10. Variational, Hybrid, Optimization, and Quantum-AI Methods

Introduction and context

Variational and hybrid quantum-classical methods occupy a central position in contemporary quantum computing. Their importance arises from a practical constraint: near-term quantum processors are noisy, have limited coherence, and cannot generally execute the deep fault-tolerant circuits assumed by many asymptotic quantum algorithms. Variational methods address this constraint by dividing computation between a parameterized quantum circuit and a classical optimizer. The quantum processor estimates quantities such as expectation values, probabilities, overlaps, or kernel entries; the classical computer updates circuit parameters in response.

This architecture is powerful but not automatically advantageous. A parameterized circuit may be expressive yet impossible to train, while a shallow circuit may be trainable but unable to represent the desired solution. Measurement noise, finite sampling, hardware connectivity, optimizer instability, data-loading costs, and classical simulation overhead can eliminate any practical benefit. Consequently, the doctoral-level study of these methods must move beyond descriptions of algorithms and examine trainability, computational cost, statistical validity, resource scaling, and the conditions under which a quantum component contributes meaningful value.

This chapter examines the variational quantum eigensolver (VQE), the quantum approximate optimization algorithm (QAOA), parameterized quantum circuits, quantum kernels, quantum generative models, quantum neural networks, barren plateaus, optimization landscapes, data encoding, and hybrid integration. The discussion treats these methods as research objects rather than as guaranteed routes to quantum advantage.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Explain the architecture and operating principles of variational and hybrid quantum algorithms.

2. Critically compare VQE, QAOA, quantum kernels, generative models, and quantum neural networks.

3. Analyze parameterized-circuit expressivity, trainability, optimization landscapes, and barren plateaus.

4. Evaluate data-encoding strategies and their computational and statistical consequences.

5. Design a hybrid quantum-classical workflow with appropriate validation and resource accounting.

6. Identify technical, organizational, ethical, and governance risks in quantum-AI research.

7. Assess claims of quantum advantage using meaningful baselines and reproducible evidence.

The variational paradigm

A variational algorithm defines a family of quantum states

|ψ(θ)⟩ = U(θ)|ψ₀⟩,

where |ψ₀⟩ is an initial state, U(θ) is a parameterized circuit, and θ is a vector of classical parameters. A cost function is estimated from measurements of the state. The classical optimizer then seeks

θ* = arg minθ C(θ).

For an observable H, a common objective is

C(θ) = ⟨ψ(θ)|H|ψ(θ)⟩.

If H is decomposed into measurable Pauli terms, H = Σᵢ cᵢPᵢ, the expectation is estimated by measuring each Pᵢ repeatedly. The estimated cost is therefore a statistical quantity rather than an exact value. This distinction affects optimizer behavior, stopping criteria, confidence intervals, and the interpretation of small improvements.

The basic loop contains four stages: parameter initialization, circuit execution, measurement and objective estimation, and classical parameter update. In a real device, additional stages include transpilation, qubit mapping, calibration, error mitigation, shot allocation, and data management. The total cost is consequently determined not only by the number of optimization iterations but also by circuit depth, number of observables, shots per observable, compilation overhead, and classical processing.

Variational methods are attractive because they can use shallow circuits and tolerate some hardware imperfections. Their weakness is that the optimization problem may be highly nonconvex and noisy. The variational principle also gives a bound only in particular settings. For ground-state energy estimation, an approximate state produces an energy no lower than the true ground-state energy, but this does not imply that the state has high fidelity or useful properties for another observable.

VQE: energy minimization as a hybrid workflow

The variational quantum eigensolver is designed primarily for estimating the ground-state energy of a Hamiltonian. It is widely discussed in quantum chemistry and materials science, although its use in these areas involves substantial modeling and measurement challenges.

A VQE workflow begins by mapping a physical problem to a qubit Hamiltonian. Fermionic operators may be transformed using mappings such as Jordan–Wigner or Bravyi–Kitaev. An ansatz, such as a hardware-efficient circuit, unitary coupled-cluster-inspired circuit, or problem-informed excitation circuit, prepares trial states. Measurements estimate the energy, and a classical optimizer changes θ.

The principal design trade-off is between expressivity and trainability. Chemistry-informed ansätze may encode useful structure but require more gates and connectivity. Hardware-efficient ansätze are easier to implement but can contain many irrelevant directions and may suffer from barren plateaus or symmetry violations. Symmetry-preserving ansätze, tapering techniques, active-space reductions, and initial states based on classical approximations can reduce the search burden.

