Quantum State Tomography

• Quantum state tomography is a method to reconstruct a quantum system's density matrix from measurement data, enabling device verification and benchmarking.
• Techniques such as compressed sensing, tensor-network methods, and maximum-likelihood estimation address the exponential growth of parameters in multi-qubit systems.
• Innovative measurement designs and learning-based algorithms offer efficient, structured approaches for both full and selective state reconstruction.

Quantum state tomography (QST) is the task of inferring the state of a quantum system by appropriate measurements, or, equivalently, reconstructing an unknown quantum state $\rho$ from measured data. In the standard formulation, QST reconstructs the density matrix by estimating measurement probabilities such as $y_m = \operatorname{Tr}(\rho E_m)$ or $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$, depending on the measurement model. It is described as the gold standard for verification and benchmarking of quantum devices, but its central difficulty is the exponential growth of parameters and measurement resources with system size: for an $N$-qubit system, the density matrix is of size $2^N \times 2^N$; full QST may require $4^n-1$ observables, $3^N$ Pauli measurement settings, or, in the most direct count, $4^n$ independent density-matrix entries (Sofi et al., 30 Jun 2025, Xin et al., 2016, Stricker et al., 2022, Binosi et al., 2024).

1. Core formulation and the scalability problem

In its conventional form, QST estimates a density operator from repeated measurements on identically prepared copies. Several equivalent formulations appear across the literature. In compressed-sensing QST, one solves

$$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$

where $\hat{\rho}$ is constrained to Hermitian, PSD, and unit-trace matrices (Sofi et al., 30 Jun 2025). In informationally complete measurement schemes, the outcome distribution itself determines the state through relations of the form

$y_m = \operatorname{Tr}(\rho E_m)$0

so that learning the measurement distribution is equivalent to learning $y_m = \operatorname{Tr}(\rho E_m)$1 (Zhong et al., 2022).

The principal obstruction is dimensional. Standard QST becomes impractical because the number of quantum measurements and the amount of computation required to process them grows exponentially in the system size, and both memory and computational bottlenecks become severe even for moderately sized systems (Cramer et al., 2011, Sofi et al., 30 Jun 2025). This is why a large part of the modern literature does not treat tomography as a single protocol, but as a family of protocols adapted to structure: low rank, matrix product structure, tensor-network compressibility, locality, restricted observability, continuous measurement records, or fixed-shot property estimation.

A further aspect of the basic formulation is physicality. Density matrices must be positive semidefinite and trace one, and many practical reconstruction methods are designed to enforce these constraints by construction rather than by a posteriori correction. Cholesky-type parameterizations, lower-triangular factorizations, and tensor-network factorizations are all used for this purpose in different settings (Ikuta et al., 2017, Sofi et al., 30 Jun 2025).

2. Informational completeness, locality, and identifiability

A central question in QST is not only how to reconstruct a state, but whether the measured data determine it uniquely. Full informational completeness is the strongest setting, but many efficient schemes rely on reduced or restricted data. One prominent example is tomography via local reduced density matrices (RDMs). Measuring all weight-1 and weight-2 Pauli operators can be much more manageable than full tomography, but this only works when the global state is uniquely determined by its local RDMs (Xin et al., 2016).

That work distinguishes two notions of uniqueness. A state is UDA if it is uniquely determined among all states by its RDMs, and UDP if it is uniquely determined among pure states only. It presents explicit 4-qubit examples of states that are UDP but not UDA for their 2-RDMs, thereby disproving the conjectured implication UDP $y_m = \operatorname{Tr}(\rho E_m)$2 UDA. The classification leads to three classes: states that are neither UDP nor UDA, states that are both UDP and UDA, and states that are UDP but not UDA. The practical consequence is precise: for Class B, 2-RDM tomography is sufficient; for Class C, it is sufficient only if global purity is assumed; for Class A, full QST is necessary (Xin et al., 2016).

Locality also appears in tensor-network tomography. For one-dimensional systems well approximated by a matrix product state (MPS) ansatz, local reductions on blocks of $y_m = \operatorname{Tr}(\rho E_m)$3 consecutive sites can determine the global state generically, and only a linear number of experimental operations is required (Cramer et al., 2011). This suggests that “efficient tomography” is often inseparable from a structural identifiability statement: without a guarantee that local or compressed data determine the global object, reduced-measurement protocols become ambiguous rather than merely approximate.

