Skip to content
Tech Interview Prep
Technical interview guide

Hybrid Classical-Quantum Systems & Quantum Machine Learning

How a quantum subroutine plugs into a larger classical pipeline, what quantum machine learning actually promises today, and how that differs from the hype.

Read
55 min
Practice MCQs
25
Interview Q&A
25
Edition
v2
Editorial status
Reviewed

Scope: IBM Quantum current hybrid HPC and QML guidance; PennyLane 0.45; classical-perspective and dequantization literature reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-14A binary variational quantum classifier uses 4 qubits: each of the 4 input features is angle-encoded as RX(x_i), then a hardware-efficient ansatz with 3 layers of 8 trainable rotation gates (24 trainable circuit angles), then ⟨Z⟩ measured on each qubit, then a classical linear layer (4 weights + 1 bias) with a sigmoid output and binary cross-entropy loss. The team computes circuit gradients with the parameter-shift rule and classical-layer gradients by backpropagation. Which parameters does the parameter-shift rule differentiate to train this classifier?
MCQ-9A team must encode 16 binary features into a quantum state for a variational classifier. They need each feature's value readable directly from one qubit's measurement, with no decoding step. Which encoding fits, and at what width?
MCQ-11A team builds a one-qubit quantum kernel classifier. Their feature map is U(x)=RY(2x), so |φ(x)⟩=cos x|0⟩+sin x|1⟩. For the training pair x_i=0 and x_j=π/4, they run the compute-uncompute circuit: prepare |φ(x_i)⟩, apply U(x_j)†, and measure the probability of outcome |0⟩ with 4096 shots. What value should they expect for the kernel entry K_ij?
MCQ-12A team fits a kernel ridge regression on 50,000 samples. The library materializes the full dense kernel matrix as a float64 array. Which statement about the matrix's storage is correct?
MCQ-7A team encodes 12 tabular features, each already scaled into [-π, π], with RY(x_i) on 12 qubits, adds no entangling gates, and reads out only qubit 0. Training converges, but held-out accuracy is 0.62 against 0.88 for logistic regression on the same 12 features. What is the most likely cause?
MCQ-6A team encodes four features into two qubits with RX(x_i) rotations applied once, then appends 20 variational layers. Training loss plateaus at 0.18; with 1,024 shots per evaluation the loss estimate is stable to ±0.01, and the target function's Fourier spectrum has a component at frequency 3. Which change addresses the plateau?
MCQ-8A team amplitude-encodes a dense classical vector of 2^20 amplitudes into a 20-qubit register. They use the standard recursive state-preparation scheme, which applies one controlled rotation per amplitude, and budget 10^6 elementary gates for the preparation circuit. Which outcome should they expect?
MCQ-21A team claims a quantum machine-learning pipeline is exponentially faster than a classical one on 10^6 classical samples of 100 features. The pipeline loads the data by amplitude-encoding each sample with a generic state-preparation circuit. Which statement about the cost of this loading step is correct?
MCQ-22A hybrid pipeline estimates each entry of a quantum kernel matrix as the fraction of '0' outcomes from N = 100 shots, then applies the transform g(k) = exp(−γ(1−k)) with γ = 5 to every entry before passing the matrix to a classical SVM. For a pair whose true kernel value is k = 0.9, what is the approximate standard deviation of the transformed feature g(k̂) that shot noise induces?
MCQ-13A team builds a 400×400 quantum kernel Gram matrix for an SVM. Each entry K_ij is estimated independently from 1,000 shots of the overlap circuit. The precomputed-kernel SVM rejects the raw matrix: it is asymmetric (max |K_ij − K_ji| ≈ 0.05) and its smallest eigenvalue is −0.03. Re-running the same circuits with the same shot count produces a different matrix with the same magnitude of asymmetry. What should the team do?
MCQ-16A team trains a 20-qubit hardware-efficient ansatz with random initialization, using a cost function that is the expectation of a single Pauli string acting on all 20 qubits. Using the same ansatz family and initialization scheme, they measure the variance of one parameter's gradient over 500 random initializations: 1.0e-3 at 8 qubits and 6.3e-5 at 12 qubits. Extrapolating from these two measurements, what gradient variance should they expect at 20 qubits?
MCQ-10A team trains a 24-qubit, 12-layer variational circuit on 500 labelled samples of a binary classification task. Train loss falls to 0.01, test accuracy is 51%, the variance of the parameter-shift gradients across 20 random initializations is about 1e-6 for every layer past the fourth, and an RBF-kernel SVM on the same features scores 88% test accuracy. Which conclusion does the evidence best support?
MCQ-15A 12-parameter variational ansatz is ported from a statevector simulator to a superconducting device whose sampler accepts only a transpiled circuit and returns shot histograms, so the framework has no autodiff through the device. One gradient step now costs 24 circuit evaluations, and a 20-parameter variant of the same ansatz costs 40; the 24 holds when the shot budget rises from 4096 to 16384, when the minibatch drops from two training samples to one, and on a rerun with error mitigation disabled. Which mechanism produces those evaluations?
MCQ-1A team profiles one gradient step of a 20-parameter VQE. The parameter-shift rule needs two circuit evaluations per parameter, so a step runs 40 circuits. Their measured profile for the step is: classical optimizer update 1 ms, transpiling the 40 shifted circuits 400 ms, submitting the 40 circuits as separate jobs and retrieving results 300 ms, QPU execution 40 ms, and classical post-processing of the returned counts 20 ms. They want the single change that gives the largest reduction in wall time per gradient step. Which change should they make?
