Variational and Hybrid Quantum Computing

Introduction

Most practical quantum algorithms are not executed entirely on a quantum processor. Instead, they combine a parameterized quantum circuit with a classical computer. The quantum processor prepares states and estimates quantities by measurement; the classical processor updates circuit parameters, evaluates progress, and decides when to stop. This arrangement is called hybrid quantum computing.

Variational methods are especially important for near-term quantum devices because they can use relatively shallow circuits and tolerate some hardware imperfections. Two central examples are the variational quantum eigensolver (VQE), commonly used for estimating ground-state energies, and the quantum approximate optimization algorithm (QAOA), used for combinatorial optimization.

These methods are powerful but not automatically superior to classical algorithms. Their performance depends on circuit design, parameter initialization, measurement accuracy, optimizer behavior, hardware connectivity, noise, and problem structure.

Learning Objectives

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

• explain parameterized quantum circuits and expectation-value estimation;
• distinguish cost functions, objective functions, gradients, and parameters;
• describe the quantum–classical optimization loop;
• compare common classical optimizers used with quantum circuits;
• construct introductory VQE and QAOA models;
• explain sampling error, shot noise, barren plateaus, and other limitations;
• select sensible practices for small hybrid quantum experiments.

1. The Variational Principle

A variational algorithm searches over a family of quantum states rather than directly solving a problem by a fixed sequence of gates. Let a parameterized circuit be represented by a unitary operator U(θ), where θ=(θ₁,θ₂,...,θ_p) is a vector of adjustable real parameters. Starting from a reference state |ψ₀⟩, the circuit prepares

For a Hermitian observable or Hamiltonian H, the objective is often its expectation value:

The expectation value is real because H is Hermitian. In quantum chemistry, H is a molecular Hamiltonian and the goal is to minimize C(θ). The variational principle states that for any normalized trial state,

where E₀ is the ground-state energy, the smallest eigenvalue of H. Thus, minimizing the measured expectation gives an upper bound on the true ground-state energy. This bound is meaningful only when the state is normalized and the Hamiltonian is correctly represented.

The variational principle does not guarantee that a chosen circuit can represent the ground state, nor that a classical optimizer will find the global minimum. It supplies an objective and a bound, not an automatic solution.

Parameterized gates and ansätze

A parameterized gate is a gate whose rotation angle is controlled by a classical parameter. For example,

The matrix form applies to a single qubit in the computational basis. A circuit may contain several such rotations and fixed entangling gates. The complete parameterized circuit is called an ansatz. An effective ansatz should balance expressiveness, depth, trainability, hardware connectivity, and physical constraints.

Two broad ansatz styles are common:

A reference state can also encode useful information. For example, a computational-basis state representing a feasible classical solution may be preferable to |0...0⟩, provided that the encoding is correct.

2. Cost Functions and Measurement

The classical optimizer does not see a quantum state directly. It receives a numerical cost, usually estimated from repeated measurements. If the Hamiltonian can be decomposed as

where c_j are real coefficients and P_j are Pauli strings such as Z₁Z₂ or X₁, then

Each Pauli expectation is estimated by preparing the state repeatedly, measuring in a suitable basis, and calculating the average of the outcomes. For a Pauli observable, each shot produces an outcome of +1 or −1. If N_+ is the number of +1 results and N_- is the number of −1 results among N=N_++N_- shots, then

The hat indicates an estimate. Under independent identical measurements, the standard error of this mean is approximately

The exact unknown expectation appears in this expression; in practice it is replaced by the measured estimate. The formula assumes independent shots and outcomes restricted to ±1. Increasing the number of shots generally reduces standard error in proportion to 1/√N, but it increases execution cost.

For a sum of Pauli terms, measurement errors propagate. If estimates are approximately independent, a useful approximation is

Covariances may exist when measurements are grouped or processed jointly, so this expression is not universally exact. It is nevertheless useful for understanding why large coefficients and poorly estimated terms can dominate uncertainty.

