbert spaces to improve the separability of complex datasets, potentially offering exponential advantages over classical kernels for certain problems. In these methods, classical data points x and y are encoded via a quantum feature map \phi, which embeds them as quantum states |\phi(x)\rangle and |\phi(y)\rangle. The resulting quantum kernel function is then given by K(x, y) = |\langle \phi(x) | \phi(y) \rangle|^2, capturing inner products in the implicit high-dimensional feature space without explicitly computing the mapping, analogous to classical kernel tricks but leveraging quantum superposition and entanglement for richer representations.^[16] Common quantum feature maps include amplitude encoding, which loads data directly into the amplitudes of a quantum state using O(\log n) qubits for n-dimensional vectors, and time-evolution under a problem-specific Hamiltonian, such as the ZZ-feature map that applies controlled-phase gates to encode pairwise interactions suitable for graph-structured data. Kernel values are estimated on quantum hardware using circuits like the swap test, which measures the overlap between two prepared states with success probability proportional to | \langle \phi(x) | \phi(y) \rangle |^2, requiring O(1/\epsilon^2) repetitions for precision \epsilon. These maps and estimation protocols enable the integration of quantum kernels into classical algorithms, preserving compatibility with frameworks like support vector machines while exploiting quantum dimensionality.^[16]^[17] A primary application is the quantum support vector machine (QSVM), where the quantum kernel replaces the classical one to achieve potential exponential speedup in separating high-dimensional data, as demonstrated for parity classification problems where classical kernels require exponentially many features but quantum kernels succeed with polynomial resources. Similarly, quantum principal component analysis (qPCA) uses quantum phase estimation on a time-evolved density matrix to extract dominant eigenvectors, providing quadratic speedup in sample complexity over classical PCA for low-rank covariance matrices. The foundational framework for these quantum kernel approaches was established in the seminal work by Havlíček et al. (2019), which experimentally validated QSVM on superconducting processors for tasks like graph state discrimination.^[16]^[18] Recent advances in quantum kernel-based nearest neighbors (quantum KNN) classifiers, incorporating distance metrics in the quantum feature space, have explored classification on benchmark datasets like Iris and Wine on noisy intermediate-scale quantum devices. These developments, including noise-resilient implementations, highlight the practical scalability of quantum kernels toward near-term applications in pattern recognition.

Quantum Neural Networks

Quantum neural networks (QNNs) represent a class of quantum machine learning models that emulate the structure and functionality of classical neural networks using quantum circuits. These models leverage parameterized quantum circuits (PQCs) to process data and learn representations, where input data is encoded into quantum states, and the circuit applies a series of quantum gates to evolve the state. The architecture typically consists of multiple layers of variational quantum circuits, each incorporating trainable parameters in gates such as single-qubit rotation gates, for example, the Y-rotation gate R_y(\theta) = \begin{pmatrix} \cos(\theta/2) & -\sin(\theta/2) \\ \sin(\theta/2) & \cos(\theta/2) \end{pmatrix}. During the forward pass, the quantum state evolves under the unitary operation of the full circuit, and outputs are obtained via projective measurements of specific observables, yielding expectation values that serve as the network's predictions. This setup allows QNNs to model complex, high-dimensional functions potentially inaccessible to classical neural networks due to quantum superposition and entanglement.^[19]^[20] Training QNNs involves optimizing the parameters to minimize a loss function, often through hybrid quantum-classical optimization. Unlike classical backpropagation, which relies on chain rule differentiation, quantum gradients are computed using the parameter-shift rule, an exact method for Pauli-eigenvalue-based observables. For a parameter \theta_i in a gate, the gradient of the expectation value \partial \langle O \rangle / \partial \theta_i is given by

\frac{\langle O(\theta_i + \pi/2) \rangle - \langle O(\theta_i - \pi/2) \rangle}{2},

where the shifted circuits are evaluated separately on quantum hardware or simulators. This rule requires only two circuit evaluations per parameter, making it efficient for near-term devices, though it scales linearly with the number of parameters. Early proposals for QNNs, such as those introducing classification via layered PQCs, demonstrated their potential for supervised learning on near-term processors.^[21]^[19] Variants of QNNs include quantum convolutional neural networks (QCNNs), which incorporate convolutional and pooling layers adapted to quantum settings for tasks like image recognition or quantum state classification. In QCNNs, convolutional layers apply parameterized unitaries to subsets of qubits to extract local features, while pooling is achieved through measurements that project states onto symmetric subspaces, reducing dimensionality without classical post-processing. These models can be fully quantum, processing quantum data end-to-end, or hybrid, where classical preprocessing encodes data and post-processing interprets measurements. QNNs integrate with broader variational quantum algorithms by using ansatz circuits optimized iteratively, but their layered design draws direct inspiration from neural network topologies. Recent developments from 2023 to 2025 have extended QNN architectures to quantum transformers, enabling efficient handling of sequence data through quantum attention mechanisms that exploit entanglement for parallel processing.^[22]^[23]^[24]

Algorithms Based on Quantum Search and Sampling

Algorithms based on quantum search and sampling leverage the principles of Grover's algorithm and quantum sampling techniques to address machine learning tasks such as classification, pattern retrieval, and generative modeling. These methods exploit quantum superposition and interference to achieve potential speedups over classical counterparts, particularly for unstructured search problems and probabilistic distributions. Quantum search algorithms, rooted in amplitude amplification, enable efficient identification of optimal solutions in high-dimensional spaces, while sampling approaches facilitate the generation of complex data distributions using quantum thermal states or adversarial training.

Grover-Based Algorithms

Grover's algorithm provides a quadratic speedup for unstructured search problems by amplifying the amplitude of target states in a superposition. In quantum machine learning, this is adapted for binary classification tasks, where the algorithm searches for optimal decision boundaries by marking states corresponding to correct classifications. For instance, a Grover-search based quantum learning scheme encodes classical data into quantum states and uses iterative amplitude amplification to identify classifying hyperplanes, with applications demonstrated for datasets like Iris or synthetic binaries. Similarly, quantum discriminators employ Grover's oracle to flip phases of misclassified samples, enabling binary classification with reduced query complexity compared to classical methods. The core operation in these classifiers is the Grover iterate, defined as

