Overlapping State Tomography

• Overlapping state tomography is a method that reconstructs collections of overlapping reduced density matrices from an n-partite state using reusable global product measurements.
• It reduces the exponential measurement demands of full tomography by targeting local marginals, thus efficiently capturing key local quantum properties.
• The approach leverages combinatorial designs and convex optimization to enable precise, scalable, and experimentally realizable state reconstruction.

Overlapping state tomography denotes a family of tomographic protocols in which one reconstructs a collection of mutually overlapping reduced states, or experimentally accessible state overlaps, without performing full tomography of the global system. In the canonical quantum formulation, the objective is to determine all $k$-body reduced density matrices (RDMs) of an $n$-partite state by designing global product measurements that are simultaneously informative for many subsystems. The defining resource is measurement reuse: one parallel measurement round contributes data to multiple marginals because the target subsystems overlap in support. This places overlapping tomography between full-state tomography and independent local tomography, and it has developed into a distinct area spanning combinatorial measurement design, locality-aware protocols, convex post-processing, and several experimentally realized overlap-based optical and weak-measurement schemes (Cotler et al., 2019, Araújo et al., 2021).

1. Formal task and conceptual scope

In its standard many-body form, overlapping tomography starts from an unknown $n$-qubit mixed state $\rho_{1,2,\dots,n}$ and asks for approximations $\sigma_S$ to all reduced states

$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$

such that

$\|\rho_S-\sigma_S\|_1<\epsilon$

for every target subsystem $S$, with success probability at least $1-\delta$ (Yu, 2020). The task is therefore a form of partial tomography: it targets local marginals rather than the full density operator.

The motivation is the severe scaling of full tomography. For qubits, a generic $n$-qubit density matrix has $n$0 real parameters, and standard Pauli-basis full state tomography (FST) requires $n$1 measurement settings because each qubit is measured in $n$2, $n$3, or $n$4, while identity factors are inferred by marginalization (Araújo et al., 2021, 2207.14488). Overlapping tomography exploits the fact that many physically relevant quantities—local correlation functions, energies of local Hamiltonians, entanglement structure, and nearby-particle reduced states—depend only on small subsystems. Measuring each $n$5-body subsystem independently is still redundant, because overlapping subsystems reuse much of the same local information.

A central structural observation is that global product measurements preserve local marginal statistics. If each qubit is measured by the same single-qubit informationally complete POVM, then for any subset $n$6, the distribution of the corresponding outcome coordinates is exactly the distribution obtained by measuring $n$7 with the restricted product POVM on $n$8 (Yu, 2020). This is the formal basis of overlap reuse.

A common misconception is that overlapping tomography reconstructs the full global state. In general it does not. Its primary output is the family of chosen local marginals, and any global reconstruction requires additional assumptions or an auxiliary model.

2. Perfect-hash constructions and the original qubit protocols

The first systematic formulation of quantum overlapping tomography (QOT) introduced a combinatorial measurement design based on perfect hash families (Cotler et al., 2019). An $n$9 perfect hash family is a set of functions

$n$0

such that for every $n$1-element subset $n$2, at least one $n$3 is injective on $n$4. Operationally, each $n$5 colors the $n$6 qubits into $n$7 classes. For a fixed coloring, one performs all $n$8 assignments of Pauli bases $n$9 to the color classes, measuring all qubits in parallel. Whenever a target $\rho_{1,2,\dots,n}$0-subset is split into distinct colors by some $\rho_{1,2,\dots,n}$1, those $\rho_{1,2,\dots,n}$2 settings suffice to determine all Pauli coefficients of that $\rho_{1,2,\dots,n}$3-qubit marginal (Cotler et al., 2019).