Cost-function design

A cost function should reflect the actual task. Examples include:

• energy expectation for VQE;
• a penalty-augmented objective for constrained optimization;
• an approximation ratio for a combinatorial problem;
• a classification loss for a variational quantum classifier.

For a constraint g(x)=0, a classical penalty formulation may use

where f(x) is the original objective and λ>0 controls the penalty. The units and scale of λ should be chosen consistently with f. Excessively large penalties can make optimization numerically difficult, while weak penalties may produce infeasible solutions.

3. The Quantum–Classical Hybrid Loop

A generic variational workflow is:

1. choose an ansatz, reference state, and initial parameter vector;
2. bind parameters to the circuit;
3. execute the circuit repeatedly on a simulator or quantum processor;
4. estimate the cost from measurement results;
5. use a classical optimizer to propose new parameters;
6. repeat until a stopping condition is met;
7. verify the final state and report uncertainty.

The optimizer may use the update rule

where t is the iteration number, η_t>0 is a learning rate, and ∇C(θ_t) is the gradient. This rule describes gradient descent; many practical optimizers use more sophisticated updates or do not require gradients.

A stopping condition might be a small change in cost,

or a small parameter change, a maximum number of evaluations, or an application-specific accuracy target. A small change does not prove global optimality: the algorithm may have stalled at a local minimum, saddle point, noise floor, or barren plateau.

Classical optimizers

No optimizer is best for every hardware and cost landscape. Gradient-free methods may be attractive when measurement noise is substantial, whereas gradient methods may be efficient when reliable derivatives are available.

4. Gradients and Parameter-Shift Evaluation

For many rotation gates, a derivative can be estimated without numerically differentiating the entire circuit. If a parameter enters through a generator with two eigenvalues ±1/2, the parameter-shift rule takes the form

Here, θ_k±π/2 means that only parameter θ_k is shifted while the other parameters remain fixed. The rule applies under the stated generator convention; other generator eigenvalue spacings require a modified shift and coefficient. It also assumes the measured quantity has the appropriate differentiable gate dependence.

Parameter-shift gradients are exact in the ideal, infinite-shot setting, but finite sampling makes each cost estimate noisy. For p parameters, a straightforward gradient evaluation may require roughly 2p additional circuit evaluations per iteration. Grouping measurements, analytic simulation, stochastic gradients, and adaptive shot allocation can reduce this burden.

5. Variational Quantum Eigensolver

VQE estimates the lowest eigenvalue of a Hamiltonian. Its components are:

• a Hamiltonian mapped to qubit operators;
• a reference state;
• a parameterized ansatz;
• a measurement procedure for Pauli terms;
• a classical minimizer.

The VQE objective is

and the reported estimate is the smallest observed or final optimized value, subject to statistical and hardware uncertainty.

Worked example: one-qubit VQE

Consider the one-qubit Hamiltonian

whose eigenvalues are +1 and −1. Use the reference state |0⟩ and ansatz R_y(θ)|0⟩. The resulting state is

Since Z|0⟩=|0⟩ and Z|1⟩=−|1⟩,

The minimum is E_min=−1, reached when θ=π modulo 2π. The corresponding state is |1⟩ up to a physically irrelevant global phase. In an experiment, measuring Z gives +1 for a 0 and −1 for a 1. With 800 shots and 600 outcomes equal to −1, the estimate is

The exact value at this parameter may differ because of finite sampling. This small example demonstrates the VQE loop, but it does not represent the difficulty of molecular Hamiltonians, which contain many Pauli terms and require careful encoding and error control.

6. Quantum Approximate Optimization Algorithm

QAOA addresses discrete optimization problems by alternating two types of evolution. A classical objective is converted into a cost Hamiltonian H_C, whose computational-basis eigenvalues represent solution costs. A mixer Hamiltonian H_M moves amplitude among candidate states.

For depth p, a common QAOA state is

where γ=(γ₁,...,γ_p) and β=(β₁,...,β_p) are classical parameters. The objective is typically

For a maximization problem, the sign or optimization direction must be chosen consistently. Measuring the final state produces candidate bit strings; the best feasible string found may be more important than the expectation value alone.

