Google's Digital Error Model

• Google's Digital Error Model is a framework that models output distributions as mixtures of ideal behavior and uniform noise based on independently characterized hardware errors.
• It employs a multiplicative ansatz to predict circuit fidelity using calibrated error rates from gate operations and readouts, validated in studies like the Sycamore experiment.
• Empirical validations reveal limitations such as overestimated fidelities, nonstationary noise, and ignored correlations, prompting ongoing refinements and alternative formulations.

Google’s digital error model refers to a class of mathematically streamlined models developed and deployed by Google for error budgeting and fidelity estimation in both quantum computing and digital hardware platforms. It is foundational in the analysis of Google’s 2019 Sycamore quantum supremacy experiment, and has been adapted for classical error analysis in fixed-function accelerators such as TPUs. The model posits output distributions as explicit mixtures of idealized behavior and uniform (random) noise, using parameters estimated from independently characterized hardware error rates. The essential formulation is a product model for circuit fidelity under assumptions of component-level statistical independence and dominant single-event failure. Its empirical performance, strengths, and limitations have been systematically interrogated in subsequent independent and collaborative investigations.

1. Digital Error Model in Google’s Sycamore Quantum Processor

The central instance of Google’s digital error model is the description and prediction of output statistics in random quantum circuits under noise. For an $n$-qubit, $m$-depth random quantum circuit $C$, the ideal output probability distribution for bit-strings $x \in \{0,1\}^n$ is

${\bf P}_C(x) = |\langle x | U_C | 0^n \rangle|^2$

where $U_C$ is the unitary transformation specified by the circuit.

Digital noise model ("one-parameter depolarizing mixture"):

$${\bf N}_C(x) = \phi \, {\bf P}_C(x) + (1 - \phi) 2^{-n} \tag{1}$$

where $\phi \in [0,1]$ is the circuit fidelity, interpreted as the probability of no error on any circuit component; $(1-\phi)$ reflects the probability of an “error event” that randomizes the output (drawn uniformly).

A Priori Fidelity Estimation (“Formula (77)”)

To generate predictions for $\phi$ prior to experiment, Google introduced a multiplicative ansatz for the survival probability:

$$\hat\phi = \prod_{g \in \mathcal{G}_1} (1 - e_g) \prod_{g \in \mathcal{G}_2} (1 - e_g) \prod_{q \in \mathcal{Q}} (1 - e_q) \tag{77}$$

where

• $\mathcal{G}_1$: all single-qubit gates in $C$;
• $\mathcal{G}_2$: all two-qubit gates in $C$;
• $\mathcal{Q}$: all qubits subject to final readout;
• $e_g$: error rate for gate $g$;
• $e_q$: readout error probability for qubit $q$.

Practical evaluation often uses averaged values:

$$\hat\phi \approx (1-\bar e_1)^{|\mathcal{G}_1|} (1-\bar e_2)^{|\mathcal{G}_2|} (1-\bar e_{ro})^n$$

with typical values $\bar e_1 \approx 0.0016$, $\bar e_2 \approx 0.0062$, $\bar e_{ro} \approx 0.038$ for Sycamore-53 (Kalai et al., 2023, Kalai et al., 11 Dec 2025).

2. Assumptions, Statistical Structure, and Calibration

The underlying assumptions of the model as formalized in Sycamore include:

• Statistical independence: Each gate and measurement fails independently; no crosstalk.
• No error propagation: Any failure—gate or readout—renders the full sample undistillable.
• Calibration quality: Estimated error rates for each element are unbiased and accurate to within $\pm 20\%$ (Kalai et al., 11 Dec 2025).

Prior to measurement, extensive calibration is performed:

• Each two-qubit gate is tuned by adjusting pre/post $Z$-rotations and interaction parameters via benchmarking routines specific to its neighborhood.
• Local calibration exerts global impact: simulating removal of even a single gate calibration can sharply drop the observed fidelity; in the Sycamore-12 device, fidelity collapses to near zero absent calibration (Kalai et al., 2023).

The model's remarkable empirical agreement with global cross-entropy benchmarks arises despite these extremely simplified independence and locality assumptions.

3. Empirical Validation and Breakdown

Comprehensive statistical analysis and independent re-examination have revealed systematic deviation between the digital error model and real device data, particularly as system size and depth increase (Kalai et al., 2023, Kalai et al., 11 Dec 2025):

• Goodness-of-fit failures: For small circuits ($n=12$ or $14$), full-distribution $\chi^2$ tests yield values far above theoretical expectations. Other metrics ($\ell_1, \ell_2$, KL, Pearson correlation) likewise show the empirical distribution is farther from the digital model than from uniform noise for $n > 16$.
• Overestimation of fidelity: Predictions based on (77) with either average or component-specific error rates systematically overestimate observed fidelity at large depth.
• Temporal and spatial drift: Empirically measured distributions demonstrate nonstationary behavior within long sampling runs, and patch circuits (non-overlapping device halves) display notable systematic bias.
• Ignored correlations: The model excludes all forms of gate-dependent, spatial, or temporal error correlations.

