WorksPhysics and Astronomy
Quantum state tomography with measurement-conditioned diffusion models
Abstract
We introduce Diff-QST, a measurement-conditioned diffusion model for quantum state tomography that learns from an ensemble of input states and reconstructs a physical density operator from noisy, possibly incomplete Pauli measurements in two or three denoising steps. In simulations of two to five qubits, Diff-QST has higher mean fidelity than maximum-likelihood estimation (MLE) and a semidefinite-programming least-squares estimator (SDP) in all 120 tested conditions at nominal noise. After training, GPU reconstruction is faster than both in condition-averaged time at every system size and, at four and five qubits, in every condition. At four and five qubits, though not at two and three, Diff-QST also has higher fidelity than an architecture-matched regression baseline trained without diffusion. At five qubits, over the tested numbers of observables and shots per observable, it reaches a specified fidelity with 25%-67% fewer total shots than this baseline. Under state and noise distributions that differ from those used in training, Diff-QST retains its fidelity advantage over MLE and SDP in most tested conditions, although it can reverse. Diff-QST is therefore suited to fast tomography from noisy, incomplete Pauli measurements of states represented by its training ensemble.
Introduction
Quantum state tomography (QST) is a standard method for determining the state that a device has actually prepared [1, 2]. The state is reconstructed from measurement statistics obtained on identically prepared copies. Because full QST estimates the entire density operator, from which the statistics of any measurement follow, it serves as a reference point for more resource-efficient characterization methods [3, 4, 5]. An arbitrary -qubit density operator has independent real parameters, so full reconstruction becomes increasingly demanding as the system grows [6, 7]. At fixed accuracy, even optimal tomography requires a number of copies that grows with the Hilbert-space dimension and the state rank [8]. On present hardware, these scaling challenges are compounded by imperfect data: expectation values are estimated from a finite number of repetitions, and measurement outcomes can be misassigned [9, 10]. Decoherence can also make the prepared state mixed [2].
Maximum-likelihood estimation (MLE) selects the density operator that maximizes the likelihood of the measurement data [11, 12, 13], whereas physical least-squares estimates can be obtained by projecting a linear reconstruction onto the set of density operators [14]. For the Pauli data considered here, we use a regularized weighted least-squares estimator solved by semidefinite programming, denoted SDP and defined in Sec. II C. MLE and SDP both enforce physicality and incorporate an assumed statistical and measurement model into their objectives. Both must be rerun for each new data set
and become more costly as the Hilbert-space dimension grows. A separate line of work instead reduces the number of measurements required by exploiting structure in the state, as in compressed sensing [7] and tensor-network tomography [6, 15], or by estimating many properties from randomized measurements without reconstructing the full density operator [5].
Neural-network-based QST includes per-instance methods, which fit a neural representation of the state or its measurement statistics separately to each tomographic data set [16, 17, 18], and methods that learn a single map for an entire ensemble of states. Neural-shadow tomography combines per-instance fitting with an infidelity objective estimated from classical shadows [19]. The second category takes several forms: supervised networks trained on simulated states map measurement data directly to density operators [20, 21, 22], with reconstruction from incomplete measurements also demonstrated [20, 21]; generative query networks instead predict the statistics of unperformed measurements from any number of measurement-outcome pairs [23]; neural adaptive tomography replaces the Bayesian update of adaptive tomography with a trained network [24]; and supervised neural filtering mitigates state-preparation-and-measurement errors before reconstruction [25]. Conditional generative adversarial tomography also supports direct reconstruction after pretraining on related states [26]. More recent quantum-aware transformers encode measurement operators together with their outcome frequencies to reconstruct physical density operators [27]. Masked autoencoder pretraining further learns representations for state reconstruction from incomplete and noisy frequency data [28].
Learned denoisers have also been combined with standard reconstruction methods, either to refine their output or to act as priors within their iterations. Attention-based networks refine estimates obtained by linear inversion or MLE from informationally complete measurements [29]. Hybrid-denoising QST (HD-QST) alternates measurement-residual updates with a combination of a neural denoiser pretrained on natural images and a low-rank denoiser, weighted by reinforcement learning, and was tested with incomplete Pauli measurements on several quantum processors [30].
Incomplete measurement data do not generally identify a unique density operator. Prior information can reduce the measurements needed to identify a state [31] and enters Bayesian estimators through the prior distribution [32, 33]. Diffusion models learn a distribution by training a network to remove Gaussian noise of varying strength from its samples and generate new samples by repeated denoising [34, 35]; for inverse problems, their denoisers can supply such prior information, either conditioned directly on the data [36] or guided during reconstruction by a measurement likelihood [37]. Related quantum diffusion models generate pure- and mixed-state ensembles [38, 39]. Classical diffusion models have also learned the measurement-outcome distribution of a fixed state from finite samples [40] and generated measurement records conditioned on Hamiltonian parameters to estimate ground-state properties [41].
Here, diffusion is used not to generate states or measurement records but to reconstruct states: unlike the direct maps and masked-frequency pretraining above [27, 28], the network learns to remove Gaussian noise of varying strength from a target-state representation, conditioned on the measurement data. Our estimator, Diff-QST, learns from an ensemble of input states to map pairs of Pauli observables and their empirical expectation values (Figure 1) to a physical density operator; one network per system size and input-state ensemble handles different observable subsets and shot counts. Reconstruction starts from a Gaussian sample rather than an initial tomographic estimate [29] and uses a fixed sequence of two or three denoising steps without a convergence test. Unlike HD-QST, it uses learned conditioning instead of explicit measurement-residual updates. We investigate whether this approach, with its short reconstruction process, can improve fidelity over MLE and SDP while reducing reconstruction time after training.
Figure 1. Overview of the state reconstruction task. The shaded blocks show the simulated preparation and measurement used to generate the data sets. Preparation noise maps the input state to the target state , and a subset of Pauli observables is measured with asymmetric readout error (orange arrows) and shot noise. The white blocks show the resulting measurement data , which pair each observable with its empirical expectation value (error bars, one shot-noise standard deviation), and the state estimate of reconstructed from by Diff-QST, MLE, or SDP.
Compared with MLE and SDP, which use no information learned from an input-state ensemble, Diff-QST achieves higher mean fidelity in all 120 nominal conditions tested, 96 of which use incomplete Pauli sets. These conditions span two to five qubits, pure and mixed input-state ensembles, 20%–100% of the nonidentity Pauli observables, and – shots per observable, at fixed nominal preparation-channel strengths and readout-error rates. At batch size 1, Diff-QST on GPU is faster than both references in condition-averaged time at every system size; at four and five qubits, Diff-QST is faster in every condition, even on the same CPU resources. Batching further reduces the GPU time per state.
As a supporting comparison, an architecture-matched regression baseline, a single-pass deterministic regressor trained without diffusion [36], tests the benefit of diffusion beyond direct learned reconstruction. Diff-QST has higher condition-averaged fidelity at four and five qubits; the regression baseline has higher fidelity at two and three qubits and reconstructs faster at every tested size. Section III C and Appendix E examine the roles of repeated denoising and measurement budgets in this comparison. Section III D repeats both comparisons under changes in the state distribution, readout calibration, and preparation noise, testing how far the nominal results extend.
State reconstruction from Pauli data
Our target is the physical state after preparation noise and before readout (Figure 1). Accuracy is quantified by the squared quantum state fidelity [42]. All comparisons use simulated data and the known target .
Input-state ensembles and measurement model
For qubits with Hilbert-space dimension , the pure input states are Haar-random and the mixed input states are depolarized pure states [43], with equal eigenvalues. For the mixed input states, the normalized purity is uniform on . Local depolarizing, amplitude-damping, and phase-damping channels, with nominal per-qubit depolarizing probability and damping strengths , act on to produce the targets . The labels pure and mixed refer to ; both target ensembles are mixed. Channel strengths vary by up to during training and take their nominal values for validation and the main comparison. Over the training states, preparation noise lowers the median target purity for pure input states to 0.828, 0.730, 0.642, and 0.564 at , respectively.
For each system size and input-state ensemble, 100,000 training, 10,000 validation, and 300 test states are drawn independently; Appendix A specifies the distributions and noise channels and gives the target-purity ranges of the training states.
The target state admits the Pauli expansion
where runs over the nonidentity elements of , which form the complete Pauli set. Their exact expectation values determine an arbitrary state. We measure of these observables, drawn uniformly at random without replacement, where is the observable sampling ratio. For an incomplete Pauli set (), the expectation values are generally consistent with multiple physical states. Readout is modeled by independent asymmetric bit flips [9]: at the nominal rates, a true 1 is observed as 0 with probability , and a true 0 as 1 with probability . This replaces with the effective observable , obtained by substituting for each nonidentity single-qubit factor of , where and (Appendix A). For each observable, the observed parity is the product of the local outcomes at its nonidentity factors. We sample it times; with the number of outcomes,
The reconstruction input consists of the measurement data and the shot count . Each observable-value pair contains a Pauli observable , rather than its effective counterpart , and its empirical expectation value . Instead of the full distribution of multi-qubit outcomes, we retain only the mean parity , which contains both readout error and shot noise. Each observable is thus acquired separately, as in compressed-sensing tomography with randomly chosen Pauli observables [7]; measurements in local Pauli bases would instead yield the parities of several observables from each shot.
Conditional diffusion estimator
The reconstruction map, shown schematically in Figure 2, is a denoising diffusion probabilistic model (DDPM) [34, 35] conditioned on the measurement data. In the forward process of a DDPM, Gaussian noise of controlled strength is added to a target representation, and the network is trained to remove it. This supplies denoising tasks over a range of noise strengths, distinct from the shot noise and readout errors in the conditioning data. Reconstruction, the reverse process, applies the learned map repeatedly from an initial Gaussian sample. The network operates on a real representation, in which the target becomes

Figure 2. Measurement-conditioned diffusion for quantum state reconstruction. (a) The denoising map takes observable-value pairs, the noise level , and the shot count. Blue dashed arrows denote measurement or shot-count paths, orange dash-dotted arrows denote the noise level or Gaussian noise, and the green dotted arrow denotes the measurement features entering at the coarsest resolution. The output is a physical density operator, Eq. (4). (b) Training minimizes the loss between the state estimate and the target state , given the measurement data and the noisy sample . (c) Reconstruction starts from an initial Gaussian sample and uses the same measurements at each of validation-selected denoising steps. Density operators are displayed through their real parts; , , and the reverse-process variable are real arrays, not density operators.
, an array of two real channels. The forward process produces the noisy sample
where is standard Gaussian noise, indexes the diffusion steps, , called the noise level as in Ref. [36], decreases from near unity to near zero as more noise is added, following the cosine schedule of Ref. [44], and is the variance of the noise added at step . The array is not a density operator, since Eq. (3) perturbs the real and imaginary parts independently and does not preserve Hermiticity, positivity, or unit trace.
The denoiser , with trainable parameters , estimates the target representation from , , , and . The observable-value pairs retain which physical quantities were measured, and the conditioning is invariant to their ordering. Figure 2(a) shows the U-Net denoiser [45], which processes multiple resolutions. A measurement encoder maps any number of observable-value pairs to measurement features, which modulate the U-Net and enter its coarsest resolution, the bottleneck, through attention. Appendix B specifies layer dimensions and training settings.
The output layer returns two channels, taken as the real and imaginary parts of a matrix , from which the state estimate at diffusion step is formed as
For , Eq. (4) yields a Hermitian, positive-semidefinite, unit-trace operator [1]. Every state estimate is therefore physical without a penalty or projection step. Physicality alone does not make reconstruction from an incomplete Pauli set unique; the estimator also relies on structure learned from the input-state ensemble.
