Sparse Pure-State Tomography Overview

• Sparse pure-state tomography exploits a low-rank model and sparsity assumptions to reduce the measurement count compared to full quantum state tomography.
• Approaches such as compressed sensing, matrix completion, and circuit-based methods enable robust state recovery even with noise and sparse outliers.
• Techniques including minimal Pauli measurement designs, ADMM optimization, and tensor-train parameterizations illustrate practical trade-offs between resource efficiency and reconstruction fidelity.

Searching arXiv for the cited works to ground the article and verify metadata. Sparse pure-state tomography denotes a family of quantum state tomography protocols that exploit structural promises stronger than generic low-noise reconstruction. In the broadest usage found in the literature, the target is assumed to be pure or nearly pure, so that the relevant model class is low rank; in narrower usages, the state is promised to have only a small number of nonzero computational-basis amplitudes, or the measurement corruption is modeled by a sparse outlier term. For an $n$-qubit system with $d=2^n$, full tomography needs $4^n-1$ real parameters, whereas a normalized pure state depends on $2d-2$ real parameters, and this gap motivates measurement-minimal, compressed-sensing, matrix-completion, and circuit-based schemes (Ma et al., 2016).

1. Conceptual scope and uniqueness notions

A central distinction is between generic mixed-state tomography and tomography restricted to pure states. In the pure-state setting, one measures a collection of Hermitian observables $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$, producing outcomes

$\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$

Two uniqueness notions are standard. A pure state $|\psi\rangle$ is uniquely determined among all states (UDA) if no other density operator reproduces the same data, and it is uniquely determined among pure states (UDP) if no different pure state does so. If $\mathcal S(\mathbf A)$ is the real span of the measured observables and $\mathcal S(\mathbf A)^\perp$ its orthogonal complement in the Hermitian operator space, then $\mathbf A$ is UDA for every pure state if and only if every nonzero $d=2^n$0 has at least two positive and two negative eigenvalues (Ma et al., 2016).

This literature uses “sparse” in several technically distinct ways. One line treats purity itself as a rank-1 sparsity prior on the density matrix. Another assumes that only $d=2^n$1 computational-basis amplitudes are nonzero. A third augments a low-rank density matrix with a sparse outlier matrix in the measurement model. This suggests that “sparse pure-state tomography” is not a single algorithmic paradigm but a cluster of related structural-recovery problems (Li et al., 2014).

Approach	Structural assumption	Key claim
ADMM compressive tomography	$d=2^n$2 low rank, $d=2^n$3 sparse outliers	Recover pure or near pure states corrupted by sparse noise (Li et al., 2014)
Pauli UDA design	Arbitrary pure $d=2^n$4-qubit states	11 Pauli measurements for two qubits; 31 for three qubits (Ma et al., 2016)
MC-QST	Rank-1 density matrix	$d=2^n$5 local Pauli settings and direct matrix filling (Farooq et al., 2021)
Sparse-entry circuit tomography	$d=2^n$6 nonzero amplitudes	Identity plus at most $d=2^n$7 further measurements (Li et al., 2024)
Phase-estimation tomography	Unknown $d=2^n$8-sparse superposition	Support finding via phase estimation on a designed unitary (Gulbahar, 2021)
Two-basis algebraic tomography	$d=2^n$9-sparse algebraic pure state	Two measurement bases suffice asymptotically (Feng et al., 28 Jan 2025)

2. Convex compressed-sensing formulations and robust ADMM

In the low-rank-plus-outlier formulation of Li and Cong, one seeks a density matrix $4^n-1$0 and a sparse matrix $4^n-1$1 that jointly explain the measurements: $4^n-1$2 Here $4^n-1$3 is the nuclear norm, $4^n-1$4 is the entry-wise $4^n-1$5 norm, and $4^n-1$6 may be added if desired. The scaled augmented Lagrangian introduces a dual variable $4^n-1$7,

$4^n-1$8

with $4^n-1$9, and then splits the updates over $2d-2$0 and $2d-2$1. The $2d-2$2-step uses a least-squares solve followed by projection onto $2d-2$3 with low-rank shrinkage, while the $2d-2$4-step uses a least-squares solve followed by entry-wise soft thresholding. The dual update is

$2d-2$5

If the true rank $2d-2$6 is known, the algorithm keeps only the top $2d-2$7 positive eigenvalues in the shrinkage step (Li et al., 2014).

