QML Algorithms Explained: A Researcher’s Guide to 12 Essential Approaches

From variational quantum classifiers to quantum kernels, generative models, and quantum linear solvers — exploring the algorithms, mathematical foundations, applications and limitations of quantum machine learning.

Topics: Quantum Computing + Machine Learning

What Happens When Quantum Computing Meets Machine Learning?

Machine learning has transformed how we recognize patterns, classify information, generate content, and make predictions. Yet many computational problems remain challenging, particularly when the underlying data structures or mathematical relationships become difficult to represent and process efficiently.

Quantum machine learning (QML) explores whether quantum computing can introduce new computational strategies for solving selected machine learning problems.

Classical computers process information using bits, conventionally represented as 0 or 1. Quantum computers use quantum bits, or qubits, which can exist in superpositions of computational basis states. Quantum algorithms use operations such as quantum gates, interference and entanglement to process quantum information.

These properties create opportunities for new approaches to computation. However, they do not automatically make every machine learning algorithm faster, more accurate or more efficient. The central research question is therefore not simply whether quantum computing can be combined with machine learning.

The real question is: Under what conditions can a quantum learning algorithm provide a measurable advantage over a suitable classical alternative?

To answer this, we need to understand the algorithms, their mathematical foundations, their potential applications, and the challenges that remain.

In this article, we explore 12 important approaches to quantum machine learning and examine where each fits into the broader research landscape.

1. Foundations of Quantum Machine Learning

Before examining individual algorithms, let us establish three essential concepts.

1.1 Quantum states and qubits

A single qubit can be represented as: |ψ⟩ = α|0⟩ + β|1⟩. Here, α and β are complex probability amplitudes satisfying: |α|² + |β|² = 1. When a measurement is performed in the computational basis, the probability of obtaining 0 is |α|², while the probability of obtaining 1 is |β|².

Multiple qubits can represent quantum states with correlations that cannot always be described as independent single-qubit states. Entanglement is one of the features that distinguishes quantum information processing from conventional classical representations.

However, quantum measurement does not reveal an entire quantum state in a single operation. Extracting useful information from a quantum computation can require repeated measurements and careful experimental design.

1.2 Quantum feature maps

Many quantum machine learning algorithms begin by encoding classical input data into a quantum state.

The transformation can be written as: |φ(x)⟩ = Uφ(x)|0⟩

Here:

• x represents the input data.
• Uφ(x) represents a quantum circuit that encodes the input.
• |0⟩ represents the initial reference state.
• |φ(x)⟩ represents the resulting quantum feature state.

The purpose of a quantum feature map is to transform input data into a representation that a subsequent quantum algorithm can process.

The effectiveness of this representation depends on the circuit structure, the encoding strategy, the dataset, and the learning objective.

A complicated feature map is not necessarily a useful one. The representation must capture relevant information while remaining computationally practical.

1.3 Hybrid quantum-classical learning

Many near-term quantum machine learning approaches combine quantum circuits with classical optimization.

A typical training process follows these steps:

1. Encode the input data into a quantum circuit.
2. Apply parameterized quantum gates.
3. Measure selected observables.
4. Calculate a classical loss function.
5. Update the circuit parameters using a classical optimizer.
6. Repeat until a stopping criterion is reached.

The trainable parameters are often represented by θ.

The model can be described abstractly as: f(x; θ) = QuantumCircuit(x, θ). The optimizer adjusts θ to minimize a task-specific loss. This hybrid structure is widely used in variational quantum classifiers and quantum neural network research. However, it introduces practical challenges involving measurement costs, circuit noise, optimization stability, and execution time.

2. Twelve Essential Quantum Machine Learning Algorithms

2.1 Variational Quantum Classifier (VQC)

Category: Supervised learning

A Variational Quantum Classifier uses a parameterized quantum circuit to classify input data. The circuit transforms an input state, performs trainable quantum operations, and produces measurements that are interpreted as model predictions.

