Quantum state tomography turns repeated measurement data into an estimated density matrix. The difficult part is not writing down enough equations; it is making sure the reconstructed matrix is physically valid, statistically justified and informative for the question that motivated the experiment.
A quantum state is not directly observable as a list of amplitudes. We infer it from measurement statistics collected over many nominally identical preparations. Quantum state tomography is the family of methods that performs this reconstruction.
The subject sits at the meeting point of linear algebra, probability, estimation and quantum measurement. Earlier guides introduced density matrices, POVMs and expectation values. Here we reverse that pipeline: observed probabilities become constraints on an unknown density operator. Informationally complete measurements determine whether the data can, in principle, identify the state. Statistical reconstruction determines what estimate should be reported from finite noisy samples.
Prepare repeatedly → measure in a complete set of ways → estimate probabilities → solve for ρ → enforce physicality → quantify uncertainty → validate on held-out or predicted observables.
1. What tomography is trying to reconstruct
A d-dimensional quantum state is represented by a density matrix ρ satisfying ρ=ρ†, ρ≥0 and Trρ=1. Hermiticity gives real expectation values. Positivity prevents negative probabilities. Unit trace normalises total probability.
A Hermitian d×d matrix has d² real parameters. The trace-one constraint removes one, leaving d²−1 independent real parameters for a generic state.
For n qubits, d=2ⁿ, so the generic parameter count is 4ⁿ−1. Full tomography therefore scales exponentially with qubit number unless additional structure is known and exploited.
2. One qubit: the Bloch-vector representation
Every qubit state can be written
ρ=(I+r_xX+r_yY+r_zZ)/2.
The Bloch vector r=(r_x,r_y,r_z) satisfies |r|≤1. Pure states lie on the unit sphere; mixed states lie inside.
The components are expectation values:
r_x=Tr(ρX)=⟨X⟩r_y=Tr(ρY)=⟨Y⟩r_z=Tr(ρZ)=⟨Z⟩
Thus measuring X, Y and Z on separate ensembles is enough to reconstruct any qubit state.
3. Worked qubit reconstruction
Suppose ideal expectation values are measured as
⟨X⟩=0.6⟨Y⟩=0.0⟨Z⟩=0.8
The Bloch-vector length is √(0.36+0.64)=1, so the state is pure. The density matrix is
ρ=(1/2)[[1.8,0.6],[0.6,0.2]]=[[0.9,0.3],[0.3,0.1]].
Its determinant is 0.09−0.09=0, confirming rank one.
4. From counts to expectation values
If a Pauli observable has outcomes +1 and −1, an empirical mean from m shots is
r_hat=(n_+−n_-)/m.
For example, if X measurement gives 790 plus outcomes and 210 minus outcomes in 1000 shots, then r_x≈0.58.
The estimator fluctuates from one dataset to another. Its variance in the ideal independent-shot model is (1-r_x²)/m. Tomography therefore produces an estimate with uncertainty, not an exact matrix simply because the algebraic inversion is exact.
5. Informational completeness
A measurement scheme is informationally complete if its outcome probabilities uniquely determine every state in the chosen model class.
For a POVM {E_k}, probabilities satisfy p_k=Tr(ρE_k). Informational completeness means the linear map from ρ to the probability vector has no nonzero Hermitian trace-zero operator in its kernel.
Equivalently, the measurement effects span the space of Hermitian operators. A d-dimensional system therefore requires enough linearly independent probability constraints to determine d²−1 state parameters after normalisation.
6. Minimal informational completeness versus convenient overcompleteness
A minimal informationally complete POVM can use d² outcomes. But minimal outcome count is not always optimal experimentally.
Overcomplete measurement sets can provide redundancy useful for noise averaging, calibration checks and robust estimation. Measuring the three Pauli bases on a qubit gives six raw outcomes, though each two-outcome basis contributes only one independent probability after normalisation.
The correct design question is not “What is the smallest possible number of settings?” It is “Which measurements give stable reconstruction under the actual detector, noise model and quantity of interest?”
7. Linear inversion
Choose an operator basis {B_j} and write
ρ=Σ_j θ_jB_j.
The measurement probabilities become linear equations
p_k=Σ_j M_{kj}θ_j
where M_{kj}=Tr(E_kB_j). If M has full column rank on the state-parameter subspace, solve for θ using an inverse or pseudoinverse.
This is linear inversion. It is transparent and fast, making it excellent for diagnostics and small teaching examples.
8. Why linear inversion can produce an impossible state
Finite-sample frequencies are not exact probabilities. Independent estimation error in different settings can produce a reconstructed Bloch vector with length greater than one.
Suppose the empirical estimates are (0.8,0.6,0.4). Their squared length is 0.64+0.36+0.16=1.16. The corresponding linear-inversion matrix has a negative eigenvalue.
