Quantum Channel Tomography Overview

• Quantum channel tomography is a technique to reconstruct the full CPTP map of quantum processes by experimentally determining the process (Choi) matrix.
• Standard protocols use informationally complete measurements and maximum-likelihood estimation to achieve physically valid reconstructions while mitigating nonphysical artifacts.
• Recent advances integrate tensor networks, unsupervised learning, and device-independent frameworks to scale tomography for high-dimensional systems and complex networks.

Quantum channel tomography is the experimental and computational procedure for reconstructing the complete input–output map (the "quantum channel") describing a physical process acting on quantum states. In modern language, a quantum channel is a completely positive, trace-preserving (CPTP) linear map between operator spaces, and channel tomography provides the empirical Choi or process matrix characterizing this map in full, up to statistical and systematic uncertainties. Quantum channel tomography underpins fields such as quantum processor benchmarking, high-dimensional communication, quantum network certification, error-correction code design, and foundational quantum-test experiments. Cutting-edge developments address informationally incomplete and device-independent scenarios, scalable algorithms based on tensor networks and unsupervised learning, and query-complexity reductions by exploiting channel structure or operational constraints.

1. Mathematical Framework and Representations

A quantum channel $\mathcal{E}$ is a CPTP map:

• Kraus representation: $\mathcal{E}(\rho) = \sum_k E_k \rho E_k^\dagger$, $\sum_k E_k^\dagger E_k = I$.
• Choi–Jamiołkowski isomorphism: $$\mathcal{E} \mapsto J_\mathcal{E} = d_\mathrm{in} (I \otimes \mathcal{E}) (|\Psi_+\rangle\langle \Psi_+|)$$, with $$|\Psi_+\rangle = \sum_{i=1}^d |i\rangle \otimes |i\rangle / \sqrt{d}$$. Channel action is recovered as $$\mathcal{E}(\rho) = \mathrm{Tr}_\mathrm{in}[J_\mathcal{E} (\rho^T \otimes I)]$$.
• Process (χ) matrix representation: $$\mathcal{E}(\rho) = \sum_{mn} \chi_{mn} E_m \rho E_n^\dagger$$ in a fixed operator basis $\{E_m\}$, with $\chi$ positive semidefinite, Hermitian, and subject to trace-preservation constraints (Varga et al., 2018, Huang et al., 2018, Bouchard et al., 2018).

Key constraints:

• Complete positivity: $J_\mathcal{E} \ge 0$, $\chi \succeq 0$.
• Trace preservation: $$\mathrm{Tr}_\mathrm{out} J_\mathcal{E} = I_\mathrm{in}$$, or $\sum_{m,n} \chi_{mn} E_n^\dagger E_m = I$.

2. Standard Quantum Channel Tomography Protocols

Informationally Complete Tomography

The standard protocol (Teo et al., 2013, Varga et al., 2018, Bouchard et al., 2018):

• Prepare a tomographically complete set of $d_\text{in}^2$ input states $\{\rho_\ell\}$.
• Pass each $\rho_\ell$ through $\mathcal{E}$; perform quantum state tomography (QST) on each output $\mathcal{E}(\rho_\ell)$, using an informationally complete measurement.
• Set up the linear system: measured outcome probabilities $$p_{\ell m} = \text{Tr}[\Pi_m \mathcal{E}(\rho_\ell)]$$, for POVM elements $\{\Pi_m\}$.
• Solve for the process matrix $\chi$ (or Choi matrix $J$) by linear inversion or maximum-likelihood estimation (MLE), subject to $\chi \ge 0$ and the TP constraint.

Maximum likelihood approaches (Teo et al., 2013, Huang et al., 2018):

• Optimize $$\mathcal{L}(\chi) = \sum_{\ell,m} n_{\ell m}\log p_{\ell m}(\chi)$$ under the CPTP constraints, using iterative algorithms or convex optimization.
• The physical solution is globally optimal, eliminating spurious negativity in $\chi$ that can arise from raw inversion.