The training objective is
which combines the mean squared Hilbert-Schmidt (Frobenius) distance to the target with the infidelity weighted by . Here returns the real representation of in Eq. (4), and the expectation is taken over the training states, uniformly sampled diffusion steps , the Gaussian noise , and the measurement data. The training targets are fixed, whereas the diffusion noise and the measurement data are redrawn throughout training. During training, the observable sampling ratio varies between 20% and 100%, and the shot count is drawn uniformly from , where denotes the infinite-shot limit , which retains readout error but has no shot noise. All nominal test conditions therefore lie within the training range (Appendix B 1).
Reconstruction runs the reverse process from an initial Gaussian sample through denoising steps spanning the full diffusion range [Figure 2(c)]. Between scheduled indices , is sampled from the forward process’s Gaussian conditional distribution given and , with replaced by the predicted state’s real representation. This strided posterior update is ancestral sampling [35] restricted to of the diffusion steps [44], and the final denoiser output gives . The number of denoising steps is selected on validation data from , subject to the constraint that it does not decrease with within each input-state ensemble (Appendix B 2). The selected values, at and at , are used for all measurement conditions and distribution-shift tests. Excluding , which returns only the denoiser prediction from the initial Gaussian sample, guarantees at least one reverse-process update, and is evaluated separately as a control (§ 0.3 C). Appendix B 3 compares this sampler with a deterministic one and with other initial samples and injected-noise scales; none of these is better in every condition for all eight networks or raises the condition-averaged fidelity by more than 0.0053, and some lower it by more than 0.05, so ancestral sampling is retained.
Maximum-likelihood and least-squares estimation
The reference estimators, MLE and SDP, use no learned information about the input-state ensemble. They receive the same data as Diff-QST and use the nominal readout model through the effective observables , as in readout-aware tomography with calibrated measurement effects [46]; Diff-QST instead incorporates nominal readout error through training. MLE maximizes the likelihood with the diluted iteration [13], treating each observable as a two-outcome measurement with effects and counts recovered from .
The SDP estimator solves
where is the Frobenius norm. The weights are inverse estimates of the sampling variance of , normalized to unit mean, with the outcome probability taken as the posterior mean under the Jeffreys prior [47], so that each weight is finite even when all outcomes of an observable agree. For unit-trace , the regularization term in Eq. (6) is the purity penalty [48], with . It makes the objective strictly convex, so the estimate is unique even for incomplete Pauli sets, but it can also alter the measurement fit. Appendix C 2 gives the weights, solver settings, and numerical checks.
Architecture-matched regression baseline
To assess the benefit of diffusion beyond direct learned reconstruction, we train an architecture-matched regression baseline for each system size and input-state ensemble, as in evaluations of conditional diffusion models [36]. This single-pass deterministic regressor retains the corresponding denoiser architecture, measurement encoder, physical output parametrization of Eq. (4), and composite training loss, but makes a single prediction with its noisy-sample input set to zero and its noise level set to one, the value at which no Gaussian noise is added; it is trained without Gaussian perturbation at randomly sampled diffusion steps and reconstructs without a reverse process. To make the use of diffusion the only difference between the two estimators, each regression network is trained with the data, optimizer, epoch limit, and early-stopping rules of the corresponding Diff-QST network and with its learning rate, ratio curriculum, and infidelity weight , which were chosen on validation data for Diff-QST. The weights of the two estimators are selected independently on validation data. The regression baseline uses one network evaluation, whereas Diff-QST uses one for each of its two or three denoising steps.
Results
For each system size and each input-state ensemble, the estimators are compared at three shot counts, , , and , and at five observable sampling ratios from to in steps of . This gives 120 nominal conditions, of which 96 with use incomplete Pauli sets. In each condition, all estimators reconstruct the same 300 test states (Sec. II A) from the same measurement outcomes. For , each state’s fidelity is averaged over three observable subsets shared by all test states of the condition, and the mean fidelity averages these per-state values over the test states. Uncertainties of fidelity differences are 95% bootstrap confidence intervals (CIs) over test states [49], conditional on the trained networks and observable subsets; “both CIs” refers to those relative to MLE and to SDP (Appendix C 1).
Fidelity relative to MLE and SDP
In all 120 nominal conditions, Diff-QST has higher mean fidelity than both MLE and SDP, with both CIs above zero (Table I). After Holm’s correction for the 240 comparisons [50], all differences remain significant at family-wise level 0.05 (Appendix C 1). Figure 3 shows this comparison at shots per observable. This ordering is not limited to mean fidelities: at four and five qubits, Diff-QST has higher fidelity than both MLE and SDP for every test state in 59 of the 60 conditions and for 299 of the 300 states in the remaining one ( pure input states, shots, ). At two and three qubits, the fraction of such states is at least 0.62 and 0.83, respectively, with the lowest values at .
Figure 3. Mean fidelity of Diff-QST, MLE, and SDP. The top row, (a)–(d), corresponds to pure input states, the bottom row, (e)–(h), to mixed input states, and the columns to , , , and . All results use shots per observable with nominal preparation-channel strengths and readout-error rates, where denotes the measured fraction of the nonidentity Pauli observables. Markers give over 300 test states (§ 0.3), and vertical bars mark standard deviation across these states, truncated at 0 and 1.
Table I summarizes across all 120 nominal conditions. Equivalently, is the smaller of the differences from MLE and from SDP in each condition. For the 96 incomplete-Pauli conditions, , with medians of 0.120 for pure and 0.158 for mixed input states. For each input-state ensemble, the median and minimum of over the 15 conditions of a system size increase with , although individual conditions follow no single monotonic trend in . Two other quantities change with : the target purity decreases (Sec. II A), and at fixed and the total number of shots, , increases. The trend in therefore does not isolate a system-size effect at fixed target purity or total measurement budget.
At fixed , the amount of data per state parameter also grows with . The number of measured observables
| Pure input states | Mixed input states | |
| 2 | 0.0603 [0.0019] | 0.0770 [0.0007] |
| 3 | 0.1161 [0.0119] | 0.1243 [0.0074] |
| 4 | 0.1585 [0.0321] | 0.1912 [0.0386] |
| 5 | 0.1752 [0.0631] | 0.2294 [0.0894] |
Table 1. Fidelity difference of Diff-QST from the better of MLE and SDP over the full nominal grid. Entries give the median and, in brackets, the minimum over the 15 shot-ratio conditions of each system size and input-state ensemble. Both CIs lie above zero in all 120 conditions.
per real parameter of the input-state family, , exceeds 3 throughout the five-qubit grid. Appendix D 1 therefore analyzes against , with four- and five-qubit networks separately retrained to reach ratios below the 20% training bound. For these retrained networks at , the largest observed fidelity difference occurs at –2.02 in seven of eight combinations of system size, input-state ensemble, and shot count, corresponding to measurement counts approximately between the dimensional lower bound for identifying every member of the family and the roughly that topological obstructions can require [31].
With the complete Pauli set, the fidelity difference narrows as the number of shots per observable increases. From to shots, the mean fidelity rises by 0.068–0.287 for MLE and 0.069–0.284 for SDP, but by only 0.010–0.043 for Diff-QST. decreases from 0.060–0.333 to 0.0007–0.089, and every minimum in Table I occurs at shots with the complete Pauli set. Three shot counts do not separate statistical error from reconstruction bias, nor do they show whether the fidelity of Diff-QST saturates below unity at larger shot counts.
At and , the advantage is not robust to the choice of observables. With only three observables measured, averaging over all 455 possible subsets leaves a difference of 0.004–0.010 for pure input states, and both CIs include zero at and shots. Some individual subsets reverse the ordering (Appendix C 1).
A sensitivity check with hedged MLE, which multiplies the likelihood by a power of to keep the estimate full rank [51], narrows the fidelity differences, with Diff-QST retaining higher mean fidelity in 91 of the 96 incomplete-Pauli and all 24 complete-Pauli conditions (Appendix D 2). With the trace distance [2] instead of fidelity, Diff-QST is closer to the target than both references, with both CIs below zero, in 112 of the 120 nominal conditions; the references are closer, with both CIs above zero, in four (Appendix D 3): three with the complete Pauli set at shots and , and one at , shots, and , where the rank-deficient MLE and SDP estimates are penalized more by fidelity than by trace distance.
Reconstruction time relative to MLE and SDP
For a fixed network, number of measured observables , and , Diff-QST applies the same operations to every state, whereas MLE and SDP iterate to state-dependent convergence: for five-qubit pure input states at shots and , MLE takes 1.83–12.0 s per state and SDP 0.657–0.776 s. We time the estimators on the same 15 test states and observable subsets at batch size 1, excluding training and measurement acquisition (Appendix C 3). Diff-QST uses one NVIDIA H200 NVL GPU or the same CPU resource configuration as the references: two AMD EPYC 9135 processors.
When condition-level mean times are averaged equally over the 30 nominal conditions at each system size, GPU Diff-QST is faster than both references for . Figure 4 shows the distribution of the condition-level ratios. At five qubits, the median ratios relative to MLE and SDP are 482 and 125, with minima of 135 and 71.9. At four and five qubits and shots, GPU
Figure 4. Reconstruction time of MLE and SDP relative to Diff-QST at batch size 1. For reference , the ratio divides its mean time over the same 15 states and observable subsets by Diff-QST’s mean time per state. Diff-QST runs on (a) the GPU and (b) the CPU, and MLE and SDP use the 32-core CPU configuration in both panels. Open symbols, offset horizontally for visibility, are the 30 nominal conditions per system size. Filled symbols and connecting lines show their medians, short horizontal bars mark the minima, and the line at unity indicates equal times.
Diff-QST takes 4.70–5.08 ms per state for incomplete Pauli sets. The speed advantage holds in every four- and five-qubit condition, including when Diff-QST uses the same CPU resources [Figure 4(b)]. At two and three qubits, GPU Diff-QST is faster than SDP throughout, but slower than MLE in 11 and 2 conditions, respectively [Figure 4(a)]. These ratios describe the tested implementations and hardware. Accelerated projected-gradient algorithms reach the maximum-likelihood estimate faster than the diluted iteration used here [52, 53], and a faster solver could likewise shorten the SDP times; such algorithms would not change the converged estimates wherever these are unique (Appendix C 3).
In 12 four- and five-qubit conditions at shots, stopping MLE or SDP early removes neither the fidelity nor the time advantage (Appendix C 3): even the best tested stopping point, selected with knowledge of the targets, leaves their mean fidelity 0.09–0.20 below that of Diff-QST, and stopping each once its mean fidelity comes within 0.001 of the converged value leaves them 13–676 times slower than GPU Diff-QST. When many states require reconstruction, GPU batching further reduces Diff-QST’s time per state, by a factor of 9.3–87.1 at batch size 256 relative to batch size 1 across the tested networks and observable sampling ratios.
