Quantum Simulation Techniques

Introduction

Quantum simulation is the computational study of quantum circuits, states, and measurements using a classical computer. It is indispensable when developing algorithms, debugging circuits, validating hardware results, estimating resource requirements, and teaching quantum programming. A simulator does not reproduce quantum computation by physically creating the same quantum system; instead, it stores or manipulates a mathematical representation of the system.

The choice of representation strongly affects memory use, execution time, numerical fidelity, and the largest circuit that can be studied. A small circuit with general entanglement may require a full state-vector simulator, while a noisy circuit may require a density-matrix or quantum-trajectory method. Clifford circuits can often be simulated efficiently with stabilizer techniques, and circuits with limited entanglement may be handled by tensor networks. Shot-based execution is appropriate when the goal is to reproduce the statistical output of a quantum device rather than obtain exact probabilities directly.

This chapter compares these techniques and develops a practical method for selecting and evaluating a simulator.

Learning Objectives

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

1. Explain the representations used by major quantum simulation methods.
2. Distinguish exact state-vector, unitary, density-matrix, stabilizer, tensor-network, and shot-based simulation.
3. Estimate memory and computational requirements for common simulators.
4. Select a simulator according to circuit structure, noise, fidelity, runtime, and scalability.
5. Compare exact probabilities with finite-shot estimates.
6. Design and interpret a small simulation experiment.
7. Recognize numerical, statistical, and modelling limitations.

1. The Simulation Problem

An n-qubit pure state is represented by a vector in a complex Hilbert space of dimension 2ⁿ:

Here, |x⟩ denotes a computational-basis state, aₓ is its complex amplitude, and normalization requires

The probability of measuring bit string x is |aₓ|². A conventional double-precision complex number uses 16 bytes, so storing a general state vector requires approximately

This estimate excludes indexing, temporary arrays, communication buffers, and software overhead. For n = 30, the amplitudes alone require approximately 16 × 2³⁰ bytes, or 16 GiB. At n = 40, the requirement is approximately 16 TiB. Thus, exponential memory growth is the central limitation of general-purpose state-vector simulation.

A quantum circuit applies gates to the state. If the circuit contains unitary gates U₁, U₂, …, Uₖ, its total unitary is

and the final state is |ψ_f⟩ = U|ψ₀⟩. The order is significant: the rightmost operation acts first.

For a one-qubit gate G acting on qubit j, a simulator updates pairs of amplitudes whose basis labels differ only in bit j. Although a full matrix multiplication would be wasteful, optimized simulators exploit this local structure. A generic dense state-vector simulation of a circuit with G gates commonly has time approximately proportional to G × 2ⁿ, with the exact constant depending on gate type, parallelism, and implementation.

2. State-Vector Simulation

2.1 Principle

State-vector simulation stores all amplitudes of a pure quantum state. It is the most direct classical representation and supports arbitrary unitary circuits, including non-Clifford gates and highly entangled states. Measurement is usually deferred until the end or performed by sampling from the calculated probability distribution.

For a computational-basis measurement, the probability of outcome x is

For an observable A, the expected value is

provided |ψ⟩ is normalized and A is a Hermitian operator.

2.2 Strengths and limitations

State-vector simulation provides exact amplitudes up to floating-point round-off. It is suitable for:

• debugging gate order and qubit indexing;
• inspecting intermediate states;
• calculating exact probabilities and expectation values;
• testing variational circuits;
• studying arbitrary pure-state algorithms.

Its main limitation is exponential memory. It also becomes expensive when many gates act on a large state, particularly when gates create extensive entanglement.

2.3 Small example

Consider a two-qubit circuit that begins in |00⟩, applies a Hadamard gate to the first qubit, and then applies a controlled-NOT with the first qubit as control:

The exact measurement probabilities are p(00) = 1/2 and p(11) = 1/2; all other outcomes have probability zero. A state-vector simulator reveals these probabilities immediately. A shot-based simulator, by contrast, estimates them through repeated measurements.

3. Unitary and Operator Simulation

The phrase “unitary simulation” is used in two related ways. First, a simulator may apply gates as unitary transformations to a state vector. Second, it may explicitly construct the complete circuit unitary U.