VQE should be evaluated using more than final energy. Important measures include energy error, state fidelity, conservation of particle number or spin, number of circuit evaluations, total shots, wall-clock time, robustness to noise, and comparison with classical methods. A small energy error may conceal an incorrect wavefunction if the spectral gap is small or if the desired observable is not energy itself.

QAOA and structured combinatorial optimization

QAOA applies variational principles to discrete optimization. A problem is represented by a cost Hamiltonian C whose computational-basis eigenvalues correspond to objective values. A mixing Hamiltonian B generates transitions among candidate solutions. Starting from an initial state, QAOA applies alternating layers:

|ψ(γ,β)⟩ = ∏ₚ U_B(βₚ)U_C(γₚ)|+⟩ⁿ.

The parameters γ and β are optimized to maximize expected cost, minimize a related loss, or increase the probability of high-quality solutions. At depth p, the circuit has p pairs of problem and mixer operators. Greater depth can improve representational capacity, but it also increases noise, parameter dimension, and optimization difficulty.

QAOA is not simply a universal optimizer with a quantum label. Its performance depends on graph structure, encoding, mixer choice, depth, initialization, and the quality of classical baselines. Constraint-preserving mixers may be preferable to penalty terms because they restrict evolution to feasible states. Warm starts can incorporate a classical approximate solution. Parameter transfer can use solutions from related problem instances or lower-depth circuits.

For applications such as scheduling, routing, portfolio selection, and maximum cut, the objective should include feasibility, solution quality, sampling diversity, and end-to-end cost. A quantum circuit that produces a good sample but requires thousands of expensive evaluations may not outperform a classical heuristic.

Parameterized quantum circuits and optimization landscapes

Parameterized quantum circuits (PQCs) are the common computational substrate of VQE, QAOA, quantum neural networks, and many quantum machine-learning models. A PQC consists of fixed gates and trainable rotations, often arranged in repeated layers. Its behavior is determined by topology, parameter initialization, entangling pattern, gate generators, depth, and measurement observable.

Three properties must be separated:

• Expressivity concerns the range of states or functions the circuit can represent.
• Trainability concerns whether an optimizer can discover useful parameters with feasible resources.
• Generalization concerns whether performance extends beyond the training instances or data.

Increasing depth can improve expressivity while reducing trainability. Excessive entanglement may produce highly complex states whose gradients become statistically uninformative. Conversely, insufficient depth creates underfitting or approximation bias. The appropriate circuit is therefore problem-dependent rather than universally “deep” or “hardware efficient.”

Optimization landscapes can contain local minima, saddle regions, narrow valleys, flat directions, and noise-induced distortions. Gradient-based methods use parameter-shift rules or related estimators, but each gradient component may require many circuit evaluations. Gradient-free methods such as COBYLA, Nelder–Mead, SPSA, and evolutionary optimizers can be more tolerant of noise, although they may scale poorly or require careful tuning. Natural-gradient methods use an information-geometric metric, often approximated by the quantum Fisher information, to account for the geometry of the state manifold.

A rigorous experimental protocol should vary initialization, optimizer, shot budget, circuit depth, and noise model. Reporting only the best run is inadequate because stochastic optimization can produce highly variable outcomes.

Barren plateaus and trainability

A barren plateau is a region in which gradients become extremely small, often with variance that decreases rapidly as the number of qubits or circuit depth increases. In such a region, a measured gradient is dominated by shot noise and hardware error. The issue is not merely that the optimizer moves slowly; it may be statistically impossible to determine a useful direction without an impractical number of measurements.

Barren plateaus can arise from highly expressive random circuits, global cost functions, excessive depth, or noise. Local cost functions, structured ansätze, layerwise training, problem-informed initialization, symmetry restrictions, and correlated parameter initialization can mitigate—but not universally eliminate—the problem. Noise-induced barren plateaus are particularly concerning because error channels can make gradients vanish even when the noiseless circuit would be trainable.

Trainability analysis should examine gradient means and variances across system sizes, parameter distributions, observables, and noise levels. A few successful small experiments do not establish scalable trainability. Researchers should distinguish a formal asymptotic barren plateau from practical flatness caused by poor scaling, inadequate precision, or an unsuitable optimizer.

