Skip to content
Tech Interview Prep
Technical interview guide

Variational Algorithms & NISQ Computing

How hybrid classical-quantum loops like VQE and QAOA are designed to extract value from today's noisy, error-uncorrected hardware.

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

Scope: IBM Quantum current variational, VQE, Estimator, mitigation, runtime, and transpiler guidance; foundational VQE, QAOA, and barren-plateau literature reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-11A team implements QAOA for a 12-node MaxCut instance with p = 3 layers, optimizing the 2p angles with COBYLA. A reviewer proposes replacing the mixer in every layer with a second copy of the cost unitary, arguing that doubling the problem term should sharpen the solution. Which description of what QAOA alternates is correct?
MCQ-13A team computes analytic gradients of a 12-qubit variational circuit on a superconducting QPU. Each trainable gate is a Pauli rotation exp(−iθP/2). What does the parameter-shift rule require to get ∂⟨H⟩/∂θ for one such parameter?
MCQ-7A team estimates ⟨H⟩ for a 4-qubit Hamiltonian whose Pauli terms include `X0X1`, `Z0Z1`, `Y0Y1` and `Z0Z2`, all with comparable coefficients. Shot noise dominates their error budget, so they want the fewest measurement circuits per energy evaluation. Which statement about grouping these terms is correct?
MCQ-6A 12-qubit Hamiltonian has 40 Pauli terms. Grouping qubit-wise commuting terms yields 6 groups, and a pilot run gives each group's estimator variance per shot. The team needs an unbiased estimate of ⟨H⟩ with the smallest total shot budget for a target error. Which procedure should they use?
MCQ-4A team runs VQE on a 4-qubit Hamiltonian H = Σ_{i=1}^{15} c_i P_i. Nine of the Pauli terms are diagonal in the Z basis; six involve X or Y. They define the cost function as the expectation value of H but measure only the nine Z-basis terms, arguing that the ansatz's final layer consists of Z-rotations. The optimizer converges to -0.90 Ha; the exact ground state energy is -1.20 Ha. Which statement explains the 0.30 Ha gap?
MCQ-5A team is debugging a VQE loop for the 2-qubit Hamiltonian H = Z0Z1. Each energy evaluation calls Estimator with shots=1024, and the optimizer computes its gradient by central finite differences with step h = 0.01. Ten runs from the same initial parameters converge to final energies spread over about 0.3. What is the most likely cause of the spread?
MCQ-10A team tries to reproduce a published VQE result for a 4-qubit H2 Hamiltonian. They match the paper's Hamiltonian, hardware-efficient ansatz (2 layers), initial point (seed 0), SLSQP optimizer, and stopping rule (gradient norm < 1e-6). Their final energy is -1.087 Ha; the paper reports -1.137 Ha. The paper's methods section specifies 4096 shots per Pauli term; the team's estimator was configured for 100 shots per term. What is the most likely explanation for the 0.05 Ha gap?
MCQ-8A team runs a 4-qubit VQE on a superconducting device. The ansatz has 12 parameters, optimized once and frozen. Each evaluation prepares the same circuit and estimates ⟨H⟩ by measuring 8192 shots per Pauli term. Across 20 repeats with identical parameters and calibration, the reported energy varies by ±0.03 Ha, while the true energy should be fixed. The optimizer sometimes stops at different parameter values. What is the most likely cause and appropriate response?
MCQ-9A VQE run for H2 at 0.735 Å uses a 4-qubit Hamiltonian whose exact ground-state energy is -1.137 Hartree. With a normalized UCCSD ansatz, 10^5 shots per Pauli term, and readout-error mitigation, the optimizer reports -1.142 Hartree. The Hamiltonian and ansatz are unchanged. What does this result most likely indicate?