The algebra solved the measured equations, but the answer is not a valid density matrix. This is a statistical consistency problem, not a failure of quantum positivity itself.
Exact inversion of noisy frequencies can leave the physical state space.
9. Physical reconstruction as constrained estimation
A physical estimator restricts the answer to positive semidefinite trace-one matrices. One common strategy is maximum likelihood: choose the state ρ that maximises the probability of the observed counts under the measurement model.
For independent multinomial data, the log-likelihood has the form
ℓ(ρ)=Σ_k n_k log Tr(ρE_k)
up to additive constants, with measurement settings handled in their respective groups.
The optimisation is subject to ρ≥0 and Trρ=1. This produces a physical state by construction.
10. Parameterising positivity
A common numerical parameterisation is
ρ=T†T/Tr(T†T)
for a complex triangular or otherwise parameterised matrix T. Positivity is automatic because T†T is positive semidefinite.
This changes an optimisation over constrained matrices into an optimisation over unconstrained matrix entries plus the nonlinear normalisation. It is convenient but can introduce numerical geometry different from direct convex optimisation over ρ.
Do not confuse a convenient parameterisation with a unique physical representation. Many T matrices can correspond to the same ρ.
11. Maximum likelihood can sit on the boundary
If data are close to a pure state, the likelihood-maximising physical estimate can have one or more zero eigenvalues. The state space has a boundary consisting of rank-deficient density matrices.
Standard asymptotic Gaussian error formulas can become unreliable near such boundaries because the parameter space is no longer locally an unconstrained Euclidean region.
This is one reason uncertainty quantification in tomography deserves its own analysis rather than being inferred mechanically from the inverse Hessian of an optimiser.
12. Bayesian tomography
A Bayesian approach combines a prior distribution over physical states with the likelihood of the observed data. The posterior is proportional to prior times likelihood.
Posterior means, credible regions and predicted observables can then incorporate parameter uncertainty directly. Results depend on the prior, especially with limited data.
Bayesian reconstruction is not automatically “more physical” than maximum likelihood. Both can respect positivity. They answer different inferential questions and encode assumptions differently.
13. Two-qubit Pauli tomography
Any two-qubit state can be expanded as
ρ=(1/4)Σ_{P,Q∈{I,X,Y,Z}} c_{PQ} P⊗Q
with c_{PQ}=Tr(ρP⊗Q) and c_{II}=1.
Thus fifteen nontrivial Pauli expectations determine a generic two-qubit state. In practice, local basis settings such as XX, XY, XZ and so on can estimate several compatible correlators from the same recorded bit strings.
The count grows exponentially: n-qubit Pauli tomography contains 4ⁿ−1 nontrivial coefficients.
14. Worked Bell-state signatures
For |Φ+⟩=(|00⟩+|11⟩)/√2, important Pauli expectations are
⟨XX⟩=1⟨YY⟩=-1⟨ZZ⟩=1- single-qubit X, Y and Z expectations equal 0
The density matrix can be written
ρ=(I⊗I+X⊗X−Y⊗Y+Z⊗Z)/4.
These few correlators suffice because we already know the target family. Full tomography is unnecessary if the scientific question is only whether a particular Bell-state fidelity exceeds a threshold.
15. Tomography versus targeted estimation
Full tomography reconstructs a general state. Targeted estimation asks for a smaller set of observables or properties. The second problem can require dramatically fewer measurements.
If the only objective is energy of a local Hamiltonian, measure the Hamiltonian terms. If the only objective is fidelity with one stabilizer state, exploit its stabilizers. If the objective is to predict many observables chosen after measurement, classical-shadow methods may be appropriate.
Guide 23 develops that alternative. Full reconstruction should be justified by the need for a full state model, not selected by default because it sounds comprehensive.
16. Sample complexity and dimensionality
A generic d-dimensional density matrix contains d²−1 parameters, so full tomography cannot evade dimensional growth without structural assumptions.
Required sample counts depend on the loss function, desired confidence, rank, measurement design and estimator. Estimating every matrix entry to a fixed coordinate-wise precision is a different problem from estimating trace distance or fidelity.
Statements such as “tomography needs 4ⁿ measurements” should therefore be interpreted carefully: 4ⁿ describes parameter or setting scales in common schemes, while the number of repeated shots needed for a statistical accuracy target adds another layer.
17. Low-rank and compressed tomography
If prior knowledge suggests the state has low rank, fewer parameters may be relevant. Compressed-sensing approaches exploit low-rank structure under suitable measurement conditions.
This is a model assumption. If the true state has significant spectral weight outside the assumed low-rank subspace, aggressive compression can bias the reconstruction.
A sensible workflow compares rank-restricted and unrestricted or weakly regularised models and checks predictive performance on observables not used to fit the state.
18. Detector calibration and SPAM errors
Tomography assumes a measurement model. If the detector effects Ek are miscalibrated, the reconstructed ρ can absorb detector errors and become a biased description of the source.
