Blind calibration of a quantum computer

Technical Summary: Blind Calibration of a Quantum Computer

Problem Statement
The calibration of quantum measurement apparatuses is fundamentally limited by the "calibration problem": accurately characterizing measurement errors requires high-quality input data, which in turn requires accurate state preparation. Traditionally, solving this requires a high-quality component (either state preparation or measurement) to serve as a reliable probe. This creates a dependency where improving one aspect of the system necessitates precise knowledge of the other, often requiring multiple bespoke, separate calibration experiments for different error sources (e.g., state preparation and measurement (SPAM) errors, crosstalk, overrotations). The authors identify the need for a data-economical, post-processing-light technique that can characterize measurement errors without relying on perfectly prepared states.

Methodology: Blind Calibration
The paper proposes a protocol termed "blind calibration," which infers measurement calibration parameters ($\zeta$) from tomographic data collected on a noisy quantum state ($\rho$) without prior knowledge of the state's exact preparation fidelity.

1. Theoretical Framework: The method models the measured data $y$ as a function of the true state $\rho$ and calibration parameters $\zeta$: $y(\zeta, \rho) = \text{tr}[M(\zeta)\rho]$. The measured operator $M(\zeta)$ is decomposed into a linear combination of ideal operators $N_j$ weighted by calibration coefficients $\xi_j(\zeta)$.
2. Optimization: The protocol employs an alternating minimization algorithm. It iteratively optimizes the estimate of the quantum state $\rho$ (constrained to be a rank-1 pure state) and the calibration coefficients $\xi$ to best fit the experimental data. Once the coefficients are determined, the calibration parameters $\zeta$ are solved for via a linear system.
3. Error Models: The method is designed to simultaneously calibrate multiple error mechanisms in a single experiment. The specific errors modeled for the trapped-ion system include:
   • Over-rotation: Small angular deviations in single-qubit basis rotations.
   • Beam Crosstalk: Unintended rotations on neighboring ions caused by laser beam leakage.
   • Readout Errors: "Dark" errors (false $|0\rangle$ detection) and "bright" errors (false $|1\rangle$ detection) due to off-resonant decay.
   • Detector Spillover: Signal leakage between adjacent detector channels.
4. State Selection: While the method is "blind" to the specific state, the authors note that sensitivity varies. They demonstrate that random product states can yield accurate recoveries, but for native system calibration, they optimize specific probe states to maximize parameter recovery precision.

Key Contributions and Results

• Experimental Demonstration: Using a 171Yb+ trapped-ion quantum computer, the authors demonstrated blind calibration on a three-qubit system. They successfully identified and quantified native system errors (on the order of $\lesssim 1\%$) and intentionally injected errors (up to 5%).
• Simultaneous Multi-Error Calibration: The protocol calibrated four distinct error types (overrotation, crosstalk, readout, spillover) simultaneously from a single set of projective Pauli measurement data, eliminating the need for separate experiments for each error source.
• Accuracy and Robustness:
  • Recovery of Injected Errors: When intentional miscalibrations were introduced, blind calibration recovered the parameters with high accuracy, matching the performance of direct, custom-tailored calibration methods.
  • State Tomography Improvement: Using the recovered parameters to correct tomographic data resulted in state estimates with trace distances to the target state comparable to those obtained via direct calibration, and significantly better than standard (uncalibrated) tomography.
  • Insensitivity to State Preparation Noise: Numerical benchmarks showed that the recovery of calibration parameters is only weakly sensitive to depolarizing noise in state preparation. In contrast, direct calibration methods (specifically for SPAM recovery) showed a linear dependence on depolarizing strength, making blind calibration roughly an order of magnitude more robust in noisy conditions.
• Resource Efficiency: For a three-qubit system, blind calibration required $2.7 \times 10^5$ total measurement shots. Direct calibration (combining SPAM correction and beam crosstalk calibration) required $6.3 \times 10^5$ shots. The authors propose a scaling strategy for larger systems by breaking ion chains into overlapping three-qubit subgroups, reducing the scaling from exponential to linear.

Significance and Claims
The paper claims to solve the measurement calibration problem in practice, not just in principle, using a single "blind" experiment. The primary significance lies in the ability to:

1. Decouple Calibration from State Preparation: Accurately characterize measurement errors without requiring a perfectly prepared probe state.
2. Consolidate Calibration: Replace multiple bespoke calibration experiments with a single, unified protocol that handles multiple error mechanisms simultaneously.
3. Enable Post-Hoc Analysis: The method supports a "measure-first-think-later" approach. Experimentalists can collect tomographic data and, in post-processing, determine which error models best fit the data, allowing for flexible identification of dominant error sources without prior assumptions.

The authors maintain a modest tone regarding future implications, noting that while the method is general, future work is needed to establish the limits on the number and types of errors that can be simultaneously recovered and to develop a detailed theory for identifying optimal probe states for specific error models. They emphasize that their current work establishes the efficacy of the method on a trapped-ion platform with performance competitive to traditional direct techniques.