An n-qubit unitary is a 2ⁿ × 2ⁿ complex matrix. Storing it densely requires approximately

This is much more demanding than storing one state vector. For this reason, practical simulators generally apply gates directly rather than form the complete matrix. Explicit unitary construction is useful for very small circuits, process verification, quantum compiling, and studying operator properties, but it is rarely scalable.

A unitary matrix satisfies

where U† is the conjugate transpose and I is the identity matrix. This condition guarantees preservation of inner products and state normalization. It does not describe noise, measurement collapse, or irreversible processes.

4. Density-Matrix Simulation

4.1 Mixed states

A density matrix represents both pure and mixed quantum states. It is written as ρ and satisfies

For a pure state |ψ⟩, ρ = |ψ⟩⟨ψ|. A classical mixture of states |ψᵢ⟩ occurring with probabilities qᵢ is represented by

where qᵢ ≥ 0 and ∑ᵢqᵢ = 1.

The probability of measuring outcome x is

and the expectation of observable A is

4.2 Evolution and noise

For a unitary operation U,

A general quantum channel can be expressed using Kraus operators Kᵢ:

The operators must satisfy the trace-preserving condition

This representation supports decoherence, relaxation, depolarization, readout models, and other noise processes.

4.3 Cost

A density matrix has 2ⁿ × 2ⁿ complex entries. Its memory requirement is approximately

Thus, density-matrix simulation becomes impractical at substantially smaller qubit counts than pure-state simulation. It is valuable when noise modelling is more important than maximum system size. Some simulators instead use quantum trajectories: each trajectory follows a state vector under stochastic noise, and many trajectories are averaged. This can reduce memory per trajectory but introduces sampling error and may require many repetitions.

5. Stabilizer Simulation

5.1 Clifford circuits

Stabilizer simulation exploits the special structure of Clifford circuits. Clifford gates map Pauli operators to Pauli operators under conjugation. Common Clifford gates include X, Y, Z, Hadamard, phase S, and controlled-NOT. Computational-basis preparation and measurement are also compatible with the stabilizer framework.

The Gottesman–Knill result shows that circuits composed only of Clifford operations, computational-basis inputs, and computational-basis measurements can be simulated efficiently on a classical computer, despite potentially generating entanglement.

A stabilizer state is specified by n independent commuting Pauli operators that leave the state unchanged. For example, |00⟩ is stabilized by Z₁ and Z₂. The Bell state (|00⟩ + |11⟩)/√2 is stabilized by Z₁Z₂ and X₁X₂.

5.2 Advantages and boundaries

Stabilizer methods can simulate very large Clifford circuits using polynomial rather than exponential resources. They are particularly useful for:

• error-correcting-code experiments;
• syndrome extraction;
• Clifford benchmarking;
• fault-tolerance research;
• large circuits dominated by stabilizer operations.

Their limitation is that arbitrary non-Clifford gates, such as a general T gate or arbitrary rotation, cannot be represented exactly by the basic stabilizer formalism. Extensions such as stabilizer-rank methods can handle some non-Clifford structure, but their cost generally grows with the number and type of non-Clifford resources.

6. Tensor-Network Simulation

6.1 Representation

A tensor network represents a quantum state or circuit as a collection of tensors connected by indices. A product state has little internal correlation and can be represented compactly. Entangling gates increase the dimensions of connecting indices, known as bond dimensions.

For a one-dimensional chain, a matrix-product state may be written schematically as

The indices iⱼ represent physical qubits, while contracted matrix indices represent correlations between neighbouring sites. The storage cost depends on bond dimension χ rather than directly on 2ⁿ. For a chain with local dimension 2, a rough storage scale is O(nχ²), although boundary conditions and tensor shapes affect the exact cost.

6.2 Entanglement and contraction

Tensor-network efficiency depends on circuit geometry, contraction order, and entanglement. Low-entanglement circuits, shallow local circuits, and one-dimensional systems are often favourable. Highly entangled random circuits can cause bond dimensions to grow rapidly, removing the advantage.

Truncation is commonly used: small singular values are discarded after a decomposition. If a normalized bipartite state has Schmidt coefficients sᵢ and values below a selected threshold are removed, the resulting state is approximate. The discarded weight is

assuming the Schmidt coefficients satisfy ∑ᵢsᵢ² = 1. This quantity helps quantify truncation error, but it does not automatically equal the total error of a long simulation.