Data encoding and the quantum data-loading problem

Quantum machine learning requires a mapping from classical data x to a quantum state or circuit. Common approaches include basis encoding, angle encoding, amplitude encoding, and feature-map circuits. Basis encoding represents discrete values in computational-basis states. Angle encoding places features into rotation angles and is often hardware-friendly. Amplitude encoding can represent many values in a relatively small number of qubits, but state preparation may be expensive and can dominate the algorithm. Feature maps may encode nonlinear relationships through entangling gates and repeated rotations.

Encoding is not a neutral preprocessing step. It determines circuit depth, parameter count, noise exposure, periodicity, inductive bias, and the effective geometry of the data. If a classical dataset requires costly state preparation for every sample, any speedup claimed for the quantum model must include that cost. Feature scaling is also essential because rotation angles are periodic; unscaled inputs can create aliasing or unstable optimization.

A sound study compares several encodings and includes a classical model receiving the same information. It should report data-loading time, circuit evaluations per sample, measurement shots, preprocessing, and inference cost. The relevant question is not whether a feature map is mathematically expressive, but whether it produces a measurable improvement under a complete resource accounting.

Quantum kernels and quantum neural networks

A quantum kernel method constructs a similarity measure from quantum states. Given a feature map U(x), a kernel may be defined through state overlap, for example

K(x,z) = |⟨0|U†(z)U(x)|0⟩|².

The resulting kernel matrix can be used by a classical support-vector machine or another kernel learner. Quantum kernels avoid some parameter-training difficulties because the feature map may be fixed, but they introduce their own concerns: kernel estimation requires repeated measurements, the matrix may become nearly constant, and the classical learner may not benefit from a high-dimensional feature space if the kernel is poorly aligned with the task.

Kernel-target alignment, condition number, effective dimension, generalization performance, and robustness to finite-shot estimation are more informative than raw circuit depth. A quantum kernel should be compared with strong classical kernels, including radial-basis, polynomial, and task-specific kernels.

Quantum neural networks generally refer to trainable PQCs used as classifiers, regressors, generative models, or differentiable quantum layers. A typical classifier encodes x, applies trainable operations U(θ), measures an observable, and maps the expectation to a prediction. Hybrid neural networks may place a quantum layer between classical feature extraction and a classical output layer. This can reduce encoding demands but may also mean that the quantum layer contributes little beyond a classical nonlinear transformation.

The central research questions are representational advantage, trainability, sample efficiency, robustness, and computational efficiency. Accuracy alone is insufficient, especially on small datasets where classical models can match or exceed quantum models with far lower cost.

Quantum generative models

Quantum generative models seek to learn a probability distribution, data-generating state, or latent representation. Quantum circuits can generate samples directly from measurement outcomes. Variational quantum circuits may be trained using likelihood-inspired objectives, adversarial objectives, moment matching, or distributional distances. Quantum circuit Born machines are an important example: the probability of a bit string is determined by the squared amplitude generated by a circuit.

Generative training is challenging because likelihoods may be difficult to compute, gradients may be unstable, and sampling noise can obscure distributional differences. Adversarial formulations introduce the additional instability familiar from classical generative adversarial networks. Evaluation should use held-out likelihood proxies where available, distributional distances, coverage, mode collapse indicators, sample quality, and sampling cost. A model that reproduces a few common modes but misses rare events may be unsuitable for scientific or financial applications.

Quantum generative models are potentially relevant to chemistry, materials design, anomaly detection, and synthetic-data generation. However, generated data must not be treated as reliable merely because it originates from a quantum device. Scientific validity, privacy, bias, and downstream decision risk remain classical and institutional responsibilities.

Integrating quantum and classical computation

Hybrid integration can be organized at several levels. In a sequential architecture, a classical algorithm preprocesses data and a quantum subroutine evaluates a feature, objective, or sample distribution. In a co-processing architecture, classical and quantum components repeatedly exchange parameters and measurements. In a differentiable architecture, a quantum layer is embedded in a machine-learning pipeline and trained jointly with classical layers. In a decomposition architecture, a large problem is partitioned into quantum-evaluable subproblems.

A practical architecture should define interfaces explicitly: data formats, parameter precision, latency expectations, shot allocation, error mitigation, failure handling, and reproducibility metadata. The quantum processor should not be invoked for operations that a classical processor performs more efficiently. Batching, asynchronous execution, parameter caching, adaptive shot allocation, and surrogate modeling can reduce communication and measurement overhead.

