Quantum Tomography Procedure Overview

• Quantum tomography procedure is a family of methods to reconstruct unknown quantum objects via designed measurement schemes, employing techniques such as linear inversion, maximum-likelihood, and Bayesian regularization.
• It applies systematic steps—state preparation, measurement, data acquisition, and inversion—to estimate density operators, CPTP maps, and temporal process tensors with high fidelity.
• Recent advances introduce measurement compression, ancilla-assisted strategies, and direct weak-measurement techniques, reducing resource requirements and addressing non-Markovian dynamics.

Quantum tomography procedure denotes the family of experimental and computational workflows used to reconstruct an unknown quantum object from measurement data. In the literature considered here, the reconstructed object may be a density operator, a CPTP map, a POVM or detector, a temporal input–output map with memory, a multi-time process tensor, or a continuous quantum field. Across these variants, the recurring structure is to choose a representation of the target object, design an informationally sufficient measurement scheme, acquire data under repeated preparations or controlled interventions, and reconstruct a physical estimator by linear inversion, maximum-likelihood methods, evidence-based Bayesian regularization, or constrained projection (Shukla et al., 2014, Keith et al., 2018, Tran et al., 2021, White et al., 2021, Steffens et al., 2014).

1. Foundational formalism

Quantum state tomography (QST) reconstructs an unknown density operator $\rho$ of a $d$-dimensional quantum system from measurement data. Quantum process tomography (QPT) reconstructs an unknown quantum process $\mathcal{E}$ acting on density operators. In the standard formulation, a quantum process is modeled as a completely positive, trace-preserving map $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$, and standard QPT proceeds by preparing a complete set of $d^2$ linearly independent input states, applying the process to each input, and performing QST on each output (Shukla et al., 2014).

Two operator representations dominate the procedure. In Kraus form,

$$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$

In a Hilbert–Schmidt orthonormal operator basis $\{E_m\}$, the same channel is written as

$$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$

with $\chi$ positive semidefinite. The Choi–Jamiołkowski isomorphism packages the process into a bipartite state

$$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$

so that reconstructing $d$0 is equivalent to reconstructing the channel. For state tomography and detector tomography, the basic measurement model is the Born rule, $d$1 or $d$2, with POVM elements constrained by positivity and normalization (Rau, 2010, Barberà-Rodríguez et al., 2024).

This formalism extends naturally beyond memoryless channels. Temporal tomography targets a time-dependent input–output law for a device that processes a stream of quantum states with memory, rather than a single time-local channel (Tran et al., 2021). Process tensor tomography reconstructs a $d$3-step comb $d$4 that maps an entire sequence of interventions to a final state, thereby capturing non-Markovian dynamics over multiple times (White et al., 2021). In continuous systems, cMPS tomography replaces finite-dimensional operator expansions by a field-theoretic ansatz

$d$5

and reconstructs the state from low-order correlators (Steffens et al., 2014).

2. Standard workflow and resource scaling

The canonical tomography workflow has four stages: state or probe preparation, application of the unknown process or measurement, acquisition of outcome statistics, and inversion or constrained estimation. For standard QPT this means $d$6 input preparations and a QST subroutine on every output state. Because QST requires measurements of non-commuting observables, the number of experimental configurations grows rapidly with system size (Shukla et al., 2014).

The scaling is explicit in several platforms. In the $d$7-qubit NMR setting, with $d$8 and $d$9, a single QST can be implemented with approximately $\mathcal{E}$0 independent NMR measurements, so total QPT measurements scale as

$\mathcal{E}$1

The stated examples are $\mathcal{E}$2 measurements for a single qubit, $\mathcal{E}$3 for two qubits, and $\mathcal{E}$4 for three qubits (Shukla et al., 2014). In polarization photonics, an overcomplete six-state tomography uses $\mathcal{E}$5 settings, whereas a dual-output three-base analyzer uses $\mathcal{E}$6 settings for state tomography; process tomography still requires individual input preparations (Hošák et al., 2018).