This yields the original asymptotic guarantee that all $\rho_{1,2,\dots,n}$4-qubit reduced density matrices of an $\rho_{1,2,\dots,n}$5-qubit state can be determined with at most

$\rho_{1,2,\dots,n}$6

rounds of parallel measurements (Cotler et al., 2019). The construction combines a perfect-hash family size

$\rho_{1,2,\dots,n}$7

with a repetition count

$\rho_{1,2,\dots,n}$8

so that the total budget scales as $\rho_{1,2,\dots,n}$9 up to the hidden $\sigma_S$0-dependent constants (Cotler et al., 2019).

The $\sigma_S$1 case admits an explicit binary-hash implementation. Writing

$\sigma_S$2

one divides the register into two color classes in $\sigma_S$3 different ways so that every pair of qubits is separated in at least one division. One then uses three same-basis global settings and six mixed-basis settings per division, for a total of

$\sigma_S$4

measurement basis sets (2207.14488). This protocol was the template used in the first photonic demonstration of QOT.

These constructions established the central methodological idea of the field: overlap is not a nuisance to be removed but a combinatorial structure to be engineered.

3. Locality, lattice geometry, and optimal measurement design

A major refinement is local quantum overlapping tomography, where the target RDMs are not all $\sigma_S$5-body subsystems but only geometrically local ones, such as neighboring clusters on a lattice (Araújo et al., 2021). In that setting the number of distinct measurement settings can be made independent of the total system size $\sigma_S$6. The underlying tiling argument is geometric: one covers the lattice by cells large enough to contain the target local shape, measures all such cells in parallel, and shifts the tiling a bounded number of times to cover boundary-crossing clusters. The resulting number of displacements is at most $\sigma_S$7 in 3D, $\sigma_S$8 in 2D, and $\sigma_S$9 in 1D, leading in general to at most

$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$0

and in many concrete geometries simply

$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$1

with no dependence on $$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$2 (Araújo et al., 2021).

Representative local-QOT setting counts are summarized below.

Geometry	Target local object	Measurement settings
1D string	$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$3 consecutive qubits	$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$4
Ring	$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$5 consecutive qubits	$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$6
Square lattice	star RDM ($$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$7)	$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$8
Cylinder	square-lattice local object	$$\rho_S=\operatorname{tr}_{\bar S}(\rho), \qquad |S|\le k,$$9
Torus	square-lattice local object	$\|\rho_S-\sigma_S\|_1<\epsilon$0
Cubic lattice	star RDM ($\|\rho_S-\sigma_S\|_1<\epsilon$1)	$\|\rho_S-\sigma_S\|_1<\epsilon$2
Honeycomb lattice	first-neighbor 2-RDMs	$\|\rho_S-\sigma_S\|_1<\epsilon$3

Later work recast Pauli-restricted overlapping tomography as an explicit optimization problem. For qubits, the two-body problem can be mapped to an edge-clique-cover or covering-array problem, with each global Pauli setting corresponding to a clique covering many pairwise Pauli observables (Hansenne et al., 2024). This produced exact or provably optimal Pauli schemes and clarified that optimality depends on the measurement model.

Two results are especially notable. First, for Pauli measurements on qubits, two-body overlapping tomography of nearest neighbours can always be performed with nine Pauli settings (Hansenne et al., 2024). Second, if arbitrary local projective measurements are allowed instead of Pauli-only measurements, then all $\|\rho_S-\sigma_S\|_1<\epsilon$4-body marginals can be reconstructed with exactly

$\|\rho_S-\sigma_S\|_1<\epsilon$5

global settings, independently of the total number of qubits (Hansenne et al., 2024). The same measurement-model distinction appears in the clique-cover framework of optimal QOT, which gives $\|\rho_S-\sigma_S\|_1<\epsilon$6 parallel observables for nearest-neighbor $\|\rho_S-\sigma_S\|_1<\epsilon$7-RDMs on a 1D chain and 9 for nearest-neighbor 2-RDMs on a 2D lattice (Wei et al., 2024).

A recurrent point of confusion is that “optimal” is not universal. Nine settings for pairwise local tomography, logarithmic-in-$\|\rho_S-\sigma_S\|_1<\epsilon$8 constructions for all pairs, and $\|\rho_S-\sigma_S\|_1<\epsilon$9 system-size-independent schemes all coexist because they refer to different target families and different admissible measurement classes.