Illustrative two-variable model

Suppose the goal is to maximize whether two binary variables disagree. Encode bits by computational basis states and choose the cost Hamiltonian

The eigenvalue is 0 when the bits are equal and 1 when they differ. Therefore the maximizing strings are 01 and 10. A QAOA experiment prepares |+⟩|+⟩, applies alternating cost and mixer operations, measures the result, and uses a classical optimizer to adjust γ and β. This is a valid model of the workflow, although a useful performance claim would require comparison with an exact classical optimum and repeated trials.

For graph problems such as Max-Cut, each edge contributes a term that rewards separated endpoint bits. The mapping must be checked carefully: sign conventions, edge weights, and whether the optimizer minimizes or maximizes all affect the interpretation.

7. Barren Plateaus and Optimization Limitations

A barren plateau is an optimization region in which gradients become extremely small, often decreasing rapidly with circuit size or qubit number. In a simplified statistical description, if

as the system grows, gradient information becomes difficult to distinguish from sampling and hardware noise. The exact scaling depends on the ansatz, cost function, circuit depth, connectivity, initialization, and problem structure. Barren plateaus are not the only source of difficulty.

Important limitations include:

• Local minima and saddle points: the optimizer may stop away from the best solution.
• Shot noise: finite measurements make the cost fluctuate.
• Device noise: gate errors, readout errors, decoherence, and crosstalk distort estimates.
• Expressibility–trainability trade-off: highly expressive circuits may be difficult to optimize.
• Parameter concentration: random initialization can place parameters in poor regions.
• Evaluation cost: each optimizer step may require many circuit executions.
• Constraint violations: QAOA or penalty methods may produce infeasible strings.
• Ansatz bias: a circuit that cannot represent a target state cannot reach the ideal value.
• Classical overhead: optimization and post-processing may dominate total runtime.

A low measured energy is not by itself proof of a correct physical answer. Results should be compared with exact diagonalization for small systems, classical baselines, symmetry checks, repeated runs, and uncertainty estimates.

8. Practical Design and Best Practices

A disciplined hybrid experiment should:

1. Begin with a classical reference. Solve small instances exactly where possible.
2. Use the smallest adequate ansatz. Excessive depth increases noise and optimization difficulty.
3. Respect symmetries and constraints. Symmetry-preserving circuits can reduce the search space.
4. Use informed initialization. Classical approximate solutions or problem-specific states may help.
5. Separate training from evaluation. Reserve independent shots or parameter settings for final reporting.
6. Track uncertainty. Report shot counts, repeated-run variation, and mitigation procedures.
7. Calibrate and validate measurements. Readout correction should be tested rather than assumed.
8. Choose optimizers experimentally. Compare gradient-based and gradient-free methods under the same evaluation budget.
9. Use adaptive resources. More shots may be allocated near convergence or to high-impact terms.
10. Report failed runs honestly. A reproducible negative result is more informative than selective reporting.

Error mitigation techniques, such as measurement-error mitigation, zero-noise extrapolation, and symmetry verification, can improve estimates but do not eliminate physical noise or guarantee unbiased results. Their computational and statistical costs must be included in an evaluation.

9. Applied Case: Illustrative Portfolio-Selection Experiment

Background

A small investment team wants to select two assets from four candidates under a simplified risk–return model. The problem is illustrative rather than financial advice. Binary variable x_i∈{0,1} indicates whether asset i is selected. A classical quadratic objective represents expected benefit, pairwise risk, and a penalty for selecting the wrong number of assets.

Problem

The team wants a quantum prototype but must determine whether it offers value beyond an exact classical solver. The instance has only four binary variables, so classical enumeration is feasible.

Analysis

The quadratic objective is converted into a cost Hamiltonian using the standard mapping of a binary variable to a qubit operator. The prototype uses a shallow QAOA circuit with a mixer and a cost Hamiltonian. The classical optimizer receives an estimated expectation value from sampled measurements. Because the desired portfolio size is a constraint, the cost includes a penalty term. The penalty is tested at several scales so that infeasible solutions do not dominate merely because of poor calibration.

