Checking what state a quantum processor actually produced, known as quantum state tomography, normally requires measuring the qubits in exponentially many combinations of measurement directions. That is $3^N$ settings for $N$ qubits, which comes to more than half a million settings at 12 qubits. A recent experiment cut this to a few hundred settings with a trick: rotate and read out all qubits at once, and choose the per-qubit measurement directions so that every small group of qubits sees every combination it needs somewhere in the schedule. The catch is that "rotate all qubits at once" is a layer of simultaneous gates. On today's noisy, error-prone processors, gates run in parallel disturb their neighbors. A qubit's effective measurement depends on what its neighbors are being rotated to at the same moment. Current reconstruction methods treat that measurement layer as perfect.
We propose to make the reconstruction aware of this. We model each qubit's measurement as depending on its own setting and its neighbors' settings on the hardware coupling graph. We calibrate that model cheaply with product states and fold it into the likelihood used for reconstruction. We also choose the measurement schedule so that every neighbor configuration that occurs is well calibrated. The same schedules come with a built-in, model-free test for crosstalk. The scheme measures each small group of qubits several times, under different surroundings. If those estimates disagree beyond shot noise, parallel execution is biasing the result. That redundancy is the core of the proposal: it lets us detect the crosstalk bias, then remove it, and check whether the removal worked.
Core Claim
Experimental sample-efficient quantum state tomography via parallel measurements (PQST) introduced the scheme described above. It assigns measurement bases using combinatorial objects called fractional perfect hash families, which guarantee that every $k$-qubit reduced density matrix (the state of a $k$-qubit subset) is covered. It represents the noisy global state with a locally purified tensor network and fits it by maximum likelihood on raw shots. On a 31-transmon chip it reported 98.68% fidelity for $|W_6\rangle$ with 75 settings, 95.07% for $|W_9\rangle$ with 99 settings, and 89.23% similarity for $|W_{12}\rangle$ with 243 settings. The paper itself leaves simultaneous-readout crosstalk uncharacterized, and it tested only tree-like subgraphs.
Our claim has three parts. Parallel measurement layers produce errors that depend on the setting chosen for the qubit's neighbors. These errors bias reduced density matrices in a structured way. Neither per-qubit readout correction nor a single global noise calibration can remove that bias. A neighbor-conditional measurement model, calibrated on product states and used inside the likelihood, removes most of it at modest calibration cost.
The model works as follows. A measurement setting $s$ assigns a basis $s_j \in \{X,Y,Z\}$ to each qubit $j$. Write $N(j)$ for the hardware neighbors of $j$. The probability of outcome string $o$ is then modeled as
$$ p(o \mid s, \rho) = \mathrm{Tr}\Big[\rho \bigotimes_j M^{(j)}_{o_j \mid s_j,\, s_{N(j)}}\Big], $$
where each $M^{(j)}$ is a single-qubit measurement operator that is allowed to depend on the neighbors' bases. Ideal PQST is the special case $M^{(j)}_{o_j\mid s_j}=|o_j\rangle\langle o_j|$ in basis $s_j$. On heavy-hex hardware the maximum degree is 3, so each qubit has at most $3^4=81$ setting configurations. That keeps the model small. We will extend it to correlated two-qubit outcome errors on edges only if the consistency test below shows the single-qubit form is insufficient.
Research Program
The first component is a crosstalk diagnostic that needs no noise model. Take any $k$-qubit subset $A$. In an overlapping schedule, the same local basis string on $A$ appears in several global settings that differ on the qubits bordering $A$. Group the shots by the configuration on that border. Then compare the $A$-marginal distributions across groups with a chi-squared statistic, which tells us how large the disagreement is relative to shot noise. Ideal measurements predict agreement within shot noise, whatever state was prepared. Significant disagreement therefore points to setting-dependent error, not to state-preparation error.
The second component is calibration. We prepare the six single-qubit basis eigenstates on every qubit in parallel and run them through the same measurement layers. We schedule these calibration circuits with the same hash-family construction, so that every (own setting, neighbor settings) configuration is covered. From these runs we fit each $M^{(j)}$ by constrained maximum likelihood. Our rough estimate is a few hundred calibration circuits, independent of $N$ at fixed degree. Verifying that estimate is part of the work.
The third component is co-design of the schedule. Among the valid hash families, we prefer schedules that spread neighbor-configuration occurrences evenly across hardware edges. This keeps the calibration well-conditioned, and it lets the estimator average the crosstalk out where full correction is not possible.
The fourth component is reconstruction. We place the conditional model inside PQST's locally-purified-state maximum-likelihood fit. Because the model only changes the measurement operators, it should also work with other backends. These include Projected Classical Shadows, which project raw shadow estimates onto physical, low-rank, or matrix-product manifolds, and the shadow-based infidelity loss of Neural-Shadow Quantum State Tomography. Both of those works assume ideal randomized measurements and do not model the gates that implement them. We test a second backend only if time allows.
Other parallel-execution work shapes the hardware side without addressing verification. PHOENIX schedules Pauli-level gates around heavy-hex connectivity and idle gaps, and Physics-Aware Compilation partitions neutral-atom arrays into isolated parallel zones while explicitly assuming uniform gate fidelity. Neither checks the output states it produces. QASMBench documents fidelity "hotspots" when entangling activity is placed carelessly on hardware. That motivates the expectation that crosstalk is non-uniform and depends on location, which is the case our conditional model is built for.