The correct unit of analysis is the complete workflow. A hybrid method may have a theoretically interesting quantum subroutine but no practical advantage because classical optimization and data transfer dominate runtime. Conversely, a modest quantum component may be valuable if it improves a bottleneck in a high-value process, even without asymptotic speedup.

Applied case study: illustrative hybrid optimization for constrained logistics

Background

Consider a fictional logistics company, Northstar Distribution, operating a regional delivery network. It must assign vehicles to delivery clusters while respecting capacity, driver-hour, and service-window constraints. The company investigates a hybrid QAOA-based method as a research prototype, not as an assumption of production superiority.

Situation and problem

The existing system uses mixed-integer programming followed by local search. It produces feasible solutions reliably, but solution time increases during peak periods. The research team proposes a QAOA formulation for a reduced weekly planning problem. The objective combines travel distance, lateness penalties, and vehicle imbalance. Constraints are represented through a combination of feasible-state mixers and penalty terms.

Analysis

The team first establishes baselines: the incumbent optimizer, simulated annealing, tabu search, and a classical QAOA-inspired circuit simulator. It discovers that a naive binary encoding requires too many qubits and creates large penalty coefficients, which distort the optimization landscape. A reduced formulation uses domain-specific clustering, symmetry removal, and a constraint-preserving mixer. Because the quantum device supports limited connectivity, the logical graph is mapped using additional routing gates, increasing circuit depth.

The study tracks feasible-solution probability, best objective value, mean objective value over samples, number of circuit executions, total shots, compilation time, and wall-clock latency. It also evaluates performance under repeated random seeds and different initializations. A warm start derived from the classical incumbent improves feasibility, but the resulting solutions do not outperform the best classical baseline within the available budget.

Approach and outcome

The prototype demonstrates that the hybrid method can generate feasible candidate solutions and provides a useful experimental platform for studying mixers, warm starts, and adaptive shot allocation. It does not establish quantum advantage. The main bottleneck is not the theoretical QAOA layer but repeated noisy objective estimation and routing overhead. Northstar therefore retains the classical optimizer for production and uses the quantum prototype for research into subproblem selection and future hardware requirements.

Learning points

The case illustrates four principles. First, formulation quality often matters more than algorithm branding. Second, feasibility must be measured separately from objective quality. Third, a fair comparison includes all classical and quantum overheads. Fourth, a negative result can be scientifically valuable when it identifies the conditions under which a quantum approach is unlikely to succeed.

Risks, limitations, and common errors

A recurrent error is equating a flexible circuit with a useful model. Expressivity without trainability is operationally ineffective. Another error is comparing a quantum model with a weak classical baseline or reporting accuracy without inference cost. Data leakage, inadequate train-test separation, and tuning on the test set are especially dangerous in quantum machine-learning experiments because datasets are often small.

Measurement noise creates biased or high-variance objective estimates. Error mitigation can reduce bias but may increase sampling cost and can amplify statistical uncertainty. Optimizers may exploit noise, producing apparent improvements that disappear under independent evaluation. Hardware drift, calibration changes, queue latency, and compiler versions further complicate reproducibility.

The no-free-lunch perspective is also important: a quantum model is not universally superior across tasks. Many proposed demonstrations use synthetic datasets selected to favor a feature map. Such demonstrations can establish implementability but not general advantage. Claims should specify the task distribution, resource model, noise assumptions, and baseline family.

Ethical, governance, and professional considerations

Quantum-AI systems may influence credit, medical, employment, security, or logistics decisions. Their experimental novelty does not exempt them from fairness, privacy, explainability, or accountability requirements. Quantum embeddings may expose sensitive data to external hardware providers, and cloud execution creates concerns about data residency and access control. Synthetic quantum-generated data can reproduce or amplify the biases of its training data.

Governance should require documented datasets, consent and lawful-use analysis, model cards or equivalent records, audit trails for parameter and hardware changes, uncertainty reporting, and human oversight for consequential decisions. Researchers should distinguish proof-of-concept claims from validated operational performance. Intellectual-property and export-control obligations may also apply to quantum hardware, cryptographic applications, and proprietary circuit designs.

