A Primer on Quantum Machine Learning

TL;DR

Quantum machine learning (QML) is analyzed across theory, variational methods, linear-algebraic primitives, and learning with quantum data. It surveys PAC/VC learning theory, provable separations, variational QML with data embeddings and quantum kernels, and linear-algebraic algorithms (e.g., $QPE$, $HHL$, $qPCA$, $QSVE$, $QSVT$), illustrating with a quantum recommendation system. The discussion emphasizes where evidence supports real benefits and where speedups rely on strong data-access assumptions, including dequantization challenges, while also surveying quantum generative models, unsupervised learning, reinforcement learning, and practical nonstandard applications like quantum sensing and compiling. The overall message is a balanced map: meaningful quantum advantages exist for quantum-data tasks and carefully scoped problems, but broad practical gains for classical-data tasks require explicit, auditable data-access models and rigorous benchmarking.

Abstract

Quantum machine learning (QML) is a computational paradigm that seeks to apply quantum-mechanical resources to solve learning problems. As such, the goal of this framework is to leverage quantum processors to tackle optimization, supervised, unsupervised and reinforcement learning, and generative modeling-among other tasks-more efficiently than classical models. Here we offer a high level overview of QML, focusing on settings where the quantum device is the primary learning or data generating unit. We outline the field's tensions between practicality and guarantees, access models and speedups, and classical baselines and claimed quantum advantages-flagging where evidence is strong, where it is conditional or still lacking, and where open questions remain. By shedding light on these nuances and debates, we aim to provide a friendly map of the QML landscape so that the reader can judge when-and under what assumptions-quantum approaches may offer real benefits.

Figures (5)

• Figure 1: Quantum machine learning (QML) is somewhat of an umbrella term spanning many settings. This figure schematically organizes the landscape along two axes—data type (classical vs. quantum) and algorithm type (classical vs. quantum)--yielding four illustrative categories. Examples in each box are indicative rather than exhaustive, and the boundaries are guidelines rather than strict rules as hybrid paradigms can straddle multiple categories. This review concentrates on schemes where the quantum device is the primary learning and considers both classical and quantum data (categories $1$ and $2$, respectively).
• Figure 2: Toy-model example and its solution using classical ML techniques. a) We consider a one-dimensional classification problem with input domain $\mathcal{X}=[-\pi,\pi]\subseteq\mathbb{R}$ and labels $\mathcal{Y}=\{0,1\}$ assigned as per the target function $f(x)$ of Eq. \ref{['eq:toy-model-target']}. b) Since there is no linear classifier over the input domain which solves the task, one can map the data onto the high-dimensional space $\mathbb{R}^2$ via the feature map $\phi(x)=(x,\ x^2)$ where a classifying hyperplane does exist (dashed colored line). All points above the classifying hyperplane are assigned a label $h(x)=1$, and all point below a label $h(x)=0$. c) We show the kernel $K(x,x')=\left(1+\frac{xx'}{\pi^2}\right)^2$ for the dataset ordered per increasing value of $x$. This kernel is the backbone of the SVM trained for classification. d) Schematic diagram of a feedforward NN with one input, two hidden, and one output neuron. After training, the output of $F_{2,1}$ (shown as a solid red line) can be used for classification.
• Figure 3: Schematic representation of an archetypal variational QML pipeline. States from a supervised learning dataset $\{\rho_i,y_i\}_{i=1}^N$ are sent through a parametrized quantum circuit (or quantum neural network) $U(\boldsymbol{\theta})$, and one can estimate the quantities $\tilde{y}_{\boldsymbol{\theta}}^i={\rm Tr}[\rho_i(\boldsymbol{\theta})O]$ for $\rho_i(\boldsymbol{\theta})=U(\boldsymbol{\theta})\rho_iU^\dagger(\boldsymbol{\theta})$. These expectation values are fed to a classical optimizer that computes an empirical loss $\widehat{R}_{\mathcal{S}}(h_{\boldsymbol{\theta}})$ (where $h_{\boldsymbol{\theta}}$ denotes the output parametrized hypothesis), and leverages classical optimizers to find the new set of parameters which solves the optimization problem $\min_{\boldsymbol{\theta}} \ \widehat{R}_{\mathcal{S}}(h_{\boldsymbol{\theta}})$.
• Figure 4: Toy-model example and its solution using QML techniques on a single qubit. a) A feature map encodes data as gate-rotation angles, embedding points into the qubit Hilbert space $\mathcal{H}=\mathbb{C}^2$ (a higher-dimensional feature space relative to the 1D input). A classifying hyperplane is then trained by measuring a rotated Pauli observable with trainable parameters. We show the encoded data for $\boldsymbol{v}=(0,1,0)$. b) One can also leverage the encoded data to estimate—via the SWAP-test circuit—the quantum kernel $K_q(x,x')=|\langle\phi(x)\vert\phi(x')\rangle|^2$. c) A QNN is implemented as a sequence of data encoding unitaries and parametrized trainable gates. After training, the expectation value of $\tilde{y}_{\boldsymbol\theta}(x)=\mathrm{Tr}[\rho(x,\boldsymbol{\theta})\,O]$ for $O=Z$ (shown as a solid red line) can be used for classification.
• Figure 5: a) Example of a binary tree $\mathsf{B}_{i}$ to store a classical state $\boldsymbol{x}$ and to prepare the data structure for a two-qubit state. The algorithm on the right illustrates how the classical vector stored in the binary tree (left) is encoded into a quantum state. b) Circuit for the quantum phase estimation algorithm applied to a unitary operator $U$ via $n$ ancilla qubits. Here, QFT denotes the quantum Fourier transform nielsen2000quantum and QFT$^{-1}$ its inverse. c) Circuit for using the Harrow-Hassidim-Lloyd (HHL) algorithm to tackle the quantum linear system problem to solve $A \boldsymbol{x} = \boldsymbol{b}$. The first register is used to load the input vector $\boldsymbol{b}$ and the resulting solution state, while the second scratch register provides additional qubits required for QPE algorithm. Both QRAM and the ability to perform quantum simulation on $A$ to implement $e^{iA}$ are basic ingredients of this technique.