Another route to informational completeness replaces many settings with a single generalized measurement. Symmetric informationally complete (SIC) POVMs are described as mathematically optimal measurements for tomography, requiring only $y_m = \operatorname{Tr}(\rho E_m)$4 outcomes for a $y_m = \operatorname{Tr}(\rho E_m)$5-dimensional quantum system—the absolute minimal informationally complete measurement. In that setting, complete tomographic information is contained in every single experimental shot (Stricker et al., 2022).

3. Reconstruction principles and estimators

Reconstruction procedures in QST range from direct linear inversion to constrained optimization and likelihood-based inference. Maximum-likelihood estimation (MLE) remains a standard approach because linear inversion can yield unphysical density matrices. A representative parameterization is

$y_m = \operatorname{Tr}(\rho E_m)$6

with $y_m = \operatorname{Tr}(\rho E_m)$7 lower triangular, so positivity is enforced by construction (Ikuta et al., 2017). In photonic time-bin tomography, MLE is used together with a likelihood

$y_m = \operatorname{Tr}(\rho E_m)$8

where $y_m = \operatorname{Tr}(\rho E_m)$9 are measured counts and $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$0 are model expectations (Ikuta et al., 2017).

For realistic experiments, the measurement process may not be described by an instantaneous POVM. In that case, trajectory-based MaxLike tomography replaces static POVM elements by trajectory-dependent effective operators $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$1 computed from the adjoint quantum filter. The generalized likelihood takes the form

$$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$2

and confidence intervals for observables are obtained from an asymptotic expansion of multidimensional Laplace integrals (Six et al., 2015). This extends QST to continuous-time signals, measurement imperfections, and decoherence, while keeping the inferential object an initial quantum state.

Statistical reliability is a separate issue from point estimation. Confidence-region tomography constructs subsets of state space that contain the true state with probability at least $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$3, independently of any prior assumption on the distribution of possible states. The construction proceeds through a data-dependent distribution

$$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$4

followed by a “fattening” step using fidelity balls of radius $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$5, where

$$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$6

The resulting regions are operational, prior-free, and shrink as $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$7 (Christandl et al., 2011).

Neural-network reconstructions introduce a different estimator design. In Neural-Shadow Quantum State Tomography, the training loss is the infidelity

$$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$8

estimated from classical shadows rather than basis-dependent cross-entropy. The stated motivation is that cross-entropy does not correlate well with true distance, especially for phases, whereas infidelity is a natural loss for quantum states (Wei et al., 2023).

4. Structured-state tomography: low rank, tensor networks, and matrix completion

Many efficient QST protocols are efficient only on restricted state classes. Compressed sensing assumes that the unknown state is low rank or close to pure. In one formulation, only $$P(\boldsymbol{a}) = \mathrm{Tr}(M^{\boldsymbol{a}}\rho)$$9 measurements are needed for rank-$N$0 states, using random Pauli measurements and convex recovery (Sofi et al., 30 Jun 2025). Single-observable tomography uses the same low-rank principle but compensates for a lack of informational completeness by introducing an ancilla and random linear mixing, so that the recovery problem becomes low-rank matrix recovery under physical constraints (Oren et al., 2017).

Tensor-network methods replace the full density matrix by a compact ansatz. In one-dimensional systems, states well approximated by MPS of small bond dimension admit tomography from local measurements with classical postprocessing polynomial in system size, and one of the two main schemes comes with rigorous certification without any a priori assumptions (Cramer et al., 2011). A more recent tensor-network development parameterizes the density matrix using a low-rank Block Tensor Train (Block-TT) decomposition,

$N$1

where $N$2 is a Block-TT tensor with a shared index of size $N$3. When TT-rank $N$4, the parameter count scales as $N$5, and for states well approximated by low TT-rank the effective scaling is $N$6 rather than $N$7. The parameterization also inherently preserves PSD (Sofi et al., 30 Jun 2025).

Pure-state structure allows even more specialized reconstructions. Matrix-Completion QST reconstructs an $N$8-qubit pure state from only $N$9 local Pauli measurement settings. The method uses the fact that a rank-1 density matrix can be completed from $2^N \times 2^N$0 entries, together with the rule

$2^N \times 2^N$1

derived from vanishing $2^N \times 2^N$2 minors. The protocol measures all diagonal elements in the computational basis and selected off-diagonal elements via local $2^N \times 2^N$3 and $2^N \times 2^N$4 basis measurements, then fills the rest algebraically (Farooq et al., 2021).