Current developments and research directions

Current research is moving toward error-aware ansätze, adaptive circuit growth, symmetry preservation, problem-specific mixers, error mitigation integrated with optimization, and resource-aware benchmarking. Layerwise learning and transfer of parameters across related instances aim to improve trainability. Quantum natural-gradient methods and gradient-free stochastic optimizers remain active areas of investigation.

Quantum kernels are increasingly studied through generalization theory, kernel alignment, and data-dependent resource analysis rather than through circuit complexity alone. In generative modeling, researchers are exploring distribution learning, quantum-assisted sampling, and hybrid latent-variable systems. Error-corrected quantum computing may eventually permit deeper and more expressive models, but fault tolerance changes the resource analysis rather than removing the need for sound modeling.

A particularly important direction is the development of standards for quantum advantage claims. Such standards should include reproducible code, fixed evaluation protocols, transparent shot and latency accounting, strong classical baselines, scaling experiments, and uncertainty estimates. The field will mature when negative and inconclusive results are reported as carefully as positive demonstrations.

Best-practice framework for research and implementation

A defensible project should proceed through the following logic. Define the decision problem and success metric before selecting a quantum algorithm. Establish a strong classical baseline and a trivial reference model. Choose an encoding and ansatz based on hardware, data, and domain structure. Pilot trainability at small scale by examining gradient statistics and optimizer sensitivity. Estimate total quantum resources, including state preparation and measurements. Use validation protocols that prevent leakage and report distributions across repeated runs. Test robustness under realistic noise and hardware conditions. Finally, determine whether the quantum component improves the complete workflow sufficiently to justify its complexity.

This framework shifts attention from isolated circuit demonstrations to evidence-based system evaluation. It also encourages researchers to treat hybrid quantum computing as an engineering discipline involving algorithms, hardware, statistics, software, governance, and domain knowledge.

Chapter Summary

Variational and hybrid quantum methods use a classical optimizer to control a parameterized quantum circuit. VQE minimizes Hamiltonian expectation values, while QAOA targets structured combinatorial objectives through alternating cost and mixer operators. PQCs must be evaluated simultaneously for expressivity, trainability, and generalization. Barren plateaus, noisy gradients, measurement cost, and hardware constraints can prevent useful optimization. Data encoding determines computational overhead and inductive bias, making the data-loading problem central to quantum-AI evaluation. Quantum kernels, quantum neural networks, and quantum generative models offer distinct approaches but require strong classical comparisons and complete resource accounting. Effective hybrid integration depends on interfaces, latency, shot allocation, error handling, and workflow-level performance. Responsible research reports uncertainty, limitations, governance implications, and negative results rather than equating novelty with advantage.

Key Terms

Parameterized quantum circuit: A quantum circuit containing adjustable classical parameters.

Variational quantum eigensolver: A hybrid algorithm for estimating low-energy states of a Hamiltonian.

Ansatz: A parameterized family of quantum states selected for a computational task.

Hamiltonian: An operator representing the energy or objective structure of a quantum system.

QAOA: A variational algorithm using alternating cost and mixer evolutions for combinatorial optimization.

Mixer Hamiltonian: An operator that creates transitions among candidate computational states.

Expressivity: The range of states or functions representable by a model.

Trainability: The practical ability to optimize a model with available computational resources.

Barren plateau: An optimization region in which gradient magnitudes or variances become extremely small.

Optimization landscape: The relationship between parameter values and the resulting objective function.

Parameter-shift rule: A method for estimating derivatives of quantum expectation values through shifted circuits.

Quantum kernel: A similarity function computed from quantum-encoded data states.

Feature map: A circuit that transforms classical input data into a quantum state.

Quantum neural network: A trainable parameterized quantum circuit used for learning tasks.

Quantum generative model: A model that learns or produces a probability distribution using quantum states or circuits.

Shot: One execution and measurement of a quantum circuit.

Error mitigation: Techniques that reduce the effect of noise without fully correcting errors.

Warm start: Initialization using information from a classical solution or related optimization instance.

Quantum advantage: Demonstrated performance superiority under a clearly specified resource and comparison model.

Review Questions

Short Answer Questions

1. What is the role of the classical optimizer in a variational quantum algorithm?

2. Distinguish expressivity from trainability in a parameterized quantum circuit.

3. Why can amplitude encoding create a data-loading problem?

4. What is a barren plateau?

5. Why must feasibility and objective quality be measured separately in QAOA applications?

Long Answer Questions

1. Explain the VQE workflow, including Hamiltonian construction, ansatz selection, measurement estimation, and classical optimization. Critique its principal resource costs.