Approach and solution

The team first enumerates all 16 bit strings and records the exact best feasible portfolio. It then runs QAOA from several initial parameter settings, with identical shot budgets and a fixed evaluation limit. For each run it records the final expectation, best feasible bit string, number of circuit evaluations, and variation across repetitions. A simulator with noise is used before hardware execution. On hardware, readout calibration and symmetry or feasibility checks are applied where appropriate.

Outcome

The prototype finds the classical optimum in some runs but not consistently. Its expectation value is affected by sampling and readout noise, and the classical exact method is much faster for this four-variable instance. The prototype is therefore useful as an educational and engineering experiment, not as evidence of quantum advantage.

Learning points

The case demonstrates that a hybrid algorithm must be judged against a relevant baseline, not only by whether it produces a plausible bit string. Constraint encoding, shot budgets, initialization, optimizer choice, and reproducibility all affect the conclusion. Larger problem sizes may change the balance, but they also introduce greater circuit and optimization challenges.

10. Contemporary Directions and Professional Relevance

Current research explores problem-informed ansätze, warm-start QAOA, adaptive circuit growth, error mitigation, quantum natural-gradient methods, and application-specific measurement strategies. These directions aim to improve trainability or reduce resource requirements, but they remain active research topics rather than universally established solutions.

For practitioners, the most transferable skills are not memorizing one optimizer. They include translating a domain problem into a valid Hamiltonian, checking conventions, designing controlled experiments, estimating uncertainty, interpreting optimization traces, and communicating limitations. Governance is also important: claims of quantum advantage should specify the hardware, classical baseline, data-processing cost, accuracy target, and total resource measurement. In finance, chemistry, logistics, and healthcare, domain validation and responsible use are required in addition to algorithmic performance.

Chapter Summary

Variational and hybrid quantum computing combine parameterized quantum circuits with classical optimization. A circuit prepares |ψ(θ)⟩, measurements estimate a cost such as ⟨ψ(θ)|H|ψ(θ)⟩, and a classical optimizer updates parameters. Pauli decomposition makes Hamiltonian expectations measurable, while finite shots introduce statistical uncertainty that decreases approximately as 1/√N.

VQE minimizes a Hamiltonian expectation to obtain a ground-state-energy upper bound. QAOA alternates cost and mixer evolutions to search for good solutions to discrete optimization problems. Both methods depend strongly on ansatz design, initialization, optimizer choice, measurement allocation, and hardware quality. Barren plateaus, local minima, sampling noise, device noise, constraint violations, and classical overhead limit performance. Reliable practice requires classical baselines, small validated experiments, uncertainty reporting, fair resource accounting, and cautious claims about advantage.

Key Terms

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

Ansatz: A chosen parameterized family of quantum states.

Hybrid quantum computing: Computing that alternates quantum processing with classical processing.

Cost function: A numerical quantity optimized by the classical controller.

Expectation value: The probability-weighted average outcome of an observable measurement.

Hamiltonian: A Hermitian operator representing energy or a problem objective.

Pauli string: A tensor product of single-qubit Pauli operators and identity operators.

Shot: One repeated execution and measurement of a quantum circuit.

Shot noise: Statistical fluctuation caused by a finite number of measurements.

Optimizer: A classical procedure that proposes new parameter values.

Gradient: The vector of partial derivatives of a cost with respect to parameters.

Parameter-shift rule: A method for estimating certain circuit derivatives using shifted evaluations.

VQE: Variational quantum eigensolver for estimating a Hamiltonian’s lowest eigenvalue.

QAOA: Quantum approximate optimization algorithm using alternating cost and mixer operations.

Mixer Hamiltonian: An operator that moves amplitude among candidate computational states.

Barren plateau: A region with gradients too small to guide practical optimization.

Ansatz expressibility: The range of states a parameterized circuit can represent.