G = -A S_{\chi} A^{-1} S_{0},

where A prepares the initial uniform superposition, S_{\chi} is the oracle that flips the phase of marked (target) states, A^{-1} inverts this preparation, and S_{0} performs diffusion about the mean by inverting amplitudes relative to the average. This iterate is applied \mathcal{O}(\sqrt{N/M}) times, where N is the search space size and M the number of solutions, to amplify target probabilities. Quantum amplitude amplification, generalizing Grover's search, was formalized by Brassard et al. in 2000, allowing flexible amplification of subspaces defined by partial reflections. An optimized variant uses partial diffusion operators to enhance convergence for binary quantum neural networks, reducing iterations by up to 30% in simulations on 4-qubit systems. Quantum associative memories extend search principles for pattern storage and retrieval, using amplitude encoding to represent multiple patterns in a single quantum state. In this encoding, classical patterns \mathbf{x}_i \in \mathbb{R}^d are mapped to amplitudes of a normalized quantum state |\psi\rangle = \sum_i \alpha_i |i\rangle, where \sum_i |\alpha_i|^2 = 1 and \alpha_i encodes features of \mathbf{x}_i. Retrieval involves applying a Grover-like search to amplify the amplitude of the closest stored pattern to a query input, enabling robust recall even with noisy queries. High-capacity models, inspired by Hopfield networks, store exponentially many patterns (up to $2^n for n qubits) with fidelity preserved via entanglement, as shown in theoretical frameworks that compare open quantum system dynamics for storage efficiency. A quantum set intersection algorithm applies Grover search to find overlapping patterns between query and memory states, achieving retrieval probabilities scaling as \sin^2((2k+1)\theta) after k iterations, with applications to image recognition tasks.

Sampling Methods

Quantum sampling algorithms draw from Gibbs distributions or adversarial frameworks to generate data for machine learning, particularly in unsupervised settings. Quantum Boltzmann machines (QBMs) model generative distributions using quantum Hamiltonians, where the probability of a configuration is p(\mathbf{x}) = \langle \mathbf{x} | e^{-\beta H} | \mathbf{x} \rangle / Z, with H a parameterized Hamiltonian and \beta the inverse temperature. Training involves variational optimization to minimize the free energy, enabling QBMs to learn classical data distributions like bars-and-stripes patterns with fewer parameters than classical Boltzmann machines. The Quantum Approximate Optimization Algorithm (QAOA) approximates Gibbs sampling by alternating applications of mixing and cost Hamiltonians, U(\gamma, \beta) = e^{-i\beta B} e^{-i\gamma C}, iterated p times to sample from thermal states, providing a hybrid approach for learning latent representations in datasets. Quantum generative adversarial networks (qGANs) incorporate quantum discriminators and generators to learn probability distributions implicitly defined by data. In qGANs, the generator produces quantum states encoding synthetic samples, while the discriminator, often a quantum circuit, distinguishes real from fake data using amplitude estimation or kernel methods; this setup converges to the Nash equilibrium via minimax optimization, \min_G \max_D V(D,G) = \mathbb{E}_{\mathbf{x} \sim p_{\text{data}}}[\log D(\mathbf{x})] + \mathbb{E}_{\mathbf{z} \sim p_z}[\log(1 - D(G(\mathbf{z}))))]. Early proposals in 2018 demonstrated qGANs loading distributions into quantum states for loading onto quantum RAM, with applications to financial time-series generation showing reduced sample complexity. These sampling techniques have been briefly explored for enhancing reinforcement learning by generating diverse state transitions, though full integration remains nascent.

Hybrid and Classical Approaches to Quantum Problems

Classical Learning on Quantum Data

Classical learning on quantum data refers to the application of classical machine learning algorithms to process and analyze datasets generated from quantum systems, such as measurement outcomes from quantum states, processes, or evolutions. This approach leverages the efficiency of classical computational resources to reconstruct or predict quantum properties without requiring full quantum simulation, which is often intractable for large systems. Key motivations include reducing the exponential scaling of traditional quantum tomography, where the number of measurements grows with system size, and enabling scalable inference for noisy intermediate-scale quantum (NISQ) devices. A prominent method is neural-network-based quantum state tomography (QST), where deep neural networks are trained on measurement data to reconstruct high-dimensional quantum states. In this framework, the network learns a parameterized representation of the density matrix from projective measurements in random bases, enforcing physical constraints like positivity and unit trace during optimization. This technique has demonstrated high fidelity for highly entangled states with over 100 qubits in simulations, outperforming compressed sensing methods by handling highly entangled states with fewer measurements. For instance, restricted Boltzmann machines or variational autoencoders can parameterize the state. Experimental validations on photonic and superconducting platforms confirm its robustness to noise.^[25]^[26] Another foundational technique is classical shadow tomography, which constructs a classical ensemble of "shadows" from randomized single-qubit measurements to estimate multiple linear properties of an unknown quantum state. Each shadow is an invertible transformation of a measurement outcome, enabling unbiased estimation of expectation values for observables like Pauli strings or entanglement witnesses with sample complexity independent of system size—typically O(log M / ε^2) for M properties with precision ε. This method achieves near-optimal scaling, matching information-theoretic bounds, and has been applied to predict fidelities, Rényi entropies, and correlation functions from just hundreds of measurements on 20-qubit states. Extensions to process tomography and Hamiltonian learning further allow inference of quantum channels and time-evolution operators from time-series data. In Hamiltonian learning, classical algorithms infer the parameters of an unknown quantum Hamiltonian from observed dynamics or equilibrium data. Bayesian methods combined with Gaussian processes have experimentally reconstructed Hamiltonians for spin chains using time-resolved measurements on ion traps, achieving parameter errors below 1% with O(1) measurements per parameter in low-noise settings. Neural networks enhance this by modeling time evolution directly, as in neural differential equations that fit trajectories to learn many-body interactions, demonstrating accuracy on Ising models up to 8 sites. These approaches are particularly valuable for quantum simulation validation and control, where they bypass full state reconstruction.^[27]^[28]

Quantum Annealing for Optimization

Quantum annealing represents a hardware-based approach in quantum machine learning that leverages quantum effects to solve optimization problems central to machine learning tasks, such as minimizing energy-based objectives or selecting optimal model parameters.^[29] In this paradigm, the quantum annealer evolves a physical system from an initial easy-to-prepare state toward the ground state of a problem Hamiltonian encoding the machine learning objective, exploiting quantum tunneling to navigate complex energy landscapes more efficiently than classical methods in certain cases.^[30] This process is particularly suited for problems formulated as Ising models or quadratic unconstrained binary optimization (QUBO) problems, which are ubiquitous in machine learning for tasks like training probabilistic models and feature engineering.^[31] The annealing process begins with an initial Hamiltonian H_s, typically a transverse-field driver that places the system in its ground state of uniform superposition, and gradually transitions to the problem Hamiltonian H_p via a time-dependent schedule s(t), where s(0) = 0 (full driver) and s(1) = 1 (full problem).^[30] The total Hamiltonian is given by