Comparison with the regression baseline
Over the 12 finite-shot conditions with incomplete Pauli sets (Table II), Diff-QST has higher condition-averaged fidelity than the regression baseline at four and
TABLE II. Comparison with the regression baseline (Reg) over the 12 finite-shot conditions with incomplete Pauli sets. Entries are condition-averaged mean fidelities; denotes Diff-QST with denoising steps. Diff-QST uses , 2 at and 3 otherwise; the control evaluates the same trained networks with one network evaluation at . Differences are computed from unrounded means and given with 95% CIs over test states. The CIs quantify test-state uncertainty for the reported trained networks; sensitivity to training initialization is assessed separately using three additional training runs per estimator and input-state ensemble at each system size (Appendix E).
| Regression | [95% CI] | [95% CI] | [95% CI] | ||||||
| Pure input states | |||||||||
| 2 | 0.8689 | — | 0.8612 | −0.0077 | [−0.0111, −0.0044] | — | — | ||
| 3 | 0.9007 | 0.8693 | 0.8925 | −0.0081 | [−0.0102, −0.0060] | −0.0313 | [−0.0327, −0.0300] | +0.0232 | [+0.0210, +0.0254] |
| 4 | 0.9041 | 0.9296 | 0.9303 | +0.0262 | [+0.0251, +0.0273] | +0.0255 | [+0.0245, +0.0266] | +0.0007 | [+0.0001, +0.0012] |
| 5 | 0.8405 | 0.9142 | 0.9080 | +0.0675 | [+0.0642, +0.0708] | +0.0737 | [+0.0704, +0.0770] | −0.0062 | [−0.0066, −0.0057] |
| Mixed input states | |||||||||
| 2 | 0.8952 | — | 0.8891 | −0.0061 | [−0.0081, −0.0042] | — | — | ||
| 3 | 0.9173 | 0.8935 | 0.9118 | −0.0054 | [−0.0071, −0.0037] | −0.0237 | [−0.0249, −0.0225] | +0.0183 | [+0.0165, +0.0201] |
| 4 | 0.9106 | 0.9353 | 0.9326 | +0.0220 | [+0.0210, +0.0230] | +0.0247 | [+0.0239, +0.0255] | −0.0027 | [−0.0032, −0.0022] |
| 5 | 0.8976 | 0.9217 | 0.9179 | +0.0203 | [+0.0187, +0.0219] | +0.0241 | [+0.0226, +0.0256] | −0.0038 | [−0.0043, −0.0033] |
Table 2.
five qubits for both input-state ensembles, by 0.0203–0.0675, whereas the regression baseline has higher fidelity at two and three qubits; all eight CIs exclude zero. At four and five qubits, the advantage is largest at : averaged over shot counts, it is 0.028–0.134 at but at most 0.014 with the complete Pauli set (Appendix E). Retraining both estimators from three further initializations at reproduces the two- and three-qubit ordering in all 12 further pairs; at four qubits, three of the four pairs per input-state ensemble, including the reported one, favor Diff-QST. At five qubits, all three further pairs per input-state ensemble favor Diff-QST, with CIs above zero (Appendix E). The regression baseline, however, reconstructs faster. At batch size 1 on the same GPU, the four- and five-qubit Diff-QST networks take about 2.2–2.3 times as long (Appendix C 3).
A separate control (Appendix E 1) evaluates the same trained networks with a single denoising step. It retains the four- and five-qubit fidelity advantage over the regression baseline, so this advantage does not require repeated network evaluation. Because the estimator and the regression baseline still differ in training and in their noisy-sample and noise-level inputs, this comparison does not identify which element of diffusion training produces the advantage. Relative to , the selected raises the condition-averaged fidelity of both three-qubit networks, by 0.0183 and 0.0232. At four and five qubits it changes the fidelity little: a gain of 0.0007 for the four-qubit network for pure input states and losses of 0.0027–0.0062 for the other three networks. Because selection excludes (Sec. II B), is not an optimum over all step counts.
At five qubits, we also compare the smallest evaluated total shot budget, , at which each estimator attains a specified mean fidelity, over 21 combinations of and per input-state ensemble, including and 500, which lie between the shot counts used in training (Appendix E 2). For , Diff-QST requires for both input-state ensembles, whereas the regression baseline requires 204,600 for pure and 306,500 for mixed input states, so Diff-QST uses 50% and 67% fewer total shots; for , it uses 40% and 25% fewer. A finer grid than the evaluated one could change these minima.
Changes in the state and noise distributions
The fidelity advantage over MLE and SDP is tested next under changes in the state or noise distribution. The networks are applied without retraining and with unchanged, and MLE and SDP retain the nominal readout model. In each test, all estimators use the same targets, observable subsets, and measurement outcomes.
The regression baseline requires 204,600 for pure and 306,500 for mixed input states, so Diff-QST uses 50% and 67% fewer total shots; for , it uses 40% and 25% fewer. A finer grid than the evaluated one could change these minima.
Table 3 summarizes the results.
| Change | Input states | Cond. | ||
| Degeneracy lifted | mixed | 60 | 59 | 1 |
| Lower purity range | mixed | 60 | 58 | 2 |
| Readout | pure, mixed | 120 | 120 | 0 |
| Readout | pure, mixed | 120 | 120 | 0 |
| Canonical structured | pure | 300 | 233 | 30 |
| Rotated structured | pure | 180 | 171 | 9 |
| Preparation | pure | 36 | 26 | 10 |
| mixed | 36 | 35 | 1 | |
| Preparation | pure | 36 | 35 | 1 |
| mixed | 36 | 36 | 0 |
Table 3. TABLE III. Summary of the tests with shifted state or noise distributions. All counts are summed over . For each test, the last two columns give the number of conditions in which both CIs relative to MLE and SDP lie above or below zero. The spectral and readout tests (top) use the 15 shot-ratio conditions per system size. The structured-state tests (middle) apply these conditions to each of five canonical representatives or three rotated ensembles. The preparation-noise tests (bottom) use , and at three shot counts. The grids differ, so the counts do not rank the tests.
With nominal noise, we test two shifts of the distribution of mixed input states: (i) replacing the degenerate spectra with randomly generated nondegenerate spectra that have the same distribution of normalized purity, and (ii) lowering the purity range to (Appendix F 1). For each shift, both CIs lie above zero in all 48 incomplete-Pauli conditions. The exceptions on the full grids occur for complete Pauli sets at shots and small system sizes, with a most negative difference of . For incomplete Pauli sets, the fidelity advantage therefore generalizes beyond the degenerate spectral structure used in training. Device readout error rates drift between calibrations [54, 55, 56]; to test sensitivity to the resulting readout miscalibration, we halve or double the actual readout rates. For each rescaling, without retraining or updating the assumed readout model, Diff-QST has higher mean fidelity than both references in all 120 conditions, with both CIs above zero (Appendix F 2).
The networks for pure input states, whose training and tuning include no structured-state examples, are next applied to Greenberger-Horne-Zeilinger (GHZ) [57], W [58], and linear cluster states [59] under nominal noise and the 15 measurement conditions (Appendix F 3). Each family is tested in two forms: canonical representatives and locally rotated ensembles. The canonical representatives are the GHZ states with relative phases 0, , and (GHZ, GHZ, GHZ), the W state, and the linear cluster state.
The rotated ensembles are generated by applying independent Haar-random single-qubit rotations to GHZ, W, and the linear cluster state, which keeps the entanglement of each ideal input but randomizes its orientation relative to the Pauli measurement basis. As an unstructured control, a separate ensemble of Haar-random pure input states, drawn from the same distribution as the test states of the main comparison, is evaluated in the same way as the rotated ensembles.
Every tested canonical representative and rotated ensemble at has a positive median over the 15 conditions, although some three-qubit canonical conditions have both CIs below zero. At , the canonical GHZ representatives have negative medians, and the rotated ensembles have both CIs below zero at for every shot count. The Haar-random ensemble also has both CIs below zero at , with its own three observable subsets, which differ from those of the main comparison, but above zero after averaging over all 455 three-observable subsets (Appendix C 1); the rotated ensembles remain below zero after this averaging (Appendix F 3), so their reversal is not explained by subset choice. Figure 5 shows the five-qubit mean fidelities at shots: at , Diff-QST attains 0.683–0.867 for the canonical states, with the global utility even though the difference from MLE and SDP is positive.
Figure 5. Reconstruction fidelity for five-qubit structured input states at shots per observable, using the network for pure input states: (a)–(c) canonical GHZ, W, and linear cluster states and (d)–(f) their locally rotated ensembles. Markers show mean fidelities. Vertical bars mark standard deviation across 100 independent acquisitions for each canonical state or across the subset-averaged fidelities of 300 states for each rotated ensemble, truncated at 0 and 1. Figure 9 shows fidelity differences over the full shot-ratio grids for .
Scaling all preparation-channel parameters to one-half or one-quarter of their nominal values, with readout nominal, shifts the target distribution beyond the training range. Both CIs remain above zero in 36 (mixed) and 35 (pure) of the 36 conditions at half strength, and in 35 and 26 conditions at quarter strength. The 10 quarter-strength reversals for pure input states occur at every system size, in 2, 2, 2, and 4 of the nine conditions at , 3, 4, and 5. They include incomplete Pauli sets, with the most negative difference, , at , , and shots. The nominal advantage therefore does not extend uniformly to weaker preparation noise. The shifted targets are purer than those in training: at , the median target purity for pure input states is 0.87 and 0.75 at quarter and half strength, compared with 0.56 nominally and at most 0.64 over the training states. Because these tests do not isolate purity from other properties of the targets, the reversals cannot be attributed to purity alone (Appendix F § 1.0.4).
Applied without retraining to the same shifted data, the regression baseline has lower mean fidelity than Diff-QST at four and five qubits in 238 of the 252 spectral, readout, and preparation-noise conditions and, on average, for the structured states other than the canonical GHZ representatives; at two and three qubits, it has the higher mean fidelity in 178 of the corresponding 252 conditions (Appendix E 3). Under weaker preparation noise, it also reverses relative to MLE and SDP more often than Diff-QST: for pure input states, both CIs lie below zero in 12 and 4 of 36 conditions at quarter and half strength, compared with 10 and 1 for Diff-QST.
Conclusion
We have introduced Diff-QST, a measurement-conditioned diffusion estimator that learns state information from an input-state ensemble and reconstructs a density operator from noisy, possibly incomplete Pauli measurements in two or three denoising steps. In simulations of two to five qubits, it has higher mean fidelity than MLE and SDP, which use no learned state information, in all 120 nominal conditions. Because the number of steps is fixed, reconstruction involves no state-dependent convergence iterations. Once trained, Diff-QST reconstructs individual states on a GPU at a batch size of 1, achieving a lower condition-averaged time than both references at every system size, and is faster than them in all four- and five-qubit conditions, even on the same CPU resources.
Two conditions limit these advantages. The first is the training distribution, from which Diff-QST learns its state information. For incomplete Pauli sets, the advantage over MLE and SDP persists under the tested changes in mixed-state spectra and readout calibration but can reverse under weaker preparation noise and for some two- and three-qubit structured states. The second is the system size, as the advantage of diffusion over learned regression appears only in the larger tested systems. At four and five qubits, Diff-QST has higher condition-averaged fidelity than the regression baseline, an advantage retained in most of the tested distribution shifts and training initializations. At nominal noise, it does not require repeated denoising, as it is retained with a single denoising step. At five qubits, over the tested numbers of observables and shots per observable, Diff-QST also needs 25%–67% fewer total shots to reach specified fidelities. At two and three qubits, the regression baseline has higher fidelity. It is also faster at every tested size.
These results are limited to simulations of two to five qubits with separately trained networks. They do not establish performance on experimental data, whose preparation and readout noise can differ from those assumed in training, or beyond five qubits, because the number of measured Pauli observables at fixed sampling ratio and the density-operator representation both grow as . Within this scope, Diff-QST is useful where fast reconstruction is required. Its speed advantage over MLE and SDP grows with the number of qubits over the tested sizes and, through GPU batching, with the number of states to be reconstructed.
Acknowledgement
We acknowledge the use of AI tools to refine the English phrasing and grammar of this manuscript. This work was partly supported by the Institute for Information & communications Technology Planning & Evaluation (IITP) grant (RS-2025-02292999), the IITP-ITRC (Information Technology Research Center) grant (RS-2026-25519864), and Creation of the Quantum Information Science R&D Ecosystem (RS-2023-NR068116) through the National Research Foundation of Korea (NRF), funded by the Korean government (MSIT).
