Confidence Regions in Quantum Tomography

• Confidence regions in quantum state tomography are statistically defined subsets of the quantum state space that guarantee inclusion of the true state with a specified probability.
• Both frequentist methods, such as likelihood-ratio tests and polytope bounds, and Bayesian techniques are applied to balance computational efficiency with rigorous error certification.
• Advanced approaches using matrix concentration inequalities, adaptive stopping rules, and quantum error bars enhance the operational utility and tightness of these regions in experimental settings.

Confidence regions in quantum state tomography are statistically rigorous subsets of quantum state space that capture, with a prescribed probability, the true unknown state based on measurement data and a reconstruction algorithm. They are essential for quantifying uncertainty in quantum tomography, providing not just point estimates but regions that encode how much the measurement data constrain the true state, and how much ambiguity remains due to finite sample size and experimental imperfections. Methods for constructing such regions in quantum tomography have advanced substantially, integrating both frequentist and Bayesian statistical theory, and expanding beyond traditional error bars to address the complex geometry and constraints inherent to the set of quantum states.

1. Formalism and Definition of Confidence Regions

A confidence region in quantum tomography is a data-dependent subset $\Gamma(D) \subseteq \mathcal{S}(H)$ (the set of density matrices for Hilbert space $H$) such that, with probability at least $1-\alpha$, the unknown true state $\rho$ is included in $\Gamma(D)$: $$\Pr_{\text{data}|\rho}[\rho \in \Gamma(D)] \geq 1-\alpha$$ for all $\rho$ (Christandl et al., 2011, Wang et al., 2018). The construction of $\Gamma(D)$ depends on the measurement model, the data analysis protocol, and the desired statistical coverage guarantee.

Measurement data typically take the form of frequencies $\{n_i\}$ for POVM elements $\{\Pi_i\}$, leading to a likelihood function

$$L(\rho) = \prod_i [\operatorname{Tr}(\Pi_i \rho)]^{n_i}$$

and, equivalently, a log-likelihood function

$$\ell(\rho) = \sum_i n_i \ln[\operatorname{Tr}(\Pi_i \rho)]$$

(Glancy et al., 2012, Blume-Kohout, 2012). Point estimation (e.g., maximum likelihood estimation, MLE) yields $\hat{\rho}$, but confidence regions require characterizing the distribution—exact or approximate—of possible $\rho$ compatible with data.

2. Frequentist and Bayesian Region Construction

Frequentist Approaches

Frequentist confidence regions are defined such that they contain the true state with high probability, independently of any prior. Two principal strategies are:

• Likelihood-Ratio (LR) Regions: The LR region for confidence level $1-\alpha$ is

$$R_{1-\alpha}(D) = \left\{ \rho \geq 0, \operatorname{Tr} \rho = 1 : \lambda(\rho) \leq c_\alpha \right\}$$

with $\lambda(\rho) = -2 \ln[L(\rho)/L(\hat{\rho})]$. The threshold $c_\alpha$ can be chosen using the asymptotic $\chi^2_{d^2-1}$ law or finite-sample bounds:

$\Pr_{D|\rho}[\lambda(\rho) > c_\alpha] \leq \alpha$

(Blume-Kohout, 2012, Glancy et al., 2012). Wilks' theorem ensures that for large sample sizes and interior points, $\lambda(\rho)$ is approximately $\chi^2$-distributed with $f$ degrees of freedom.

• Polytope Regions (Clopper–Pearson generalized): Confidence polytopes are constructed by bounding the outcome probabilities using the Clopper–Pearson method,

$$\Gamma(D) = \{ \rho : L_i \leq \operatorname{Tr}(\Pi_i \rho) \leq U_i, \forall i \}$$

where $L_i$ and $U_i$ are distribution-free (binomial) lower and upper bounds, allocated via Bonferroni's inequality to ensure global coverage (Wang et al., 2018).