H(t) = A(t) H_s + B(t) H_p,

where A(t) decreases and B(t) increases over time, with H_s = -\sum_i \sigma_x^i as the transverse-field driver promoting quantum fluctuations, and H_p = \sum_{i<j} J_{ij} \sigma_z^i \sigma_z^j + \sum_i h_i \sigma_z^i encoding the Ising model for the optimization objective.^[31] By slowly evolving the system adiabatically, quantum annealing aims to find the low-energy configurations of H_p, corresponding to optimal solutions for machine learning problems mapped to ground-state searches.^[29] Commercial implementations, such as those from D-Wave Systems since the release of the 128-qubit D-Wave One in 2011, have enabled practical experimentation with this approach on real hardware.^[32] In machine learning applications, quantum annealing has been applied to train Boltzmann machines by sampling from the quantum Boltzmann distribution to approximate the partition function and update model parameters, as demonstrated in the development of quantum Boltzmann machines where the transverse-field Ising model captures quantum correlations beyond classical restricted Boltzmann machines.^[33] For feature selection, optimization problems are reformulated as QUBO tasks to identify sparse subsets of informative features that minimize classification error while reducing dimensionality, with quantum annealing providing solutions that can outperform classical heuristics in high-dimensional datasets.^[34] These QUBO formulations allow direct embedding onto annealer hardware, enabling efficient exploration of the combinatorial search space inherent in feature interactions.^[34] Recent studies have explored quantum annealing for clustering tasks, which are NP-hard optimization problems involving partitioning data into groups that minimize intra-cluster variance.^[35] In 2025 benchmarks, quantum annealing on D-Wave processors demonstrated advantages over classical simulated annealing in solving combinatorial optimization problems, achieving better solution quality for certain problem scales where classical methods plateau due to local minima trapping.^[29] This performance edge highlights quantum annealing's potential for scaling machine learning workflows on combinatorial problems, though hybrid classical-quantum solvers are often employed to handle embedding and post-processing.

Quantum-Enhanced Reinforcement Learning

Quantum-enhanced reinforcement learning (QRL) leverages quantum computing to augment classical reinforcement learning (RL) by representing policies and value functions using quantum states, potentially enabling speedups in exploration, policy optimization, and handling of high-dimensional environments. In the QRL framework, an agent's state is encoded as a quantum state |s⟩ in a Hilbert space, while actions are accessed via quantum oracles that prepare superpositions over action sets for parallel evaluation. Amplitude encoding maps classical Q-learning tables—representing state-action values—onto the amplitudes of a quantum register, allowing efficient storage of exponentially large tables using logarithmic qubits and facilitating interference-based updates. This approach exploits quantum superposition to evaluate multiple actions simultaneously, reducing the computational overhead in value iteration.^[36]^[37] Key algorithms in QRL achieve quantum speedups through techniques like Grover's search for action selection, which provides a quadratic improvement (O(√|A|) queries) over classical exhaustive search when selecting high-value actions from large discrete action spaces in partially observable environments. Additionally, variational quantum circuits (VQCs) parameterize policy networks, where quantum gates encode policy distributions and are optimized via classical gradients to approximate optimal policies, often requiring fewer parameters than classical neural networks for certain tasks. These methods integrate seamlessly with hybrid quantum-classical loops, such as quantum approximate optimization for policy refinement.^[38]^[39]^[40] The foundational Bellman equation in QRL adopts a quantum form to compute the value function under policy π:

V^\pi(s) = \max_a \langle s,a | R + \gamma P V^\pi | s,a \rangle

Here, |s,a⟩ denotes the joint quantum state-action encoding, R is the reward operator, P the transition operator, and γ the discount factor; the inner product evaluates quantum matrix elements, enabling parallel computation of expectations across superposed trajectories. This operator formulation supports non-Abelian dynamics and entanglement in quantum environments. Early seminal work by Dong et al. (2008) introduced QRL by using quantum amplitudes to balance exploration and exploitation, proving convergence to optimal policies in finite Markov decision processes via simulated experiments on grid worlds.^[41]^[42] In 2025, advances in quantum deep RL have demonstrated practical enhancements in robotics simulations, where hybrid quantum Soft Actor-Critic models for humanoid navigation in MuJoCo environments achieve 8% higher average returns and 92% fewer training steps than classical baselines, by using VQCs to encode continuous action spaces and mitigate curse-of-dimensionality issues in stochastic settings.^[43]

Quantum Learning Theory

Quantum Generalization and Complexity

Quantum machine learning models leverage the geometry of Hilbert space to achieve generalization from training data, where the high-dimensional structure enables effective interpolation between data points represented as quantum states. In this framework, the overlap and entanglement properties inherent to quantum states facilitate a form of kernel-based separation that classical Euclidean geometry cannot replicate, allowing quantum classifiers to capture complex decision boundaries with fewer parameters relative to the exponential dimensionality of the space. This geometric advantage is particularly evident in quantum kernel methods, where the feature map embeds data into the Hilbert space, promoting smoother generalization curves compared to classical nonlinear embeddings.^[44] Extensions of the no free lunch theorem to quantum settings reveal that no quantum learning algorithm can outperform others on average across all possible quantum data distributions, underscoring the need for task-specific architectures in QML. Specifically, the quantum NFL theorem demonstrates that the expected performance of any quantum learner over a uniform ensemble of unitaries is identical, limiting universal advantages and emphasizing the role of prior knowledge about the data's quantum structure. This result implies stricter bounds on learnability for entangled datasets, where classical NFL assumptions break down due to quantum correlations.^[45] In terms of computational complexity, quantum machine learning operates within the BQP complexity class, which encompasses problems solvable in polynomial time on a quantum computer with bounded error probability, contrasting with classical P (deterministic polynomial time) and the presumed intractability of NP-complete problems. While no unconditional separations are proven, oracle constructions establish relativized separations showing BQP ≠ P, and quantum learning tasks can separate BPP from more powerful classes like QMA in specific query models. For instance, learning certain Boolean linear functions over product distributions requires only O(poly(n)) quantum circuit depth, enabling efficient training for linearly separable quantum data.^[46]^[47] Recent theoretical advances highlight separations in learning capabilities, with quantum learners achieving exponential speedups over classical ones for specific function classes under cryptographic hardness assumptions, as shown in analyses of PAC-style frameworks. Additionally, studies on quantum overparameterization have revealed double descent phenomena, where increasing model parameters beyond the data size leads to improved generalization rather than overfitting, mirroring classical deep learning trends but rooted in quantum circuit expressivity.^[48] An analog to the VC dimension in quantum models measures the expressive power through the complexity of hypothesis classes in Hilbert space \mathcal{H}.^[49]