State-preparation and measurement errors are often abbreviated SPAM. Ordinary state tomography cannot cleanly distinguish source-state error from measurement-model error when the measurement calibration is treated as fixed and possibly wrong.
Gate-set tomography and self-consistent methods address parts of this identifiability problem by estimating gates, states and measurements together, but at the cost of a larger model and gauge freedoms.
19. Bootstrap and resampling
A practical uncertainty method is parametric bootstrap. Fit a physical state, simulate new datasets from the fitted measurement model, reconstruct each simulated dataset and examine the spread of the resulting quantities.
This captures estimator nonlinearity better than a simple linear error propagation in many small problems, though it assumes the fitted model is an adequate generator of repeated data.
Bootstrap intervals are computational estimates, not exact coverage guarantees under every boundary or model-mismatch condition.
20. Validate the reconstruction by prediction
A reconstructed state should predict measurement statistics not used trivially in the inversion. Hold out some settings, or compare additional observables with their independently measured frequencies.
If a state fits the tomography settings but fails badly on independent checks, possible causes include drift, model mismatch, detector error or non-identical preparations across time.
Tomography should therefore return to the experiment. A matrix that is positive and trace one is necessary, but physical admissibility alone does not certify model adequacy.
21. Common misconception: tomography measures the wavefunction directly
No. It estimates a state model from repeated measurement statistics. A single unknown system cannot be fully read out without being disturbed, and a global phase of a pure state is never an observable state parameter.
22. Common misconception: a physical density matrix must be the true state
Positivity and trace one only define admissibility. Many physical states can fit finite noisy data nearly equally well. Uncertainty and model checking remain necessary.
23. Common misconception: more settings always improve tomography
Additional settings can improve identifiability and redundancy, but they consume experimental time. If total shots are fixed, spreading them across many weakly informative settings can increase variance. Experimental design should allocate effort according to the target loss and model.
24. Worked synthesis problem
A qubit experiment returns empirical Pauli means x=0.72, y=0.10, z=0.75.
Step 1: Linear inversion. Form ρ_lin=(I+0.72X+0.10Y+0.75Z)/2.
Step 2: Physicality check. The Bloch length is √(0.72²+0.10²+0.75²)=√1.0909≈1.0445, so the estimate is outside the Bloch ball.
Step 3: Diagnose. This does not prove the experiment violated quantum mechanics. Finite sampling, calibration error or drift can create inconsistent independent empirical means.
Step 4: Reconstruct physically. Use a constrained likelihood or another physical estimator instead of reporting the negative-eigenvalue matrix as a state.
Step 5: Validate. Compare the fitted state against held-out measurement frequencies and report uncertainty on any downstream fidelity or expectation value.
25. Practice set
- How many independent real parameters does a generic d-dimensional density matrix contain?
- Write the Bloch-vector form of a qubit state.
- How are Bloch components estimated from Pauli measurements?
- What does informational completeness mean?
- Why can linear inversion produce a non-positive matrix?
- What constraints define a physical density matrix?
- What is the purpose of maximum-likelihood tomography?
- Why does full n-qubit tomography scale exponentially in the generic case?
- Give one reason to prefer targeted observable estimation over full tomography.
- What are SPAM errors?
- Why can low-rank tomography be biased?
- Give one way to validate a fitted state beyond checking positivity.
Answers
d²−1.ρ=(I+r·σ)/2.- By empirical means of ±1 outcomes in the X, Y and Z bases.
- The measurement probabilities uniquely determine the state within the chosen model class.
- Independent finite-sample frequency errors may be inconsistent with positivity.
- Hermiticity, positive semidefiniteness and unit trace.
- To choose a physical state that best explains the observed counts under the specified measurement model.
- The generic parameter count is
4^n−1. - If only a few observables matter, estimating them directly can require far fewer samples.
- State-preparation and measurement errors.
- The true state may have significant weight outside the assumed rank.
- Predict held-out measurement settings or independently measured observables.
Sources and further study
[1] D. F. V. James, P. G. Kwiat, W. J. Munro and A. G. White, Measurement of qubits. A widely used treatment of practical density-matrix reconstruction and maximum-likelihood methods.
[2] Matthias Paris and Jaroslav Řeháček, editors, Quantum State Estimation. A collection covering statistical and physical aspects of quantum-state reconstruction.
[3] Artur Czerwinski, Quantum state tomography with informationally complete POVMs generated in the time domain. An example of informational completeness engineered through measurement dynamics.
Continue through Quantum Mathematics
Guide 22: Quantum Process Tomography, Choi Matrices and Channel Verification reconstructs transformations rather than states. Guide 23: Classical Shadows, Randomised Measurements and Observable Prediction avoids full reconstruction when many selected observables are the real target. Guide 24: Randomised Benchmarking, Gate Fidelity and Error Budgets measures operational performance without reconstructing an entire process matrix.