Resource scaling: For $d$-dimensional systems, full QPT generically requires $O(d^4)$ experimental settings.

3. Advances in Tomographic Methodologies

Informationally Incomplete and Regularized Tomography

Incomplete data (e.g., insufficient input–output pairs) yields a convex set of compatible CPTP maps (Teo et al., 2013). Regularization techniques select solutions with maximal process entropy—a principle of least-bias—via the MLME (maximum likelihood-maximum entropy) approach:

$$S_\mathrm{proc}(J) = -\mathrm{Tr}\left(\frac{J}{d_\text{in}}\ln\frac{J}{d_\text{in}}\right).$$

Convex Optimization and Physicality Enforcement

Convex optimization methods, e.g., least-squares or log-likelihood minimization over $\chi \succeq 0$, $\mathrm{Tr}_{\rm out} \chi = I$, guarantee physical reconstructions and minimize deviations from measured data (Huang et al., 2018). They overcome nonphysical artifacts arising from statistical error in linear inversion.

Projective and Analytical Projection Approaches

Recent analytic projection schemes, such as the Cholesky-based analytic (CBA) method, optimally project arbitrary Hermitian matrices onto CPTP Choi matrices in closed form, dramatically improving numerical precision and efficiency, especially when combined with iterative projection algorithms (e.g., Dykstra's algorithm) (Barberà-Rodríguez et al., 2024).

Scalability and Tensor-Network/ML Approaches

For large quantum processors, representing the Choi matrix as a locally purified density operator (LPDO/MPO) and fitting via unsupervised learning algorithms (Adam/minibatch optimization on negative log-likelihood) enables channel tomography on circuits up to 10 qubits with polynomial classical resources (Torlai et al., 2020). This approach leverages tensor-network contraction and automatic differentiation to achieve process fidelities >0.99 for moderate sample sizes.

Compressive, Shadow, and Local Test Tomography

Shadow process tomography generalizes "classical shadows" to quantum channels, allowing targeted estimation of up to $M$ linear functionals of the channel with $\text{poly}(n,\log M,1/\epsilon)$ sample complexity for $n$-qubit channels, exponentially outperforming full process reconstructions for many tasks (Kunjummen et al., 2021). Query complexity for global tomography under diamond-norm or trace-norm error can be reduced to $O(r d_1 d_2 / \epsilon^2)$ if the Kraus rank is $r$ (Chen et al., 15 Dec 2025), and in favorable cases (unitary or isometric channels) achieves Heisenberg scaling $O(1/\epsilon)$ (Gutoski et al., 2013, Chen et al., 15 Dec 2025).

4. Specialized and Novel Tomography Regimes

Device-Independent Channel Tomography

Device-independent (DI) frameworks (Agresti et al., 2018) address the circularity of trusting preparation and measurement devices by treating them as uncharacterized, inferring accessible sets of input–output statistics compatible with a hypothesized CPTP map. DI protocols allow falsification of candidate process matrices by checking consistency of observed correlations, or identify minimal channels compatible with empirical data, up to an unavoidable equivalence class.

High-Dimensional and Network Tomography

Recent implementations extend standard and DI methods to photonic spatial qudits ($d>2$) (Varga et al., 2018, Bouchard et al., 2018). Tomography protocols have been devised for network settings, e.g., inferring individual Pauli channel parameters on star networks via only end-node measurements (Andrade et al., 2023, Andrade et al., 2022). Entanglement-assistance and adaptive protocol selection further improve parameter identifiability and sample efficiency.