Input-state ensembles and noise model
For each system size and each input-state ensemble, we independently draw 100,000 training states, 10,000 validation states, and 300 test states. The pure input states are Haar-random. The mixed input states are depolarized pure states [43],
with Haar-random, normalized purity uniform on , and . Each state in this family has one large eigenvalue and equal eigenvalues, so the family does not cover arbitrary mixed states.
Preparation noise applies local depolarizing, amplitude-damping, and phase-damping channels in that order to every qubit [2]. This is a phenomenological channel applied once after ideal preparation, not a gate-resolved model of a particular device. The depolarizing channel is
The damping channels have Kraus operators
The nominal channel parameters are and ; during training, each strength is sampled independently per state within of its nominal value. Over the training states, which include this variation, the target purity ranges over 0.763–0.926, 0.664–0.860, 0.584–0.751, and 0.510–0.636 at for pure input states and over 0.532–0.912, 0.428–0.833, 0.354–0.741, and 0.300–0.620 for mixed input states. At , the purity of mixed input states has median 0.783 before the channel and 0.445 afterward.
Independent asymmetric readout errors replace each Pauli string by
where is the factor of acting on qubit , , and . Finite-shot data follow Eq. (2).
Diffusion estimator
Network and training
The denoiser is a U-Net [45] that predicts directly. The measurement encoder embeds each observable-value pair as a feature vector, or token, retaining the local Pauli symbols and qubit positions. Learned latent vectors collect information from these tokens through attention and are then refined by Transformer layers [60, 61], without positional encoding over the observable list. Their average gives the pooled measurement features, which, together with the noise level and the shot count, control channelwise scales and offsets of the U-Net through feature-wise linear modulation [62]. Bottleneck cross-attention lets the U-Net features at the coarsest resolution draw on the individual latent representations. Table IV specifies the dimensions and training settings.
| Input state | U-Net channels | Parameters | Dim. | Latents | Heads | Layers | Learning rate () | Curriculum | ||||
| 2 | pure | 60 | 64, 128 | 1,656,755 | 128 | 32 | 4 | 1 | 1 | 3.65 | 20 | 2 |
| 3 | pure | 90 | 128, 256, 512 | 23,312,659 | 256 | 64 | 8 | 2 | 4 | 1.48 | 80 | 3 |
| 4 | pure | 120 | 128, 256, 512 | 30,771,731 | 512 | 96 | 8 | 3 | 4 | 0.732 | 120 | 3 |
| 5 | pure | 180 | 192, 384, 768 | 59,006,163 | 512 | 256 | 4 | 4 | 4 | 0.567 | 280 | 3 |
| 2 | mixed | 50 | 128, 256 | 5,264,563 | 128 | 64 | 4 | 1 | 2 | 3.49 | 40 | 2 |
| 3 | mixed | 70 | 128, 256, 512 | 31,139,027 | 512 | 32 | 8 | 2 | 2 | 1.20 | 160 | 3 |
| 4 | mixed | 130 | 128, 256, 512 | 31,837,715 | 512 | 128 | 4 | 3 | 2 | 0.670 | 160 | 3 |
| 5 | mixed | 180 | 192, 384, 768 | 59,006,163 | 512 | 256 | 4 | 4 | 2 | 0.348 | 360 | 3 |
Table IV. Network dimensions and training settings. U-Net channel counts run from the input resolution to the bottleneck, and Parameters is the number of trainable parameters. Encoder columns specify token dimension, maximum latent-vector count, attention heads, and Transformer layers. is the number of diffusion steps, is the infidelity weight in Eq. (5), the learning rate is the initial AdamW value, curriculum is the number of epochs over which the lower bound of the training ratio decreases to 20%, and is the number of denoising steps selected on validation data under the nondecreasing constraint described in Appendix B 2.
The observable sampling ratio used in training is bounded below, and a curriculum lowers this bound from 100% to 20% over the initial epochs listed in Table IV. The shot count is drawn uniformly from , where denotes the infinite-shot limit , which retains readout error but has no shot noise. All reported tests use finite shot counts and ratios within the training range.
The PyTorch [63] implementation uses AdamW [64] with minibatches of 512; and the other network-specific settings in Table IV are chosen on validation data. The networks are trained from random initialization for at most 3500 epochs, with validation-based learning-rate reduction and early stopping. The reported weights are those with the highest mean fidelity on 512 fixed validation states over the 15 finite-shot conditions. No structured-state examples or test data enter training or selection; the weights and are fixed during testing.
Number of denoising steps
Reconstruction uses evenly spaced indices over the diffusion steps and the strided posterior update of Appendix B3 [35, 44]. The selected has the highest mean validation fidelity over 240 states per measurement condition, under two conditions. First, the candidates are , so that every reconstruction includes at least one reverse-process update (Sec. II B); is evaluated as a separate control (Appendix E1). Second, is constrained not to decrease with within each input-state ensemble. This constraint is a design assumption: the denoised array, with real entries, grows with , and a larger system is not assigned fewer denoising steps than a smaller one.
Without the constraint, validation would select the same for seven of the eight networks; for the four-qubit network for mixed input states, the unconstrained maximum occurs at , whose fidelity exceeds that at by . Adding to the candidates leaves all eight values unchanged under the constraint, because is still selected at . Without the constraint, however, the four-qubit network for mixed input states and the five-qubit networks for pure and mixed input states would select , whose validation fidelities exceed those at by , , and , respectively. The constraint therefore does not establish that the unconstrained optimum grows with system size or that the selected value is optimal for each measurement condition.
Reverse sampler and initialization
For successive scheduled indices , let be the real representation of the state estimate obtained from . The strided posterior mean and variance are
where dependence on and is omitted from the notation. Ancestral sampling adds Gaussian noise with this variance to the mean; we refer to it as the injected noise. Updates occur only between scheduled indices, and the final denoiser output is the reported state estimate.
The initial Gaussian sample and the injected noise are separate sources of randomness. We vary each while keeping the networks, , test states, measurement outcomes, and observable subsets of each comparison; the sampler of the main comparison was fixed beforehand and is not changed by these controls. The injected noise is scaled by ,
so that recovers ancestral sampling and removes the injected noise. The sampler does not yield the Bayesian posterior mean of the density operator and differs from the deterministic sampler of the denoising diffusion implicit model (DDIM), which uses [65]
where is computed from the direct prediction of rather than predicted by a separately trained noise-prediction network. Its stochasticity parameter is , and ancestral sampling is algebraically equivalent to DDIM with on this schedule [65]. Fidelities are averaged over three replicates of the Gaussian samples and over the 12 finite-shot conditions with incomplete Pauli sets.
Removing the injected noise () raises the condition-averaged fidelity of six networks, with CIs excluding zero for five of them, but lowers it by 0.0035 and 0.0515 for the four- and five-qubit networks for mixed input states (Table V). For the two- and three-qubit networks, no intermediate scale exceeds by more than , and the fidelity decreases toward . Each four- and five-qubit network instead has an intermediate scale, between and 0.75, whose fidelity exceeds that of the better endpoint by 0.0014–0.0037, with CIs above zero. The best tested scale exceeds ancestral sampling by 0.0026–0.0053; because this scale is selected on the test data, the gain is an optimistic estimate of what tuning would achieve. For every network, the fidelity range over the tested scales is largest at , where it is 0.006–0.022, or 0.157 for five-qubit mixed input states, whose fidelity decreases steadily from to ; the range narrows as increases, to at most 0.0012 at . Ancestral sampling has higher fidelity than deterministic DDIM for the six networks with , by 0.0006–0.0026 with CIs above zero; the two-qubit CIs include zero. Because DDIM also changes the deterministic update coefficients, this difference does not isolate the effect of the injected noise.
| Input state | Ancestral | Best | DDIM | Ancestral − DDIM [95% CI] | |||
| 2 | pure | 2 | 0.8609 | 0.8647 | 0.8647 (0.1) | 0.8602 | +0.0008 [−0.0012, +0.0027] |
| 3 | pure | 3 | 0.8915 | 0.8960 | 0.8960 (0) | 0.8889 | +0.0026 [+0.0011, +0.0042] |
| 4 | pure | 3 | 0.9302 | 0.9322 | 0.9336 (0.1) | 0.9293 | +0.0010 [+0.0006, +0.0013] |
| 5 | pure | 3 | 0.9075 | 0.9077 | 0.9114 (0.5) | 0.9068 | +0.0007 [+0.0004, +0.0011] |
| 2 | mixed | 2 | 0.8893 | 0.8920 | 0.8920 (0) | 0.8893 | +0.0000 [−0.0013, +0.0013] |
| 3 | mixed | 3 | 0.9101 | 0.9153 | 0.9153 (0) | 0.9086 | +0.0015 [+0.0002, +0.0027] |
| 4 | mixed | 3 | 0.9326 | 0.9291 | 0.9356 (0.25) | 0.9313 | +0.0013 [+0.0010, +0.0017] |
| 5 | mixed | 3 | 0.9185 | 0.8670 | 0.9211 (0.75) | 0.9178 | +0.0006 [+0.0003, +0.0010] |
Table V. Exploratory reverse-sampler comparison at the validation-selected number of denoising steps . Entries are condition-averaged mean fidelities over the 12 finite-shot conditions with incomplete Pauli sets, averaged over three state-specific initial Gaussian samples . Ancestral sampling corresponds to in Eq. (14), and is the exact strided posterior update with only its injected noise removed. Best gives the highest mean over the six tested scales, with that scale in parentheses; it is selected on the test data. The last column is ancestral minus DDIM with , computed from unrounded means, with its 95% CI. Because fidelities are averaged over , the ancestral column differs slightly from the single-sample Diff-QST means in Table II.
In the main comparison, each state starts from its own Gaussian sample . Sharing one standard Gaussian array across all states and conditions changes the condition-averaged fidelity of each network by at most at and and by at most with deterministic DDIM. The zero array, the mean of , keeps the starting noise level , so it serves as the noisy-sample input rather than as an initial estimate of the density operator. At , starting from it changes the fidelity by at most , except for a decrease of 0.031 for the two-qubit network for pure input states. At , it lowers the fidelity by 0.033 and 0.048 for the two- and five-qubit networks for pure input states and changes it by at most 0.002 for the others; with deterministic DDIM, it raises the fidelity of four networks and lowers that of the other four. Starting from the mean of the initial distribution therefore does not, in general, give the fidelity averaged over Gaussian initial samples. These controls give no reason to replace ancestral sampling from state-specific Gaussian samples in the main comparison: no tested alternative has higher fidelity in all 120 combinations of network and condition, the best alternative differs between networks, and the largest gain in any incomplete-Pauli condition, 0.017, is below the smallest incomplete-Pauli fidelity difference from MLE and SDP, 0.021 (§ 0.3 A). For the comparison with the regression baseline, however, the condition-averaged gain of the best tested alternative for each two- and three-qubit network, 0.0027–0.0053 when selected on the test data, is comparable with the magnitudes of their fidelity differences from the regression baseline, 0.0054–0.0081 (Table II). These comparisons show how sensitive the fidelity is to the sampler settings; they do not test whether the sampled estimates represent calibrated posterior uncertainty.
Statistical and numerical protocol
Paired comparisons
In the main comparison, all estimators receive the same states, observable subsets, and measurement outcomes in each condition. For test state , estimator , and subset , define . We first average over subsets and then over states:
Here is the number of test states, and for and for the complete Pauli set. For each reference , the paired differences are , and their mean is . Its 95% confidence interval (CI) uses 10,000 percentile bootstrap resamples of the shared state indices, so that all estimators are resampled together [49, 66]; CIs for the mean of a single estimator use the same resampling. These CIs condition on the selected subsets, trained network, and . They are pointwise, not simultaneous confidence intervals across the grid. The multiplicity correction in § 0.3 A applies Holm’s step-down procedure to two-sided bootstrap -values from resamples per comparison. Other appendices state only departures from this procedure.
