Twirled Readout Error Extinction (T-REx)

• T-REx is a measurement-side error mitigation technique that applies Pauli twirling to symmetrize readout errors, effectively reducing bias in Z-type observables.
• The method replaces expensive calibration matrices with simple per-qubit correction factors, integrating efficiently into VQE loops and consensus protocols.
• Empirical results in both VQE and Detectable Byzantine Agreement show that T-REx improves energy estimates and enhances protocol success on NISQ devices.

Searching arXiv for the cited T-REx papers and related context. Twirled Readout Error Extinction (T-REx) is a measurement-side error-mitigation technique for noisy quantum hardware that uses Pauli twirling at readout to suppress systematic measurement bias and improve task-level performance. In the literature considered here, T-REx appears in two complementary forms. In distributed quantum computing, it is used via IBM’s Qiskit Runtime Service as “EM,” where it is described as exploiting Pauli twirling to mitigate noise during quantum measurement “without necessitating a particular noise form,” and is combined with dynamical decoupling (DD) in a Detectable Byzantine Agreement (DBA) protocol on IBM Nairobi (Prest et al., 2023). In variational quantum simulation, it is formulated explicitly as a readout-mitigation method based on random $X$ flips before measurement and XOR post-processing, which symmetrize the readout channel and reduce correction of $Z$-type observables to division by a per-qubit scalar contraction factor $\lambda$ (Belaloui et al., 20 Aug 2025).

1. Conceptual definition and scope

T-REx targets classical readout errors at the end of quantum circuits, particularly the bias between misclassifying $0$ as $1$ and misclassifying $1$ as $0$ on superconducting devices. In the VQE setting, this bias distorts expectation values of $Z$ and products of $Z$ (Pauli strings), which directly enter the VQE objective. In the DBA setting, the stated role of T-REx is to reduce measurement-induced scrambling of correlations needed by the consensus protocol, thereby increasing the probabilities of the desired outcome strings, such as the “Retreat” or “Attack” bit-strings (Belaloui et al., 20 Aug 2025).

Relative to conventional measurement error mitigation (MEM), T-REx is presented as a lighter-weight alternative for the class of observables it targets. Conventional MEM models the readout channel by a confusion matrix calibrated on basis states and uses matrix inversion to debias outcome histograms; for $n$ qubits, this is a $Z$0 matrix, which can be accurate but becomes expensive and fragile as $Z$1 grows. T-REx instead uses simple Pauli twirling with $Z$2 bit-flips inserted before measurement and classical post-processing to symmetrize the readout channel, turning asymmetric flip errors into an effective binary-symmetric channel. Under this twirl, the effect of readout error on $Z$3-type observables collapses to a single multiplicative factor $Z$4 per qubit, estimated via a lightweight calibration (Belaloui et al., 20 Aug 2025).

The DBA work emphasizes a different aspect of the method. There, the novelty is not a formal estimator but the combination of T-REx with DD and the empirical demonstration that the combined mitigation significantly boosts performance in the DBA protocol. The paper does not provide a side-by-side technical comparison to calibration-matrix inversion (M3), Bayesian/maximum-likelihood readout mitigation, or randomized compiling; the emphasis is on generic applicability on NISQ devices and on empirical gains in a consensus task (Prest et al., 2023).

2. Readout model, twirling operation, and estimators

A formal readout-error model is developed explicitly in the VQE study. Let $Z$5 be the probability that the true bit $Z$6 is reported as $Z$7, and $Z$8 the probability that true $Z$9 is reported as $\lambda$0. The single-qubit confusion matrix in the computational basis $\lambda$1 is

$\lambda$2

If qubit-wise errors are assumed independent, the readout channel is

$\lambda$3

and for a true distribution $\lambda$4 over $\lambda$5 bitstrings, the observed distribution is

$\lambda$6

Under standard MEM with matrix inversion, one would estimate

$\lambda$7

subject to normalization and non-negativity constraints (Belaloui et al., 20 Aug 2025).