Tomography in High-Energy and Other Exotic Regimes

Quantum process tomography has been applied to high-energy contexts, e.g. collider experiments, where quantum channels and instruments encode the transformation and measurement of particle spin–flavour densities. Complete Choi matrix reconstruction from collider data enables precision tests of the Standard Model and probes for beyond-quantum dynamics (Altomonte et al., 2024). Methodologies extend to devices transmitting quantum and classical information simultaneously, e.g., fibre channels with Raman background—here, spectrally resolved Bayesian techniques yield wavelength-dependent depolarizing channel models for realistic network modeling (Chapman et al., 2022).

5. Practical Implementations and Experimental Considerations

• Experimental QPT has been performed with high fidelity (>97%) on polarization qubits (Shaham et al., 2011), photonic qudits (Varga et al., 2018), and high-dimensional quantum communication channels (Bouchard et al., 2018).
• Physicality of the reconstructed channel (CPTP constraints) is enforced throughout reconstruction via MLE (Teo et al., 2013), convex optimization (Huang et al., 2018), or analytic projection (Barberà-Rodríguez et al., 2024).
• In optical and free-space transmission, classical light with nonseparable (vector) modes can be used to efficiently reconstruct channel χ-matrices, owing to the equivalence of quantum and classical process evolution for linearly acting channels (Ndagano et al., 2016).
• The tomographic representation translates quantum channels into (possibly nonclassical) kernels acting on state tomograms; for bosonic Gaussian channels, the tomographic kernel is a positive, normalized convolution mapping (i.e., a classical stochastic process), whereas for qubits it is generally not (Amosov et al., 2017).
• Bayesian estimation and Monte Carlo error analysis provide robust uncertainty quantification (Chapman et al., 2022, Bouchard et al., 2018).

6. Limitations, Scaling, and Pathways Forward

• Full process tomography resource cost scales as $O(d^4)$ for $d$-dimensional systems; compressive and structure-exploiting protocols reduce the scaling to $O(r d^2)$ (low Kraus rank) or $O(d^2)$ (unitary) (Gutoski et al., 2013, Chen et al., 15 Dec 2025).
• Informationally incomplete data necessitates regularization or entropy-maximization to avoid ambiguity (Teo et al., 2013).
• Device-independent and network-aware tomography frameworks trade off information completeness for robustness to calibration errors or adversarial settings (Agresti et al., 2018, Andrade et al., 2022).
• For quantum processors beyond 10–12 qubits, classical post-processing memory and runtime become prohibitive; tensor-network and "shadow" methods, as well as direct ML-like approaches, are essential (Torlai et al., 2020, Kunjummen et al., 2021).
• Applications include quantum benchmarking, error-correction code optimization, cryptographic security analysis, channel characterization in noisy environments, real-time error correction, and foundational tests in high-energy physics (Altomonte et al., 2024, Bouchard et al., 2018, Chapman et al., 2022).

7. Summary Table: Core Quantum Channel Tomography Methodologies

Protocol Type	Sample Complexity	Key Features
Full QPT	$O(d^4)$	Tomographically complete; exponential in $n$; convex optimization (MLE) (Teo et al., 2013, Varga et al., 2018)
Compressed Sensing	$O(r d\,\mathrm{polylog}(d))$	Assumes low rank; incomplete data (Kunjummen et al., 2021, Chen et al., 15 Dec 2025)
Shadow/Targeted	$O(\mathrm{poly}(n, \log M, 1/\epsilon))$	Efficient for specific figures of merit; not full reconstruction (Kunjummen et al., 2021)
Tensor Network + ML	$\ll d^4$ (for moderate $n$)	MPO/LPDO ansatz, scalable optimization (Torlai et al., 2020)
Device-independent	Data-driven; set-testing	No trust in devices; falsification and minimal characterization (Agresti et al., 2018)
Convex Optimization	As above	Enforces physicality; globally optimal fit (Huang et al., 2018, Barberà-Rodríguez et al., 2024)

Recent research demonstrates the convergence of information theory, machine learning, experimental physics, and foundational quantum mechanics in the advancement of quantum channel tomography across system sizes, operational regimes, and application domains.