A simplified mathematical representation is: f(x; θ) = ⟨0|U†(x, θ) O U(x, θ)|0⟩

Here:

• x is the input data.
• θ represents trainable circuit parameters.
• U(x, θ) represents the data-encoding and parameterized circuit.
• U† is the conjugate transpose of U.
• O is an observable whose expectation value is measured.

The resulting value can be used directly as a prediction score or transformed into class probabilities through an appropriate model design.

How it works: The circuit produces predictions for training examples. A classical loss function measures the difference between the predictions and the target labels. A classical optimizer then updates the circuit parameters. This process repeats until the model meets a chosen stopping condition.

Best use cases

• Small-scale classification experiments.
• Quantum-state classification.
• Research into hybrid quantum-classical learning.
• Investigations of trainable quantum representations.

Advantages: VQCs are flexible and can be adapted to different input representations and learning objectives. They can also be implemented on quantum simulators before being tested on physical hardware.

Limitations: Training can be difficult when gradients become very small, a problem associated with barren plateaus in certain circuit families and training conditions. Noise, circuit depth, and measurement costs can further limit performance.

Research perspective: A VQC is a useful experimental starting point, but it should be compared with well-tuned classical models before any claim of improved performance is made.

2.2 Quantum Support Vector Machine (QSVM)

Category: Supervised learning

A classical Support Vector Machine uses a kernel function to measure relationships between data points and construct a decision boundary.A Quantum Support Vector Machine can use quantum circuits to estimate a kernel that represents similarities between encoded quantum states.

One common fidelity-based quantum kernel is: K(x, z) = |⟨φ(x)|φ(z)⟩|². Here, |φ(x)⟩ and |φ(z)⟩ are the quantum feature states corresponding to inputs x and z. The kernel value measures the squared magnitude of their inner product.

How it works

1. Encode each input using a quantum feature map.
2. Estimate similarities between selected pairs of inputs.
3. Construct a kernel matrix.
4. Supply the matrix to a classical kernel-based learning algorithm.
5. Evaluate the resulting classifier on unseen data.

Best use cases

• Small-data classification.
• Research into quantum feature representations.
• Selected pattern-recognition problems.
• Experiments using quantum-generated data.

Advantages: Quantum kernels provide a framework for investigating feature spaces that may be difficult to reproduce efficiently using particular classical methods.

Limitations: Estimating kernel entries can require many circuit executions. Kernel quality can also be poor if the feature map fails to capture useful structure. The full computational pipeline matters: data encoding, kernel estimation, classical optimization, and prediction must all be considered.

Research perspective: A quantum kernel is not automatically superior to a classical kernel. The relevant question is whether its representation and computational cost provide a meaningful advantage for a specific task.

2.3 Quantum Neural Networks (QNNs)

Category: Primarily supervised learning

Quantum Neural Networks commonly use parameterized quantum circuits as trainable models. Although their name is inspired by classical neural networks, quantum circuits do not simply reproduce conventional neurons and activation functions. Their computational operations and information representations are different.

A typical QNN contains:

• An input-encoding circuit.
• Trainable quantum gates.
• Optional entangling operations.
• A measurement stage.
• A classical procedure for computing predictions and updating parameters.

An abstract model can be written as: f(x; θ) = Measurement(U(x, θ)|0⟩). The measurement function converts information from the quantum state into a classical output.

Best use cases

• Small-scale classification.
• Regression experiments.
• Quantum-state recognition.
• Hybrid quantum-classical learning research.

Advantages: QNNs offer a flexible framework for studying trainable quantum representations and different circuit architectures.

Limitations: Optimization can be sensitive to circuit initialization, parameterization, noise, and the chosen loss function. Greater circuit expressivity can sometimes make training harder rather than easier.

Research perspective: A QNN should be evaluated according to its task, architecture, and empirical performance rather than the assumption that quantum models inherently outperform classical neural networks.