Approximate tomography can also exploit sparsity in a chosen basis. Threshold Quantum State Tomography first measures all diagonal elements $2^N \times 2^N$5, then measures only those off-diagonal elements $2^N \times 2^N$6 for which $2^N \times 2^N$7, using the PSD inequality $2^N \times 2^N$8. With $2^N \times 2^N$9, the protocol reduces to standard QST; for $4^n-1$0, many off-diagonal elements are omitted. The paper gives a lower bound on the fidelity between the true and thresholded reconstructions: $4^n-1$1 This suggests a resource-accuracy tradeoff controlled directly at the level of matrix entries (Binosi et al., 2024).

The following summary organizes the structured-state approaches described in the literature.

Approach	Structural assumption	Resource statement
Compressed sensing / low rank	Low rank or nearly pure state	$4^n-1$2 measurements for rank-$4^n-1$3 states
MPS / tensor networks	Small bond dimension, local correlations	Linear number of experimental operations; polynomial postprocessing
Block-TT QST	Low TT-rank, TT-compressible density matrix	$4^n-1$4 effective scaling
Matrix-completion QST	Pure state, rank-1 density matrix	$4^n-1$5 local Pauli settings
Threshold QST	Small or negligible off-diagonals in a chosen basis	Basis-dependent measurement reduction controlled by threshold $4^n-1$6

5. Measurement design innovations

A major research direction in QST concerns measurement design: reducing the number of settings, adapting to hardware constraints, or extracting full information from unconventional observables. Several proposals replace large setting families by a single setting or a fixed small number of settings.

Single-setting SIC tomography implements a local SIC POVM on trapped ions by mapping each qubit SIC POVM to orthogonal states embedded in a higher-dimensional system, read out using repeated in-sequence detections and without ancilla qubits. Combined with classical shadows, this yields a fixed measurement basis with full tomography information in every shot and enables estimation of arbitrary polynomial functions of the density matrix orders of magnitudes faster than standard methods (Stricker et al., 2022).

Single-observable tomography pursues a related objective from a different angle. A single observable is generally insufficient, so the protocol enlarges the system by an ancilla and applies random linear mixing. Recovery then uses either rank minimization heuristics or constrained least squares under Hermiticity, positivity, and trace-one constraints (Oren et al., 2017). A later auxiliary-system method reduces the number of settings to two by entangling the target system with a quantum auxiliary system or correlating it with a probabilistic classical auxiliary system. The stated sampling complexity is $4^n-1$7, and the method includes two purity-measurement schemes, one of which achieves measurement precision at the Heisenberg limit (Zhao et al., 22 Jul 2025).

Measurement design can also be tailored to experimental access restrictions. In a qubit-qutrit system, QST is formulated as a numerical optimization problem over a quorum of rank-3 projectors, with the geometric quality measure

$4^n-1$8

as objective. This framework accommodates subsystem-only measurements and numerically approximates mutually unbiased subspaces in dimension six with a deviation described as irrelevant for practical applications (Ivanova-Rohling et al., 2020).

Some measurement schemes are architecture-specific. Time-bin qudit tomography uses cascaded Mach-Zehnder interferometers with binary delays, where the number of settings scales as $4^n-1$9 for $3^N$0. For $3^N$1, the implementation used 16 measurement settings and reconstructed a maximally entangled time-bin state with average fidelity $3^N$2 (Ikuta et al., 2017). In NV centers, a fast protocol measures only a time-independent observable—the electron spin population—and maps off-diagonal density-matrix elements into directly measurable population differences via selective unitary operations, thereby removing the need for measurements at many evolution times (Zhang et al., 2021).

There are also measurement models in which the raw data are not conventional projective outcomes. Transport-based two-qubit tomography reconstructs the density matrix of an open system from currents, current cross-correlations, and current time-derivatives within a Markovian Lindblad framework. Populations are recovered analytically from currents and noise, while coherences are encoded in current derivatives (Bourgeois et al., 28 Jan 2025).

6. Learning-based, selective, and unconventional formulations

Recent work increasingly treats QST as a problem of statistical learning rather than explicit inversion of measurement equations. In one formulation, QST is recast as a language-modeling task. The unknown quantum state is treated as an unknown language, quantum correlations as semantic information, and IC-POVM measurement outcomes as token sequences

$3^N$3

modeled autoregressively by a customized GPT-2-style decoder-only transformer. The method learns $3^N$4 instead of $3^N$5 directly, and a single trained model can reconstruct a family of similar quantum states simultaneously (Zhong et al., 2022).

Selective Quantum State Tomography (SQST) changes the objective more radically. Rather than reconstructing the full state first, it estimates arbitrary elements of an unknown quantum state using a fixed measurement record. For a single matrix element, the stated sample complexity is

$3^N$6

independent of the dimension $3^N$7, using measurements derived from mutually unbiased bases and estimators built from phase factors $3^N$8. The same fixed and dimension-independent data sample can be used to estimate mean values from a continuous class of bounded operators on demand (Morris et al., 2019).