The same resource issue appears in spin and solid-state platforms. In NV-center spin tomography, standard spin-qubit QST commonly measures transverse magnetization as a function of time and therefore requires many readout points per configuration. The alternative NV protocol based on time-independent observables and unitary pre-rotations reduces the measurements to the minimal informationally complete set, $\mathcal{E}$7 for a $\mathcal{E}$8-dimensional system, with one measurement per operator coefficient (Zhang et al., 2021). In CQED microwave tomography, the raw observable is not a quadrature directly but the number $\mathcal{E}$9 of atoms detected in the upper working state in a series of $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$0 atoms; quadrature information is extracted statistically from repeated series and then deconvolved from the instrumental function (Miroshnichenko, 2015).

A related operational bottleneck is reconfiguration time rather than informational completeness. In photonic tomography, the sequential arrangement of wave-plate settings can be optimized as a traveling salesman problem. For the six-state scheme, the speedup factor grows monotonically with the number of qubits and is approximately $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$1 from $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$2; experimentally, full quantum process tomography of a three-qubit controlled-controlled-phase gate was reduced from approximately $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$3 hours to less than $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$4 hours without changing the estimator (Hošák et al., 2018).

3. Measurement compression, ancillas, and direct parameter access

Several procedures reduce the number of settings by trading them for ancillas, weak couplings, phase retrieval, or tailored unitary mappings.

Procedure	Core mechanism	Stated resource property
Standard QPT	$\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$5 inputs plus QST on each output	$\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$6 in $\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$7-qubit NMR
AAPT	Encode the channel into a joint system–ancilla state	$\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$8
SSPT	Combine AAPT with AAQST	$\mathcal{E}:\rho \mapsto \mathcal{E}(\rho)$9
Direct weak-measurement QPT	Two weak measurements and one post-selection	Each parameter is determined from only five experimental values
FQPT	Near-field/far-field distributions plus phase retrieval	7 measurements total; 5 minimal-in-principle

The ancilla-assisted route is built around the Choi state. In AAPT, the system and an ancilla of the same dimension are prepared in a maximally entangled state, the process acts on the system only, and a single QST of the joint state yields $d^2$0. AAQST further maps non-commuting observables of a target density matrix onto a commuting set in a larger Hilbert space, allowing all required expectations to be acquired in a single collective measurement. SSPT combines the two: one state preparation, one application of $d^2$1, and one collective measurement of commuting observables on system plus ancillas produce complete process characterization (Shukla et al., 2014).

The direct weak-measurement scheme replaces full output tomography by a parameter-local relation. For each tuple $d^2$2, the process parameter $d^2$3 is obtained from two weak measurements, one post-selection, four joint pointer correlations $d^2$4, and the post-selection probability $d^2$5. The key intermediate quantity is

$d^2$6

and the paper’s central claim is that each parameter of the process is directly determined from only five experimental values. The number of experimental setups is $d^2$7, because one input setting yields all $d^2$8 parameters associated with that input in parallel (Zhang et al., 2013).

A different compression mechanism appears in single-qubit reaped tomography. There the system is coupled to a single-qubit pointer initialized in $d^2$9, the system is measured in a computational basis $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$0, and standard qubit tomography on the pointer yields the data needed for reconstruction. The scheme requires the measurement of only three observables acting jointly on system and pointer, regardless of system size, and the exact pure-state reconstruction reduces to a linear system for amplitude ratios; for finite data and mixed states, the paper gives an iterative maximum-likelihood algorithm that scales via MPS/MPO structure (Choi, 2022).

Fourier Quantum Process Tomography achieves compression in a different regime. For a space-dependent SU(2) process, it measures probability distributions in two Fourier-conjugate spaces and uses phase retrieval to recover the local unitary. The practically recommended set is 7 measurements total: 3 near-field distributions, 3 far-field distributions, and 1 far-field distribution without polarization projection. The paper reports this as a near-minimal set, benchmarked against standard maximum-likelihood QPT (Colandrea et al., 2023).

4. Temporal, dynamical, and non-Markovian generalizations

When the device has memory, the object to be reconstructed is no longer a single channel. Three distinct generalizations appear in the cited literature.

The first is dynamical tomography with known evolution and a fixed measurement setup. The procedure subjects the unknown state to a known time evolution for different times and performs the same measurement each time. For a suitable unitary dynamics on an $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$1-dimensional system, any two states can be discriminated by a measurement with $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$2 outcomes at $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$3 points in time. If prior information restricts the state to a semi-algebraic subset $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$4, then $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$5 time steps suffice when $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$6; beyond unitary dynamics, feasible CPTP dynamics reduce the measurement setup to the minimal number of two outcomes (Kech, 2016).