Representative Quantitative Data (Kalai et al., 11 Dec 2025)

	Sycamore-53, d=20	Patch I (d=14)	Patch II (d=14)
Product (77) pred.	0.0055	0.048	0.042
Individual-rates (77)	0.0053	--	--
Empirical XEB	0.0074	0.054	0.047

For “patch circuits,” per-component error rates and corrections could not explain $\sim$10–15% discrepancies.

4. Extensions and Technical Refinements

Several refinements and alternatives extend the digital error model to incorporate missing effects or to fit empirical trends:

• Average-rate ("Formula (77)*"): Uses global averages rather than per-component rates.
• Effective two-gate-cycle error ("Formula (77)**"): Absorbs single-qubit errors into an effective two-qubit gate error rate.
• Idle and preparation fidelities: Later adopted in the USTC 83-qubit experiment, explicit fidelity factors are introduced for qubits idle during gates and for imperfect initial state preparation.
• Multi-term mixture models: Incorporates independent readout error, gate error, and an unknown large-variance “noise term” to approximate the observed distributions:

$${\bf N}^*_C = \phi {\bf P}_C + \phi_{ro} {\bf N}_{ro} + (1-\phi_g) {\bf N}_g + {\bf N}_T$$

where $\phi = \phi_g - \phi_{ro}$, and ${\bf N}_T$ absorbs unexplained variance.

• Variance-based estimators: The $T^2$ statistic is introduced, which is unbiased for $\phi^2$ under the Porter–Thomas mixture, revealing excess variance unaccounted for by the digital model.
• Fourier-level fidelity estimators: Expansion in Walsh–Fourier basis permits isolation of contributions by Hamming weight and identification of coherent error terms.

5. Applications Beyond Quantum: Digital Error Modeling for TPUs

In classical digital hardware, Google’s digital error model has been instantiated for the analysis of permanent manufacturing faults in systolic-array-based Tensor Processing Units (TPUs) (Kundu et al., 2020):

• Markov Chain Formalism: Permanent stuck-at-0/1 faults in weight registers, accumulators, or multipliers are modeled using a composed discrete-time Markov chain (DTMC) that tracks input selection, MAC computation steps, and error propagation across network layers.
• Probabilistic model checking: The model supports formal PCTL queries for the probability of inference error and mean time to failure as a function of fault bit position, MAC location, and network depth.
• Empirical validation: DTMC-based predictions for classification accuracy as a function of injected faults closely match observed accuracy depletions in real MNIST DNN experiments.

6. Critique, Limitations, and Ongoing Adaptation

The digital error model is favored for its analytic simplicity and interpretability, but its empirical validity is presently restricted to the zeroth or first moment statistics of small-to-moderate depth circuits. Documented limitations comprise:

• No explicit cross-talk modeling: Completely factorized error structure omits all spatial and temporal correlations.
• Non-unitarity and leakage: Systematic amplitude damping, readout asymmetry, and leakage phenomena alter observed bitstring statistics in ways not addressed in the model.
• Model drift and context dependence: Error rates characterize the qubits and gates only under specific calibrations and operational contexts; errors are not stationary.
• Potential overfitting: Agreement with global XEB metrics does not test for model adequacy at the full distributional level—models may be retrofitted to match expected behaviors.

Refined models—such as per-qubit asymmetric measurement error channels, Fourier-based analysis, and multi-component mixture models—aim to repair some of these deficiencies, but no universal digital error model valid for large, deep NISQ circuits currently exists (Kalai et al., 2023, Kalai et al., 11 Dec 2025).

7. Future Directions and Statistical Validation

Advancing the predictive and diagnostic power of the digital error model hinges on several directions:

• Publication of full bitstring data and gate calibration logs to allow independent validation and high-order statistical analysis.
• Benchmarks moving beyond fidelity: harnessing full output distributions, high-order moment tests, and targeted circuit classes (e.g., Clifford or IQP circuits).
• Incorporation of correlated-error and non-unital noise channels in product formulas, dynamic calibration tracking, and incorporation of idle/mid-circuit effects.
• Integration with data-driven and microscopic noise processes such as Lindblad dynamics, control electronics contamination, and physical environment couplings.

The Google digital error model remains a reference baseline for the first-order characterization of hardware-induced noise in both quantum and classical circuits. However, high-resolution statistical validation and iterative refinement are essential for robust performance benchmarking and for interpreting claims at or beyond the quantum supremacy threshold (Kalai et al., 11 Dec 2025, Kundu et al., 2020, Kalai et al., 2023).