Gate Set Tomography Overview

• Gate set tomography is a self-consistent, calibration-free method for characterizing quantum operations, including state preparations, quantum gates, and measurements.
• It employs informationally complete fiducial and germ sequences to amplify and accurately estimate both coherent and stochastic error modes.
• GST utilizes advanced performance metrics like diamond-norm distance and average gate fidelity, and is validated on diverse quantum platforms.

Gate set tomography (GST) is a self-consistent, calibration-free protocol for the comprehensive characterization of quantum logic operations on a quantum information processor. GST estimates an entire set of quantum processes—including state preparations, logic gates, and measurements—simultaneously, explicitly accounting for state-preparation and measurement (SPAM) errors and quantifying all error modes compatible with physical constraints. Unlike standard process tomography, GST achieves high precision, supports rigorous error analysis, and can certify quantum gates below stringent fault-tolerance thresholds. Its algorithmic framework, performance metrics, and extensions have been validated across a variety of quantum platforms, including superconducting circuits, trapped ions, donor spin qubits, and quantum dots.

1. Fundamental Principles and Mathematical Structure

GST models a quantum device via a gate set $\mathcal{G} = \{\rho, \{G_i\}, \{M_j\}\}$, where:

• $\rho$ is a state-preparation map (density matrix in $d$-dimensional Hilbert space),
• $G_i$ are logic gates, each represented as a completely positive, trace-preserving (CPTP) superoperator ($d^2 \times d^2$ matrix in Liouville representation),
• $M_j$ are measurement effects forming a positive-operator-valued measure (POVM), $\sum_j M_j = I$.

GST is fundamentally self-consistent: all elements—preparation, measurements, and gates—are estimated jointly, and none are assumed perfect or pre-calibrated. Circuit probabilities are computed as

$$p(j | C) = \langle\langle E_j | G_{n} \cdots G_{1} | \rho \rangle\rangle$$

where $C$ denotes the circuit sequence and all superoperators act in the Liouville (or Pauli-transfer) basis (Blume-Kohout et al., 2013, Nielsen et al., 2020, Greenbaum, 2015).

A central feature is gauge freedom: all probabilities are invariant under similarity transformations $T$ acting as

$$\rho \rightarrow T\rho, \quad G_i \rightarrow T G_i T^{-1}, \quad E_j \rightarrow E_j T^{-1}$$

implying that estimates are only meaningful up to gauge (i.e., a change of operator basis). Metrics such as average infidelity, diamond-norm distance, and process generators must be reported relative to a gauge-optimized alignment.

2. Experiment Design and Data Acquisition

GST employs informationally complete sets of "fiducial" sequences (for input and output preparation) and "germs"—short sequences of gates designed to amplify all physically relevant error modes when repeated. The full experimental set includes all circuits of the form

$$\text{prep-fiducial} \to \text{germ}^\ell \to \text{meas-fiducial}$$

with germ-power $\ell$ spanning logarithmically up to the system's decoherence-limited circuit depth (Nielsen et al., 2020, Rudinger et al., 2023).

In two- and multi-qubit systems this results in a rapidly growing number of distinct circuits, as each unique fiducial-germ-fiducial triple is needed for full parameter identifiability. Modern GST experiment design now deploys principled circuit-selection algorithms to greatly reduce the measurement burden while preserving Heisenberg-like scaling (Rudinger et al., 2023, Ostrove et al., 2023).

3. Estimation, Likelihood Maximization, and Performance Metrics

Parameter estimation in GST proceeds via:

• Linear-inversion GST (LGST): a closed-form estimator based on short circuits, immune to local minima but potentially unphysical;
• Maximum-likelihood estimation (MLE): iterative nonlinear convex optimization of the log-likelihood

$\mathcal{L}(\theta) = \sum_c n_c \log p_c(\theta)$

over the full gate-set parameterization $\theta$, enforcing CPTP constraints. The MLE fit incorporates all long-sequence amplifying circuits to extract tiny coherent errors and stochastic noise (Blume-Kohout et al., 2013, Greenbaum, 2015).

• Gauge Optimization: post-processing to align the estimated gate set with an ideal target set by minimizing the distance between them under all allowed gauge transformations.

Performance is quantified via:

• Diamond-norm distance ($\|G-\tilde{G}\|_\diamond$): the worst-case trace-norm error on an entangled input (Blume-Kohout et al., 2016),
• Average gate fidelity (in the $\chi$-matrix or PTM representation),
• Process infidelity and "error generator" analysis, which decomposes estimated errors into Hamiltonian (coherent) and stochastic (Pauli) components,
• Classical Fisher information matrix and Cramér–Rao bounds for estimator uncertainty (Greenbaum, 2015).