The second is temporal quantum tomography via quantum reservoir computing. Here the target is a temporal map acting on a stream of quantum states with finite memory. The reservoir obeys

$$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$7

local observables define a feature vector $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$8, and a linear readout $$\mathcal{E}(\rho)=\sum_k K_k \rho K_k^\dagger,\qquad \sum_k K_k^\dagger K_k = I.$$9 reconstructs the output density matrix. The method is explicitly approximate, but in the reported benchmark tasks it achieved RMSF above $\{E_m\}$0 with $\{E_m\}$1 up to $\{E_m\}$2 qubits, $\{E_m\}$3 up to $\{E_m\}$4, training lengths $\{E_m\}$5, and evaluation horizons $\{E_m\}$6 (Tran et al., 2021).

The third is process tensor tomography. Rather than infer an effective recurrence, it reconstructs the full multi-time object $\{E_m\}$7 satisfying positivity and recursive causality constraints such as

$\{E_m\}$8

This fully characterizes non-Markovian dynamics over a time frame and predicts outcomes for arbitrary intervention sequences within the control span. Because the setting count scales as $\{E_m\}$9 for an $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$0-element basis over $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$1 steps, the paper also develops a finite-Markov-order variant in which the experimental scaling reduces from $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$2 to $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$3 when the effective memory length is $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$4 (White et al., 2021).

These extensions shift the role of tomography from static certification to dynamical diagnostics and control. A stated implication is that memory-aware reconstruction can be used to optimize subsequent circuits: on superconducting devices, full process-tensor-based predictive control improved average final-state fidelity by approximately $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$5, with maxima around $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$6, and finite-$$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$7 models performed better when neighbor-induced ZZ crosstalk increased effective memory (White et al., 2021).

5. Inference, physicality, and uncertainty quantification

Tomography procedures differ as much in their estimators as in their measurement designs. The supplied literature covers at least six inference strategies.

Evidence-based tomography introduces a prior expectation state $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$8 and an unknown confidence hyperparameter $$\mathcal{E}(\rho)=\sum_{m,n}\chi_{mn} E_m \rho E_n^\dagger,$$9, with prior

$\chi$0

The hyperparameter is selected by maximizing the evidence, and the MAP estimator becomes an entropic interpolation between the MaxEnt state consistent with measured expectations and the anticipated state. The method is efficient when the measured observables are not necessarily informationally complete but are numerous enough and broadly compatible with the prior; it yields error bars and avoids zero-eigenvalue artifacts of small-sample maximum likelihood (Rau, 2010).

Reliable tomography in the frequentist sense replaces a single estimate by a confidence region. The Christandl–Renner construction defines a likelihood-proportional density $\chi$1 over states and forms a region $\chi$2 by taking a high-probability set in $\chi$3 and thickening it in fidelity by

$\chi$4

The resulting confidence region has coverage at least $\chi$5, independently of any prior assumption on the distribution of possible states, and the construction applies to arbitrary measurements including fully coherent collective ones (Christandl et al., 2011).

A more recent line identifies iterative maximum-likelihood tomography with retrodictive recovery maps. For POVM tomography, the log-likelihood gradient coincides with the Petz recovery map of the measurement channel, and the update

$\chi$6

is exactly the symmetric R$\chi$7R step. In the commuting-output case, repeated Petz updates monotonically increase the likelihood; for channel tomography, the same construction lifts to the Choi matrix and is followed by a TP retraction (Murk et al., 22 Jun 2026).

Physicality enforcement is treated algorithmically in boosted projective methods. Starting from a Hermitian linear-inversion estimate $\chi$8, the Frobenius projection onto the CPTP set is approximated by alternating exact affine projections with positivity projections and a final Cholesky-based correction. For QPT up to four qubits, the combined Dykstra + CBA method achieved up to approximately three orders of magnitude reduction in Frobenius error to the exact SDP projection relative to earlier projective methods, at comparable runtime (Barberà-Rodríguez et al., 2024).