The stopping criteria monitor the primal residual $2d-2$8 and dual residuals $2d-2$9 against tolerances $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$0. Typical $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$1, $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$2, and 30–50 iterations suffice. The per-iteration cost is dominated by a least-squares solve and a $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$3 eigendecomposition in $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$4, with $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$5 implying eigendecomposition cost $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$6. The compressive advantage appears when $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$7 (Li et al., 2014).

The numerical findings reported for $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$8 qubits and rank $\mathbf A=\{A_0=I,A_1,\dots,A_{m-1}\}$9 are specific. For relative error $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$0, least squares requires $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$1 for 10% error, the Dantzig method needs $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$2, and ADMM needs $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$3. With sparse outliers affecting 1% of entries, Dantzig fails with error approaching 1, whereas ADMM still recovers with error $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$4 using $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$5. The stated interpretation is that the $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$6 penalty on $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$7 isolates heavily corrupted entries while preserving the low-rank $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$8 (Li et al., 2014).

3. Minimal Pauli measurement sets and pure-state identifiability

For Pauli tomography, the problem is to find a minimal subset of $\boldsymbol{\alpha}=\bigl(\Tr(\rho A_0),\Tr(\rho A_1),\dots,\Tr(\rho A_{m-1})\bigr).$9-qubit Pauli operators that is UDA for all pure states. On two qubits, full tomography uses 15 nontrivial Pauli operators, while Ma et al. showed that the 11-element set

$|\psi\rangle$0

suffices to UDA any two-qubit pure state, and no smaller Pauli set can do so for all two-qubit pure states. On three qubits, full tomography uses 63 nontrivial Paulis, and a set of 31 Paulis suffices to UDA every three-qubit pure state; again, one cannot use fewer than 31 Paulis for UDA of all pure states (Ma et al., 2016).

Known dimension-dependent bounds are sharper in the non-Pauli setting than in the Pauli setting. For a $|\psi\rangle$1-dimensional pure state, UDP requires $|\psi\rangle$2 observables, whereas UDA requires $|\psi\rangle$3 observables. For Pauli measurements on general $|\psi\rangle$4-qubit systems, no closed form is known. Compressed sensing shows that $|\psi\rangle$5 random Paulis suffice to identify almost all pure states, but this fails on a measure-zero set. The conjecture reported in the three-qubit work is that the minimal Pauli count scales linearly in $|\psi\rangle$6, with $|\psi\rangle$7 and $|\psi\rangle$8 for $|\psi\rangle$9; a constructive algorithm for general $\mathcal S(\mathbf A)$0 remains open (Ma et al., 2016).

The same work also analyzes robustness. Under the depolarizing channel

$\mathcal S(\mathbf A)$1

simulations on 100 Haar-random pure states show average fidelity remaining $\mathcal S(\mathbf A)$2 for $\mathcal S(\mathbf A)$3, with roughly linear decay as $\mathcal S(\mathbf A)$4 increases. In liquid-state NMR experiments, the two-qubit sparse protocol achieved typical fidelity $\mathcal S(\mathbf A)$5 relative to full tomography, and the three-qubit GHZ experiment achieved $\mathcal S(\mathbf A)$6. Random Pauli subsets of the same size performed worse: mean fidelity $\mathcal S(\mathbf A)$7 for 11 random Paulis in the two-qubit case and $\mathcal S(\mathbf A)$8 for 31 random Paulis in the three-qubit case, which the paper attributes to “failing sets” whose complements admit Hermitians with the wrong eigenvalue signature (Ma et al., 2016).

4. Matrix completion from local Pauli data

Matrix-Completion Quantum State Tomography (MC-QST) specializes pure-state tomography to the rank-1 case $\mathcal S(\mathbf A)$9. The algebraic basis is that all $\mathcal S(\mathbf A)^\perp$0 minors of a rank-1 matrix vanish: $\mathcal S(\mathbf A)^\perp$1 The measurement scheme uses only $\mathcal S(\mathbf A)^\perp$2 local settings: one computational-basis measurement and, for each qubit $\mathcal S(\mathbf A)^\perp$3, single-qubit measurements of

$\mathcal S(\mathbf A)^\perp$4

The computational basis yields all $\mathcal S(\mathbf A)^\perp$5 diagonal entries $\mathcal S(\mathbf A)^\perp$6, and the $\mathcal S(\mathbf A)^\perp$7 and $\mathcal S(\mathbf A)^\perp$8 measurements recover off-diagonal entries between basis states that differ in exactly one bit. The directly measured entries therefore total

$\mathcal S(\mathbf A)^\perp$9

which matches the spanning-tree style criterion quoted from Strang: any $\mathbf A$0 entries forming an acyclic row-column graph uniquely determine the full rank-1 matrix (Farooq et al., 2021).