Quantum PAC Learning and Sample Complexity

In quantum probably approximately correct (PAC) learning, a learner aims to identify a hypothesis from a quantum concept class that approximates an unknown target concept with high probability, using access to quantum examples. A quantum example consists of a quantum state encoding the input and label, allowing the learner to perform quantum measurements or circuits on it before collapsing the state. This framework extends the classical PAC model to quantum settings, where the realizable case assumes the target is in the concept class, while the agnostic case allows the target to be arbitrary, seeking the best approximation within the class. Quantum PAC learning has been formalized to capture scenarios where quantum data provides richer information than classical bits, potentially leading to efficiency gains in sample usage.^[50] Sample complexity in quantum PAC learning refers to the number of quantum examples required to achieve an error rate ε with confidence 1-δ. For learning rank-r quantum states in a d-dimensional Hilbert space to fidelity 1-ε, information-theoretic lower bounds establish that Ω(r d / ε) samples are necessary, even for tomography tasks underlying many learning problems, as fewer samples cannot distinguish between sufficiently separated states. However, quantum data offers advantages over classical counterparts by encoding exponential information in superposition, enabling algorithms to extract more utility per sample; for instance, in agnostic quantum PAC learning of Boolean functions, the sample complexity matches classical bounds up to constants, but specific classes exhibit quadratic speedups. Recent results demonstrate provable advantages, such as a square-root improvement in sample complexity for general concept classes with VC dimension d, requiring O((d / √ε) polylog(1/ε, 1/δ)) quantum samples compared to classical Θ(d / ε).^[51]^[52] The quantum adaptation of the classical PAC generalization bound relies on the Vapnik-Chervonenkis (VC) dimension of the quantum hypothesis class. For a class with quantum VC dimension VC, the error ε satisfies

\varepsilon \leq \frac{\mathrm{VC}}{m} + \sqrt{\frac{\log(1/\delta) + \mathrm{VC} \log m}{m}},

where m is the number of quantum samples, mirroring the classical form since measuring quantum examples yields classical data, though quantum processing can tighten bounds for certain classes.^[50] Key results include exact learning algorithms for specific quantum concepts; for example, quantum examples enable exact identification of Boolean linear functions under product distributions using O(1) to O(log n) samples, far fewer than the classical Ω(n) requirement. For quantum kernel methods, sublinear sample algorithms achieve classification with guarantees using O(√(n d / ε)) queries for n data points in d dimensions, leveraging quantum linear algebra for kernel estimation. In 2025 analyses extending beyond binary classification, quantum PAC sample complexities for multiclass settings remain Θ((Ndim(H) + log(1/δ))/ε) in realizable cases, with no broad sublinear gains but targeted improvements in online and agnostic multiclass learning.^[47]^[53]^[54]

Implementations and Experiments

Hardware Platforms and Frameworks

Quantum machine learning (QML) implementations rely on diverse hardware platforms that provide the foundational qubits and gates for quantum operations, primarily operating in the noisy intermediate-scale quantum (NISQ) era. Superconducting qubit systems, developed by companies like IBM and Google, use Josephson junctions to create transmon qubits cooled to millikelvin temperatures, enabling fast gate operations with single-qubit fidelities exceeding 99.9% and two-qubit gate fidelities around 99.5% as of 2025. IBM's 2025 roadmap features the Nighthawk processor with 120 qubits and high connectivity, supporting up to 5,000 gates, while aiming to scale fault-tolerant systems to 2,000 qubits for deeper circuits. Trapped-ion platforms, led by IonQ, employ electromagnetic traps to confine ions like ytterbium or barium, achieving some of the highest gate fidelities, with single-qubit operations above 99.99% and two-qubit gates exceeding 99.99%, due to their long coherence times exceeding seconds.^[55] Photonic quantum hardware from Xanadu utilizes photons in integrated silicon chips for room-temperature operation, demonstrating gate fidelities over 99.9% in linear optical setups, which facilitate scalable networking but face challenges in single-photon sources; in 2025, Xanadu introduced the networked Aurora system for modular photonic QML applications.^[56] Quantum annealing processors from D-Wave, such as the Advantage2 system introduced in 2025, specialize in optimization tasks with over 4,400 qubits, offering energy-based sampling advantages for machine learning applications like clustering, though limited to adiabatic evolutions.^[57] Software frameworks bridge these hardware platforms with classical machine learning ecosystems, enabling hybrid QML workflows where quantum circuits interface with gradient-based optimizers. Qiskit, IBM's open-source library, supports circuit design, simulation, and execution on superconducting hardware, integrating with TensorFlow and PyTorch via extensions like Qiskit Machine Learning for variational quantum algorithms. Google's Cirq focuses on NISQ devices with customizable gates and noise models, compatible with TensorFlow Quantum for end-to-end hybrid training on their Sycamore processors. PennyLane, developed by Xanadu, emphasizes differentiable quantum programming for photonic and other backends, seamlessly integrating with PyTorch, TensorFlow, and JAX to compute quantum gradients for QML models like quantum neural networks. These frameworks allow users to deploy variational circuits—parameterized quantum circuits optimized classically—across cloud-accessible hardware, democratizing QML development without requiring physical quantum devices. Scalability in the NISQ era remains constrained by noise and connectivity, with practical QML applications typically limited to 50-100 usable qubits due to decoherence times on the order of microseconds for superconducting systems and error accumulation in multi-qubit operations. While hardware advances push toward 1,000+ physical qubits by late 2025, effective circuit depths for QML are often below 100 gates, necessitating error mitigation techniques integrated into frameworks like Qiskit and PennyLane. Access to these platforms is facilitated through cloud services, such as IBM Quantum Network, IonQ's Harmony systems, and Xanadu's Strawberry Fields, promoting collaborative QML research.

Key Experimental Demonstrations

In more recent superconducting implementations, researchers demonstrated quantum continual learning on a programmable processor in 2024, attaining 95.8% accuracy on sequential tasks while effectively mitigating catastrophic forgetting—a common challenge in classical neural networks—through variational quantum algorithms that adapt parameters across learning phases.^[58] Complementing this, quantum kernel experiments have evaluated resource-efficient kernels for classification on IBM's superconducting hardware, with ongoing work exploring improved feature mapping fidelity, though specific speedups remain under investigation for larger scales beyond classical simulation. A landmark demonstration of verifiable quantum advantage came from Google's October 2025 experiment with the Quantum Echoes algorithm on their Willow processor, which provided a 13,000x speedup over classical supercomputers in measuring quantum out-of-time-order correlators relevant to machine learning tasks like error detection and pattern recognition in noisy data.^[59] Earlier, in 2023, quantum generative adversarial networks (qGANs) were explored through simulations to generate synthetic financial time series data, showing potential for capturing correlations in stock indices via quantum circuits for future applications in risk assessment, though experimental hardware demonstrations are pending.^[60] These experiments, often employing variational quantum algorithms, highlight practical speedups in sampling (e.g., 10x for high-dimensional distributions) but underscore that many current demonstrations remain within the reach of classical simulation due to noise and scale limitations.