T-REx modifies this picture by randomizing the readout frame. For each measured qubit $\lambda$8, a twirl bit $\lambda$9 is sampled; $0$0 is applied immediately before measurement; and if $0$1 is the device’s reported bit after the twirl, the classical post-processing is $0$2. With random $0$3 twirls and the corresponding XOR post-processing, the asymmetry is averaged out, yielding an effective binary-symmetric channel with flip probability

$0$4

For $0$5-type observables, the measured expectation becomes

$0$6

The “extinction” in T-REx refers to the cancellation of the bias term, which depends on $0$7, by symmetrization, leaving only a single multiplicative contraction to be estimated (Belaloui et al., 20 Aug 2025).

This leads to simple estimators. If $0$8 denotes the twirled-and-postprocessed empirical expectation of $0$9 on qubit $1$0, then

$1$1

For a Pauli string $1$2 measured under independent per-qubit readout channels,

$1$3

so the T-REx estimator is

$1$4

By contrast, the DBA paper does not present estimators, correction formulas, or inversion procedures for T-REx. It does not provide proofs or sketches demonstrating unbiasedness or variance reduction for its T-REx usage; support there is empirical within the DBA use case (Prest et al., 2023).

3. Calibration and in-loop deployment for VQE

In the VQE implementation, T-REx is integrated tightly into the optimization loop rather than being applied only as a final post-processing step. The paper studies BeH$1$5 at a fixed geometry, with an active space of 2 electrons and 3 active orbitals, mapped by parity transformation with qubit tapering from 6 to 4 qubits. The hardware ansatz is Efficient SU2; UCCSD was also studied in simulation but was too deep/noisy on hardware. The optimizer is SPSA with 250 iterations and 50-cost-function calls for initial calibration of $1$6; each iteration uses 2 energy measurements for gradient estimation and 4000 shots per energy measurement. The Hamiltonian is decomposed into 7 qubit-wise commuting sets (Belaloui et al., 20 Aug 2025).

The calibration-and-measurement workflow is explicit. For $1$7 measured qubits, the calibration stage prepares the $1$8 computational-basis states. For each calibration circuit, a twirl $1$9 is sampled, $1$0 is applied before measurement, and the output is post-processed by XOR. The contraction factors $1$1 are then estimated on each qubit from the contraction of $1$2 expectations across the calibration set. In the BeH$1$3 experiment, where 4 qubits are measured, the calibration set comprised 16 circuits, giving a stable estimate of each $1$4 (Belaloui et al., 20 Aug 2025).

For each commuting measurement group $1$5, the implementation uses per-qubit $1$6 twirls sampled independently, with $1$7 twirled circuits per commuting measurement group and a fixed budget $1$8 shots per group, so twirling does not increase shots for the energy measurement itself; only the calibration adds overhead. The corrected group expectations are assembled into the VQE objective

$1$9

with T-REx applied as

$0$0

where $0$1 is the subset of qubits on which $0$2 has $0$3’s. The energy supplied to SPSA is formed from these T-REx-corrected group expectations (Belaloui et al., 20 Aug 2025).

The reported overhead is modest relative to the baseline shot budget. For the BeH$0$4 configuration, the calibration uses 16 circuits and 8192 shots per iteration. The total shots per iteration are

$0$5

and the total circuits per iteration are 240, including calibration. The baseline shots per iteration without twirl and without calibration are

$0$6

so T-REx adds 8192 calibration shots per iteration, approximately 14.6% overhead. The reported total QPU runtime was 7667 s for 551 measurements, including SPSA calibration (Belaloui et al., 20 Aug 2025).

4. Use in Detectable Byzantine Agreement and synergy with dynamical decoupling

In the DBA application, T-REx is incorporated through IBM Qiskit Runtime Service on IBM Nairobi. The protocol prepares an entangled resource state $0$7 distributed from a Quantum Source Device (QSD), with the commander receiving the first two qubits and each lieutenant receiving one qubit. After entanglement distribution, the protocol adopts a verification method reminiscent of Quantum Key Distribution schemes. The commander then issues orders encoded in specific quantum states, such as Retreat or Attack. When received orders diverge, lieutenants engage in structured games to reconcile discrepancies (Prest et al., 2023).

