Characterizing Quantum Gates via Randomized Benchmarking

I. INTRODUCTION

• Process tomography is infeasible for large systems, motivating scalable methods to partially characterize noise affecting quantum systems.Complete noise characterization can help determine error-correction schemes.Source: §I · p. 1
• The paper gives a full analysis of a scalable multi-qubit randomized benchmarking protocol for Clifford gates, addressing shortcomings of earlier protocols.Earlier approaches lacked validity conditions for realistic gate- and time-dependent noise, scalable multi-qubit extensions, or explicit treatment of state-preparation and measurement errors.Source: §I · p. 1
• The analysis derives zeroth- and first-order fitting models for fidelity decay and discusses their applicability to time- and gate-dependent noise.The zeroth-order model is exponential for time- and gate-independent errors; the first-order model accounts for first-order corrections.Source: §I · p. 2

A. Diamond Norm, Average Gate Fidelity and Minimum Gate Fidelity

• The paper discusses how the diamond norm distance and other measures relate to channel fidelity, including the operational connection of the diamond norm to distinguishing channels.For Pauli errors, it reports a connection between the benchmarking average error rate and the diamond norm error measure.Source: A. Diamond Norm, Average Gate Fidelity… · p. 2, §I · p. 2
• The channel fidelity compares outputs of two quantum operations on an input state, while gate fidelity is its unitary-operation case.The minimum channel fidelity is attained at a pure state.Source: A. Diamond Norm, Average Gate Fidelity… · p. 3

B. The Clifford Group and t-Designs

• The Clifford group is the normalizer of the Pauli group and is generated by phase, Hadamard, and controlled-NOT gates.The paper notes its role in universality, stabilizer-code theory and fault tolerance, and noise estimation.Source: B. The Clifford Group and t-Designs · p. 4
• The uniform distribution over the Clifford group forms a unitary 2-design, satisfying the twirling condition for quantum channels and states.A unitary t-design reproduces the corresponding Haar-measure averages for the specified polynomials.Source: B. The Clifford Group and t-Designs · p. 4
• Clifford twirling produces a depolarizing channel with the same average fidelity as the original operation.The average fidelity is invariant under the twirling operation.Source: B. The Clifford Group and t-Designs · p. 4
• For gate-independent noise, repeated twirls compose as repeated applications of the depolarizing channel, so average fidelity decreases exponentially with sequence length.The χ-matrix's (0, 0) entry is invariant under 2-design twirling and its depolarized-channel value decreases exponentially with sequence length.Source: B. The Clifford Group and t-Designs · pp. 4–5

III. RANDOMIZED BENCHMARKING

• The section presents the randomized benchmarking protocol and derives its fidelity-decay fitting models.The paper identifies the parameter r as the average error rate and, for a Pauli channel, as its infidelity.Source: §III · p. 5
• The model allows each Clifford gate's error operator to vary with both gate and time, and defines an average error operator and average error rate for the Clifford set.The average fidelity of the average error operator is invariant under Clifford twirling.Source: §III · p. 5
• The experimentally relevant goal is estimating an error parameter while relaxing assumptions about state preparation, measurement, and ancillary states or operations.Tomography-based schemes require exponential time resources in the number of qubits and assume negligible state-preparation and measurement errors.Source: §III · pp. 5–6

A. Protocol

• For each sequence length, the protocol samples uniformly random Clifford sequences, measures survival probabilities, and averages results across sequences.The final gate is chosen as the inverse of the first m gates' composition; state-preparation and measurement errors are represented in the prepared state and measurement operator.Source: A. Protocol · p. 6
• The averaged sequence fidelity is fit across sequence lengths to a decay model whose coefficients absorb state-preparation, measurement, and final-gate errors.In the first-order model, q − p^2 measures the degree of gate-dependence, and p determines the average error rate.Source: A. Protocol · p. 6
• The protocol's exact average is the uniform average over sequences, approached as the number of sampled sequences tends to infinity.The fitting functions model the sampled sequence fidelity in terms of this exact average.Source: A. Protocol · p. 6

2. Define Di2 uniquely by the equation Ci2 = Di2 ◦D†

• The change of variables makes the final gate independent of the preceding sequence.The transformed Clifford elements are independent across steps, and averaging each index runs over every Clifford element once.Source: §2 · p. 7

B. Perturbative Expansion and the Fitting Models

• The perturbative approach expands each gate- and time-dependent error around average noise to model the fidelity decay.The deviations δΛi,j are treated as small perturbations; the resulting fit parameters estimate average error per gate and separate combined state-preparation and measurement contributions.Source: B. Perturbative Expansion and the… · p. 7
• The expansion defines terms order by order, providing the basis for the fitting models.The supplied passages introduce the higher-order terms but do not specify their formulas.Source: B. Perturbative Expansion and the… · p. 8

1. Zeroth Order Model

• With gate- and time-independent noise, the zeroth-order model exactly describes the fidelity decay.Averaging independent twirls depolarizes each factor of the noise channel.Source: §1 · p. 8

2. First Order Model

• The first-order model’s additional C1(m −1)(q(m) −p2)pm−2 term measures gate dependence in the noise.The model is written to resemble the zeroth-order model while including this correction.Source: §2 · p. 11
• The coefficients A1(m), B1(m), and C1 absorb state-preparation, measurement, and final-gate errors.Their dependence on m and q(m) disappears when errors are time-independent.Source: §2 · p. 11