Error mitigation: Techniques intended to reduce the effect of noise in estimates.

Approximation ratio: A measure comparing an obtained optimization value with a reference optimum.

Quantum advantage: A demonstrated performance benefit over an appropriate classical baseline.

Review Questions

5 Short Answer Questions

1. What is the role of the classical optimizer in a variational algorithm?
2. Why is a Hamiltonian often decomposed into Pauli strings?
3. What does the variational principle imply about a trial-state energy?
4. Distinguish a shot from a circuit parameter.
5. What is a barren plateau?

5 Long Answer Questions

1. Explain the complete quantum–classical loop, including initialization, measurement, optimization, and stopping criteria.
2. Describe VQE and discuss how ansatz choice and measurement uncertainty affect its result.
3. Explain the construction and purpose of the QAOA cost and mixer Hamiltonians.
4. Compare gradient-based and gradient-free optimizers for noisy quantum hardware.
5. Discuss why a low measured cost does not by itself establish quantum advantage.

3 Analytical / Application Questions

1. Design a two-qubit VQE experiment for a Hamiltonian containing Z₁ and Z₁Z₂. Specify the ansatz, measurements, optimizer inputs, and validation tests.
2. A QAOA experiment produces feasible solutions inconsistently across repeated runs. Analyze at least four possible causes and propose a controlled diagnostic plan.
3. For a Pauli expectation estimated from 900 shots, explain how increasing the shot count affects statistical uncertainty and why it may not remove systematic hardware error.

Multiple Choice Questions with Answers

Question 1

What does VQE minimize?

A. The number of qubits

B. A Hamiltonian expectation value

C. The circuit’s clock frequency

D. The number of classical variables

Correct Answer: B. A Hamiltonian expectation value

Question 2

For independent ±1 measurement outcomes, how does standard error generally scale with the number of shots N?

A. N

B. N²

C. 1/√N

D. √N

Correct Answer: C. 1/√N

Question 3

Which component proposes new values of variational parameters?

A. The classical optimizer

B. The measurement basis alone

C. The readout detector alone

D. The reference Hamiltonian

Correct Answer: A. The classical optimizer

Question 4

In QAOA, what is the main role of the mixer Hamiltonian?

A. To store measurement results

B. To move amplitude among candidate states

C. To remove all sampling noise

D. To calculate classical derivatives only

Correct Answer: B. To move amplitude among candidate states

Question 5

If a trial state is normalized, its expectation value of a Hamiltonian is generally what relative to the ground-state energy?

A. Always lower

B. Equal only for every trial state

C. Greater than or equal to it

D. Unrelated to it by definition

Correct Answer: C. Greater than or equal to it

Question 6

What is a barren plateau?

A. A circuit with no measurements

B. A region where gradients become extremely small

C. A classical optimizer with too many variables

D. A Hamiltonian with only identity terms

Correct Answer: B. A region where gradients become extremely small

Question 7

Which expression estimates a Pauli expectation from N+ positive and N− negative outcomes?

A. (N_++N_-)/2

B. N_+N_-

C. (N_+−N_-)/(N_++N_-)

D. N_+−N_-

Correct Answer: C. (N_+−N_-)/(N_++N_-)

Question 8

What is a principal limitation of a highly expressive ansatz?

A. It cannot prepare entanglement

B. It may be difficult to optimize

C. It requires no classical computation

D. It always gives the exact ground state

Correct Answer: B. It may be difficult to optimize

Question 9

What does the parameter-shift rule provide?

A. A way to estimate certain circuit gradients

B. A method to increase qubit coherence directly

C. A replacement for all measurements

D. A proof of global optimality

Correct Answer: A. A way to estimate certain circuit gradients

Question 10

Why should a hybrid quantum experiment use a classical baseline?

A. To guarantee a barren plateau

B. To compare performance and assess whether quantum benefit is credible

C. To eliminate the need for sampling

D. To ensure every optimizer finds a global minimum

Correct Answer: B. To compare performance and assess whether quantum benefit is credible