Challenges and Limitations

Barren Plateaus and Trainability Issues

One of the primary challenges in training variational quantum algorithms (VQAs) for quantum machine learning is the barren plateaus phenomenon, where the loss landscape becomes exponentially flat as the number of qubits increases, making gradient-based optimization ineffective.^[61] This issue was first systematically analyzed in the context of quantum neural networks, demonstrating that random parameter initializations lead to vanishing gradients across the parameter space.^[61] Specifically, the variance of the gradient of the cost function with respect to variational parameters scales exponentially with the system size, given by

\mathrm{Var}\left( \frac{\partial C}{\partial \theta} \right) \sim 2^{-n}

for n qubits, which hinders the ability to find meaningful updates during training.^[61] The root cause of barren plateaus lies in the concentration of measure phenomenon inherent to high-dimensional Hilbert spaces, where the vast majority of states concentrate around highly entangled configurations with near-constant expectation values for local observables. This leads to exponentially small fluctuations in the cost function, exacerbated by the random nature of quantum circuit initializations and the expressivity of deep parameterized circuits.^[62] Beyond barren plateaus, other trainability issues include noise-induced errors, where hardware imperfections flatten the cost landscape further and suppress gradient signals, limiting the depth and scale of trainable models.^[63] Overfitting also poses challenges, particularly in overparameterized quantum models that interpolate training data perfectly but exhibit high variance, though some quantum architectures demonstrate benign overfitting with preserved generalization due to unitary constraints.^[48] To address these trainability barriers, several mitigation strategies have emerged, including layerwise training approaches that optimize circuits incrementally by depth to avoid premature gradient vanishing in deep variational quantum circuits (VQCs).^[64] Hardware-efficient ansatze, designed to match near-term device topologies, have been shown to eliminate barren plateaus under specific parameter constraints, enabling trainability for arbitrary circuit depths without altering the architecture.^[65] Recent advances in adaptive circuits, such as the cyclic variational quantum eigensolver introduced in 2025, employ measurement-driven feedback to dynamically expand the variational space in promising directions, effectively escaping barren plateaus through staircase-like optimization paths while maintaining chemical precision.^[66]

Skepticism and Theoretical Critiques

Skepticism surrounding quantum machine learning (QML) primarily stems from the phenomenon of dequantization, where purported quantum advantages in machine learning tasks can be replicated or approximated by classical algorithms with comparable efficiency. For instance, quantum kernel methods, often touted for exponential speedups in similarity computations, have been dequantized through classical approximations that achieve similar performance without quantum hardware, as demonstrated in analyses of recommendation systems and linear algebra routines. This has led to critiques that many QML proposals rely on idealized assumptions, such as fault-tolerant quantum computers or quantum random access memory (qRAM), which remain impractical, resulting in no proven exponential speedups for real-world machine learning problems beyond small-scale demonstrations.^[67]^[68]^[69] Theoretical critiques, exemplified by Scott Aaronson's longstanding skepticism, emphasize that QML's foundational claims often overlook the "fine print" of quantum computing limitations, such as the need for massive overhead in error correction and the absence of non-trivial quantum speedups for general learning tasks. Aaronson's analyses highlight how early QML optimism, including proposals for quantum support vector machines, has been tempered by dequantization results showing classical algorithms can match quantum performance in polynomial time for most instances. These views underscore a broader philosophical doubt: without structures inherently leveraging quantum superposition for data processing, QML may not transcend classical methods in expressive power or scalability.^[70]^[67] Recent debates in 2025, following Google's Quantum Echoes announcement on the Willow chip, have intensified scrutiny over QML-specific gains, with researchers questioning whether verifiable quantum advantages in physics simulations translate to machine learning applications amid persistent noise and decoherence issues. While Quantum Echoes demonstrated a 13,000× speedup in out-of-time-order correlator measurements, critics argue it reinforces NISQ-era hype without addressing QML's core challenges, such as trainability barriers like barren plateaus that hinder optimization.^[71]^[59] The hype-reality gap in QML is further exacerbated by NISQ limitations, which delay the arrival of fault-tolerant systems through high error rates and qubit instability, and economic barriers to scaling, including exorbitant costs for cryogenic infrastructure and specialized expertise that could exceed billions for viable prototypes. Despite these critiques, a balanced perspective acknowledges QML's potential in niche domains like quantum chemistry, where quantum-enhanced models can efficiently learn molecular properties from inherently quantum data, offering targeted advantages over classical approaches in simulation-heavy tasks.^[72]^[73]^[74]^[75]

Emerging Directions

Explainable and Fully Quantum Models

Efforts to enhance explainability in quantum machine learning (QML) have focused on adapting classical interpretability techniques to quantum circuits, addressing the inherent opacity of quantum processes. One prominent approach involves Quantum Shapley values, which extend the game-theoretic Shapley values from classical machine learning to attribute the contribution of individual quantum gates or components within a circuit to the overall model performance. This method quantifies the importance of gates by averaging their marginal contributions across all possible subsets of the circuit, enabling users to understand why specific quantum operations drive successful task outcomes, such as classification accuracy. For instance, in variational quantum circuits, Quantum Shapley values have been applied to identify redundant gates, facilitating circuit optimization and interpretability without sacrificing predictive power.^[76] Circuit visualization tools complement these attribution methods by providing intuitive graphical representations of quantum model internals. Tools like VIOLET offer visual analytics for quantum neural networks, allowing users to explore parameter influences, superposition states, and entanglement patterns through interactive diagrams and heatmaps. These visualizations help demystify how quantum interference affects decision boundaries, making complex models more accessible to researchers and practitioners. Such tools are particularly valuable in debugging QML pipelines, where traditional debugging is hindered by the non-intuitive nature of quantum evolution.^[77] Fully quantum models represent a paradigm shift toward end-to-end quantum processing, where both inputs and outputs remain in the quantum domain, eliminating classical intermediaries that can introduce bottlenecks or information loss. These models process quantum data directly, such as density matrices or quantum states generated by sensors in applications like precision metrology or quantum imaging. For example, quantum sensor networks produce entangled states as inputs, which fully quantum learners can classify or generate new states from without measurement collapse until the final output stage, preserving quantum advantages like exponential state spaces. This approach learns patterns in quantum correlations to predict environmental parameters with enhanced sensitivity. In 2024, frameworks for explainable QML (XQML) emerged to integrate interpretability directly into fully quantum architectures, such as those combining quantum representation learning with post-hoc explanation modules. These frameworks, like QRLaXAI, employ quantum autoencoders for feature extraction followed by variational classifiers. Additionally, quantum decision trees, accelerated by Grover's search algorithm, enable interpretable splitting of quantum feature spaces by querying oracles in superposition to find optimal decision nodes quadratically faster than classical counterparts. This leverages Grover's amplitude amplification to evaluate tree branches efficiently, providing a tree-like structure that traces decision paths transparently.^[78]^[79] Despite these advances, challenges persist due to the black-box nature of quantum measurements, which collapse superpositions into probabilistic outcomes, obscuring the causal links between inputs and decisions. This measurement-induced irreversibility complicates feature attribution, as post-measurement statistics may not fully capture the underlying quantum dynamics, leading to incomplete explanations. Ongoing research aims to mitigate this through measurement-free protocols or advanced tomography, but the fundamental no-cloning theorem limits direct access to quantum states, underscoring the need for hybrid interpretability strategies.