MCQ-2A team runs a 200-iteration VQE loop against one fixed 27-qubit backend. Every iteration re-transpiles the parameterized ansatz from scratch, 12 s, then executes it: 0.3 s at 4096 shots. The ansatz structure and the backend's coupling map are identical across iterations; only the 12 rotation parameters change. The whole run has to finish inside a 10-minute wall-clock window. Which change gets it there?
MCQ-25A team trains a variational algorithm (VQE/QAOA) with 50 variational parameters using gradient descent via parameter-shift rules on remote quantum hardware. Cloud execution incurs a 3-minute queue wait and a 1.5-second network round-trip per submitted job batch, while each circuit run takes 10 ms. How should the team architecture the hybrid optimization loop to minimize overall wall-clock convergence time?
MCQ-23A team must choose between two hybrid QML pipelines for the same classification task, both run on the same 8-qubit QPU with 800 shots per circuit and a 240 ms queue wait per submitted job. Pipeline A (variational classifier) runs 150 optimizer iterations, each submitting one job of 6 circuits; gate time is 1.2 ms per circuit, the classical optimizer step is 20 ms per iteration, and compilation is a one-off 300 ms. Pipeline B (quantum kernel) runs 400 kernel evaluations, each submitting one job of 1 circuit; gate time is 0.9 ms per circuit, the classical SVM fit is a one-off 200 ms, and compilation is a one-off 300 ms. Readout costs 1 µs per shot in both. Which cost figure and conclusion are correct?
MCQ-3An industrial robot cell has three feedback loops: a 1 kHz motor current loop with a 100 µs deadline, a 30 Hz vision grasp correction with a 10 ms deadline, and a weekly grasp-model retraining job. The team measures a cloud Kubernetes path at 20 ms p99 round trip and proposes running all three there. Which placement should they choose?
MCQ-4A hybrid workflow submits a variational job to a provider's cloud QPU: the classical optimizer sends a parameter set, the provider queues it, runs it, and returns an energy. The client process crashes after the provider has accepted the job but before the result is written to disk. On restart the client re-submits the same parameter set. The provider bills per submitted job and the QPU is shared with other tenants. What design keeps the workflow correct and avoids paying twice for that parameter set?
MCQ-5A job scheduler assigns each circuit to the QPU with the fewest queued jobs. A 20-qubit circuit whose interaction graph is complete, every pair of qubits must interact, is submitted. Two eligible devices, with hypothetical figures: | device | queued jobs | queue wait | qubits | coupling map | median 2-qubit error | 2-qubit gates after routing | | --- | --- | --- | --- | --- | --- | --- | | QPU-A | 2 | ~8 min | 27 | heavy-hex | 3×10⁻³ | ~1,100 | | QPU-B | 9 | ~36 min | 27 | all-to-all | 3×10⁻³ | ~190 | Which assignment should the scheduler make, and why?
MCQ-17A team trains a churn model on 12 months of subscriber data. Before splitting, an engineer fits a StandardScaler on all 500,000 rows and also runs a variance-threshold feature selector on the full matrix, then splits 80/20 into train and test and reports 0.94 AUC on the held-out set. In production the model scores 0.71 AUC. Which change most directly removes the cause of that gap?
MCQ-18A team tunes a quantum kernel classifier on 1,200 labelled samples. They evaluate 8 feature-map configurations (ZZFeatureMap with reps 1–4 crossed with two entanglement patterns) against 6 values of the SVM's C, run 5-fold cross-validation on the same 1,200 samples for all 48 combinations, pick the combination with the best mean fold accuracy (0.93), and report 0.93 as the expected accuracy on unseen data. A 400-sample holdout collected separately scores 0.79. What does the 0.14 gap indicate, and what should they report?
MCQ-19A team compares a 6-qubit variational classifier with a logistic-regression baseline on 12,000 rows of 200 heavily correlated sensor columns (70/30 split). The quantum branch feeds the circuit 8 components from a frozen standardize-then-PCA pipeline and scores 0.913 test AUC (five seeds, 0.905-0.918). The baseline is trained on the raw 200 columns with no scaling and scores 0.781. The slide claims the circuit adds 0.13 AUC. They then train the same logistic regression on the identical frozen 8 components: 0.908 (five seeds, 0.901-0.912). Which reading does this ablation support?
MCQ-24A team claims quantum advantage for a variational quantum classifier. On a held-out test set of 200 samples it reaches 92% accuracy; a classical gradient-boosted tree reaches 91%. The classical model was left at library defaults, the quantum model used 8 qubits and 6 hours of QPU time, and the classical model trained in 4 seconds. No confidence intervals are reported. Which conclusion does the evidence support?
MCQ-20A team claims an exponential quantum speedup for a recommendation task on a 10^6 × 10^6 user–item matrix of rank 10. Their quantum algorithm runs in O(polylog(nm)) ≈ 10^3 operations and assumes sample-and-query access to the matrix (one operation per entry query or row-index sample) plus the ability to sample from the output distribution. Their classical baseline reads the full matrix and runs in O(nm) = 10^12 operations. A reviewer asks them to justify the comparison. Which conclusion about the claimed speedup is correct?