2.4 Quantum k-Nearest Neighbors (QKNN)

Category: Supervised learning

The classical k-nearest neighbors algorithm predicts a label by examining the labels of nearby training examples. Quantum k-nearest neighbors approaches investigate whether quantum procedures can help estimate similarities, distances, or neighborhood relationships. The specific implementation varies. Some proposals use quantum-state fidelity, while others investigate distance estimation or quantum subroutines for particular data-access models.

For quantum states, one possible similarity measure is: Similarity(x, z) = |⟨φ(x)|φ(z)⟩|². This measures fidelity-based similarity between two pure quantum states.

Best use cases

• Quantum-state classification.
• Experimental similarity search.
• Research into quantum distance estimation.

Advantages: These approaches provide a way to investigate whether quantum subroutines can contribute to similarity-based learning.

Limitations: Preparing quantum states, estimating similarities, and returning useful classical predictions can introduce substantial overhead. The complexity depends on the implementation and assumptions about data access.

Research perspective: Quantum similarity estimation does not automatically make nearest-neighbor search faster. The entire algorithm must be compared with classical alternatives.

2.5 Quantum Principal Component Analysis (QPCA)

Category: Dimensionality reduction and quantum data analysis

Classical Principal Component Analysis identifies directions of high variance in a dataset. It is commonly used for dimensionality reduction, visualization, and feature extraction. Quantum PCA refers to quantum approaches for extracting spectral information from quantum-encoded data, including certain methods based on density matrices and quantum phase estimation.

A density operator can be represented as: ρ = Σᵢ λᵢ |vᵢ⟩⟨vᵢ|

Here:

• ρ is a density operator.
• λᵢ are its eigenvalues.
• |vᵢ⟩ are the corresponding eigenvectors.

Quantum algorithms can, under suitable assumptions, estimate information about this spectral structure.

Best use cases

• Quantum-state analysis.
• Research involving quantum-generated data.
• Selected spectral-analysis problems.
• Investigations of quantum dimensionality-reduction methods.

Advantages: Quantum spectral algorithms provide a different way to study certain matrix and state representations.

Limitations: State preparation, access to copies of the relevant quantum state, circuit depth, and the extraction of classical outputs can significantly affect the overall cost.

Research perspective: QPCA should not be treated as a direct replacement for ordinary PCA on classical datasets. The input representation and output requirements are fundamentally important.

2.6 Quantum Clustering and Quantum k-Means

Category: Unsupervised learning

Classical k-means divides data into clusters by repeatedly assigning points to centroids and updating those centroids. Quantum clustering approaches investigate whether quantum procedures can assist with distance estimation, similarity calculations, or particular parts of the clustering process.

A clustering algorithm must still define how data points are assigned to clusters and how the cluster representatives are updated.

Best use cases

• Research involving quantum data.
• Experimental distance-based clustering.
• Investigations of quantum-assisted similarity computations.

Advantages: Some quantum subroutines may offer interesting computational trade-offs under specific assumptions about data representation and access.

Limitations: Preparing data, estimating distances, updating centroids, and returning cluster assignments can all be expensive. These costs must be included in a meaningful comparison.

Research perspective: Quantum clustering is a family of approaches rather than one universally defined algorithm. Its value depends on the specific implementation and task.

2.7 Quantum Generative Adversarial Networks (QGANs)

Category: Generative learning

Classical Generative Adversarial Networks use two models: a generator and a discriminator. The generator attempts to produce samples resembling a target distribution. The discriminator attempts to distinguish generated samples from reference samples.

Quantum GANs adapt this general idea using quantum circuits or quantum-state representations in one or both components.

How they work: A quantum generator prepares or samples from a quantum-based representation. A discriminator evaluates the generated output, potentially using classical or quantum operations. The training procedure updates the model components according to a defined objective.

Get Sarah Rachel’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Best use cases

• Small-scale distribution learning.
• Quantum-state generation.
• Exploratory generative modeling research.
• Experiments involving synthetic quantum data.