GST achieves Heisenberg scaling: for sequences of length $L$, error in estimated parameters decreases as $1/L$, assuming errors are amplifiable via long germs and not dominated by decoherence, an exponential improvement over standard tomography (Nielsen et al., 2020, Rudinger et al., 2023, Ostrove et al., 2023).

4. Extensions: Context-Dependence and Non-Markovianity

Standard GST assumes gates are fixed CPTP maps (Markovian noise). This assumption is violated by:

• Context dependence: where gate errors depend on neighboring gate identities (crosstalk, pulse collisions) or memory effects (idle history) (Moueddene et al., 2021, Viñas et al., 3 Jul 2025).
• Non-Markovianity: where environment retains memory or time-correlated noise is significant, leading to dynamics that cannot be captured by a sequence of independent CPTP maps (Li et al., 2023).

Context-aware GST extends model parametrization to include explicit context labels or microscopic models (e.g., phonon population for ion gates), allowing for polynomial scaling in parameter count with context size rather than exponential (Viñas et al., 3 Jul 2025). Instrument Set Tomography (IST) provides a general operational framework to account for non-Markovian system-environment correlations, estimating "instruments" at each time step and the effective process tensor (Li et al., 2023).

Recent work also employs microscopic filter-function models to reduce the parameter count in noisy gate models, leveraging knowledge of specific correlated noise, e.g., colored phase noise (Viñas et al., 2024, Viñas et al., 3 Jul 2025).

5. Applications: Multi-Qubit GST, Resource Reduction, and Quantum Instrument Tomography

GST has been experimentally demonstrated on superconducting transmons (Rudinger et al., 2021), donor qubits (Dehollain et al., 2016, Mądzik et al., 2021), quantum dots (Steinacker et al., 2024), and qutrits (Cao et al., 2022), with average gate fidelities exceeding 99.9%, diamond-norm errors certified below rigorous FTQEC thresholds ($6.7 \times 10^{-4}$), and SPAM infidelities below 1% (Blume-Kohout et al., 2016, Mądzik et al., 2021, Steinacker et al., 2024).

Key innovations in resource scaling include:

• Fiducial-pair and germ-set reduction algorithms that identify and remove redundant circuits by algebraic analysis of parameter amplification directions (using commutant structure and Fisher information) (Rudinger et al., 2023, Ostrove et al., 2023).
• Compressive GST leveraging low Kraus-rank structure and manifold optimization to perform reliable tomography with random gate sequences and an order of magnitude fewer experiments (Brieger et al., 2021).
• Randomized Linear GST, providing instantaneous, robust, and SPAM-insensitive estimates in the weak-noise regime using a linearized expansion, without circuit tailoring (Gu et al., 2020).
• Transformer-based machine learning models for GST, reformulating the task as a sequence-to-parameter regression problem, achieving estimation accuracy competitive with pyGSTi in one- and two-qubit systems (Yu et al., 2024).

Quantum Instrument Linear GST (QILGST) further generalizes GST to joint quantum-classical objects (quantum instruments), enabling the full tomography of mid-circuit measurements, including resolution of their post-measurement state disturbance and error channels in the presence of nontrivial quantum output (Rudinger et al., 2021). This enables precise fault-tolerance benchmarking for syndrome measurement primitives in quantum error correction circuits.

6. Impact, Limitations, and Future Directions

GST has fundamentally changed the landscape of quantum device characterization by enabling:

• Calibration-free, SPAM-robust estimates for all device operations,
• Precision metrology for coherent and incoherent errors, supporting direct comparison against quantum error-correction thresholds,
• Diagnostic utility—enabling the identification and mitigation of non-Markovian drift, coherent over/under-rotations, and correlated errors otherwise invisible to randomized benchmarking,
• Efficient scaling to two- and three-qubit devices through combinatorial circuit set reduction and low-dimensional physical models.

Limitations include the exponential growth in parameter and circuit count with qubit number for fully generic black-box GST, which is mitigated—but not overcome—by model reduction and context-aware parameterizations. Ongoing research seeks to harness machine learning, further compressive sensing, and platform-specific modeling to extend GST to larger registers.

GST has enabled rigorous error budgets and hardware benchmarking for the next generation of fault-tolerant and NISQ devices, with its theoretical and practical toolkit now foundational in quantum characterisation, verification, and validation (QCVV) protocols (Nielsen et al., 2020, Rudinger et al., 2023, Ostrove et al., 2023, Viñas et al., 3 Jul 2025, Rudinger et al., 2021).