7. Shot-Based Simulation

A quantum measurement produces samples, not a complete probability table. In shot-based simulation, the circuit is executed repeatedly, with each execution producing one outcome. If outcome x occurs kₓ times in N shots, its empirical frequency is

For a fixed outcome with true probability p, k follows a binomial distribution under independent identical shots. The estimator has

and

The standard error is therefore √[p(1 − p)/N]. Since p is usually unknown, p̂ is often substituted for planning purposes. The largest standard deviation occurs near p = 1/2, where it is 1/(2√N).

For an observable with measurement values in [−1,1], the standard error is generally of order 1/√N. Increasing shots improves statistical precision slowly: reducing standard error by a factor of two requires approximately four times as many shots.

Shot-based simulation is useful for:

• modelling hardware-like results;
• testing measurement post-processing;
• estimating expectation values;
• studying sampling noise and confidence intervals;
• validating algorithms that ultimately run on quantum devices.

It should not be confused with an exact simulator. A noiseless shot-based simulator may sample exactly from ideal probabilities, but the reported frequencies remain random estimates.

8. Comparative Selection Framework

Selection should be based on the circuit, not on qubit count alone. Important questions include:

1. Is the state pure or mixed?
2. Are non-Clifford gates frequent?
3. How much entanglement is expected?
4. Is exact amplitude information required?
5. Is the objective probability estimation or device-like sampling?
6. Is noise part of the scientific question?
7. What memory and runtime are available?

A practical rule is to use state-vector simulation for small-to-medium arbitrary pure circuits, stabilizer simulation for large Clifford-dominated circuits, tensor networks for structured low-entanglement circuits, density matrices for small noisy systems, and shot-based execution when statistical sampling is itself part of the experiment.

9. Exact and Sampled Experiment

Consider a circuit preparing the Bell state

An exact simulator predicts P(00) = P(11) = 0.5. Suppose 1,000 shots produce 487 instances of 00 and 513 instances of 11. Then

and p̂(11) = 0.513. The deviation from 0.5 is expected from finite sampling. Using p = 0.5 as a planning value, the standard error for either outcome is

The observed deviation, 0.013 for outcome 00, is less than one standard error. It is therefore consistent with ordinary sampling fluctuation.

A useful experiment is:

1. Construct a circuit with a Hadamard and controlled-NOT.
2. Run an exact state-vector simulation and record probabilities.
3. Run 100, 1,000, and 10,000 shots.
4. Compare empirical frequencies with exact probabilities.
5. Repeat with a depolarizing or readout-noise model.
6. Compare the noisy distribution with the ideal distribution.

The experiment demonstrates three distinct sources of discrepancy: floating-point round-off, finite-shot uncertainty, and physical or modelled noise. These must not be conflated.

10. Applied Case: Choosing a Simulator for Error-Correction Training

Background

This illustrative case concerns a teaching laboratory studying a small repetition code. Students prepare encoded states, apply stabilizer measurements, introduce bit-flip noise, and estimate logical error rates.

Problem

The laboratory has limited memory and wants to study circuits ranging from 20 to 1,000 qubits. The circuits use preparation, CNOT, Pauli gates, computational-basis measurement, and stochastic bit-flip errors. Students also need sampled results resembling hardware data.

A full state-vector simulator is unsuitable for the largest circuits because memory grows as 2ⁿ. A density-matrix simulator is even less suitable because its storage grows as 2²ⁿ. However, the core circuits are Clifford circuits with Pauli noise.

Analysis and approach

The laboratory selects a stabilizer simulator for large ideal and Pauli-noise experiments. It uses a state-vector simulator for small circuits when students need to inspect amplitudes. It uses shot-based execution for estimating logical error rates and demonstrating confidence intervals.

For each physical error probability q, the experiment is repeated over N shots. If L shots produce a logical failure, the estimated logical error rate is

The binomial standard error is approximately

when N is sufficiently large and the independent-shot assumption is reasonable.

Outcome

The stabilizer method permits large code distances to be explored. The state-vector method provides transparent demonstrations for small examples. Shot-based execution reveals that an estimated logical error rate is not identical to its exact model value. The laboratory therefore records simulator type, noise assumptions, number of shots, random seed where relevant, and uncertainty estimates.

Learning points

The best simulator is determined by circuit structure and experimental purpose. One project may legitimately use several simulators: an exact method for validation, an efficient structural method for scale, and sampled execution for measurement statistics.