Bayesian Approaches

Bayesian credible regions are subsets of the state space containing $1-\alpha$ of posterior probability under a prior $\pi_0(\rho)$. Given a locally Gaussian posterior, the credible region can be approximated by an ellipsoid: $$(\theta - \hat{\theta})^T F (\theta - \hat{\theta}) \leq \chi^2_d(1-\alpha)$$ where $F$ is the Fisher information matrix at the MLE or posterior mode (Teo et al., 2018, Oh et al., 2019, Six et al., 2016). For small or boundary data, more sophisticated Laplace expansions or truncations are required.

The key distinction is that Bayesian regions quantify $P(\rho \in \Gamma | \text{data})$, while frequentist regions guarantee $P_{\text{data} | \rho}[\rho \in \Gamma]$ uniformly for any $\rho$. In the asymptotic regime, they coincide for smooth, noninformative priors and regular models (Teo et al., 2018).

3. Statistical and Computational Methodologies

Method	Key Features / Equation	Reference
Likelihood-ratio	$$R_\alpha(D) = \{ \rho : \lambda(\rho) \leq c_\alpha \}$$	(Blume-Kohout, 2012, Glancy et al., 2012)
Polytope (CP)	$$\{ \rho: L_i \leq \operatorname{Tr}(\Pi_i \rho) \leq U_i \}$$	(Wang et al., 2018)
Bayesian ellipsoid	$$(\theta-\hat{\theta})^T F (\theta-\hat{\theta}) \leq \chi^2$$	(Teo et al., 2018)
User-friendly ellipsoid	Adaptively scaled $\|\rho-\hat{\rho}\|_{HS}$ ball via Bernstein	(Gois et al., 2023)
Quantum error bars (QEB)	Marginal $\mu(f)$ of arbitrary figure of merit	(Faist et al., 2015)

Implementation is sensitive to the physicality constraints $\rho \geq 0, \operatorname{Tr} \rho = 1$, requiring convex optimization (for MLEs or region boundaries) or Monte Carlo, especially in moderate-to-large $d$. Region “size” can be measured in Hilbert–Schmidt, trace, or Bures distance, with various region types providing tightness guarantees.

Non-asymptotic approaches based on matrix concentration inequalities (e.g., Tropp’s matrix Bernstein for spectral norm deviation) yield confidence balls with explicit finite-sample coverage (Guta et al., 2018, Almeida et al., 2023). Bayesian sampling approaches (metropolis–Hastings, HMC, accelerated hit-and-run) allow calculation of region size and credibility even in high dimension (Oh et al., 2019, Faist et al., 2015).

4. Asymptotic, Non-Asymptotic, and Optimality Results

Asymptotic Scaling

In the asymptotic (large $N$) regime, the likelihood is sharply peaked and local approximations apply:

• The confidence/credible region shrinks as $O(1/\sqrt{N})$ in the estimation metric.
• Under group symmetry, ultimate lower bounds for the region’s size scale according to representation-theoretic considerations (the Heisenberg limit versus shot-noise scaling) (Walter et al., 2013).

For full-rank MLE and uniform prior, the asymptotic credible region’s size is

$$S(C) \simeq \frac{(2\pi)^{d/2}}{\sqrt{\det[N I(\hat{\theta})]}}\, \left( \frac{\chi^2_d}{d} \right)^{d/2}$$

(Teo et al., 2018), corrected for boundary effects at small $N$ or near-pure states (Six et al., 2016).

Non-Asymptotic and Minimax

Fully non-asymptotic confidence regions (e.g., based on matrix Bernstein inequalities) guarantee, for all $N$, that

$$\Pr\left[ \|\rho - \hat{\rho}_n\|_1 \leq r \sqrt{\frac{C\,g(d)}{n}\log \frac{d}{\delta}} \right] \geq 1-\delta$$

where $$r = \min\{\operatorname{rank}(\rho),\operatorname{rank}(\hat{\rho}_n)\}$$, and $g(d)$ is measurement-type dependent (Guta et al., 2018, Almeida et al., 2023). Such regions achieve minimax optimal rates in Frobenius or trace norm for both full and low-rank cases, and enable adaptive stopping rules in sequential protocols (Carpentier et al., 2015).