MCQ-1A team runs VQE on a 6-qubit transverse-field Ising Hamiltonian with a 24-parameter hardware-efficient ansatz. Each energy evaluation uses 4096 shots and shows a run-to-run spread of σ_E ≈ 0.02 Ha; doubling to 8192 shots halves the spread, as expected for shot noise. Gradients are estimated by central finite differences with ε = 0.001 rad, and COBYLA stalls after ~30 iterations at an energy 0.15 Ha above the exact ground state. A noiseless statevector evaluation at the stalled parameters gives |∂E/∂θ_i| ≈ 0.02 Ha/rad for several components. What should the team change?
MCQ-2In a hypothetical 12-qubit VQE benchmark, two ansätze use the same optimizer, 1024 shots, and 200 iterations. Ansatz A has 24 parameters and reaches 2 mHa error on the noiseless simulator but 40 mHa on hardware. Ansatz B has 96 parameters and reaches 0.1 mHa on the noiseless simulator but 60 mHa on hardware. Both ansätze use the same two-qubit gate count per layer, but B has four times as many layers. Which conclusion about the ansatz is best supported?
MCQ-3A team benchmarks 1-layer QAOA on 40 MaxCut instances. Random initialization reaches a mean approximation ratio of 0.76 in a median of 200 optimizer iterations. Warm-starting the mixer from the rounded Goemans–Williamson solution reaches 0.81 in a median of 80 iterations on the same 40 instances. Split by the GW ratio: on the 25 instances with GW ≥ 0.85 the warm start wins (0.88 vs 0.78), and on the 15 instances with GW < 0.75 random init wins (0.72 vs 0.68). What does this pattern indicate?
MCQ-25A team running a variational quantum algorithm on a noisy intermediate-scale quantum (NISQ) processor observes systematic estimation bias caused by coherent gate errors (such as calibration drift and cross-talk unitary rotations). Because physical qubit availability is strictly bounded by hardware topology, the pipeline cannot allocate additional ancilla qubits. Which technique mitigates these coherent errors into stochastic Pauli channels without adding ancillas?
MCQ-17A 6-qubit VQA with a 24-parameter hardware-efficient ansatz is trained on a noisy superconducting device, and every energy evaluation uses 4096 shots. Over 10 seeds the unmitigated run converges to 0.021 Ha above the exact-diagonalization ground state, with a run-to-run spread of 0.003 Ha. The team then adds zero-noise extrapolation with scale factors {1, 1.5, 2}, keeping 4096 shots per scale factor, so each evaluation now executes three circuits. The mitigated runs land 0.006 Ha above exact, but the spread across the same 10 seeds rises to 0.014 Ha, and 3 of the 10 seeds stall at parameters whose mitigated energy is worse than the unmitigated one. Which conclusion do these measurements support?
MCQ-16A team runs a 10-qubit VQE with a hardware-efficient ansatz on a superconducting device. At 4 layers, the median parameter-shift gradient magnitude is 1.2e-2 and the optimizer reaches a cost within 3% of the noiseless simulation. At 12 layers, the median gradient magnitude falls to 4e-5, and the optimizer stalls at a cost 40% above the noiseless value. A density-matrix simulation using the device's daily calibration noise model reproduces both the gradient collapse and the stalled cost. Which conclusion does this evidence support?
MCQ-24A team trains an n-qubit alternating layered ansatz and measures the variance of the cost-function gradient at initialization. They replace a global observable with a sum of few-body local observables, leaving the ansatz and its parameter initialization unchanged. How does the variance's scaling with n change as the circuit depth grows from O(log n) to poly(n)?
MCQ-14A team trains a 20-qubit variational quantum circuit with a hardware-efficient ansatz and a global cost function. They measure gradient variance across system sizes: 4 qubits ~1e-2, 8 ~1e-3, 12 ~1e-4, 16 ~1e-5, 20 ~1e-6. With Adam (lr=0.01) and 10,000 shots per parameter, after 500 iterations the cost is nearly flat and gradient norms are ~1e-6. What is the most defensible next step?
