Quantum Machine Learning

1. 11. EncodeMap the input data into the initial quantum state using a fixed feature-map circuit.
2. 22. Apply the ansatzRun the parameterized circuit whose gate angles are the trainable parameters.
3. 33. MeasureEstimate the expectation value of a chosen observable by repeating the circuit for many shots.
4. 44. Compute loss and gradientTurn the expectation into a loss; get exact gradients on hardware with the parameter-shift rule.
5. 55. Update classicallyA classical optimizer nudges the angles, and the loop repeats until the loss stops improving.

The gradient step deserves attention because you cannot backpropagate through physical hardware. The parameter-shift rule gives exact analytic gradients from the same circuit run at shifted angles. For a parameter entering through a Pauli-generated gate (eigenvalues $\pm 1$), the gradient of an expectation $f(\theta) = \langle O\rangle$ is

$\frac{\partial f}{\partial \theta} = \tfrac{1}{2}\Bigl[f\bigl(\theta + \tfrac{\pi}{2}\bigr) - f\bigl(\theta - \tfrac{\pi}{2}\bigr)\Bigr].$

Note the shift is $\tfrac{\pi}{2}$, not an infinitesimal, this is an exact identity, not a finite-difference approximation.

Worked example: a gradient by hand

Take one qubit in $|0\rangle$, apply a rotation $R_Y(\theta)$, and measure the observable $Z$. The rotation gives

$R_Y(\theta)|0\rangle = \cos\tfrac{\theta}{2}\,|0\rangle + \sin\tfrac{\theta}{2}\,|1\rangle,$

so the expectation value is

$\langle Z\rangle = \cos^2\tfrac{\theta}{2} - \sin^2\tfrac{\theta}{2} = \cos\theta.$

Now apply the parameter-shift rule. Using $\cos(\theta + \tfrac{\pi}{2}) = -\sin\theta$ and $\cos(\theta - \tfrac{\pi}{2}) = \sin\theta$,

$\frac{\partial \langle Z\rangle}{\partial \theta} = \tfrac{1}{2}\bigl[-\sin\theta - \sin\theta\bigr] = -\sin\theta,$

which is exactly the analytic derivative of $\cos\theta$. The rule recovers the true gradient with two circuit evaluations. In PennyLane this is the default:

import pennylane as qml
from pennylane import numpy as np

dev = qml.device("default.qubit", wires=1)

@qml.qnode(dev)
def circuit(theta):
    qml.RY(theta, wires=0)
    return qml.expval(qml.PauliZ(0))

theta = np.array(0.9, requires_grad=True)
print(circuit(theta))            # cos(0.9)  ~=  0.6216
print(qml.grad(circuit)(theta))  # -sin(0.9) ~= -0.7833  (parameter-shift)

Quantum kernels and HHL

Two other families come up constantly, each with a catch.

Quantum kernels / feature maps. Encode each data point $x$ into a quantum state $|\phi(x)\rangle$ via a feature-map circuit, then use the squared overlap $k(x, x') = |\langle\phi(x)|\phi(x')\rangle|^2$ as a kernel inside an otherwise classical support vector machine. The hope is that a feature space hard to simulate classically yields better separation. The catch: a feature map being hard to compute classically does not imply it generalizes better. Follow-up work found classical methods frequently match quantum kernels, and expressivity without the right inductive bias just overfits.

HHL for linear systems. The Harrow–Hassidim–Lloyd algorithm solves $Ax = b$ in time polynomial in $\log$ of the dimension, an apparent exponential speedup. The fine print removes most of the win: you need efficient preparation of the state encoding $b$ (loading classical data can cost as much as you hoped to save, the input problem); $A$ must be sparse and well-conditioned, since the cost scales with the condition number $\kappa$; and the output is a quantum state $|x\rangle$, not the classical vector, so you can only extract summary statistics, not read out every entry (the output problem). HHL is not a drop-in fast linear solver.

Barren plateaus

The deepest obstacle to training VQCs is the barren plateau. For broad classes of randomly initialized, sufficiently expressive circuits, the gradient of the loss vanishes exponentially with the number of qubits: its variance shrinks like $O(2^{-n})$. The loss landscape becomes almost perfectly flat, so a gradient-based optimizer has no reliable direction to move, and the number of measurement shots needed to resolve a real gradient against sampling noise blows up exponentially. This is a fundamental scalability wall, not a tuning inconvenience, and mitigations (shallow or structured ansätze, local cost functions, careful initialization) are active research rather than a solved problem.

Dequantization: when the speedup evaporates

A sobering line of results shows that some celebrated QML speedups were never really quantum. Starting with Ewin Tang's 2018 work on recommendation systems, researchers dequantized several algorithms: they built classical algorithms, using quantum-inspired sampling techniques, that match the quantum poly-logarithmic scaling for the same problems (recommendation systems, low-rank linear algebra, some HHL-based ML). The claimed exponential separation vanished. The lesson is methodological: an apparent quantum speedup is a conjecture until someone proves no comparable classical algorithm exists, and several did not survive that test. "Quantum-inspired" classical results are not quantum-hardware results, and both must be distinguished from genuine, provable separations.

Honest status

To be explicit about confidence levels. KNOWN: the machinery is real, variational circuits, the parameter-shift rule, quantum kernels, and HHL are well-defined, and small proofs of concept run on today's hardware. EMERGING: whether any of them beats classical ML on a useful task is open and unproven; barren plateaus, noise, data loading, and dequantization each undercut the case. SPECULATIVE: claims of broad, near-term quantum advantage for machine learning. Study QML as a research frontier; do not deploy it expecting speedups.

Further reading

• Biamonte et al., "Quantum Machine Learning," Nature 549, 195 (2017), arXiv:1611.09347, the canonical survey of the field's ambitions and methods.
• Preskill, "Quantum Computing in the NISQ Era and Beyond," arXiv:1801.00862 (2018), sober framing of what near-term hardware can and cannot do.
• McClean et al., "Barren Plateaus in Quantum Neural Network Training Landscapes," arXiv:1803.11173 (2018), the original barren-plateau result.
• Aaronson, "Read the Fine Print," Nature Physics 11, 291 (2015), the definitive caution on HHL and QML speedup claims.
• PennyLane documentation and QML demos, runnable tutorials on variational circuits, the parameter-shift rule, and quantum kernels.