11. Fidelity, Error, and Validation

Simulation fidelity has several meanings. Numerical fidelity concerns how accurately the computer represents the mathematical model. Model fidelity concerns whether the noise and device assumptions represent the intended physical system. Sampling fidelity concerns whether enough shots were used to estimate probabilities reliably.

For two pure states |ψ⟩ and |φ⟩, a common state fidelity is

For a density matrix ρ and pure reference |ψ⟩, the corresponding expression is

These formulas assume normalized states. Fidelity equals 1 for identical states and 0 for orthogonal pure states. Other definitions exist for two mixed states, so the chosen convention must be reported.

Good validation practice includes:

• testing known one- and two-qubit circuits;
• checking normalization and probability sums;
• comparing two independent implementations;
• checking that unitary evolution preserves norm;
• comparing exact and high-shot results;
• recording precision, truncation thresholds, noise channels, and seeds;
• using convergence studies when increasing shots or bond dimension.

12. Practical Best Practices and Limitations

A simulator result is meaningful only with a clear specification of the model. Report qubit ordering, gate convention, measurement basis, numerical precision, noise model, shot count, and whether measurements are sampled or analytically calculated.

Do not infer hardware performance from an ideal simulator. Do not interpret finite-shot fluctuations as algorithmic failure without uncertainty analysis. Do not assume tensor-network truncation is harmless merely because discarded singular values are small locally. Do not use a density matrix when a trajectory or stabilizer method answers the question more efficiently, but do not use an approximate method when exact noise correlations are central to the conclusion.

Reproducibility also matters. Randomized sampling should use recorded seeds where possible, while recognising that a seed reproduces a computational experiment rather than proving physical certainty. Software versions and simulator settings should be documented.

Contemporary Developments

Modern simulation increasingly combines methods. Hybrid tensor-network and state-vector approaches exploit circuit structure; stabilizer decompositions handle Clifford-dominated non-Clifford circuits; tensor-network contraction is accelerated by GPUs; and distributed state-vector simulators divide amplitudes across many processors. Quantum-trajectory techniques reduce the memory burden of density-matrix simulation. Approximate simulators use truncation, sampling, or low-rank representations, but their error must be measured rather than assumed.

The practical direction is therefore not a single universally superior simulator. Instead, simulation is becoming a workflow in which exact, approximate, structural, and sampled methods cross-check one another.

Chapter Summary

Quantum simulation replaces a physical quantum system with a classical mathematical representation. State-vector simulation stores 2ⁿ amplitudes and supports arbitrary pure-state circuits, while explicit unitary simulation is far more memory-intensive because it stores a 2ⁿ × 2ⁿ matrix. Density matrices represent mixed states and noise but require approximately 2²ⁿ complex entries. Stabilizer methods efficiently simulate Clifford circuits and are particularly valuable for error-correction studies. Tensor networks exploit limited entanglement and favourable circuit geometry. Shot-based simulation estimates probabilities through repeated outcomes and therefore includes statistical uncertainty.

Simulator selection should consider purity, gate set, entanglement, noise, required outputs, memory, runtime, and desired fidelity. Exact and sampled results answer different questions. Reliable experiments distinguish numerical error, approximation error, model error, and finite-shot error, and document all relevant settings.

Key Terms

State vector: A complex amplitude vector representing a pure quantum state.

Amplitude: A complex coefficient whose squared magnitude determines measurement probability.

Unitary operator: A reversible operator U satisfying U†U = I.

Density matrix: An operator representing pure or mixed quantum states.

Mixed state: A probabilistic ensemble of possible quantum states.

Kraus operator: An operator used to represent a general quantum channel.

Quantum channel: A mathematically valid transformation of a density matrix, including noise.

Stabilizer: A Pauli operator that leaves a stabilizer state unchanged.

Clifford circuit: A circuit composed primarily of operations that map Pauli operators to Pauli operators.

Tensor network: A collection of tensors connected by contracted indices.

Bond dimension: The dimension of an internal tensor-network index.

Entanglement: Nonclassical correlation between quantum subsystems.

Shot: One complete circuit execution producing one measurement outcome.

Sampling error: Statistical difference between empirical frequencies and true probabilities.

Fidelity: A measure of similarity between quantum states.

