Practical Fidelity Estimation for Noisy Quantum Devices Using Direct and Survival-Probability-Based Protocols

AI summary

Overview

Research area: Quantum computing device characterization — fidelity estimation for noisy intermediate-scale quantum (NISQ) hardware, specifically direct fidelity estimation (DFE) and a survival-probability-based reconstruction method. Technical level: Intermediate. The paper is dense with Liouville-space formalism and polynomial approximation theory, but its two protocols are presented as software-implementable procedures with explicit sampling and shot-count formulas, making the core ideas accessible to readers with a working knowledge of quantum channels. Scope (1 sentence): The paper formulates, analyzes, and numerically benchmarks two complementary tomographic-free fidelity estimators — an importance-sampled Pauli protocol with attenuation-aware readout correction, and a return-probability protocol with $L_2$-optimized polynomial coefficients — across depolarizing, dephasing, and amplitude-damping noise.

What This Paper Is About

Full quantum state tomography becomes impractical as system size grows, so researchers need estimators that report a single operational figure of merit — fidelity — without reconstructing the entire density matrix. This paper presents two such estimators in software-ready form: direct fidelity estimation, which samples Pauli observables weighted by the known target state's Pauli coefficients, and a survival-probability method, which extracts an adjoint-cycle fidelity proxy from repeated forward–inverse evolution cycles. The goal is to compare where each protocol is appropriate, how each degrades under noise, and how to correct or optimize accordingly.

Key Contributions

1. A software-level formulation of direct fidelity estimation (DFE). The paper writes out the full protocol — importance-weighted Pauli index sampling with $\ell := \lceil(\epsilon^2\delta)^{-1}\rceil$, per-index shot counts $m_i$, the estimator $\tilde{X}_i$, and the final average $\tilde{Y}$ — in a form suitable for direct numerical or experimental implementation, and states the unbiasedness and accuracy guarantees ($2\epsilon$ accuracy at confidence $1-2\delta$).

2. An analysis of readout noise and attenuation-factor correction for DFE. The paper shows that independent per-qubit bit flips attenuate ideal Pauli expectation values as $\langle W_k\rangle_{\mathrm{meas}} = \beta_k \langle W_k\rangle_{\mathrm{ideal}}$ with $\beta_k = (1-2\alpha)^{w(k)}$, where $w(k)$ is the Pauli weight, and demonstrates that dividing by the calibrated attenuation factor substantially improves agreement between the mean estimated fidelity and the true fidelity.

3. A survival-probability reconstruction of an adjoint-cycle fidelity proxy with transparent, optimizable coefficients. The paper defines $\mathcal{Q} = \mathcal{K}^\dagger\mathcal{K}$, expresses the proxy $F_{\mathcal{Q}}$ as a polynomial in return probabilities $R_m$, and replaces naive Taylor coefficients with coefficients obtained by minimizing an $L_2$ approximation error of $\lambda^{1/2}$ over an interval $[g,1]$, yielding the linear system $\sum_n G_{mn}a_n^M = b_m$ with $G_{mn} = (1-g^{m+n+1})/(m+n+1)$ and $b_m = (1-g^{m+3/2})/(m+3/2)$.

4. A unified comparison of the two philosophies of benchmarking, with explicit validity conditions. The paper frames DFE as state-structured and the survival-probability method as dynamics-structured, and states precisely when the latter reconstructs $F_{\mathcal{Q}}$ (cycle matching $\mathcal{K}I = \mathcal{K}^\dagger$ plus a real spectrum in the approximation interval) and when $F{\mathcal{Q}}$ equals the physical fidelity $F_{\mathrm{phys}}$ (when the interaction-picture error map $\mathcal{E}$ is positive semidefinite).

Main Findings

• Depolarizing fidelity has a closed form independent of the target state. For a pure target in dimension $d$, global depolarizing noise of strength $p$ gives $F(\rho,\sigma) = 1 - p(1 - 1/d)$, i.e. $F = 1 - p(1 - 2^{-n})$ for $n$ qubits, independent of the particular pure state or any preceding ideal unitary. The four-qubit noise-sweep simulation uses this formula as its reference curve with $d = 16$.

• DFE estimates track the true fidelity closely in simulation. For a four-qubit Haar-random state under $10%$ depolarizing noise, the true calculated fidelity was $0.90626$ and the mean of the plotted estimates was $0.907213$.