2. Compare VQE and QAOA in terms of objective structure, circuit architecture, parameter optimization, and suitable applications.

3. Analyze how circuit depth, entanglement, initialization, and cost-function locality affect trainability.

4. Discuss quantum kernels, quantum neural networks, and quantum generative models as alternative quantum-AI paradigms.

5. Develop a rigorous experimental protocol for evaluating a claimed quantum advantage in a hybrid learning system.

Analytical / Application Questions

1. A quantum classifier achieves higher test accuracy than a classical baseline but requires 10,000 circuit evaluations per sample. What additional measurements and comparisons are needed before claiming practical advantage?

2. Design an encoding and ansatz strategy for a constrained scheduling problem where feasibility is more important than unconstrained objective minimization. Justify your choices.

3. In the Northstar Distribution case, explain why the QAOA prototype was scientifically useful despite failing to outperform the classical production solver.

22. Multiple Choice Questions with Answers

Question 1

Which statement best describes the central trade-off in selecting a parameterized quantum-circuit ansatz?

A. More expressive circuits are always easier to optimize

B. Hardware-efficient circuits eliminate measurement noise

C. Greater expressivity can be accompanied by reduced trainability

D. Shallower circuits always produce more accurate solutions

Correct Answer: C. Greater expressivity can be accompanied by reduced trainability

Question 2

In VQE, the objective function is most commonly estimated by measuring:

A. The number of classical optimizer iterations only

B. Expectation values of terms in a qubit Hamiltonian

C. The circuit compilation time

D. The dimension of the classical dataset

Correct Answer: B. Expectation values of terms in a qubit Hamiltonian

Question 3

What is a principal concern associated with a barren plateau?

A. The circuit contains no quantum gates

B. Gradients become too small to estimate useful optimization directions

C. The Hamiltonian has no eigenvalues

D. The classical optimizer becomes exact

Correct Answer: B. Gradients become too small to estimate useful optimization directions

Question 4

Which feature most directly distinguishes QAOA from a generic variational circuit?

A. It uses alternating cost and mixer evolutions

B. It never requires classical optimization

C. It can only process continuous data

D. It avoids all hardware connectivity constraints

Correct Answer: A. It uses alternating cost and mixer evolutions

Question 5

Why should state-preparation cost be included when evaluating amplitude encoding?

A. It may dominate the total cost of using the quantum model

B. It guarantees a barren plateau

C. It removes the need for measurements

D. It converts every dataset into a classical kernel

Correct Answer: A. It may dominate the total cost of using the quantum model

Question 6

A quantum kernel matrix whose entries are nearly identical is problematic primarily because it may:

A. Contain too many classical labels

B. Provide poor task-relevant discrimination

C. Guarantee perfect generalization

D. Require no circuit executions

Correct Answer: B. Provide poor task-relevant discrimination

Question 7

Which evaluation practice is most appropriate for a claim of quantum advantage?

A. Compare only with an unsophisticated classical method

B. Report the best random-seed result without resource information

C. Include strong baselines and complete quantum and classical resource accounting

D. Measure accuracy while excluding data-loading time

Correct Answer: C. Include strong baselines and complete quantum and classical resource accounting

Question 8

In the illustrative logistics case, why was a constraint-preserving mixer valuable?

A. It restricted the search to feasible states more directly

B. It eliminated all routing gates

C. It made the classical baseline unnecessary

D. It guaranteed superiority over integer programming

Correct Answer: A. It restricted the search to feasible states more directly

Question 9

Which statement about a ground-state energy estimate from VQE is most accurate?

A. An approximate state always has higher fidelity than its energy accuracy suggests

B. The variational principle can provide an upper bound on the ground-state energy under appropriate conditions

C. Energy minimization proves every other observable is accurate

D. Measurement noise cannot affect the estimate

Correct Answer: B. The variational principle can provide an upper bound on the ground-state energy under appropriate conditions

Question 10

Why is a complete hybrid workflow the proper unit of analysis?

A. Quantum circuits operate independently of classical computation

B. Classical optimization, communication, data loading, and measurement may dominate total cost

C. Hybrid systems never use hardware

D. It prevents the use of classical baselines

Correct Answer: B. Classical optimization, communication, data loading, and measurement may dominate total cost