Subset dependence is largest relative to the fidelity difference at , , where . Averaging over all 455 three-observable subsets for the same test states and measurement outcomes, Diff-QST retains higher mean fidelity than both references in all six conditions. The differences are 0.004–0.010 for pure and 0.012–0.014 for mixed input states, smaller than for the three main-comparison subsets. The reversal for the independent Haar-random ensemble at , (§ 0.3 D) reflects subset choice: applied to the main test states, the three subsets used for that ensemble also give negative differences, and per-subset differences across the two state sets correlate at 0.92–0.93. Averaged over all 455 subsets, the differences for that ensemble are 0.013, 0.009, and 0.006 at , , and shots, with both CIs above zero.
Reference estimators
MLE starts from the positive-semidefinite projection of the minimum-norm linear least-squares estimate, mixed with at weight . With the Frobenius-norm change and the absolute log-likelihood change per iteration, it stops when and either or , or when , within iterations. SDP is solved by SCS [67, 68] through CVXPY [69] at tolerance . The SDP weights in Eq. (6) are
where is the number of outcomes in Eq. (2) and is the posterior mean of the outcome probability under the Jeffreys prior; because , each is positive. Outside the early-stopping test of Appendix C 3, every MLE reconstruction satisfies the stopping criteria except one in the readout test of Appendix F 2, which reached the iteration limit; for some structured states, the limit is raised. Outside the early-stopping test, every SDP solution is optimal, with one structured-state problem solved by Clarabel [70] instead.
In many incomplete-Pauli conditions, SDP fits the data almost exactly: the optimized measurement-residual term of Eq. (6), averaged over states and subsets, is below in 30 of the 96 incomplete-Pauli conditions, including every system size and input-state ensemble at , shots. In such conditions, the regularizer mainly selects among data-consistent states, although it need not select the lowest-purity member of the unregularized solution set.
Reconstruction time and early stopping
For timing, MLE and SDP use the same objectives and stopping criteria as in the accuracy comparison, and reference times are mean serial reconstruction times per state. Timings exclude training, network-parameter loading, measurement acquisition, data loading, construction of the effective observables, and fidelity evaluation; GPU times include synchronization and transfer of reconstructed states to the CPU. For Diff-QST and the regression baseline, time per state is the elapsed time divided by the number of reconstructed states, on the same GPU and software settings. For all eight networks across the five ratios at shots, the regression baseline takes 1.5–2.4 ms per state at batch size 1, and Diff-QST takes 1.7–2.4 times as long; at batch size 256, the factor is 1.1–1.8. For four and five qubits over incomplete-Pauli ratios at shots, Diff-QST takes 4.70–5.08 ms per state versus 2.05–2.31 ms for the regression baseline.
Reference times are measured in every nominal shot-ratio condition, whereas each Diff-QST and regression network is timed at every ratio at shots, and that time is used for all shot counts because only input values change. Figure 4 and Table VI summarize the condition-level ratio of mean reference time to Diff-QST’s mean time per state, with each of the 30 conditions at a given system size weighted equally. Averaging the condition-level mean times over the 30 conditions instead gives ratios of averaged reference time to averaged GPU Diff-QST time of 3.85, 20.7, 111, and 495 for MLE and 3.79, 5.53, 21.7, and 127 for SDP at , respectively. At batch size 256, Diff-QST’s GPU time per state is shorter than the mean MLE and SDP times in every nominal condition.
| Diff-QST on GPU | Diff-QST on CPU | |||
| MLE/Diff-QST | SDP/Diff-QST | MLE/Diff-QST | SDP/Diff-QST | |
| 2 | 1.55 (0.107) | 3.67 (3.07) | 0.977 (0.0584) | 2.63 (1.68) |
| 3 | 9.76 (0.449) | 5.53 (2.80) | 1.74 (0.0829) | 0.963 (0.518) |
| 4 | 70.8 (14.9) | 21.1 (13.6) | 10.2 (2.25) | 2.98 (2.04) |
| 5 | 482 (135) | 125 (71.9) | 31.4 (9.22) | 8.25 (4.83) |
Table VI. Median reference/Diff-QST reconstruction-time ratios at batch size 1, with minima in parentheses, across 30 equally weighted conditions per system size. MLE and SDP use the 32-core CPU configuration; Diff-QST runs on the GPU or uses 32 CPU cores as indicated. A minimum above one means shorter Diff-QST mean times throughout the evaluated conditions.
On the 32-core CPU configuration at batch size 1, Diff-QST takes 31.04–35.51 ms per state at four qubits and 74.57–78.23 ms at five. It has lower condition-averaged time than MLE at all four sizes and than SDP at two, four, and five qubits; at three qubits, SDP’s condition-averaged time is 0.971 times that of Diff-QST. These results describe the tested implementations and hardware, not hardware-independent speedups.
MLE and SDP times vary across test states, as in the five-qubit example of § 0.3 B. Across the five-qubit grid, MLE mean times and their spread across states are largest at ; for pure input states at shots and this ratio, one reconstruction takes 16.8 s. SDP varies less but has occasional long reconstructions: for pure input states at shots and , its iteration counts range from 250 to 5575, and one reconstruction takes 6.06 s. The per-state reference times also include fluctuations from a shared server, so their spread does not isolate the contribution of solver iterations.
For both input-state ensembles at and 5 and at and , we also stop MLE after 0, 10, 30, 100, 300, , , , or iterations and solve SDP at SCS tolerances from to ; iteration 0 is the MLE starting estimate. MLE times are the elapsed times at each iteration limit within one run per state, and each SDP tolerance is a separate solve; all are measured on the 15 states and observable subsets of the CPU reference timing (Figure 6). Projected least squares (PLS), the density operator closest in Frobenius norm to the least-squares estimate [14], is included as a noniterative estimator.
Figure 6. Mean fidelity versus reconstruction time per state for pure input states at , (a)–(c) and (d)–(f) at , , and . Filled circles show Diff-QST and open diamonds the regression baseline, both at their GPU times per state at batch size 1; crosses show projected least squares (PLS). Open squares show MLE stopped after 0 to iterations and filled squares the converged MLE; open triangles show SDP at SCS tolerances from to and filled triangles at . All times are measured on the 15 states and observable subsets of the reconstruction-time comparison. Vertical scales differ between panels. Mixed input states, which complete the 12 conditions of the early-stopping comparison, are not shown.
Early stopping can raise the mean fidelity above the converged value, by up to 0.045 for MLE and 0.142 for SDP, but in no condition to that of Diff-QST (§ 0.3 B). Stopping each estimator when its mean fidelity first comes within 0.001 of its converged value leaves MLE and SDP 15–676 and 13–145 times slower than GPU Diff-QST. No tested stopping point of MLE or SDP, including iteration 0 and SCS tolerance , is faster than GPU Diff-QST in any of the 12 conditions. PLS has mean fidelity 0.40–0.88, and its time depends on the least-squares implementation. Faster MLE and SDP algorithms (§ 0.3 B) would reach the same converged estimates wherever these are unique: for SDP, whose objective is strictly convex, and for MLE with the complete Pauli set.
Further accuracy comparisons
Observables per state parameter and ratios below the training range
Let denote the number of real parameters of the input-state family, for pure and for mixed input states. Expressing the measurement count as distinguishes the fraction of the complete Pauli set from the number of observables available per state parameter (Figure 7). The fixed nominal preparation channel acts as an invertible linear map on density operators, so each target family has the same dimension as its input-state family. To determine whether the location of the largest fidelity difference scales with below the 20% lower bound of the training ratio, we retrained the four- and five-qubit networks with this bound reduced to 10% and 5%, respectively. The retrained networks retain the architecture and training settings of the corresponding networks in Table IV and use ; at , the valida-
Figure 7. Fidelity difference versus , the number of measured observables per real parameter of the input-state family, for (a)–(c) pure and (d)–(f) mixed input states at ; along each curve, is fixed and varies. Original networks (open markers) were trained with observable sampling ratios down to 20%; retrained networks (filled markers), down to 10% for and 5% for . Most error bars, 95% CIs relative to the better of MLE and SDP, are smaller than the markers. Vertical lines mark , the dimensional lower bound on the number of measured quantities for identifying every member of the input-state family, and , about the number that topological obstructions can require [31]; the horizontal line marks .
tion maxima at , within of , are excluded by the nondecreasing constraint of Appendix B 2. They are evaluated on the 300 test states of the main comparison at the nominal ratios and at to for and to for , in steps of , with MLE and SDP computed from the same observable subsets and measurement outcomes. These networks are used only for this test.
For the retrained networks, both CIs lie above zero in 119 of their 120 conditions; at , pure input states, shots, and , the positive mean difference has a CI relative to MLE that includes zero. Retraining changes fidelity in both directions at ; at , the mean fidelities of the retrained networks lie within 0.006 of those of the original networks, except for the five-qubit network for mixed input states, whose means are 0.002–0.020 lower.
The exception to the range – given in § 0.3 A for is five-qubit mixed input states at shots, for which the largest observed fidelity difference occurs at . At shots per observable, all four retrained networks attain their largest observed fidelity differences at higher ratios within the evaluated grid, –. This shift is consistent with larger shot noise weakening the constraints supplied by each observable, so that more measurements are needed before the references narrow the fidelity difference. The location of the largest observed fidelity difference on the axis therefore depends on .
From the observed maxima to the complete Pauli set, the reference estimators show larger overall fidelity increases than Diff-QST, narrowing the fidelity difference. This decrease does not imply that Diff-QST has already saturated. For example, for five-qubit pure input states at shots, increasing the number of measured observables from to the complete Pauli set raises the retrained network’s mean fidelity from 0.707 to 0.993, while MLE and SDP improve from 0.464 to 0.932 and from 0.363 to 0.923, respectively. In this example, the smallest evaluated measurement counts attaining are for Diff-QST and for both references.
Hedged maximum-likelihood reference
Hedged MLE (HMLE) maximizes [51], where is the binary likelihood used by MLE. The determinant term enforces full rank, and the exponent gives the add-1/2 (Jeffreys) frequency rule for an ideal projective measurement in one orthonormal basis. HMLE uses the same starting state and stopping criteria as MLE, and all its reconstructions satisfy them. Table VII summarizes the incomplete-Pauli comparison; in the main text remains relative to the better of MLE and SDP.
| Diff-QST − HMLE [95% CI] | Positive | median [max] | ||
| Pure input states | ||||
| 2 | 0.8440 | +0.0173 [+0.0138, +0.0207] | 9/12 | −0.0398 [+0.0066] |
| 3 | 0.8504 | +0.0421 [+0.0398, +0.0444] | 12/12 | −0.0681 [+0.0057] |
| 4 | 0.8514 | +0.0789 [+0.0780, +0.0798] | 12/12 | −0.0879 [−0.0184] |
| 5 | 0.8021 | +0.1058 [+0.1051, +0.1066] | 12/12 | −0.0817 [+0.0184] |
| Mixed input states | ||||
| 2 | 0.8751 | +0.0140 [+0.0120, +0.0160] | 10/12 | −0.0497 [+0.0032] |
| 3 | 0.8727 | +0.0391 [+0.0374, +0.0408] | 12/12 | −0.1054 [−0.0038] |
| 4 | 0.8539 | +0.0787 [+0.0777, +0.0797] | 12/12 | −0.1455 [−0.0385] |
| 5 | 0.7958 | +0.1221 [+0.1205, +0.1236] | 12/12 | −0.1540 [−0.0756] |
Table VII. Sensitivity of the comparison to the reference estimator and the accuracy measure. The HMLE columns cover the 12 finite-shot conditions with incomplete Pauli sets: the condition-averaged mean fidelity of HMLE, the difference computed from unrounded means with its 95% CI, and the number of conditions with a positive condition-level difference. The last column gives the median and, in brackets, the maximum over the 15 shot-ratio conditions of , the trace-distance difference of Diff-QST from the closer of MLE and SDP; negative values favor Diff-QST. The Diff-QST mean fidelities are those in the column of Table II.