• Monte Carlo extension preserves the estimator distribution at low cost. Extending 1000 directly simulated DFE estimates to 9000 samples via a fitted parametric distribution gave an absolute mean difference of $0.00102$ between the original and extended estimation distributions for that case; the two histograms are displayed separately so the approximation's effect remains visible.

• Attenuation-aware correction improves DFE robustness. Applying the effective binary readout-flip correction (dividing by $1 - 2p_{\mathrm{flip}}$, with $p_{\mathrm{flip}} = 0.05$ in the noise sweep) substantially improved agreement between mean estimated and true fidelity, even for measurement error rates up to roughly $5%$. The paper notes the correction can amplify statistical noise when $|\beta_k|$ is small.

• $L_2$-optimized coefficients beat Taylor truncation at high noise. In the survival-probability reconstruction, the $L_2$-optimized expansion substantially outperforms naive Taylor truncation in moderate- and strong-noise regimes. Taylor coefficients are only locally optimal around the weak-noise limit and can fail badly when the noise action is not perturbatively small.

• Validity conditions are stricter than they may first appear. The reconstructed quantity is $F_{\mathcal{Q}}$ only when the implemented inverse realizes $\mathcal{K}^\dagger$; a generic physical approximate inverse yields moments of $\mathcal{C} = \mathcal{K}I\mathcal{K}$ and requires separately justified reconstruction. Moreover, $F{\mathcal{Q}} \neq F_{\mathrm{phys}}$ in general, coinciding only when $\mathcal{E}$ is positive semidefinite.

• A spectral bound supports the choice of approximation interval. Using the interaction-picture propagator and Gronwall's inequality, the paper bounds $s_{\min}(\mathcal{K}) \geq e^{-\eta}$ with $\eta = \int_0^T \lVert\mathcal{L}(t)\rVert,dt$, hence $\lambda_{\min}(\mathcal{Q}) \geq e^{-2\eta}$, so the interval in the $L_2$ problem may be chosen with $g \leq e^{-2\eta}$ provided $\lambda_{\max}(\mathcal{Q}) \leq 1$ or the spectrum is rescaled.

• The reported survival-probability results use exact moments, not finite-shot data. The simulations for a two-qubit Haar-random state under global depolarizing noise, independent local amplitude damping, and collective amplitude damping used exact $R_m = \langle\langle\rho_0|\mathcal{Q}^m|\rho_0\rangle\rangle$ with $m = 0,\dots,M$ for $M = 2,3,4$ and $g = 0.05$. The paper explicitly states these do not establish robustness to finite-shot or explicit SPAM errors.

• Noise sweeps were run at fixed, reported settings. DFE used four-qubit Haar-random, GHZ, and $W$ targets with $\epsilon = \delta = 0.05$, $\ell = 8000$ importance-sampled Pauli indices per trial, shot counts from the paper's Eq. (10), noise strength varied from $0.10$ to $0.25$ in increments of $0.01$, and 1000 independent DFE trials per value. Survival-probability simulations sampled noise strength at 100 uniformly spaced points in $[0,1]$.

Methodology in Plain English

The authors took two different routes to the same question — how good is my quantum state, without doing full tomography?

For the first route, they exploit the fact that if you know what state you were trying to prepare, you can use its structure to decide which measurements matter. Instead of measuring all $4^n$ Pauli observables, they randomly pick observables with probability proportional to how much each one contributes to the target state, measure each one a number of times set by a formula tied to the desired accuracy, and average the results into a fidelity estimate. They then simulated this on a classical computer: prepare a four-qubit target (Haar-random, GHZ, or $W$), push it through a noise channel, sample measurement outcomes from a binomial distribution, and compare the estimate to the exactly computed overlap. To model imperfect readout, they flipped each effective binary outcome with probability $0.05$ and then divided by the known attenuation factor to see whether the correction rescued the estimate.

For the second route, they used a completely different signal: instead of measuring many observables, run the noisy circuit forward, run an approximate inverse, and measure how often the system returns to where it started. Repeating this cycle $m$ times gives return probabilities. The quantity of interest turns out to be a square root of an eigenvalue-like parameter, so they approximate that square root with a polynomial in the return probabilities. The key move is choosing the polynomial coefficients not by Taylor expansion around zero noise (which only works when noise is tiny), but by minimizing the squared error of the approximation over a whole interval of possible noise strengths — a small linear system that can be solved once and reused.