4. Sample complexity, optimality bounds, and qudit generalization

The main information-theoretic upper bound for general qubit overlapping tomography obtained via Pauli measurements is

$S$0

copies of the unknown state (Yu, 2020). For a selected list of $S$1 target subsystems, the corresponding bound is

$S$2

The same work proves the lower bound

$S$3

for $S$4, even when arbitrary joint measurements are allowed. For constant $S$5, this shows that joint, highly entangled measurements are not asymptotically more efficient than Pauli measurements (Yu, 2020).

Local-QOT analysis distinguishes the number of settings $S$6 from the number of repetitions $S$7. For full $S$8-RDM tomography one obtains

$S$9

whereas for local tomography of neighboring reduced states the repetition count improves to

$1-\delta$0

so the price of keeping $1-\delta$1 independent of $1-\delta$2 is only a logarithmic growth in repetitions (Araújo et al., 2021).

The qudit extension replaces Pauli observables by generalized Gell-Mann (GGM) measurements. For an $1-\delta$3-qudit density operator

$1-\delta$4

full tomography requires $1-\delta$5 local settings, while naive independent tomography of all $1-\delta$6-body marginals requires $1-\delta$7 settings (Ma et al., 15 Jan 2026). Overlapping qudit tomography is then equivalent to constructing covering arrays over an alphabet of size $1-\delta$8, with minimum setting count $1-\delta$9.

Two explicit optimal constructions are known in this framework. The zero-sum construction proves

$n$0

and Bush’s construction gives

$n$1

when $n$2 and $n$3 is a prime power (Ma et al., 15 Jan 2026). For qutrits, pairwise overlapping tomography obeys the explicit bound

$n$4

The same work also optimizes the order in which the measurement settings are executed, modeling switching overhead by Hamming distance between settings and reporting an approximately 50% reduction in switching costs relative to the worst-case ordering (Ma et al., 15 Jan 2026).

5. Experimental realization and noise-aware reconstruction

The first direct experimental demonstration of QOT used photonic four-photon and six-photon GHZ states together with Bayesian mean estimation (BME) implemented by Gibbs sampling (2207.14488). In the four-qubit test, full-state tomography yielded a fidelity

$n$5

with the reference GHZ state, while the QOT-reconstructed two-qubit marginals agreed with those obtained by partial tracing the FST estimate: the reported subsystem fidelities differed by less than 0.01 within the error bars, and the Von Neumann entropies agreed within uncertainty (2207.14488). For equal numbers of measured copies, the 95% confidence intervals obtained from FST were significantly larger than those from QOT, indicating better estimation efficiency per measurement. In the six-qubit experiment, FST would have required $n$6 basis settings and around 120 days at 700 events per setting, whereas QOT reconstructed all 15 two-qubit subsystems using only 21 basis settings in around 80 hours (2207.14488).

Overlap-based measurement reuse has also been pushed beyond marginal reconstruction. Parallel-measurement quantum state tomography (PQST), explicitly inspired by QOT, uses a carefully chosen set of global Pauli settings so that many overlapping local observables are acquired simultaneously and then fed into a locally purified state tensor-network ansatz (Hu et al., 2024). On a 31-transmon superconducting chip with adjustable-frequency couplers, this approach reconstructed 6-qubit and 9-qubit W states with fidelities of 98.68% and 95.07% after measuring 75 and 99 observables, respectively. For the 12-qubit W state, the reported largest reconstructed density matrix achieved cosine similarity

$n$7

after measuring 243 parallel observables, whereas FQST would require $n$8 observables (Hu et al., 2024).