Relative to HMLE, Diff-QST has higher mean fidelity in 91 of the 96 incomplete-Pauli conditions and all 24 complete-Pauli conditions; the CIs lie above zero in 89 and 24, respectively, and none lies entirely below zero. The five reversals occur at , , for pure input states at all shot counts and mixed input states at and shots, with differences from to .
Trace distance
The trace distance of § 0.3 A is evaluated on the reconstructions of the main comparison and enters neither training nor selection. Mean distances to are formed as in Eq. (16); negative differences favor Diff-QST. The last column of Table VII summarizes , the difference from the reference with the smaller mean distance.
The 112 conditions in which Diff-QST is closer to the target (§ 0.3 A) include 91 of the 96 incomplete-Pauli conditions. Three of the four in which the references are closer use the complete Pauli set at shots, for two- and three-qubit pure and two-qubit mixed input states, where the fidelity difference is also smallest (§ 0.3 A). The fourth is five-qubit pure input states at shots and . There, the target states are full rank, whereas the MLE and SDP estimates are often rank deficient [51]: on average, 13.3 and 19.2 of their 32 eigenvalues lie below , compared with 3.2 for Diff-QST. Because fidelity depends on the square roots of the eigenvalues, these near-zero eigenvalues affect fidelity more strongly than trace distance. The dominant eigenvectors of the MLE and SDP estimates also overlap more with those of the targets, with mean squared overlap 0.998 for both, compared with 0.949 for Diff-QST. In the remaining four conditions, all at and , both CIs include zero.
Further comparisons with the regression baseline
For the condition-averaged comparisons in Table II, CIs resample state indices after averaging subsets and conditions. The shared learning rate, ratio curriculum, and infidelity weight are listed in Table IV.
At four and five qubits, the fidelity differences are larger at low observable sampling ratios; all 16 shot-averaged four- and five-qubit differences for incomplete Pauli sets have CIs above zero. At , all four- and five-qubit networks have higher mean fidelity than the regression baseline at every shot count, with CIs above zero. Adding the three complete-Pauli conditions preserves all eight directions in Table II, with CIs still excluding zero. Table VIII averages over shot counts within each ratio; 36 of its 40 CIs exclude zero. At shots and , the regression baseline has slightly higher mean fidelity for both four-qubit networks.
| Input state | 20% | 40% | 60% | 80% | 100% | |
| 2 | pure | −0.0254 | −0.0052† | −0.0009† | +0.0008 | +0.0002 |
| 3 | pure | −0.0274 | −0.0004† | −0.0023 | −0.0024 | −0.0022 |
| 4 | pure | +0.0424 | +0.0369 | +0.0182 | +0.0073 | +0.0008 |
| 5 | pure | +0.1343 | +0.0723 | +0.0392 | +0.0241 | +0.0141 |
| 2 | mixed | −0.0108 | −0.0067 | −0.0060 | −0.0011 | −0.0004 |
| 3 | mixed | −0.0185 | +0.0002† | −0.0019 | −0.0016 | −0.0013 |
| 4 | mixed | +0.0284 | +0.0338 | +0.0180 | +0.0077 | +0.0007 |
| 5 | mixed | +0.0380 | +0.0221 | +0.0125 | +0.0085 | +0.0051 |
Table VIII. Mean fidelity difference from the regression baseline by observable sampling ratio. Each entry equally weights the three finite shot counts. Positive values favor Diff-QST. A dagger marks the entries whose 95% CI, obtained by resampling states after averaging over the three shot counts, includes zero; all other CIs exclude zero.
Table IX reports the three additional training runs per estimator for each system size and input-state ensemble, trained from further random initializations with unchanged settings; § 0.3 C summarizes their signs. In the two four-pair sets that favor the regression baseline, by 0.010 for pure and 0.008 for mixed input states, the Diff-QST network has a lower condition-averaged mean fidelity, 0.899 and 0.903, than the other three Diff-QST networks, 0.930–0.941.
| Regression | Diff-QST | Regression | Diff-QST | Regression | Diff-QST | ||||
| Pure input states | |||||||||
| 2 | 0.8697 | 0.8579 | −0.0118 | 0.8690 | 0.8562 | −0.0128 | 0.8700 | 0.8623 | −0.0077 |
| 3 | 0.8972 | 0.8924 | −0.0049 | 0.8992 | 0.8940 | −0.0052 | 0.8990 | 0.8926 | −0.0064 |
| 4 | 0.8994 | 0.9313 | +0.0319 | 0.9083 | 0.8986 | −0.0097 | 0.9015 | 0.9410 | +0.0396 |
| 5 | 0.8986 | 0.9190 | +0.0204 | 0.9084 | 0.9177 | +0.0093 | 0.8816 | 0.9189 | +0.0373 |
| Mixed input states | |||||||||
| 2 | 0.8953 | 0.8861 | −0.0092 | 0.8946 | 0.8873 | −0.0073 | 0.8950 | 0.8882 | −0.0068 |
| 3 | 0.9178 | 0.9092 | −0.0086 | 0.9166 | 0.9048 | −0.0119 | 0.9166 | 0.9094 | −0.0072 |
| 4 | 0.9208 | 0.9381 | +0.0173 | 0.9098 | 0.9392 | +0.0294 | 0.9109 | 0.9032 | −0.0077 |
| 5 | 0.8639 | 0.9267 | +0.0628 | 0.8971 | 0.9141 | +0.0169 | 0.9019 | 0.9237 | +0.0217 |
Table IX. Additional training runs for the comparison with the regression baseline. Each system size and input-state ensemble has three further runs per estimator, paired by initialization seed: 1001, 2001, and 3001 at , 3, 4 and 51001, 51002, and 51003 at . Entries are mean fidelities equally weighting the same 12 finite-shot conditions with incomplete Pauli sets as Table II; Diff-QST uses at and otherwise. is computed from unrounded means, and positive values favor Diff-QST. All 24 differences have 95% CIs that exclude zero, with half-widths 0.0009–0.0039; the CIs use 10,000 paired bootstrap resamples of the 300 test states after averaging subsets and conditions, and each is conditional on the corresponding pair of trained networks. The reported networks in Table II are not repeated here.
is the reconstruction; the reconstruction uses the same initial sample.
In the columns of Table II, compares one network evaluation per estimator, and compares the two reconstruction procedures at fixed diffusion-network parameters. The control is not evaluated under the distribution shifts of Appendix E3.
One- and three-step reconstruction
We evaluate the trained networks at and the selected for in the 15 finite-shot conditions of the main comparison, and, for the five-qubit networks, the six additional conditions of Appendix E2. No weights are retrained, and is not reselected. The two-qubit networks, whose selected value is , are outside this control. For , the denoiser is applied once at , the noise level at which reconstruction starts, to the initial Gaussian sample , and its state estimate
Total shot budgets at specified fidelity
We extend the five-qubit test to at , retaining the trained networks, , and the 300 targets of the main comparison for each input-state ensemble. This exploratory extension adds six conditions per ensemble to the 15-condition finite-shot grid of the main comparison, giving 21 evaluated allocations, that is, combinations of and per estimator and ensemble. At each specified fidelity, Table X selects the smallest evaluated whose condition-level mean fidelity meets the target. Figure 8 shows the dependence on for the three ratios evaluated at the additional shot counts.
| Target | Input state | Estimator | 95% CI | ||||
| 0.90 | pure | Diff-QST | 613 | 100 | 61,300 | 0.9145 | [0.9134, 0.9157] |
| 0.90 | pure | Regression | 1023 | 100 | 102,300 | 0.9292 | [0.9272, 0.9312] |
| 0.90 | mixed | Diff-QST | 613 | 100 | 61,300 | 0.9090 | [0.9076, 0.9104] |
| 0.90 | mixed | Regression | 818 | 100 | 81,800 | 0.9287 | [0.9272, 0.9301] |
| 0.95 | pure | Diff-QST | 1023 | 100 | 102,300 | 0.9583 | [0.9575, 0.9590] |
| 0.95 | pure | Regression | 1023 | 200 | 204,600 | 0.9581 | [0.9570, 0.9592] |
| 0.95 | mixed | Diff-QST | 1023 | 100 | 102,300 | 0.9525 | [0.9516, 0.9534] |
| 0.95 | mixed | Regression | 613 | 500 | 306,500 | 0.9547 | [0.9539, 0.9555] |
Table X. Smallest evaluated total shot budget attaining the specified mean fidelity at five qubits, over the expanded 21-allocation grid per input-state ensemble. is the number of Pauli observables and the shots per observable. The attained and its 95% CI over test states describe the selected condition, not uncertainty in the minimum budget or its ratio between estimators.
Figure 8. Five-qubit condition-level mean fidelity of Diff-QST and the regression baseline versus total shot budget for (a)–(c) pure and (d)–(f) mixed input states at (). Each curve contains . Symbols give means over 300 test states; bars are 95% bootstrap CIs over test states. Connecting lines guide the eye.
At target 0.95, both Diff-QST allocations use the complete Pauli set, so the reductions given in § 0.3 C concern total shots, not fewer measured observables. At target 0.90, all four selected allocations use 100 shots per observable, so the reductions apply to both total shots and measured observables. Using the lower CI endpoints instead of the means selects the same budgets. In all 12 additional conditions, Diff-QST has higher mean fidelity than the regression baseline, with CIs above zero and differences of 0.0080–0.0439.
The control reaches mean fidelity 0.90 at and 0.95 at in both input-state
ensembles, the same smallest evaluated budgets as , so the reductions relative to the regression baseline persist without repeated denoising. The allocation grid is nonuniform: 200 and 500 shots are not tested at and 40%.