Reconstruction proceeds in three steps. First, estimate the diagonal entries from the computational basis. Second, recover nearest-neighbor off-diagonals from the $\mathbf A$1 and $\mathbf A$2 data through

$\mathbf A$3

Third, fill in all remaining entries by rank-1 completion: $\mathbf A$4 Finally, renormalize by $\mathbf A$5 and extract the leading eigenvector. Because each fill-in is $\mathbf A$6 and there are $\mathbf A$7 entries, the classical post-processing cost is $\mathbf A$8 (Farooq et al., 2021).

The stated scaling is $\mathbf A$9 measurement settings, $d=2^n$00 estimated probabilities, and $d=2^n$01 total samples to reach mean-square error $d=2^n$02 on each probability. The method is exact in the noiseless rank-1 limit but does not provide a fully worked-out finite-sample error bound. Numerically, it outperformed contemporary pure-state tomography baselines on Haar-random states for $d=2^n$03, and on IBM “ibmq-casablanca” it reconstructed two-qubit pure states with median infidelities below 5%. For a three-qubit GHZ state, after a prerotation $d=2^n$04, the reported median fidelity was 97%. The principal limitation stated in the paper is that the method assumes exact purity; if $d=2^n$05 is slightly mixed, the minors do not vanish exactly and regularization becomes necessary (Farooq et al., 2021).

5. Computational-basis sparsity: circuit constructions and phase estimation

A different branch of sparse pure-state tomography assumes that only $d=2^n$06 computational-basis amplitudes are nonzero. In the circuit-based sparse-entry protocol, one first performs the identity measurement in the computational basis to learn the support set $d=2^n$07 and the magnitudes $d=2^n$08. One then performs $d=2^n$09 additional measurements $d=2^n$10 with $d=2^n$11, where each $d=2^n$12 consists of a short CNOT network $d=2^n$13 mapping a pair of support states $d=2^n$14 to basis states differing in exactly one bit, followed by either a Hadamard $d=2^n$15 or a phase-shifted Hadamard $d=2^n$16, $d=2^n$17, on the differing qubit. The support graph on $d=2^n$18 is weighted by Hamming distance, and a minimum-spanning tree determines the order of pairwise phase recovery. For an edge $d=2^n$19, the probabilities $d=2^n$20, $d=2^n$21, $d=2^n$22, and $d=2^n$23 determine the real and imaginary parts of $d=2^n$24, thereby recovering the unknown amplitude once the adjacent known amplitude is fixed (Li et al., 2024).

The resource counts are explicit. The protocol uses the identity measurement plus at most $d=2^n$25 further measurements, so the total is $d=2^n$26. If the minimum-spanning tree has $d=2^n$27 edges of weight $d=2^n$28, then the total CNOT count is bounded by

$d=2^n$29

The worst case is $d=2^n$30, but when $d=2^n$31 the minimum-spanning tree can be chosen entirely from weight-1 edges, so no CNOTs are needed. The single-qubit gate count is $d=2^n$32, the depth of each $d=2^n$33 is $d=2^n$34, and the overall depth is $d=2^n$35. Simulations in Qiskit/Aer with the IBM Brisbane noise model showed median fidelity $d=2^n$36 when support states were separated by Hamming distance 1, while a randomized backup strategy using $d=2^n$37 and only single-qubit gates often outperformed the unrandomized MST construction once the MST required at least two CNOTs. The same sparse-state embedding was also applied to process tomography, with reported fidelities $d=2^n$38 after a polar-decomposition correction (Li et al., 2024).

The phase-estimation approach to $d=2^n$39-sparse tomography assumes no prior knowledge of $d=2^n$40, the support bit strings, or the complex coefficients. It constructs a specially designed unitary $d=2^n$41 on $d=2^n$42 qubits whose eigenbasis encodes the computational labels, prepares the embedded state

$d=2^n$43

and uses standard quantum phase estimation with controlled powers $d=2^n$44. Measuring the phase register yields an estimate of an eigenphase with probability $d=2^n$45, and after applying $d=2^n$46 to the data register, a computational-basis measurement reveals the support string $d=2^n$47. Repeating this $d=2^n$48 times suffices, with high probability, to observe all $d=2^n$49 distinct supports, where $d=2^n$50 is the least amplitude magnitude appearing in the superposition. Once the support is known, the coefficient vector is reconstructed using conventional compressive-sensing tomography on the $d=2^n$51-dimensional reduced problem, requiring $d=2^n$52 settings with $d=2^n$53 and $d=2^n$54 a constant, independent of the ambient dimension $d=2^n$55. The paper emphasizes open problems concerning the existence of the required phase choices $d=2^n$56 and the efficient implementation of exponentially large powers $d=2^n$57 (Gulbahar, 2021).