MCQ-15A team is training a 24-qubit variational circuit for a chemistry ground-state problem. With a hardware-efficient ansatz of 12 layers and random initialization, the gradient norm of the first layer's parameters falls below 1e-6 by layer 8 and training stalls. They have a fixed shot budget and cannot increase circuit depth. Which change is most likely to restore a usable gradient signal?
MCQ-18An 8-CNOT, 4-parameter VQE ansatz is authored on a linear chain 0-1-2-3-4-5 and budgeted against a 6-qubit device whose coupling map has edges (0,1), (1,2), (3,4), (4,5) and (1,4) but no (2,3), and whose only native two-qubit gate is iSWAP at 0.8% error. The budget, built from the authored circuit, predicts a two-qubit-limited fidelity of 0.94 from 0.992^8. The transpiler reports 26 two-qubit gates and a 3.4 us duration for the layout it chose. Running the transpiled circuit with all parameters set to 0, so the ideal output is one known basis state, returns that state with probability 0.81 at 2000 shots. T2 is 90 us, and the ansatz's entangling pattern must stay fixed for comparison with last quarter's run. Which change makes the budget predict such runs?
MCQ-19A team runs a 200-iteration variational algorithm on a superconducting device. Each iteration submits one circuit with 4,000 shots; QPU execution is 30 ms, the classical optimizer step takes 0.2 s, and the job queue adds 12 s per submission. The device's two-qubit gate error drifts measurably over a 10-minute window. What does this timing imply for the run?
MCQ-23A team runs a 12-qubit VQE for a chemistry Hamiltonian. After 200 iterations the optimizer stops with gradient norm 1e-5 and estimated energy -7.842 Ha, matching a classical mean-field estimate to 0.01 Ha. The decoded state has not been checked against the Hamiltonian's particle-number or spin symmetries. What should they do before reporting -7.842 Ha as the ground-state energy?
MCQ-22In a hypothetical benchmark, a team claims quantum advantage on a 100-asset portfolio optimization problem. Their D-Wave annealer returns a portfolio with Sharpe ratio 1.8; the QPU time is 30 seconds, but embedding and calibration took 2 hours. They compare this to a tuned classical branch-and-bound solver that returns Sharpe ratio 1.8 in 10 minutes total, including all preprocessing. The team concludes the quantum device is superior because 30 seconds is less than 10 minutes. Which action best tests whether the claimed advantage is fair?
MCQ-21A team tunes a 12-parameter simulation-optimization model. The objective is evaluated by a stochastic simulator: two runs of the same parameter vector differ by roughly ±3% (1 s.d.) in the reported score. They run 40 random restarts of a local optimizer, each with a different seed, and pick the parameter vector with the best single reported score. On a held-out set of 200 fresh simulator runs, that vector's mean score is 4% worse than the score it reported during tuning, while the average of all 40 restart endpoints scores 1% better than the selected vector on the same held-out set. What does this gap indicate?
MCQ-20A team tunes an XGBoost model with 5-fold cross-validation and a Bayesian search. Before starting, they fixed the stopping rule: stop once no trial has beaten the incumbent validation logloss by more than two standard errors of the 5-fold mean for 20 consecutive trials. At trial 125 the incumbent is 0.4210, set at trial 100. The best score since is 0.4190 at trial 112. The fold-to-fold standard deviation of logloss is 0.0090. What should the stopping rule do?
MCQ-12A team tunes 30 hyperparameters of a training pipeline. Each evaluation costs 25 minutes of GPU time and returns validation accuracy. Re-running the same configuration with different seeds gives a standard deviation of 0.9 accuracy points. The best configuration found so far beats the runner-up by 0.4 points. Their budget is 300 evaluations. A TPE-based Bayesian optimizer that won on their previous 8-dimensional, low-noise benchmark now performs no better than random search. What is the most likely cause and the right next step?