Distribution shifts
We apply the trained regression networks, without retraining, to the distribution-shift tests of § 0.3 D. Table XI gives the fidelity differences of Diff-QST from the regression baseline by system size; for the structured states, the CIs resample acquisitions and states as in Appendix pendix F 3. At four and five qubits, the CI lies above zero in 235 of the 238 spectral, readout, and preparation-noise conditions in which Diff-QST has the higher mean fidelity (§ 0.3 D). For pure input states, the difference grows under weaker preparation noise: at five qubits it is 0.075 and 0.073 at quarter and half strength, compared with 0.063 for the same shot counts and ratios under nominal noise. It shrinks in the lower purity range. At two and three qubits, the regression baseline retains its nominal advantage under the readout and mixed-state distribution changes and under weaker preparation noise for mixed input states; for pure input states, its advantage reverses at quarter preparation strength and shrinks at half strength.
| Change | Input states | Cond. | ||||
| Nominal | pure | 15 | −0.0061 (4/5) | −0.0069 (1/12) | +0.0211 (14/1) | +0.0568 (15/0) |
| Nominal | mixed | 15 | −0.0050 (2/11) | −0.0046 (4/9) | +0.0177 (14/1) | +0.0172 (15/0) |
| Preparation ×0.25 | pure | 9 | +0.0043 (5/0) | +0.0030 (5/1) | +0.0277 (8/1) | +0.0753 (9/0) |
| Preparation ×0.25 | mixed | 9 | −0.0053 (2/6) | −0.0037 (1/6) | +0.0181 (7/1) | +0.0281 (8/0) |
| Preparation ×0.5 | pure | 9 | −0.0010 (3/0) | −0.0028 (4/3) | +0.0261 (8/1) | +0.0729 (9/0) |
| Preparation ×0.5 | mixed | 9 | −0.0059 (2/7) | −0.0050 (1/7) | +0.0183 (8/1) | +0.0235 (9/0) |
| Readout ×0.5 | pure | 15 | −0.0069 (4/6) | −0.0069 (2/12) | +0.0207 (14/1) | +0.0549 (15/0) |
| Readout ×0.5 | mixed | 15 | −0.0050 (1/11) | −0.0054 (2/12) | +0.0179 (13/1) | +0.0184 (15/0) |
| Readout ×2 | pure | 15 | −0.0046 (4/3) | −0.0092 (0/13) | +0.0175 (14/1) | +0.0617 (15/0) |
| Readout ×2 | mixed | 15 | −0.0053 (2/11) | −0.0035 (6/7) | +0.0186 (14/0) | +0.0221 (15/0) |
| Degeneracy lifted | mixed | 15 | −0.0062 (2/12) | −0.0043 (4/10) | +0.0189 (14/1) | +0.0166 (15/0) |
| Lower purity range | mixed | 15 | −0.0033 (4/9) | −0.0105 (0/15) | +0.0051 (10/3) | +0.0101 (15/0) |
| Canonical GHZ | pure | 45 | −0.0395 (6/26) | −0.0148 (0/19) | −0.0253 (0/39) | −0.0292 (1/34) |
| Canonical W, cluster | pure | 30 | +0.0050 (11/3) | −0.0275 (0/24) | +0.0067 (20/1) | +0.0289 (29/0) |
| Rotated ensembles | pure | 45 | −0.0140 (10/19) | −0.0110 (2/34) | +0.0195 (41/3) | +0.0599 (45/0) |
| Haar-random ensemble | pure | 15 | −0.0068 (3/4) | −0.0082 (2/12) | +0.0219 (14/1) | +0.0563 (15/0) |
Table XI. Mean fidelity difference from the regression baseline under the distribution shifts of § 0.3 D. Each entry averages the condition-level differences over the conditions of its row at one system size, whose number is given by Cond.; parentheses give the numbers of these conditions whose 95% CI lies above and below zero. Positive values favor Diff-QST. Nominal rows use the 15 finite-shot conditions of the main comparison, and preparation-noise rows use , 60%, and 100%. Structured rows pool the canonical representatives or ensembles named in the first column.
Under weaker preparation noise, the regression baseline also reverses relative to MLE and SDP more often than Diff-QST (§ 0.3 D); at , quarter strength, shots, and , the mean fidelities are 0.648, 0.798, and 0.899 for the regression baseline, Diff-QST, and MLE. Under the readout and mixed-state distribution changes, the numbers of conditions with both CIs above zero differ from those of Diff-QST by at most two. In the 39 two- and three-qubit structured conditions in which both CIs of Diff-QST lie below zero, the regression baseline has higher mean fidelity than Diff-QST in all 39, with CIs excluding zero in 34, and exceeds the mean fidelities of both MLE and SDP in 22.
Distribution-shift tests
Mixed-state spectra and purity
For the first distribution shift of mixed input states, we sample an initial spectrum uniformly on the probability simplex and adjust it to a separately and uniformly drawn . To reach the drawn , the initial spectrum is mixed with the uniform spectrum if its purity is too high, or with the pure-state spectrum concentrated on its largest component if its purity is too low. A Haar-random unitary sets the eigenvectors. The resulting spectra are not uniformly distributed over the spectra of a given purity. The second shift retains Eq. (7) but samples uniformly on .
Over the 15 conditions of each system size, the median differences from the better of MLE and SDP at , 3, 4, and 5 are 0.073, 0.117, 0.184, and 0.223 after lifting spectral degeneracy and 0.052, 0.141, 0.213, and 0.304 for the lower purity range, compared with 0.077, 0.124, 0.191, and 0.229 for the original distribution. The exceptions noted in § 0.3 D occur at shots and : after lifting spectral degeneracy, the difference from the better of MLE and SDP is −0.0036 [−0.0042, −0.0030], with the 95% CI in brackets. For the lower purity range, the differences are −0.00084 [−0.00094, −0.00074] at and −0.00058 [−0.00100, −0.00017] at .
Readout miscalibration
Actual readout rates are multiplied by or 2, while the assumed rates remain nominal. Over the 60 conditions of each input-state ensemble, the median fidelity difference from the better of MLE and SDP is 0.1001 at and 0.1427 at for pure input states, and 0.1369 and 0.1520 for mixed input states. Targets and subsets are shared across scales. Doubling the readout-error rates gives median fidelity losses, relative to nominal readout, of 0.0108, 0.0443, and 0.0421 for Diff-QST, MLE, and SDP, respectively, for pure input states, and 0.0203, 0.0257, and 0.0190 for mixed input states. Rescaling the actual rates changes both the readout error in the data and its mismatch with the assumed model, so the test does not isolate calibration bias.
have negative median differences, whereas W and linear cluster states have positive medians. The rotated ensembles have positive medians; at , where both CIs lie below zero for every shot count (§ 0.3 D), the differences are to and remain to after averaging over all 455 three-observable subsets, with both CIs below zero for every ensemble and shot count.
Structured states
The canonical GHZ states are , with , , for GHZ, GHZ, and GHZ. The W state is the equal superposition of the computational-basis strings containing exactly one 1. The linear cluster state is obtained by applying controlled-Z gates between adjacent qubits of an open chain initialized in , where .
Figure 9 shows the condition-level fidelity difference from the better of MLE and SDP for the canonical states and for the locally rotated and Haar-random ensembles. The canonical means average 100 independent measurement acquisitions per condition, each drawing its own observable subset for ; the ensemble means average 300 states, each with the three subsets shared by the ensemble. The CIs resample acquisitions for a canonical state and states for an ensemble, so the canonical CIs include subset variation, whereas the ensemble CIs are conditional on the three subsets.
Figure 9. Fidelity differences $\Delta\overline{F}=\overline{F}^{(\mathrm{Diff\mathchar-QST})}-\max\{\overline{F}^{(\mathrm{MLE})},\overline{F}^{(\mathrm{SDP})}\}$ for structured states over all measurement conditions, using the networks for pure input states, on a common color scale; positive values favor Diff-QST. Panel rows correspond to $10^2$, $10^3$, and $10^4$ shots per observable and panel columns to $n=2,\ldots,5$; within each panel, columns are observable sampling ratios, the upper five rows are the canonical GHZ$^{+}$, GHZ$^{-}$, GHZ$^{i}$, W, and linear cluster states, and the lower four rows are the locally rotated GHZ, W, and linear-cluster ensembles and the Haar-random ensemble. Each canonical-state mean uses 100 independent measurement acquisitions, and each ensemble mean 300 states. CIs are not displayed; {xref}Table XII` summarizes them for the canonical states.
Local rotations preserve ideal-input entanglement but can change target purity under the basis-dependent preparation channel. At , GHZ, W, and linear cluster states are locally equivalent, so their rotated ensembles sample the same distribution; canonical representatives can differ under basis-dependent preparation noise. For the locally rotated ensembles at , median fidelity differences range from 0.108 to 0.183, with both CIs above zero in all 15 conditions. Table XII shows the more variable canonical-state results. At , all five canonical representatives are maximally entangled and atypical of Haar-random states; the GHZ representatives
| State | |||||
| GHZ | (4, 8) | (12, 2) | (13, 0) | (15, 0) | 0.566 |
| GHZ | (5, 5) | (9, 2) | (12, 0) | (15, 0) | 0.566 |
| GHZ | (6, 7) | (12, 0) | (13, 0) | (15, 0) | 0.566 |
| W | (9, 1) | (13, 0) | (15, 0) | (15, 0) | 0.671 |
| Linear cluster | (12, 2) | (10, 3) | (13, 0) | (15, 0) | 0.553 |
Table XII. Canonical-state fidelity differences of Diff-QST from the better of MLE and SDP, summarized by the median over 15 measurement conditions. Parentheses give the numbers of conditions in which both CIs lie above zero and below zero, respectively. The final column gives the purity of the five-qubit target.
Weaker preparation noise
All three preparation-channel parameters are scaled together by or 0.5, leaving the two ideal input-state ensembles unchanged. The target is the resulting prepared state at each strength. Over the 36 conditions of each input-state ensemble, the median fidelity difference from the better of MLE and SDP is 0.0329 at and 0.0642 at for pure input states, and 0.0898 and 0.1082 for mixed input states. At the most negative difference, (§ 0.3 D), the mean fidelities are 0.7977, 0.8987, and 0.8937 for Diff-QST, MLE, and SDP, respectively. These tests change the target distribution; they neither isolate purity from other state properties nor represent temporal fluctuations of the noise during one acquisition.
References
References
- [1]D. F. V. James, P. G. Kwiat, W. J. Munro, and A. G. White, Measurement of qubits, Phys. Rev. A 64, 052312 (2001).DOI
- [2]M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information: 10th Anniversary Edition (Cambridge University Press, Cambridge, 2010).
- [3]J. Eisert, D. Hangleiter, N. Walk, I. Roth, D. Markham, R. Parekh, U. Chabaud, and E. Kashefi, Quantum certification and benchmarking, Nat. Rev. Phys. 2, 382 (2020).
- [4]S. T. Flammia and Y.-K. Liu, Direct fidelity estimation from few Pauli measurements, Phys. Rev. Lett. 106, 230501 (2011).
- [5]H.-Y. Huang, R. Kueng, and J. Preskill, Predicting many properties of a quantum system from very few measurements, Nat. Phys. 16, 1050 (2020).DOI
- [6]M. Cramer, M. B. Plenio, S. T. Flammia, R. Somma, D. Gross, S. D. Bartlett, O. Landon-Cardinal, D. Poulin, and Y.-K. Liu, Efficient quantum state tomography, Nat. Commun. 1, 149 (2010).DOI
- [7]D. Gross, Y.-K. Liu, S. T. Flammia, S. Becker, and J. Eisert, Efficient quantum state tomography via compressed sensing, Phys. Rev. Lett. 105, 150401 (2010).DOI
- [8]J. Haah, A. W. Harrow, Z. Ji, X. Wu, and N. Yu, Sample-optimal tomography of quantum states, IEEE Trans. Inf. Theory 63, 5628 (2017).