Trajectory simulation: Simulation of stochastic state-vector paths whose averages represent noisy evolution.

Truncation error: Error caused by discarding small tensor-network or approximate-representation components.

Scalability: The ability of a method to remain usable as circuit size increases.

Review Questions

5 Short Answer Questions

1. Why does state-vector memory grow exponentially with qubit count?
2. State the normalization condition for a density matrix.
3. What type of circuit is especially suitable for stabilizer simulation?
4. What does a shot represent in quantum computation?
5. Why can tensor networks be efficient for low-entanglement circuits?

5 Long Answer Questions

1. Compare state-vector and density-matrix simulation in representation, memory, noise modelling, and application.
2. Explain the Gottesman–Knill result and the limitations of stabilizer simulation.
3. Describe how tensor-network bond dimension is related to entanglement and computational cost.
4. Explain why explicit unitary simulation is less scalable than applying gates directly to a state vector.
5. Develop a simulator-selection framework for a noisy variational quantum algorithm.

3 Analytical / Application Questions

1. A 28-qubit pure-state experiment uses double-precision complex amplitudes. Estimate the minimum amplitude-storage requirement using M ≈ 16 × 2ⁿ bytes. Discuss why actual memory will be larger.
2. In 2,500 shots, an outcome occurs 625 times. Calculate its empirical probability and estimate the standard error using √[p̂(1 − p̂)/N]. Interpret the result.
3. Design an experiment that compares a state-vector simulator, a stabilizer simulator, and a shot-based simulator on the same Clifford circuit. Specify what should be compared and what conclusions would be justified.

Multiple Choice Questions with Answers

Question 1

Which representation stores 2ⁿ complex amplitudes for an n-qubit pure state?

A. State vector

B. Density matrix

C. Stabilizer tableau only

D. Measurement histogram only

Correct Answer: A. State vector

Question 2

What is the approximate memory requirement for a double-precision complex state vector of n qubits?

A. 2ⁿ bytes

B. 16 × 2ⁿ bytes

C. 16 × 2²ⁿ bytes

D. n² bytes

Correct Answer: B. 16 × 2ⁿ bytes

Question 3

Which simulator is most naturally suited to exact modelling of general noise channels on small systems?

A. Density-matrix simulator

B. Pure stabilizer simulator

C. Explicit measurement histogram

D. Classical sorting algorithm

Correct Answer: A. Density-matrix simulator

Question 4

A circuit containing only Clifford gates, computational-basis preparation, and computational-basis measurement is especially suitable for which method?

A. Density-matrix simulation only

B. Stabilizer simulation

C. Explicit unitary construction only

D. Tensor contraction with no tensors

Correct Answer: B. Stabilizer simulation

Question 5

What is the main structural factor limiting tensor-network simulation?

A. The number of classical measurement labels only

B. The amount of entanglement and resulting bond dimensions

C. The use of binary notation

D. The requirement that all gates be diagonal

Correct Answer: B. The amount of entanglement and resulting bond dimensions

Question 6

In shot-based execution, if an outcome appears 300 times in 1,000 shots, what is its empirical probability?

A. 0.03

B. 0.3

C. 3.0

D. 300

Correct Answer: B. 0.3

Question 7

For a fixed outcome with true probability p and N independent shots, which expression gives the variance of its frequency estimator?

A. pN

B. p²/N

C. p(1 − p)/N

D. N/[p(1 − p)]

Correct Answer: C. p(1 − p)/N

Question 8

Why is explicit construction of an n-qubit unitary generally less scalable than state-vector simulation?

A. It requires storing a matrix with 2ⁿ × 2ⁿ entries

B. It cannot represent reversible operations

C. It uses no complex numbers

D. It is restricted to one qubit

Correct Answer: A. It requires storing a matrix with 2ⁿ × 2ⁿ entries

Question 9

Which approach is most appropriate when the primary goal is to reproduce hardware-like measurement statistics?

A. Shot-based simulation

B. Symbolic algebra only

C. Explicit unitary storage only

D. Stabilizer compression regardless of circuit type

Correct Answer: A. Shot-based simulation

Question 10

A simulator reports a tensor-network result after discarding small singular values. What type of issue must be assessed?

A. Truncation error

B. Unit conversion error

C. Gate identity error only

D. Classical bit-width overflow only

Correct Answer: A. Truncation error