Finite-Shot Sensitivity for Moment Estimation in Quantum Metrology

Technical Summary: Finite-Shot Sensitivity for Moment Estimation in Quantum Metrology

Problem Statement
While the Quantum Cramér–Rao Bound (QCRB) provides the fundamental asymptotic limit for parameter estimation in quantum metrology, it does not specify the number of measurements required for a concrete estimator to approach this limit. Standard strategies, such as the method of moments (MoM), typically rely on the sample mean of a calibrated observable to infer a parameter. However, existing theoretical frameworks often assume large-sample limits where estimators are asymptotically unbiased. In practical scenarios with a finite number of measurements ($\nu$), nonlinear calibration curves introduce biases that affect the mean-square error (MSE) at orders beyond the leading $1/\nu$ term. The central problem addressed is determining how many repetitions are necessary before the asymptotic sensitivity of a moment-estimation protocol becomes operationally visible, accounting for finite-measurement corrections.

Methodology
The authors develop a finite-measurement theory for the method of moments that operates directly at the level of measured moments rather than the full likelihood function. The framework considers a real parameter $\theta$ encoded in a density matrix $\rho_\theta$. The protocol involves measuring a Hermitian observable $M$ independently $\nu$ times to obtain a sample mean $\bar{M}$.

1. Expansion Framework: The authors expand the inverse calibration function $g(\bar{M}) = f^{-1}(\bar{M})$, where $f(\theta) = \text{Tr}[\rho_\theta M]$, around the true mean. This expansion utilizes the central moments $\mu_n(\theta) = \text{Tr}[\rho_\theta (M - f(\theta)I)^n]$ of the measured observable.
2. Bias Analysis: They derive the bias of the standard MoM estimator $\hat{\theta}_0$, showing it scales as $O(1/\nu)$ due to the nonlinearity of the calibration curve.
3. Bias Correction: A bias-corrected estimator $\hat{\theta}_{bc}$ is constructed by subtracting terms proportional to $1/\nu$ and $1/\nu^2$ derived from the calibration curve derivatives and central moments. This suppresses the estimator's bias to $O(\nu^{-3})$.
4. MSE Expansion: The mean-square error of the bias-corrected estimator is expanded in powers of $1/\nu$, identifying coefficients for the $1/\nu$, $1/\nu^2$, and $1/\nu^3$ terms.

Key Contributions and Results

• General Density-Matrix Expansion: The paper derives a general expansion for the MSE of a bias-corrected moment estimator for arbitrary quantum statistical models. The expansion is expressed solely in terms of the calibrated mean $f(\theta)$, its derivatives, and the central moments $\mu_n(\theta)$ of the observable.
• Bias-Corrected Estimator: The authors construct an estimator $\hat{\theta}_{bc}$ that achieves a bias of $O(\nu^{-3})$. The resulting MSE is given by:
  $$\text{MSE}(\hat{\theta}_{bc}) = \frac{A}{\nu} + \frac{B_M}{\nu^2} + O(\nu^{-3})$$
  where $A = \mu_2/(f')^2$ is the leading sensitivity, and $B_M$ is a coefficient governing the $1/\nu^2$ correction.
• Condition for Vanishing $1/\nu^2$ Correction: A significant theoretical result is the identification of a general condition, $B_M = 0$, under which the entire $1/\nu^2$ correction vanishes. This condition depends only on the calibrated mean, its derivatives, and the central moments, independent of state purity or unitary dynamics.
• Unitary Examples and Higher-Rank Components:
  • Effective Qubits: For pure unitary families in two-dimensional subspaces, the authors show that the $1/\nu^2$ term can be cancelled by choosing a centered observable (specifically, setting a parameter $\lambda=0$ in the observable's matrix representation). The leading residual correction then appears at order $1/\nu^3$, governed by the calibration curvature.
  • Qudits: In dimensions greater than two, the observable may contain higher-rank components invisible at the leading order. The authors demonstrate that these components can be optimized to reduce or cancel the leading $1/\nu^3$ correction. They provide a solvable qutrit model where an optimal choice of the observable's higher-rank component ($\alpha$) cancels the $1/\nu^3$ term.
  • Continuous Variables: For models with exactly linear calibration curves, the standard MoM estimator is already unbiased, and no inverse-map corrections are present; the MSE scales strictly as $1/\nu$.
• Operational Thresholds: The paper defines finite-shot thresholds ($\nu^*_2$ and $\nu^*_3$) that quantify the number of measurements required for the asymptotic sensitivity to be reached within a specified relative tolerance $\epsilon$. These thresholds depend on the magnitude of the non-vanishing correction coefficients ($B_M$ or the $1/\nu^3$ coefficient).

Significance and Claims
The paper claims to provide a rigorous framework for understanding the operational visibility of quantum-enhanced sensitivity in finite-sample regimes. By moving beyond the asymptotic QCRB, the work quantifies the "cost" of finite measurements in terms of the number of repetitions needed to observe the theoretical advantage.

The authors emphasize that their expansion is not a new lower bound but a description of the sensitivity achieved by a specific bias-corrected protocol. They highlight that:

1. Nonlinear calibration curves inherently introduce finite-sample bias, necessitating correction for fair comparison with QCRB benchmarks.
2. The condition $B_M=0$ offers a practical optimization criterion for experimentalists to design observables that minimize finite-shot corrections.
3. In unitary models, higher-rank components of observables (often neglected in leading-order analysis) play a crucial role in suppressing higher-order errors, suggesting an order-by-order strategy for improving finite-shot performance.

The work concludes that these derived thresholds are essential for determining the measurement budget required to make quantum metrological resources experimentally visible, particularly in scenarios where the number of available measurements is limited.