Advantages: QGANs provide a framework for investigating how quantum circuits can represent and generate probability distributions.

Limitations: Training instability, sampling costs, hardware noise, and limited circuit resources can make optimization difficult. Comparing a quantum generator with a classical GAN also requires compatible data and evaluation protocols.

Research perspective: A successful demonstration on a small quantum system does not establish superiority for large-scale image, text, or molecular generation.

2.8 Quantum Reinforcement Learning (QRL)

Category: Reinforcement learning

Reinforcement learning studies how an agent selects actions to maximize cumulative reward while interacting with an environment. Quantum reinforcement learning is a broad research area rather than a single algorithm. Proposed approaches may use quantum representations of policies, quantum-enhanced computational subroutines, or quantum environments.

A policy is commonly represented as: π(a | s)

Here:

• s represents the current state.
• a represents a possible action.
• π(a | s) represents the probability of choosing action a in state s.

Quantum approaches investigate whether particular parts of policy representation, value estimation, or optimization can benefit from quantum computation.

Best use cases

• Quantum control.
• Research involving quantum environments.
• Selected optimization and sequential-decision problems.

Advantages: QRL creates opportunities to study how quantum computation can interact with sequential decision-making.

Limitations: Training costs, environment complexity, measurement requirements, and the need for meaningful classical comparisons remain significant concerns.

Research perspective: A quantum reinforcement learning method should be assessed using task-specific metrics, including reward, training cost, sample efficiency, and any additional quantum resources required.

2.9 Quantum Boltzmann Machines (QBMs)

Category: Generative and probabilistic modeling

Classical Boltzmann machines are energy-based models that define probability distributions over configurations. Quantum Boltzmann machines introduce quantum states or quantum Hamiltonians into the modeling process.

A simplified thermal-state expression is: ρ = exp(-H) / Z

where: Z = Tr(exp(-H))

Here:

• H is a Hamiltonian describing the model.
• exp(-H) is its matrix exponential under the stated convention.
• Z is the partition function.
• Tr denotes the matrix trace.

In a parameterized model, the Hamiltonian depends on trainable parameters. A physical inverse-temperature factor can also be included when appropriate.

Best use cases

• Quantum-state distribution modeling.
• Research into energy-based learning.
• Selected probabilistic modeling problems.

Advantages: QBMs connect probabilistic modeling with concepts from quantum statistical mechanics.

Limitations: Estimating model expectations, training parameters, and sampling from relevant distributions can be computationally demanding. Gate-based QBMs and quantum-annealing approaches also have different implementation requirements.

Research perspective: QBMs are an active research direction, not a universally superior alternative to classical energy-based models.

2.10 Quantum Phase Estimation (QPE) for Machine Learning

Category: Quantum algorithmic subroutine

Quantum Phase Estimation estimates the phase associated with an eigenvalue of a unitary operator. Suppose a unitary operator U acts on an eigenstate |u⟩ as follows:

U|u⟩ = exp(2πiφ)|u⟩

Here, φ is the phase to be estimated. QPE uses controlled operations and quantum measurements to estimate this phase, subject to precision and resource requirements.

In machine-learning research, phase estimation has been investigated as a component of quantum spectral and linear-algebra algorithms.

Best use cases

• Quantum spectral estimation.
• Selected linear-algebra routines.
• Research into algorithms that require eigenvalue information.

Advantages: QPE is a foundational quantum algorithm with applications extending well beyond machine learning.

Limitations: High-precision implementations can require substantial circuit depth and reliable quantum operations. Performance also depends on state preparation and the particular problem.

Research perspective: QPE is a subroutine, not a complete machine learning model. Its relevance depends on the larger algorithm in which it is used.

2.11 Harrow–Hassidim–Lloyd (HHL) Algorithm

Category: Quantum linear-system algorithm

The HHL algorithm was developed to solve particular linear systems when the input and output can be represented as quantum states.