- [9]S. Bravyi, S. Sheldon, A. Kandala, D. C. McKay, and J. M. Gambetta, Mitigating measurement errors in multiqubit experiments, Phys. Rev. A 103, 042605 (2021).DOI
- [10]L. Chen, H.-X. Li, Y. Lu, C. W. Warren, C. J. Križan, S. K. Kosen, M. Rommel, S. Ahmed, A. Osman, J. Biznárová, A. Fadavi Roudsari, B. Lienhard, M. Caputo, K. Grigoras, L. Grönberg, J. Govenius, A. F. Kockum, P. Delsing, J. Bylander, and G. Tancredi, Quantum qubit readout fidelity at the threshold for quantum error correction without a quantum-limited amplifier, npj Quantum Inf. 9, 26 (2023).DOI
- [11]Z. Hradil, Quantum-state estimation, Phys. Rev. A 55, R1561 (1997).DOI
- [12]J. Řeháček, Z. Hradil, and M. Ježek, Iterative algorithm for reconstruction of entangled states, Phys. Rev. A 63, 040303(R) (2001).DOI
- [13]J. Řeháček, Z. Hradil, E. Knill, and A. I. Lvovsky, Diluted maximum-likelihood algorithm for quantum tomography, Phys. Rev. A 75, 042108 (2007).DOI
- [14]J. A. Smolin, J. M. Gambetta, and G. Smith, Efficient method for computing the maximum-likelihood quantum state from measurements with additive Gaussian noise, Phys. Rev. Lett. 108, 070502 (2012).pubmed.ncbi.nlm.nih.gov/22401185
- [15]Y. Sunada, S. Kono, J. Iles, T. Sugiyama, Y. Suzuki, T. Okubo, S. Tamate, Y. Tabuchi, and Y. Nakamura, Efficient tomography of microwave photonic cluster states, PRX Quantum 7, 010323 (2026).DOI
- [16]G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, and G. Carleo, Neural-network quantum state tomography, Nat. Phys. 14, 447 (2018).DOI
- [17]T. Schmale, M. Reh, and M. Gärtner, Efficient quantum state tomography with convolutional neural networks, npj Quantum Inf. 8, 115 (2022).DOI
- [18]J. Carrasquilla, G. Torlai, R. G. Melko, and L. Aolita, Reconstructing quantum states with generative models, Nat. Mach. Intell. 1, 155 (2019).DOI
- [19]Y. Wei, W. A. Coish, P. Ronagh, and C. A. Muschik, Neural-shadow quantum state tomography, Phys. Rev. Research 6, 023250 (2024).DOI
- [20]S. Lohani, B. T. Kirby, M. Brodsky, O. Danaci, and R. T. Glasser, Machine learning assisted quantum state estimation, Mach. Learn.: Sci. Technol. 1, 035007 (2020).DOI
- [21]O. Danaci, S. Lohani, B. T. Kirby, and R. T. Glasser, Machine learning pipeline for quantum state estimation with incomplete measurements, Mach. Learn.: Sci. Technol. 2, 035014 (2021).arxiv.org/abs/2012.03104
- [22]D. Koutný, L. Motka, Z. Hradil, J. Řeháček, and L. L. Sánchez-Soto, Neural-network quantum state tomography, Phys. Rev. A 106, 012409 (2022).DOI
- [23]Y. Zhu, Y.-D. Wu, G. Bai, D.-S. Wang, Y. Wang, and G. Chiribella, Flexible learning of quantum states with generative query neural networks, Nat. Commun. 13, 6222 (2022).DOI
- [24]Y. Quek, S. Fort, and H. K. Ng, Adaptive quantum state tomography with neural networks, npj Quantum Inf. 7, 105 (2021).arxiv.org/abs/1812.06693
- [25]A. M. Palmieri, E. Kovlakov, F. Bianchi, D. Yudin, S. Straupe, J. D. Biamonte, and S. Kulik, Experimental neural network enhanced quantum tomography, npj Quantum Inf. 6, 20 (2020).DOI
- [26]S. Ahmed, C. Sánchez Muñoz, F. Nori, and A. F. Kockum, Quantum state tomography with conditional generative adversarial networks, Phys. Rev. Lett. 127, 140502 (2021).DOI
- [27]H. Ma, Z. Sun, D. Dong, C. Chen, and H. Rabitz, Mapping of quantum states from structured measurements via quantum-aware transformer, IEEE Trans. Cybern. 55, 2571 (2025).DOI
- [28]H. Ma, Z. Sun, D. Dong, and D. Gong, Learning informative latent representation for quantum state tomography, IEEE Trans. Emerg. Top. Comput. Intell. 10, 2199 (2026).arxiv.org/abs/2310.00518
- [29]A. M. Palmieri, G. Müller-Rigat, A. K. Srivastava, M. Lewenstein, G. Rajchel-Mieldzioć, and M. Płodzień, Enhancing quantum state tomography via resource-efficient attention-based neural networks, Phys. Rev. Research 6, 033248 (2024).DOI
- [30]D. Xue, W. Dai, Z. Zheng, C. Li, J. Zou, and H. Xiong, Generalized quantum state tomography with hybrid denoising priors, IEEE Trans. Signal Process. 73, 1532 (2025).DOI
- [31]T. Heinosaari, L. Mazzarella, and M. M. Wolf, Quantum tomography under prior information, Commun. Math. Phys. 318, 355 (2013).
- [32]R. Blume-Kohout, Optimal, reliable estimation of quantum states, New J. Phys. 12, 043034 (2010).
- [33]C. Granade, J. Combes, and D. G. Cory, Practical Bayesian tomography, New J. Phys. 18, 033024 (2016).DOI
- [34]J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, Deep unsupervised learning using nonequilibrium thermodynamics, in Proceedings of the 32nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 37 (PMLR, 2015) pp. 2256–2265.
- [35]J. Ho, A. Jain, and P. Abbeel, Denoising diffusion probabilistic models, in Advances in Neural Information Processing Systems, Vol. 33 (2020) pp. 6840–6851.arxiv.org/abs/2006.11239
- [36]C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, Image super-resolution via iterative refinement, IEEE Trans. Pattern Anal. Mach. Intell. 45, 4713 (2023).DOI
- [37]H. Chung, J. Kim, M. T. McCann, M. L. Klasky, and J. C. Ye, Diffusion posterior sampling for general noisy inverse problems, in International Conference on Learning Representations (2023).arxiv.org/abs/2209.14687
- [38]B. Zhang, P. Xu, X. Chen, and Q. Zhuang, Generative quantum machine learning via denoising diffusion probabilistic models, Phys. Rev. Lett. 132, 100602 (2024).DOI
- [39]G. Kwun, B. Zhang, and Q. Zhuang, Mixed-state quantum denoising diffusion probabilistic model, Phys. Rev. A 111, 032610 (2025).arxiv.org/abs/2411.17608
- [40]Y. Wang, S. Cheng, L. Li, and J. Chen, Learning quantum distributions with variational diffusion models, IFAQ-PapersOnline 56, 5888 (2023).DOI
- [41]Y. Tang, M. Long, and J. Yan, QuaDiM: A conditional diffusion model for quantum state property estimation, in International Conference on Learning Representations (2025).
- [42]R. Jozsa, Fidelity for mixed quantum states, J. Mod. Opt. 41, 2315 (1994).DOI
- [43]M. S. Byrd and G. K. Brennen, General depolarized pure states: Identification and properties, Phys. Lett. A 372, 1770 (2008).DOI
- [44]A. Q. Nichol and P. Dhariwal, Improved denoising diffusion probabilistic models, in Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 139 (PMLR, 2021) pp. 8162–8171.arxiv.org/abs/2102.09672
- [45]O. Ronneberger, P. Fischer, and T. Brox, U-Net: Convolutional networks for biomedical image segmentation, in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, Vol. 9351, edited by N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi (Springer, Cham, 2015) pp. 234–241.arxiv.org/abs/1505.04597
- [46]A. S. Aasen, A. Di Giovanni, H. Rotzinger, A. V. Ustinov, and M. Gärtner, Readout error mitigated quantum state tomography tested on superconducting qubits, Commun. Phys. 7, 301 (2024).DOI
- [47]H. Jeffreys, An invariant form for the prior probability in estimation problems, Proc. R. Soc. Lond. A 186, 453 (1946).pubmed.ncbi.nlm.nih.gov/20998741
- [48]B. Mu, H. Qi, I. R. Petersen, and G. Shi, Quantum tomography by regularized linear regressions, Automatica 114, 108837 (2020).DOI
- [49]B. Efron, Bootstrap methods: Another look at the jackknife, Ann. Stat. 7, 1 (1979).
- [50]S. Holm, A simple sequentially rejective multiple test procedure, Scand. J. Stat. 6, 65 (1979).DOI
- [51]R. Blume-Kohout, Hedged maximum likelihood quantum state estimation, Phys. Rev. Lett. 105, 200504 (2010).
- [52]J. Shang, Z. Zhang, and H. K. Ng, Superfast maximum-likelihood reconstruction for quantum tomography, Phys. Rev. A 95, 062336 (2017).DOI
- [53]E. Bolduc, G. C. Knee, E. M. Gauger, and J. Leach, Projected gradient descent algorithms for quantum state tomography, npj Quantum Inf. 3, 44 (2017).arxiv.org/abs/1612.09531
- [54]T. Proctor, M. Revelle, E. Nielsen, K. Rudinger, D. Lobser, P. Maunz, R. Blume-Kohout, and K. Young, Detecting and tracking drift in quantum information processors, Nat. Commun. 11, 5396 (2020).DOI
- [55]S. Dasgupta and T. S. Humble, Characterizing the stability of NISQ devices, in 2020 IEEE International Conference on Quantum Computing and Engineering (QCE) (IEEE, 2020) pp. 419–429.arxiv.org/abs/2008.09612
- [56]Y. Hirasaki, S. Daimon, T. Itoko, N. Kanazawa, and E. Saitoh, Detection of temporal fluctuation in superconducting qubits for quantum error mitigation, Appl. Phys. Lett. 123, 184002 (2023).DOI
- [57]D. M. Greenberger, M. A. Horne, and A. Zeilinger, Going beyond Bell’s theorem, in Bell’s Theorem, Quantum Theory and Conceptions of the Universe, edited by M. Kafatos (Springer, Dordrecht, 1989) pp. 69–72.arxiv.org/abs/0712.0921
- [58]W. Dür, G. Vidal, and J. I. Cirac, Three qubits can be entangled in two inequivalent ways, Phys. Rev. A 62, 062314 (2000).DOI
- [59]H. J. Briegel and R. Raussendorf, Persistent entanglement in arrays of interacting particles, Phys. Rev. Lett. 86, 910 (2001).pubmed.ncbi.nlm.nih.gov/11177971
- [60]A. Jaegle, F. Gimeno, A. Brock, A. Zisserman, O. Vinyals, and J. Carreira, Perceiver: General perception with iterative attention, in Proceedings of the 38th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 139 (PMLR, 2021) pp. 4651–4664.arxiv.org/abs/2103.03206
- [61]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all you need, in Advances in Neural Information Processing Systems, Vol. 30 (2017) pp. 5998–6008.arxiv.org/abs/1706.03762
- [62]E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, FiLM: Visual reasoning with a general conditioning layer, in Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 32 (2018) pp. 3942–3951.arxiv.org/abs/1709.07871
- [63]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köp f, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, PyTorch: An imperative style, high-performance deep learning library, in Advances in Neural Information Processing Systems, Vol. 32 (2019) pp. 8026–8037.arxiv.org/abs/1912.01703
- [64]I. Loshchilov and F. Hutter, Decoupled weight decay regularization, in International Conference on Learning Representations (2019).
- [65]J. Song, C. Meng, and S. Ermon, Denoising diffusion implicit models, in International Conference on Learning Representations (2021).arxiv.org/abs/2010.02502
- [66]B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap (Chapman & Hall, New York, 1993).
- [67]B. O’Donoghue, E. Chu, N. Parikh, and S. Boyd, Conic optimization via operator splitting and homogeneous self-dual embedding, J. Optim. Theory Appl. 169, 1042 (2016).
- [68]B. O’Donoghue, Operator splitting for a homogeneous embedding of the linear complementarity problem, SIAM J. Optim. 31, 1999 (2021).
- [69]S. Diamond and S. Boyd, CVXPY: A Python-embedded modeling language for convex optimization, J. Mach. Learn. Res. 17, 1 (2016).
- [70]P. J. Goulart and Y. Chen, Clarabel: An interior-point solver for conic programs with quadratic objectives, Math. Program. Comput. 18, 1279 (2026).