The numerics were done in Python with QuTiP for state preparation and noisy evolution, with the shot simulation vectorized for speed. Because running thousands of full shot-level simulations is expensive, they used a two-stage trick for the multi-trial distributions: run a smaller set of full simulations, fit a distribution to those results, then draw additional synthetic samples — validating the fitted distribution against the directly simulated subset.

Why This Matters

Fidelity is the number experimentalists quote when they say how well a device is working, but obtaining it rigorously at scale is hard. This paper contributes practical, implementable recipes with explicit shot budgets and explicit statements of when they are valid — the latter being unusually candid, since it spells out that the survival-probability route does not automatically give physical state fidelity. That clarity matters for anyone building automated characterization tools, because it prevents a diagnostic proxy from being reported as a fidelity.

Real-world applications (as motivated by the paper):

• Superconducting processor validation: the authors name the upgraded quantum-computing infrastructure at VTT as a planned testbed, and propose comparing these estimators against randomized benchmarking, cycle benchmarking, and sparse Pauli–Lindblad noise-learning frameworks.
• Continuous device monitoring: integrating scalable fidelity estimation into automated validation and calibration pipelines for ongoing processor performance tracking.
• Noise diagnosis and control optimization: using the dynamics-based proxy to detect departures from ideal unitary evolution and to inform data-driven optimization of quantum control.
• Benchmarking toward fault tolerance: assessing whether characterization tools remain resource-efficient as processors scale toward error-corrected operation.

Industry relevance: the protocols target the practical bottleneck of characterization cost rather than algorithmic speedup, which is where near-term hardware vendors and cloud quantum providers spend engineering effort. The GitHub data/code release lowers the barrier to adoption, and the explicit shot-count formulas translate directly into experiment-planning budgets.

Future Directions

1. Fidelity estimation for classically hard target states. The paper identifies highly magical nonstabilizer states as an open problem, noting that the sampling resources required by generic DFE protocols can grow exponentially with the stabilizer Rényi entropy $M_2(\rho)$, so the classical cost of characterizing the target becomes part of the benchmarking problem.
2. Reducing DFE sampling overhead. Suggested routes include grouped-Pauli sampling, adaptive measurement-budget allocation, and Markov chain Monte Carlo ideas for regimes where naive importance sampling remains costly.
3. Making the survival-probability protocol experimentally realistic. Deriving coefficient optimization directly from a noise model and analyzing error propagation under finite-shot fluctuations and imperfect inverses would make it comparable to randomized benchmarking and related scalable methods; the current simulations use exact moments and therefore do not demonstrate SPAM robustness.
4. Deployment on real hardware. Implementing both protocols on experimentally prepared multiqubit states and quantifying resilience to calibration drifts, leakage, correlated noise, coherent control errors, and readout imperfections.

Target Audience

Quantum information graduate students and postdocs working on device characterization; experimentalists on superconducting or trapped-ion platforms who need tomographic-free fidelity estimators with explicit shot budgets; quantum software engineers building calibration and validation pipelines; and researchers interested in the theory of polynomial reconstruction of channel quantities, particularly the distinction between physical fidelity and adjoint-cycle proxies. Readers seeking a beginner-level introduction to fidelity estimation should look elsewhere first; readers seeking a rigorous treatment of Pauli sampling will find the DFE portion familiar, while the survival-probability coefficient optimization is the more novel material.

Authors’ abstract

Reliable characterization of quantum states and quantum processes is central to the development of scalable quantum computing. Standard characterization techniques such as full quantum state tomography become rapidly impractical as system size increases. In this work we present two software-oriented fidelity estimation protocols that avoid full tomographic reconstruction while remaining experimentally relevant for noisy intermediate-scale quantum devices. The first protocol is direct fidelity estimation (DFE), where the fidelity between a target state and an experimentally prepared state is estimated from importance-sampled Pauli measurements. The second protocol is a survival-probability-based method that reconstructs fidelity from experimentally accessible moments of a noisy evolution and its approximate inverse, and is intrinsically robust against state-preparation-and-measurement (SPAM) errors. We discuss the theoretical basis of both approaches, analyze their noise sensitivity, and present numerical evidence on representative multiqubit states under depolarizing, dephasing, and amplitude-damping channels. For DFE we show that attenuation-aware correction improves robustness against measurement noise. For the survival-probability approach we show that an $L_2$-optimized expansion substantially outperforms naive Taylor truncation in high-noise regimes.