Fundamental lower bounds for confidence region size are derived using duality to quantum hypothesis testing, yielding “no region estimator can be tighter than the hypothesis testing bound,” and, in covariant scenarios, the volume scaling matches the Heisenberg limit (Walter et al., 2013).

5. Error Certification, Practical Stopping Rules, and Workflow Integration

For iterative estimators (like MLE), gradient-based stopping rules tie the halting criterion directly to the desired confidence level: compute the scaled maximum eigenvalue $r_k$ of the gradient-matrix $R(\rho_k)$ and halt when $2 r_k \leq \Delta_\alpha$ (the desired $\chi^2$ threshold). The current iterate $\rho_k$ is then guaranteed, up to the regularity assumptions, to belong to the asymptotic confidence region (Glancy et al., 2012).

In self-calibrating tomography, joint bounded-likelihood regions in the joint state–device parameter space are necessary to properly quantify correlated uncertainties (Sim et al., 2019). In sequential or online applications, “anytime-valid” confidence sequences are constructed based on likelihood-martingale inequalities, guaranteeing uniform-in-time coverage so that coverage is maintained even if the data-taking is stopped adaptively (Cumitini et al., 28 Jan 2026).

A practical summary for typical quantum tomography experiments includes:

1. Data collection—perform informationally complete (or sufficient) measurements.
2. Computation of the point estimate (MLE, least squares, or Bayesian mean).
3. Construction of the region (likelihood-ratio contour, ellipsoid, polytope, or error-bar-based interval) using either asymptotic or non-asymptotic methods as appropriate.
4. Reporting of region “size” and coverage level. Error bars on figures of merit are extracted via push-forward densities, optimization, or direct interval computation.
5. For high-dimensional systems, develop scalable sampling or convex-optimization techniques for region characterization (Teo et al., 2018, Gois et al., 2023, Faist et al., 2015).

6. Comparison of Strategies and Operational Considerations

Comparative studies show that the prefactor and geometry of the region (as well as computational complexity) vary greatly across methods:

• Likelihood-ratio regions are nearly minimax–optimal in average and worst-case volume (Blume-Kohout, 2012).
• Polytope (CP) regions are computationally efficient and conservative, with tightness determined by the number of measurement outcomes (Wang et al., 2018).
• Matrix-concentration–based ellipsoids and user-friendly HS balls yield asymptotic optimality and ease of reporting for arbitrary measurement schemes (Almeida et al., 2023, Gois et al., 2023).
• Bayesian credible regions may be substantially smaller, but their coverage guarantee depends on the prior and model regularity (Teo et al., 2018, Oh et al., 2019).
• Quantum error bars offer an efficient one-dimensional summary for figures of merit, rigorously convertible to confidence intervals (Faist et al., 2015).

Empirical studies quantify performance in terms of the number of samples needed to distinguish states, tightness of the region, and computational scalability (Almeida et al., 2023). The best choice depends on the size of the quantum system, the measurement design, the desired coverage, and available computational resources.

7. Assumptions, Limitations, and Future Directions

The validity of confidence regions depends critically on statistical assumptions, including independence of measurements, the regularity of the likelihood surface (applicability of Wilks’ theorem), and the placement of the true state with respect to the boundary of the state space. For small sample sizes or in the presence of near-pure states, asymptotic approximations may fail; in such cases, finite-sample (concentration or hypothesis-testing) bounds, Monte Carlo, or numerically certified methods should be employed (Glancy et al., 2012, Six et al., 2016).

Joint estimation of quantum states and device parameters (self-calibration), robustness to measurement imperfections, and construction of plausible regions with data-dependent thresholds represent active areas (Sim et al., 2019, Rau, 2010). Methods capable of integrating online stopping, adaptive sampling, and scalable convex optimization are important for near-term multi-qubit experiments and quantum computation testbeds (Carpentier et al., 2015, Cumitini et al., 28 Jan 2026).

As the complexity of quantum devices increases, efficient, numerically tractable, and rigorous certification of confidence regions—combined with clarity in reporting error metrics and their coverage properties—will remain a central requirement for both experimental and theoretical quantum information science.