The system is written as: A|x⟩ = |b⟩

Here:

• A represents the system matrix.
• |b⟩ represents the encoded input vector.
• |x⟩ represents the desired solution state.

The algorithm uses quantum procedures to encode information related to the matrix spectrum, apply eigenvalue-dependent transformations, and construct a quantum state associated with the solution.

Under appropriate assumptions about sparsity, conditioning, precision, and input preparation, HHL can have favorable complexity compared with particular classical methods.

Potential applications

• Research into quantum linear algebra.
• Selected regression-related algorithms.
• Scientific computing approaches that can use quantum-state outputs.

Advantages: HHL helped establish the potential importance of quantum algorithms for linear systems and influenced subsequent quantum machine learning research.

Limitations: HHL does not automatically return every component of the solution as an inexpensive classical vector. State preparation, the condition number, precision, and output extraction all affect the total computational cost.

Research perspective: Theoretical exponential improvements sometimes associated with HHL rely on specific assumptions. They should not be interpreted as a general speedup for ordinary machine learning or regression.

2.12 Quantum Feature Maps

Category: Data representation

Quantum feature maps encode classical input data into quantum states using data-dependent quantum circuits.

The transformation is: |φ(x)⟩ = Uφ(x)|0⟩

The resulting state can be used in a quantum kernel or as input to a trainable quantum model. For example, a quantum kernel may compare the feature states associated with two inputs: K(x, z) = |⟨φ(x)|φ(z)⟩|²

The feature map determines how input data is represented before the subsequent learning procedure.

Best use cases

• Quantum kernel classification.
• Hybrid quantum-classical models.
• Research into alternative data representations.

Advantages: Quantum feature maps provide a systematic framework for investigating quantum representations of classical data.

Limitations: Their usefulness depends on the encoding circuit, circuit depth, hardware noise, and whether the representation captures task-relevant structure.

Research perspective: A high-dimensional quantum feature space does not guarantee better generalization, more efficient computation, or quantum advantage.

3. Comparing the Algorithms: Which Should You Choose?

The appropriate algorithm depends on the research objective, dataset, input representation, and available computational resources.

=> For classification: Consider a VQC or quantum kernel method. Compare both against suitable classical classifiers.

=> For trainable quantum representations: Consider a QNN, while carefully evaluating optimization stability and circuit complexity.

=> For similarity-based learning: Investigate quantum nearest-neighbor proposals only when their data-access and distance-estimation assumptions are appropriate.

=> For spectral analysis: Study QPCA or QPE-based methods when the problem naturally involves quantum states or spectral information.

=> For generative modeling: Explore QGANs or QBMs when their output representation and evaluation criteria fit the research question.

=> For sequential decisions: Consider QRL when a well-defined environment and meaningful classical baselines are available.

=> For linear algebra: Study HHL when the input, output, matrix structure, and conditioning assumptions match the proposed application.

=> For data representation: Experiment with quantum feature maps and determine whether they improve a specific task compared with classical feature representations.

4. Major Challenges in Quantum Machine Learning

4.1 Barren plateaus and optimization

Some parameterized quantum circuits exhibit regions where gradients become extremely small under particular circuit structures and initialization conditions. When this happens, estimating useful parameter updates can become difficult.

Potential research directions include structured circuit design, local cost functions, and improved parameter initialization. These approaches are not universal solutions, and their effectiveness depends on the problem.

4.2 Data-loading overhead

Most practical machine learning tasks begin with classical data. Encoding that information into quantum states can require substantial resources.

A meaningful performance analysis must include the cost of data preparation rather than measuring only the quantum computation after encoding is complete.

4.3 Noise and hardware limitations

Quantum operations can be affected by errors and environmental interactions. These limitations may restrict circuit depth, reduce measurement reliability, and complicate model training. Near-term noisy quantum devices must therefore be distinguished from large, fault-tolerant quantum computers.