Recent Breakthroughs in Continual and Generative Learning

In 2024, a significant advancement in quantum continual learning was demonstrated on a programmable superconducting quantum processor, where researchers achieved 95.8% accuracy on sequential classification tasks while effectively avoiding catastrophic forgetting of prior knowledge.^[80]^[58] This experimental implementation adapted classical strategies like replay mechanisms to quantum memory buffers, storing representative quantum states from earlier tasks to replay during training on new data, thereby maintaining performance across non-stationary quantum datasets.^[58] The approach highlighted the feasibility of hardware-efficient continual learning in noisy intermediate-scale quantum systems, with the superconducting platform enabling scalable task sequencing without full retraining. Building on this, hybrid quantum-classical reinforcement learning (RL) frameworks have emerged for continual adaptation in dynamic environments. In 2025, modified hybrid agents incorporating quantum circuits for policy optimization adapted more rapidly to environmental shifts, yielding higher average rewards than classical counterparts in simulated tasks.^[81] Similarly, hybrid policy gradient methods using variational quantum circuits demonstrated improved control in adaptive scenarios, such as resource allocation in wireless networks, where quantum components enhanced exploration of state-action spaces. These developments enable quantum agents to learn continuously from streaming data, fostering applications in real-time decision-making. In generative modeling, quantum generative adversarial networks (QGANs) and variational autoencoders (VAEs) have incorporated entangled latent representations to capture quantum correlations beyond classical limits. A 2025 investigation revealed that quantum latent distributions in GANs improved sample quality and diversity on image datasets, outperforming classical GANs by leveraging superposition for more expressive encodings.^[82] Quantum VAEs with entangled latents similarly enhanced reconstruction fidelity for quantum data, such as molecular configurations, by modeling joint probability distributions via quantum circuits.^[83] Quantum diffusion models represent another key breakthrough, offering iterative denoising in quantum Hilbert spaces for high-fidelity generation. The QSC-Diffusion model, proposed in 2025, uses unitary scrambling operators followed by measurement-induced collapse to generate samples from complex distributions, achieving superior performance over hybrid QGANs on benchmarks like MNIST and Fashion-MNIST with fewer parameters.^[84] Hybrid quantum-classical latent diffusion models further extended this to medical imaging tasks, demonstrating practical utility on current hardware while surpassing VAEs in capturing rare events.^[85] In October 2025, advanced quantum simulation techniques, such as the Quantum Echoes algorithm, have been developed for modeling quantum systems.^[59] These innovations in continual and generative quantum learning point toward scalable systems capable of processing real-world data streams, from adaptive robotics to drug discovery, by combining quantum expressivity with classical efficiency for lifelong model evolution.