Two other inferential issues recur. First, SPAM bias: standard QPT assumes known state preparation and measurement, and the supplied literature includes an explicit correction procedure based on additional SPAM-calibration data, a gauge-regularized factorization of the SPAM error matrix $\chi$9, and corrected estimators such as

$$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$0

This correction applies to standard and overcomplete linear-inversion QPT and is designed to be extendable to MLE with CPTP constraints (Blume-Kohout et al., 2024). Second, incomplete identifiability: joint state-and-measurement tomography under imperfect measurements uses alternating maximum-likelihood estimation over unknown states and POVM parameters, and when the engineered measurements are not informationally complete it replaces point identification by SDP bounds on expectation values over the convex set of compatible states (Keith et al., 2018).

6. Platform realizations and operational limits

The procedures above are not tied to a single hardware class, and the cited papers document implementations or numerical protocols across NMR, trapped ions, superconducting devices, photonics, NV centers, CQED microwave cavities, and ultracold atoms.

In liquid-state NMR, SSPT was demonstrated on a three-qubit register in iodotrifluoroethylene on a $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$1 MHz spectrometer at $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$2 K. The system qubit and two ancillas enabled one-shot characterization of several single-qubit gates, with gate fidelities approximately $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$3 for NOP, NOT-X, NOT-Y, and Phase-$$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$4, $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$5 for Hadamard, and $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$6 for Phase-$$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$7; each full single-qubit SSPT experiment took less than $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$8 s from thermal equilibrium (Shukla et al., 2014). The same paper used SSPT to characterize a twirling channel and found measured $$J_{\mathcal{E}}=(I\otimes \mathcal{E})(|\Phi^+\rangle\langle\Phi^+|),$$9 and $d$00 curves tracking theory with average gate fidelity greater than $d$01 and errors below $d$02.

In NV-center tomography, the time-independent fluorescence observable

$d$03

is combined with unitary pre-rotations that map desired density-matrix components onto electron-spin population. The paper reports single-qubit fidelities $d$04, $d$05, $d$06, and $d$07 for representative states, and two-qubit fidelities $d$08, $d$09, $d$10, and $d$11 for prepared electron–nuclear states, using optimized control sequences of duration at most $d$12 (Zhang et al., 2021).

In photonic platforms, two distinct operational themes appear. One is measurement-order optimization, where full three-qubit process tomography of a controlled-controlled-phase gate was shortened from approximately $d$13 hours to less than $d$14 hours by reordering settings alone (Hošák et al., 2018). The other is FQPT, experimentally tested on space-dependent polarization transformations. In a 1D LCMS example, FQPT achieved average unitary overlap fidelity $d$15 with runtime about $d$16 minute, versus approximately $d$17 minutes for the standard maximum-likelihood comparison; in a 2D cascaded-LCMS example, the reported average overlap fidelity was approximately $d$18 (Colandrea et al., 2023).

Ultracold-atom field tomography uses matter-wave interferometry, even-order phase correlators, and a low-bond-dimension cMPS ansatz. For quenched quasi-condensates, a $d$19 reconstruction at $d$20 ms achieved mean relative deviations $d$21, $d$22, and $d$23; the degradation at $d$24 ms and $d$25 ms was interpreted as consistent with entanglement growth and the breakdown of a low-$d$26 cMPS description (Steffens et al., 2014). CQED microwave tomography is formulated differently: it reconstructs the quadrature tomogram from atomic detection statistics, but the measured result is the true quadrature distribution convolved with an instrumental Gaussian of variance $d$27; numerical simulations with $d$28, $d$29 or $d$30, and fitted $d$31 values $d$32 or $d$33 show the protocol’s effectiveness (Miroshnichenko, 2015).

Across platforms, the limitations recur with striking regularity. The cited papers name pulse imperfections, RF inhomogeneity, decoherence during the tomography sequence, spectrometer drift, imperfect gradients, finite shots, measurement back-action, ill-conditioned linear inversion, non-identifiability under incomplete measurements, model mismatch, ancilla overhead, spectral crowding, and the need for well-calibrated controls as the dominant obstacles (Shukla et al., 2014, Tran et al., 2021, Keith et al., 2018, Barberà-Rodríguez et al., 2024, Blume-Kohout et al., 2024). This suggests a unifying operational principle: quantum tomography procedures are most effective when the measurement design, estimator, and platform constraints are co-designed rather than treated as separable stages.