IV. NEGLECTING HIGHER ORDERS

• The analysis defines a time-dependent variation in the noise and also considers time-independent-noise cases.The supplied passage introducing the variation does not include its formula.Source: A. Bounding Higher Order Perturbation… · pp. 11–12
• The fidelity difference between perturbation orders k and k + 1 is bounded to guide where to truncate the expansion.The approach bounds the size of the omitted higher-order terms.Source: A. Bounding Higher Order Perturbation… · pp. 11–12
• The Hermitian-input 1→1 norm gives a tighter fidelity-difference bound than the diamond norm because its associated γ is smaller.The text identifies the 1→1 norm as weaker than the diamond norm for this bound.Source: A. Bounding Higher Order Perturbation… · pp. 11–12

B. Case Where Benchmarking Fails

• Inverse noise can produce fidelity Fg(m, ψ) = 1 at every sequence length despite substantial error on each gate, violating the sufficient condition.The example has highly gate-dependent noise.Source: B. Case Where Benchmarking Fails · pp. 12–13
• Map the input to an orthogonal measurement-basis state with each Clifford; an inverse-like noise pattern makes the original input likely to be measured.For small noise strength, the original input should almost never be measured.Source: B. Case Where Benchmarking Fails · p. 13

C. State Preparation and Measurement Errors

• State-preparation and measurement errors do not prevent estimating p when the fidelity curve is nonconstant; a constant curve is the exceptional case.The constant-curve cases are described as extreme or practically unlikely.Source: C. State Preparation and Measurement… · pp. 13–14
• A constant curve can arise when the average channel is totally depolarizing or the identity, or when A0 = 0 reflects substantial preparation or measurement error.For the identity-channel case, the constant fidelity is the overlap of the imperfect input state and POVM element.Source: C. State Preparation and Measurement… · p. 13

V. AVERAGE ERROR RATE AND THE DIAMOND NORM

• The average error rate r relates to diamond distance only in special cases; the paper gives a new proof for generalized Pauli channels and illustrates an SDP approach.The authors suggest the technique could help calculate or bound diamond distance for more general channels.Source: §V · p. 14

A. Calculating the Diamond Norm Distance Between Generalized Pauli Channels

• The SDP has strong duality, and its solution gives the diamond distance between the generalized Pauli channels.The proof uses primal and dual formulations and a feasible dual operator.Source: A. Calculating the Diamond Norm… · p. 14
• For channels with orthogonal unitary Kraus operators, the SDP calculation uses a generalized Bell basis in which the Choi matrix is diagonal.The diagonal entries are d(q_i − r_i); the derived norm equals the 1-norm of the vector of coefficient differences.Source: A. Calculating the Diamond Norm… · p. 14

• For a generalized Pauli channel Λ, the cited text states that its diamond distance from identity and randomized-benchmarking error rate r are related, although the supplied excerpt gives no formula.Source: B. Relating the Diamond Norm and Error… · p. 15

VI. SCALABILITY OF THE PROTOCOL

• The sequence-averaging procedure draws random Clifford sequences, evaluates their fidelities, and averages results across repeated samples and sequence lengths.The averaging obstacle arises because the number of possible sequences grows exponentially with length and system size.Source: §VI · pp. 15–16
• The required number of sampled sequences k is independent of both qubit count n and sequence length m, avoiding full averaging over the exponentially large Clifford group.The group size scales as 2^O(n^2), and the number of length-m sequences as 2^(mO(n^2)).Source: §VI · pp. 15–16
• The sampling-cost comparison favors randomized benchmarking over tomography: four-qubit process tomography requires 65,536 measurements, while its measurement scaling is 16^n.Source: §VI · p. 16
• The protocol samples uniformly from the full Clifford group by constructing a symplectic matrix and sign vector that preserve the generators’ commutation relations and Hermiticity.The matrix is built column by column using random solutions to linear constraints.Source: §VI · p. 16
• Every Clifford operation can be implemented as six rounds of one- and two-qubit Clifford operations, addressing the need to use an implementable generating set.Source: §VI · pp. 15, 17

4. Applying Hadamard operations

• The decomposition procedure uses O(n^3) time to obtain its intermediate operations, which can then be decomposed into Hadamard, phase, and CNOT gates.Source: §4 · p. 17

VII. DISCUSSION

• Randomized benchmarking scales to Clifford gates, accommodates time- and gate-dependent noise and state-preparation and measurement errors, and uses the first-order model to measure gate dependence.Source: §VII · p. 17
• For random Pauli errors, the protocol’s average-fidelity estimate has an exact relationship to the diamond-norm characterization of average error-operator strength.Source: §VII · p. 17
• Clifford benchmarks directly inform the robustness of encoding and decoding operations because stabilizer codes can use Clifford operations for both.Source: §VII · p. 17
• In the magic-state model described, Clifford gates, a single-qubit ancilla state, and computational-basis measurements suffice to implement any unitary operation, so only Clifford gates need benchmarking.Source: §VII · pp. 17–18
• A zeroth-order fit can match the data yet estimate p incorrectly, so the authors recommend first-order fitting unless prior noise knowledge indicates effective gate independence.Source: §VII · p. 18
• A significant difference between zeroth- and first-order estimates of p indicates that the zeroth-order fit is poor and, under the stated reasoning, the noise is strongly gate-dependent.Source: §VII · p. 18
• Correlations among physical qubits can produce large errors on Clifford elements even when errors on the generating set are controlled, complicating inference from the full-group average.Source: §VII · p. 18