Baselines
The naive reference point is PQST run exactly as published: ideal measurement operators and the same hash-family schedule. We run it on the same shots our method uses, so any difference comes only from the estimator. The published PQST numbers (98.68% on $|W_6\rangle$, 95.07% on $|W_9\rangle$, 89.23% similarity on $|W_{12}\rangle$, all from Hu et al.) come from a different chip. We use them for orientation only, not as a comparison point. Instead, we rerun the pipeline ourselves.
The strongest practical competitor is PQST with standard per-qubit readout correction. This applies a $2\times2$ assignment matrix per qubit, calibrated from $|0\rangle$ and $|1\rangle$ preparations. It is what most practitioners would use. We give it the same total calibration-shot budget as our method.
A second strong competitor is a single-calibration noise model. Here each qubit's measurement operator depends on its own basis but not on its neighbors', fitted from the same calibration data. It is the closest analogue to calibrating one fixed measurement-noise channel for shadow tomography.
The key ablation removes neighbor dependence while keeping everything else identical. It differs from our full method only in whether $s_{N(j)}$ enters the model. A second ablation keeps the conditional model but uses a randomly chosen valid hash family instead of the balanced one. This isolates how much the schedule co-design contributes.
A third baseline is serialized measurement. We split each global setting into sub-layers in which no two neighboring qubits are rotated simultaneously. This removes most crosstalk at roughly a constant-factor cost in circuit count, and serves as the "avoid the problem" reference.
Our method counts as beating these baselines only if two conditions hold. In simulation, it must achieve lower infidelity to the true noisy prepared state than all corrected baselines. On hardware, it must push the consistency residual down to the shot-noise floor where the baselines do not. Both must hold at equal total shots.
Why not just…?
Why not just apply per-qubit readout error mitigation? Per-qubit assignment matrices only capture bit flips at readout that do not depend on the measurement setting. The crosstalk we target arises during the simultaneous basis rotations and depends on the neighbors' settings, which a $2\times2$ matrix per qubit cannot represent. The consistency test settles this directly. If residuals across neighbor configurations vanish after per-qubit mitigation, the alternative is sufficient and our conditional model adds nothing. We will report that outcome if it happens.
Why not just calibrate one global measurement-noise channel, as in noise-robust shadow schemes? This is the most serious threat. If the parallel layer's noise is effectively independent of neighbor settings (for example after randomizing the rotations), a single calibration matches our method at lower cost. The no-neighbor ablation is exactly this alternative, run on identical data. We will also test a twirled variant. If it equals our full model on hardware, the honest conclusion is that twirling suffices, and the contribution narrows to the diagnostic.
Why not just serialize the measurement layers? Serialization avoids crosstalk but multiplies circuit count and wall-clock time. That erodes the main advantage PQST holds over the $3^N$ settings of full tomography. It also leaves qubits idle longer before readout, the "retention lifespan" penalty QASMBench quantifies. We measure the trade-off directly by comparing fidelity per unit of total shots and total device time against the serialized baseline.
Why not just use randomized benchmarking or cross-entropy benchmarking? These return one average error number. They cannot produce reduced density matrices, relative phases, or quantities like the entanglement negativity that PQST predicted. They answer a different question, so they are not baselines here.
Proposed Experiments
In simulation, we prepare $|W_6\rangle$, $|W_9\rangle$, $|W_{12}\rangle$ and the fully connected Hamiltonian ground states used by PQST, plus small QASMBench circuits. We place them on a heavy-hex patch and inject measurement-layer crosstalk: conditional $ZZ$ phases and neighbor-dependent rotation errors whose strength is swept from zero upward. Because the true noisy state is known, we can report infidelity to that state and negativity error for every estimator as a function of crosstalk strength and shot budget. The decisive simulated check has two parts. Our method should track the ideal-measurement result as crosstalk grows, while the per-qubit and single-calibration baselines degrade. At zero crosstalk, it should be no worse than plain PQST, which tests that the extra parameters cost little.
On hardware, we run the same states on an available heavy-hex superconducting device, the class benchmarked in QASMBench and PHOENIX. We do this on a tree subgraph, to match PQST's setting, and on a subgraph containing a hexagonal loop, which PQST did not test. The ground truth is full tomography of 4-to-6-qubit subsets with all other qubits idle; at most $3^6=729$ settings, this is affordable at that size. We compare each estimator's reduced density matrices against it. The decisive hardware checks also come in two parts. First, plain PQST should show consistency residuals clearly above shot noise, establishing that the bias is real. Second, our estimator should bring those residuals to the noise floor and match the isolated full-tomography reconstructions more closely than every baseline at equal total shots.
If the first check fails, parallel crosstalk is negligible on this hardware. That is itself a useful negative result for anyone scaling overlapping tomography. If the second check fails while the first succeeds, the single-qubit conditional model is too weak, and we move to the edge-correlated extension.
References
- Experimental sample-efficient quantum state tomography via parallel measurements
- Enhancing quantum state reconstruction with structured classical shadows
- Neural-shadow quantum state tomography
- PHOENIX: Pauli-Based High-Level Optimization Engine for Instruction Execution on NISQ Devices
- Physics-Aware Compilation for Parallel Quantum Circuit Execution on Neutral Atom Arrays
- QASMBench: A Low-level QASM Benchmark Suite for NISQ Evaluation and Simulation