References

[1]
Quantum machine learning - Nature
Sep 14, 2017 · The field of quantum machine learning explores how to devise and implement quantum software that could enable machine learning that is faster than that of ...
[2]
Challenges and Opportunities in Quantum Machine Learning - arXiv
Mar 16, 2023 · Here we review current methods and applications for QML. We highlight differences between quantum and classical machine learning, with a focus ...
[3]
The Development of Quantum Machine Learning
Jan 27, 2022 · In this article, we have discussed three waves of quantum machine learning, each harnessing a particular aspect of quantum computers and ...
[4]
[0811.3171] Quantum algorithm for solving linear systems of equations
We exhibit a quantum algorithm for this task that runs in poly(log N, kappa) time, an exponential improvement over the best classical algorithm.
[5]
[1411.4028] A Quantum Approximate Optimization Algorithm - arXiv
Abstract:We introduce a quantum algorithm that produces approximate solutions for combinatorial optimization problems.
[6]
A variational eigenvalue solver on a quantum processor - arXiv
Apr 10, 2013 · Access Paper: View a PDF of the paper titled A variational eigenvalue solver on a quantum processor, by Alberto Peruzzo and 7 other authors.
[7]
Quantum Computation and Quantum Information
This comprehensive textbook describes such remarkable effects as fast quantum algorithms, quantum teleportation, quantum cryptography and quantum error- ...
[8]
A fast quantum mechanical algorithm for database search - arXiv
Nov 19, 1996 · The quantum algorithm can find a phone number in O(sqrt(N)) steps, using quantum superposition to examine multiple names simultaneously.Missing: URL | Show results with:URL
[9]
[quant-ph/9508027] Polynomial-Time Algorithms for Prime ... - arXiv
Aug 30, 1995 · Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer. Authors:Peter W. Shor (AT&T Research).Missing: URL | Show results with:URL
[10]
[PDF] Pattern Recognition and Machine Learning - Microsoft
A companion volume (Bishop and Nabney,. 2008) will deal with practical aspects of pattern recognition and machine learning, and will be accompanied by Matlab ...
[11]
[1409.3097] An introduction to quantum machine learning - arXiv
Sep 10, 2014 · This contribution gives a systematic overview of the emerging field of quantum machine learning. It presents the approaches as well as technical details in an ...
[12]
A variational eigenvalue solver on a photonic quantum processor
Jul 23, 2014 · The quantum phase estimation algorithm efficiently finds the eigenvalue of a given eigenvector but requires fully coherent evolution.<|control11|><|separator|>
[13]
Universal expressiveness of variational quantum classifiers and ...
Feb 2, 2023 · Variational quantum circuits have been used to devise variational quantum classifiers (VQC). However, for QML to become a new computational ...
[14]
Efficient Estimation of Trainability for Variational Quantum Circuits
Dec 4, 2023 · The paper proposes mapping quantum circuits to efficiently simulated Clifford circuits, enabling polynomial-time calculation of gradients and ...
[15]
Supervised learning with quantum-enhanced feature spaces - Nature
Mar 13, 2019 · Here we propose and experimentally implement two quantum algorithms on a superconducting processor. A key component in both methods is the use of the quantum ...Missing: paper | Show results with:paper
[16]
[1307.0401] Quantum principal component analysis - arXiv
Jul 1, 2013 · Access Paper: View a PDF of the paper titled Quantum principal component analysis, by Seth Lloyd and 2 other authors. View PDF · TeX Source.
[17]
Quantum principal component analysis | Nature Physics
Jul 27, 2014 · This paper shows that the state can play a dynamic role in its own analysis. In particular, we show that multiple copies of the state ρ can ...
[18]
Classification with Quantum Neural Networks on Near Term ... - arXiv
Feb 16, 2018 · Access Paper: View a PDF of the paper titled Classification with Quantum Neural Networks on Near Term Processors, by Edward Farhi and Hartmut ...Missing: seminal 2014
[19]
Quantum Neural Networks - Qiskit Machine Learning 0.8.4
QNNs apply this generic principle by combining classical neural networks and parametrized quantum circuits.Missing: parameterized | Show results with:parameterized<|separator|>
[20]
Parameter-shift rules - PennyLane
The parameter-shift rule is a recipe for how to estimate gradients of quantum circuits.
[21]
Quantum convolutional neural networks | Nature Physics
Aug 26, 2019 · We introduce and analyse a quantum circuit-based algorithm inspired by convolutional neural networks, a highly effective model in machine learning.
[22]
[1810.03787] Quantum Convolutional Neural Networks - arXiv
Oct 9, 2018 · First, QCNN is used to accurately recognize quantum states associated with 1D symmetry-protected topological phases. We numerically demonstrate ...
[23]
Quantum Vision Transformers
Feb 22, 2024 · We introduce three types of quantum transformers for training and inference, including a quantum transformer based on compound matrices.
[24]
Neural-network quantum state tomography | Nature Physics
Feb 26, 2018 · Here we show how machine learning techniques can be used to perform QST of highly entangled states with more than a hundred qubits, to a high degree of ...
[25]
[2206.06736] Neural-network quantum state tomography - arXiv
Jun 14, 2022 · We revisit the application of neural networks techniques to quantum state tomography. We confirm that the positivity constraint can be successfully implemented.
[26]
Experimental quantum Hamiltonian learning | Nature Physics
Mar 13, 2017 · With the help of a quantum simulator and Bayesian inference it is possible to determine the unknown Hamiltonian of a quantum system.
[27]
Solving the quantum many-body Hamiltonian learning problem with ...
Solving the quantum many-body Hamiltonian learning problem with neural differential equations · 1. Introduction · 2. NDE in quantum dynamics · 3. Results · 4.<|control11|><|separator|>
[28]
Quantum annealing for combinatorial optimization: a benchmarking ...
May 16, 2025 · Quantum annealing (QA) has the potential to significantly improve solution quality and reduce time complexity in solving combinatorial optimization problems.
[29]
Quantum annealing: an overview | Philosophical Transactions of the ...
Dec 5, 2022 · The targeted ground state of the Ising Hamiltonian for 𝛤 = 0 , however, is an ordered state in the sense that it has a fixed magnetization + 1 ...
[30]
What is Quantum Annealing? - D-Wave Documentation
In quantum annealing, the system begins in the lowest-energy eigenstate of the initial Hamiltonian. As it anneals, it introduces the problem Hamiltonian, which ...
[31]
D-Wave sells first commercial quantum computer - Phys.org
Jun 1, 2011 · Last week, Burnaby, British Columbia-based company D-Wave Systems, Inc., announced that it sold its first commercial quantum computer.
[32]
Quantum Annealing for Enhanced Feature Selection in Single-Cell ...
Aug 16, 2024 · In this study, we implement quantum annealing-empowered quadratic unconstrained binary optimization (QUBO) for feature selection in scRNA-seq ...
[33]
Quantum annealing based feature selection - ScienceDirect.com
Oct 15, 2025 · This study showcased the usefulness of quantum annealers for feature selection in machine learning. The challenge of finding the optimal feature ...
[34]
Quantum annealing-based clustering of single cell RNA-seq data
Oct 24, 2023 · Quantum annealing based clustering yields the most significant promise of advantage for the relatively homogeneous clusters. As in those cases, ...
[35]
[0810.3828] Quantum reinforcement learning - arXiv
Oct 21, 2008 · In this paper, a novel quantum reinforcement learning (QRL) method is proposed by combining quantum theory and reinforcement learning (RL).
[36]
[PDF] Quantum-enhanced Reinforcement Learning - KLEE
The achievement was a quantum algorithm that exhibits a complexity independent on the number of states of the environment. This independence guarantees its.
[37]
A hybrid classical-quantum approach to speed-up Q-learning - Nature
Mar 8, 2023 · We introduce a classical-quantum hybrid approach to computation, allowing for a quadratic performance improvement in the decision process of a learning agent.
[38]
Variational Quantum Circuits for Deep Reinforcement Learning - arXiv
Jun 30, 2019 · This work explores variational quantum circuits for deep reinforcement learning. Specifically, we reshape classical deep reinforcement learning algorithms.
[39]
Variational quantum reinforcement learning via evolutionary ...
Feb 15, 2022 · We present two frameworks for deep quantum RL tasks using gradient-free evolutionary optimization. First, we apply the amplitude encoding scheme to the Cart- ...
[40]
Quantum Reinforcement Learning in Non-Abelian Environments
Apr 11, 2024 · By leveraging the quantum Bellman equation, we establish a methodology for maximizing expected cumulative reward over an infinite horizon, ...
[41]