Noisy local RDM estimation introduces an additional issue: independently reconstructed overlapping marginals need not be jointly physical. This has motivated semidefinite post-processing. One SDP-based approach imposes positivity on each local RDM together with overlap-consistency constraints

$n$9

and stronger enhanced-compatibility constraints on auxiliary larger regions (Wang et al., 30 Jan 2025). Since exact compatibility with a global state is QMA-complete, these are polynomial-size relaxations. In simulations on frustrated Hamiltonians and the 1D $n$00 chain, the method yields tighter bounds than unconstrained tomography and can reduce the required number of samples by a factor of $n$01 to $n$02 for the same precision in lower-bounding the energy (Wang et al., 30 Jan 2025).

A further extension incorporates readout errors directly into the overlapping-tomography model. In distributed regional tomography, each overlapping region is assigned both a local density operator and a local confusion matrix, neighboring regions are coupled by reduced-state consensus on their overlaps, and the resulting bilinear problem is solved by an alternating scheme with ADMM for the state update and parallel local updates for the confusion matrices (Taherpour et al., 15 Apr 2026). Simulations on Ring, Ladder, Torus, and Hub overlap graphs show that joint estimation improves state recovery over fixed-readout reconstruction and recovers a substantial portion of oracle performance, with geometry-dependent tradeoffs among estimation accuracy, communication, and computation (Taherpour et al., 15 Apr 2026).

6. Related overlap-based variants and broader analogues

The phrase “overlap tomography” is used in several related but non-identical ways across quantum optics and inverse problems. One important branch reconstructs a state from overlaps with calibrated probe states rather than from overlapping marginals. Generalized overlap quantum state tomography interferes an unknown single-mode optical state $n$03 with calibrated coherent probes $n$04 on a balanced beamsplitter and measures

$n$05

After truncation in the Fock basis, the overlaps satisfy a linear system $n$06, and the density matrix is reconstructed by semidefinite programming with positivity and unit-trace constraints (Nehra et al., 2019). This protocol was demonstrated on weak coherent and heralded single-photon states, with reported fidelities $n$07, $n$08, and $n$09 after phase averaging and loss correction (Nehra et al., 2019).

A second optical meaning of overlap is mode overlap. In tomography by the overlap, an unknown signal interferes with a coherent probe on an unbalanced beamsplitter and a single on/off detector records the zero-click probability as a function of temporal delay. The overlap

$n$10

becomes a tunable parameter; if the mode structure is known, the quantum state can be reconstructed, and if the state is known, the temporal or spectral profile can be inferred (Tiedau et al., 2017). The same apparatus therefore trades state information for mode information.

Weak-measurement tomography uses yet another type of overlap structure. There, one weakly measures projectors $n$11, performs a final projective measurement in a basis $n$12, and groups the pointer data by final outcome. The weak values

$n$13

directly determine pure-state amplitudes or mixed-state density-matrix elements, and all runs are retained rather than discarding failed post-selections (Wu, 2012). Here the overlap is not between target subsystems but between complementary pieces of information gathered in a single weakly disturbing measurement configuration.

Overcomplete path-entangled-photon tomography is similarly adjacent rather than identical. A two-photon path state in the symmetric basis $n$14 can be reconstructed from nine measured rates at two phases, while an overcomplete phase scan yielding 79 measurements improves robustness and exposes coherences hidden by sparse data (Santis et al., 2017). The overlap here is measurement redundancy rather than marginal overlap.

An instructive non-quantum analogue appears in 3D X-ray reconstruction with simultaneous emitters. When multiple rays overlap at the same detector, the measurement equation becomes

$n$15

leading to nonlinear sum-of-exponentials constraints and a sparse reconstruction problem solved by forward-backward splitting in a partially convex region (Klodt et al., 2016). Although this is not quantum state tomography, it exemplifies the same methodological shift: rather than discarding overlapping measurements, one models them explicitly and exploits their structure.

Across these variants, the unifying theme is that overlap—of supports, probe states, modes, or measurement information—is treated as an informational asset. The differences lie in the object being reconstructed, the admissible measurements, and the optimization or inversion machinery used to turn overlap into a tractable tomography protocol.