6. Two-basis algebraic tomography, tensor-network parameterizations, and current open directions

The two-basis protocol of Feng et al. works with $d=2^n$58-sparse algebraic pure states

$d=2^n$59

where at most $d=2^n$60 amplitudes are nonzero and each nonzero amplitude is an algebraic number. The first basis is the computational basis $d=2^n$61, from which one reads off $d=2^n$62, the magnitudes $d=2^n$63, and the support $d=2^n$64. The second basis is a shifted-Fourier basis $d=2^n$65 with $d=2^n$66, where $d=2^n$67 is the quantum Fourier transform and $d=2^n$68 is chosen so that the differences $d=2^n$69 are distinct algebraic numbers in absolute value. The observed probabilities are

$d=2^n$70

Reconstruction fixes $d=2^n$71 and minimizes $d=2^n$72 over the remaining phases, for example by simulated annealing or another nonlinear solver. The uniqueness proof uses the Lindemann–Weierstrass theorem to show that the distinct algebraic phase differences force all coefficients in the induced trigonometric relations to vanish, so the relative phases are uniquely determined. The scheme is asymptotically informationally complete for pure states because every pure state can be approximated arbitrarily well by algebraic states, and for GHZ-like and W-like states it admits two local product bases using $d=2^n$73. Numerical tests with up to 20 qubits showed reconstructed-state fidelity $d=2^n$74 at small residual probability error and $d=2^n$75 fidelity under depolarizing noise up to $d=2^n$76 (Feng et al., 28 Jan 2025).

A complementary large-scale direction uses tensor-train parameterizations. In Tensor Train Quantum State Tomography using Compressed Sensing, the pure state $d=2^n$77 is written as a TT tensor with cores $d=2^n$78,

$d=2^n$79

and one solves

$d=2^n$80

The proposed algorithm is projected gradient descent in TT form with TT-SVD truncation and reorthogonalization. For pure-state tomography, the stated per-measurement complexity is $d=2^n$81, and storage is $d=2^n$82. The paper points to existing TT iterative hard-thresholding guarantees of $d=2^n$83 random Gaussian measurements for low-TT-rank tensors and conjectures comparable Pauli-measurement analogues. This is not a sparsity model in the computational basis; it is a low-TT-rank model that includes pure states, nearly pure states, and ground states of Hamiltonians (Sofi et al., 30 Jun 2025).

Recent optimization work has also shifted attention from reconstruction itself to certifying when a reduced measurement set uniquely specifies a pure state. The ALM framework for investigating pure-state uniqueness formulates UDP as a constrained nonconvex optimization over pure states and UDA as either an SDP or a low-rank constrained problem. A key theorem states that, given $d=2^n$84 non-identity measurements, any compatible density operator can be replaced by another compatible operator of rank $d=2^n$85; with $d=2^n$86 added as an extra observable, the rank bound becomes $d=2^n$87. Using the ensemble parameterization

$d=2^n$88

and ALM updates

$d=2^n$89

the method classified qutrit and four-qubit symmetric states into three categories: UDA, UDP but not UDA, and neither. This suggests that the design of sparse measurement sets is inseparable from the geometry of uniqueness, not merely from sample complexity (Wu et al., 2024).

Several open questions recur across these approaches. No closed-form minimal Pauli count is known for arbitrary $d=2^n$90-qubit pure states, and a scalable constructive algorithm remains open. MC-QST assumes exact rank 1 and must be stabilized for mixedness. The phase-estimation protocol leaves unresolved the structured implementation of $d=2^n$91. The two-basis algebraic scheme is asymptotic for generic pure states rather than exact in finite dimension. TT-based methods trade exact convexity for nonconvex low-rank optimization. Taken together, these lines of work define sparse pure-state tomography as a technically diverse program: reducing measurement count, circuit depth, and classical complexity by matching the tomography model to the structure actually present in the quantum state.