https://arxiv.org/abs/2406.06531
[42]
[2509.11388] Quantum deep reinforcement learning for humanoid ...
By integrating quantum computing with deep RL, we aim to develop models that can efficiently learn complex navigation tasks in humanoid robots.Missing: simulation | Show results with:simulation
[43]
Generalization in Quantum Machine Learning
Nov 1, 2021 · Here we establish a link between quantum classification and quantum information theory, by showing that the accuracy and generalization capability of quantum ...Article Text · QUANTUM HYPOTHESIS... · BIAS-VARIANCE TRADE-OFF...
[44]
[2003.14103] No Free Lunch for Quantum Machine Learning - arXiv
Mar 31, 2020 · The ultimate limits for the quantum machine learning of quantum data are investigated by obtaining a generalisation of the celebrated No Free Lunch (NFL) ...
[45]
[PDF] Quantum Versus Classical Proofs and Advice - Scott Aaronson
Theorem 1.2 implies that it is pointless to try to prove a classical oracle separation between QMA and QCMA by proving a lower bound on the quantum query ...Missing: learning | Show results with:learning
[46]
Quantum learning Boolean linear functions w.r.t. product distributions
Apr 20, 2020 · The problem of learning Boolean linear functions from quantum examples w.r.t. the uniform distribution can be solved on a quantum computer ...
[47]
Generalization despite overfitting in quantum machine learning ...
Dec 20, 2023 · In this work, we provide a characterization of benign overfitting in quantum models. To do this, we derive the behavior of a classical interpolating Fourier ...
[48]
General Vapnik–Chervonenkis dimension bounds for quantum ...
Nov 14, 2022 · In this work, we prove a general VC dimension upper bound for quantum circuit learning including 'data re-uploading' circuits.
[49]
[PDF] Optimal Quantum Sample Complexity of Learning Algorithms
Upper bounds on sample complexity carry over from classical to quantum PAC learning, because a quantum example becomes a classical example if we just measure it ...
[50]
[2309.10887] Provable Advantage in Quantum PAC Learning - arXiv
Sep 19, 2023 · Up to polylogarithmic factors, this is a square root improvement over the classical learning sample complexity. We show the tightness of our ...
[51]
An Improved Sample Complexity Lower Bound for (Fidelity ...
Jan 3, 2023 · This paper presents a sharper lower bound on the number of copies of a quantum state needed to learn a classical description of it.
[52]
Sublinear quantum algorithms for training linear and kernel-based ...
We investigate quantum algorithms for classification, a fundamental problem in machine learning, with provable guarantees.Missing: sample complexity
[53]
Quantum Learning Theory Beyond Batch Binary Classification
Jul 29, 2025 · Can quantum computers learn more efficiently than classical ones? This is a central question in quantum machine learning. A major hope has been ...
[54]
Quantum continual learning on a programmable superconducting ...
Sep 15, 2024 · Here, we report an experimental demonstration of quantum continual learning on a fully programmable superconducting processor.
[55]
A verifiable quantum advantage - Google Research
Oct 22, 2025 · A verifiable quantum advantage. October 22, 2025. Xiao Mi and Kostyantyn Kechedzhi, Research Scientists, Google Quantum AI.
[56]
Implementing Quantum Generative Adversarial Network (qGAN) and ...
Aug 15, 2023 · We use real world financial dataset and compare models such as qGAN (quantum generative adversarial networks) and QCBM (quantum circuit Born machine) among ...
[57]
Barren plateaus in quantum neural network training landscapes
Nov 16, 2018 · We show that the exponential dimension of Hilbert space and the gradient estimation complexity make this choice unsuitable for hybrid quantum-classical ...
[58]
The barren plateaus of quantum neural networks: review, taxonomy ...
Dec 11, 2023 · 2.1 The origins of barren plateaus ... In the quantum algorithm, many factors lead to BP, including the high expressibility of the quantum circuit ...
[59]
Can Error Mitigation Improve Trainability of Noisy Variational ...
Mar 14, 2024 · Recent studies have shown that noise can severely limit the trainability of VQAs, eg, by exponentially flattening the cost landscape and suppressing the ...
[60]
Investigating and mitigating barren plateaus in variational quantum ...
Jan 31, 2025 · ... layer-wise training methods to incrementally avoid BPs in deep VQCs. ... Hardware-efficient ansatz without barren plateaus in any depth.
[61]
Hardware-efficient ansatz without barren plateaus in any depth - arXiv
Mar 7, 2024 · In this paper, we propose two novel parameter conditions in which the hardware-efficient ansatz (HEA) is free from barren plateaus for arbitrary circuit depths.
[62]
Cyclic Variational Quantum Eigensolver: Escaping Barren Plateaus through Staircase Descent
### Summary of https://arxiv.org/abs/2509.13096
[63]
Customers who liked this quantum recommendation engine might ...
Jul 12, 2018 · The new result thus, I think, significantly changes the landscape of quantum machine learning, by killing off one of its flagship applications.
[64]
[PDF] Quantum Machine Learning Without Any Quantum - Ewin Tang
On the other hand, the latter class can be “dequantized,” meaning that our classical framework produces algorithms to perform computations in this class at most ...Missing: skepticism | Show results with:skepticism<|separator|>
[65]
What are the strongest arguments for a genuine quantum computing ...
Dec 9, 2021 · In parallel to the development of quantum algorithms, there is a "dequantization" effort: find the essential feature of the quantum algorithm ...
[66]
Quantum Machine Learning — Dr. Scott Aaronson's “Read the Fine ...
This is the “fine print” of quantum machine learning. Recently, Dr. Scott Aaronson, a professor at MIT, published an essay about quantum machine learning.Missing: 2021 | Show results with:2021
[67]
Google claims 'quantum advantage' again — but researchers are ...
Oct 22, 2025 · Google's algorithm is able to detect subtle quantum links between distant parts of the computer, which otherwise get scrambled and lost owing to ...
[68]
Quantum Myth Busters: Experts Debunk Common NISQ-Era Myths
Jan 14, 2025 · A recent study evaluates quantum myths, addressing issues like error mitigation, algorithm scalability, and practical applications.
[69]
Quantum Computing Faces 3 Major Barriers Before Going Mainstream
So once those developing quantum computers are able to stabilize qubits, scale the number of qubits and offer a diverse set of tools, frameworks and ...
[70]
Supervised Quantum Machine Learning: A Future Outlook ... - arXiv
Jun 17, 2025 · Likely areas include: quantum chemistry and materials (where the data is quantum mechanical and quantum models naturally fit), optimization ...
[71]
Classical Machine Learning vs Quantum Machine Learning in ...
Oct 13, 2025 · This article surveys the current state of classical ML in computational chemistry, highlighting its extraordinary successes and real limitations ...
[72]
[2301.09138] Explaining Quantum Circuits with Shapley Values - arXiv
Jan 22, 2023 · The resulting attributions can be interpreted as explanations for why a specific circuit works well for a given task, improving the ...
[73]
VIOLET: Visual Analytics for Explainable Quantum Neural Networks
Dec 23, 2023 · We propose VIOLET, a novel visual analytics approach to improve the explainability of quantum neural networks.
[74]
Power of data in quantum machine learning | Nature Communications
May 11, 2021 · These are the largest combined simulation and analysis in digital quantum machine learning that we are aware of, and make use of the TensorFlow ...
[75]
Quantum Advantage in Learning from Experiments - Google Research
Jun 22, 2022 · Enter QML, a technology that straddles the line between quantum computers and quantum sensors. QML algorithms make computations that are aided ...
[76]
Opportunities and limitations of explaining quantum machine learning
Dec 19, 2024 · We additionally propose two explanation methods designed specifically for quantum machine learning models, as first of their kind to the best of ...
[77]
QRLaXAI: quantum representation learning and explainable AI
Feb 19, 2025 · We propose an explainable QRL framework combining a quantum autoencoder (QAE) with a variational quantum classifier (VQC) and incorporating theoretical and ...
[78]
Quantum computation and decision trees | Phys. Rev. A
Aug 1, 1998 · We devise a quantum-mechanical algorithm that evolves a state, initially localized at the root, through the tree.
[79]
Machine learning on quantum experimental data toward solving ...
Aug 30, 2024 · Generalization in quantum machine learning from few training data. ... Peer review. Peer review information. Nature Communications thanks ...
[80]
Quantum Machine Learning Advances: Researchers Achieve 95.8 ...
Sep 18, 2024 · Scientists demonstrated quantum continual learning on a superconducting quantum processor for catastrophic forgetting in machine learning.