4.4 Fair benchmarking

Quantum models should be evaluated against appropriate classical alternatives under comparable conditions. A rigorous experiment should document:

• Dataset size, feature representation, and preprocessing.
• Training and testing protocols.
• Classical baselines and hyperparameter tuning.
• Circuit depth, qubit count, and measurement shots.
• Runtime, data-loading costs, and hardware noise.
• Task-specific metrics, uncertainty, and resource consumption.

Without this information, a favorable result may not demonstrate a computational advantage.

5. A Practical Research Workflow for Beginners

Starting with all 12 approaches at once is rarely productive. A more reliable strategy is to build one controlled experiment.

Step 1: Define the problem

Choose a clearly specified task, such as a small classification or regression problem. Establish what success means before selecting the algorithm.

Step 2: Build a classical baseline

Train a suitable classical model, such as logistic regression, an SVM, or a small neural network. This baseline provides a reference for evaluating the quantum approach.

Step 3: Select a quantum method

A VQC is a reasonable starting point for hybrid classification experiments. A quantum kernel method is another option when the research question concerns feature representations.

Step 4: Start with simulation

Use a quantum simulator to test data encoding, circuit behavior, loss calculations, and evaluation code before moving to physical hardware.

Step 5: Evaluate fairly

Use appropriate data splits and comparable evaluation protocols. Record predictive performance alongside runtime, circuit resources, and measurement costs.

Step 6: Report the actual findings

Document successful experiments, unsuccessful runs, computational constraints, and limitations. Do not infer quantum advantage from theoretical expectations alone.

The goal of the first experiment is not to prove that quantum computing wins. It is to produce a reliable result that can be independently examined and reproduced.

Conclusion: The Future Belongs to Evidence-Driven Quantum Learning

Quantum machine learning is not one algorithm, one architecture, or one guaranteed path toward faster artificial intelligence. It is a broad research field encompassing variational classifiers, quantum kernels, generative models, reinforcement learning proposals, and quantum algorithms for linear algebra and spectral estimation.

Each approach addresses a different computational problem. Each also depends on assumptions that determine whether it can be implemented efficiently and compared meaningfully with classical alternatives. The most important question is not how many qubits an algorithm uses or how large its theoretical feature space appears. It is whether the complete method provides a reproducible and useful advantage for a clearly defined task.

For researchers and practitioners, progress will come from designing rigorous experiments, identifying where quantum representations contribute useful structure, and measuring the costs that theoretical analyses may overlook. Quantum computing may eventually transform selected areas of machine learning. Determining where and how that transformation occurs requires mathematics, engineering, careful experimentation, and intellectual honesty.

The future of quantum machine learning will be shaped not by the boldest promises, but by the strongest evidence.

References and Further Reading

1. Biamonte, J., et al. (2017). Quantum machine learning. Nature, 549, 195–202. https://doi.org/10.1038/nature23474
2. Schuld, M., & Killoran, N. (2019). Quantum Machine Learning in Feature Hilbert Spaces. Physical Review Letters, 122, 040504. https://doi.org/10.1103/PhysRevLett.122.040504
3. Cerezo, M., et al. (2021). Variational quantum algorithms. Nature Reviews Physics, 3, 625–644. https://doi.org/10.1038/s42254-021-00348-9
4. McClean, J. R., et al. (2018). Barren plateaus in quantum neural network training landscapes. Nature Communications, 9, 4812. https://doi.org/10.1038/s41467-018-07090-4
5. Huang, H.-Y., Broughton, M., Mohseni, M., et al. (2021). Power of data in quantum machine learning. Nature Communications, 12, 2631. https://doi.org/10.1038/s41467-021-22539-9

This article is an educational technical overview, not a report of new experiments or a claim that the listed algorithms outperform classical methods. Complexity advantages depend on the assumptions and implementation of each algorithm.

“Question the limits. Test the possibilities. Let evidence define the future.”

👏♻️💬 Enjoyed this read? Follow Sarah Rachel for more ideas at the frontier of AI, machine learning and quantum computing.