Classical-shadow-based neural tomography occupies an intermediate position between full reconstruction and selective estimation. Neural-Shadow QST uses randomized measurements to build shadow objects $3^N$9, then trains a neural-network ansatz by minimizing estimated infidelity rather than basis cross-entropy. The reported advantage is improved learning of relative phases and robustness against various types of noise without any error mitigation (Wei et al., 2023).

Another unconventional formulation is single-qubit reaped tomography. The system is coupled to a single ancillary qubit pointer via

$4^n$0

and after measuring the system in the computational basis, the pointer state contains the reaped wavefunction information. Standard two-state tomography on the pointer then yields the system state using only three joint observables,

$4^n$1

regardless of system size. An iterative maximum-likelihood algorithm is provided for statistically incomplete data (Choi, 2022).

Finally, successive-measurement tomography reconstructs finite-dimensional states from sequences of two measurements: projectors $4^n$2 and $4^n$3 onto two bases with no mutually orthogonal vectors. The protocol measures position-position and momentum-position correlations of two von Neumann meters, which encode the real and imaginary parts of a joint quasi-probability $4^n$4. The density matrix is then recovered by an explicit inversion formula (Kalev et al., 2012).

7. Experimental validation, performance criteria, and scope

Experimental studies emphasize that “tomography” is not a single benchmark but a combination of reconstruction accuracy, sample complexity, classical cost, and suitability to the underlying platform. Reported performance criteria include fidelity, trace distance, purity, Rényi entropies, and convergence of observable estimates.

Several platforms are represented in the literature. A 4-qubit NMR processor was used to test tomography via RDMs, including robustness under artificially added Gaussian noise, with reconstructed states maintaining fidelities $4^n$5 for the reconstructible classes (Xin et al., 2016). A trapped-ion implementation of SIC tomography demonstrated online QST in real time on an 8-qubit entangled state and entanglement analysis on a 5-qubit absolutely maximally entangled state (Stricker et al., 2022). Time-bin photonic tomography reconstructed a 4-dimensional maximally entangled state with average fidelity $4^n$6 over 15 trials (Ikuta et al., 2017). An NV-center protocol reported single-qubit fidelities $4^n$7, $4^n$8, $4^n$9, and $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$0 for representative states, and two-qubit fidelities $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$1, $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$2, $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$3, and $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$4 (Zhang et al., 2021). Matrix-completion tomography on IBMQ-Casablanca achieved $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$5 fidelity for 3-qubit GHZ states (Farooq et al., 2021). Threshold QST was demonstrated up to 7 qubits, with accurate reconstruction using only 170 measurements for a 7-qubit W state, compared with $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$6 in the conventional count (Binosi et al., 2024). Transport tomography is presented as analytically complete for two qubits in open solid-state systems when currents, noise, and derivative data are available (Bourgeois et al., 28 Jan 2025). Block-TT tomography remained tractable up to $$\min_{\hat{\rho}} \|\hat{\rho}\|_* \quad \text{s.t.} \quad \|\vec{y} - \mathcal{M}(\hat{\rho})\|_2 \leq \epsilon,\; \hat{\rho} \in \mathcal{S},$$7 qubits and reported computational time for the measurement contraction step growing linearly with the number of measurements (Sofi et al., 30 Jun 2025).

A recurrent theme across these results is that efficiency always depends on what is being asked of the protocol. Full, assumption-free, universally valid tomography remains constrained by exponential scaling. The practical literature therefore separates at least four regimes. One regime seeks full reconstruction with rigorous statistical guarantees, as in confidence-region tomography (Christandl et al., 2011). A second exploits physical structure such as low rank, locality, or tensor-network compressibility (Cramer et al., 2011, Sofi et al., 30 Jun 2025). A third targets restricted tasks—selected matrix elements, purities, entropies, or entanglement witnesses—using universal data samples or classical shadows (Morris et al., 2019, Stricker et al., 2022). A fourth adapts tomography to nonstandard observables or hardware limitations, including subsystem-only access, continuous measurement records, and transport data (Ivanova-Rohling et al., 2020, Six et al., 2015, Bourgeois et al., 28 Jan 2025).

This suggests that, in contemporary usage, “quantum state tomography” denotes a broader inferential program rather than a single reconstruction recipe. The common object is still the quantum state, but the operational form of the problem—full or selective, exact or approximate, direct or learned, isolated or open-system—depends on the measurement model, the accessible controls, and the structural assumptions that can be justified for the state class under study.