Within this protocol, T-REx is applied to all measurement stages where readout errors could scramble the correlations needed for DBA. In entanglement verification, after distributing copies of $0$8, selected indices are measured and correlations are checked, similar to QKD. In order realization and lieutenant reconciliation games, outcomes measured from the shared resource and the claimed orders are aggregated into joint distributions, which are then compared to the expected verification distributions. The protocol relies on bit-string outcome distributions corresponding to the commander’s two-qubit measurement, with “Retreat” $0$9 and “Attack” $Z$0, together with the lieutenants’ single-qubit measurements as determined by the correlations of $Z$1 (Prest et al., 2023).

The DBA paper frames T-REx as a measurement-side Pauli-twirling approach that does not require assuming a specific noise model, but the method is used as a built-in runtime option rather than through an explicit circuit-level derivation. The summarized workflow is: prepare the entangled resource state $Z$2 and distribute qubits; enable IBM Runtime error mitigation “EM,” which the authors equate to T-REx, and optionally enable DD; execute verification measurements on selected indices and record outcome distributions; execute the order distribution and consensus games with the same mitigation settings; and aggregate classical outcomes to compute DBA success probabilities and traitor-detection rates. The paper does not provide a gate-level description such as “apply $Z$3 according to a random mask before measurement,” nor pseudocode for the randomization (Prest et al., 2023).

The distinctive claim of that study is the synergy with DD. DD is described as injecting additional pulses to idle qubits to attenuate decoherence, although specific sequences are not named. With $Z$4 and 1000 shots, for a single randomly selected traitor on IBM Nairobi, the measured DBA success probabilities are: no mitigation, $Z$5; EM (T-REx) alone, $Z$6; DD alone, $Z$7; and EM (T-REx) + DD, $Z$8. The incremental gain of EM in the presence of DD is $Z$9, while EM alone over no mitigation gives $Z$0; the authors quantify the synergy as “a 1.663 times greater impact from EM than without DD,” computed as $Z$1 (Prest et al., 2023).

5. Empirical results and task-level benchmarks

The VQE study reports that T-REx improves both energy estimates and the quality of optimized variational parameters. The reference ground-state energy for the chosen active-space model is $Z$2 Ha. On IBMQ Belem with T-REx inside the loop, the final measured energy is $Z$3 Ha with $Z$4 Ha, and the reported single-measurement standard deviation is $Z$5 Ha over 4000 shots. Evaluating the final hardware-optimized parameters on a noiseless statevector simulator gives $Z$6 Ha, which is 22.64 mHa below the raw QPU energy. The best parameter set during the run, at iteration 121, gives $Z$7 Ha, within chemical accuracy and only 1.1 mHa from target. The paper further states that the smaller IBMQ Belem device with T-REx produced higher-quality variational parameters than the larger IBM Fez device without mitigation, leading to an “order of magnitude” improvement in energy error after statevector re-evaluation of the hardware-optimized parameters (Belaloui et al., 20 Aug 2025).

That paper also compares T-REx to zero-noise extrapolation (ZNE) applied only to the final energy. With the simulated noise model, an exponential ZNE fit gives $Z$8 Ha with $Z$9 Ha; on the QPU, the exponential fit gives $n$0 Ha with $n$1 Ha. The stated interpretation is that T-REx’s principal advantage lies in improving the optimization landscape seen by SPSA during the loop, rather than only polishing a final noisy estimate (Belaloui et al., 20 Aug 2025).

The DBA study reports task-level improvements in both entanglement verification and consensus success. In entanglement verification for $n$2, using 100 iterations and 8192 shots per iteration on IBM Nairobi, enabling EM (T-REx) and DD increases the probability of the bins corresponding to “Retreat” and “Attack” compared to unmitigated measurements, and the combined effect makes the probability mass function closer to the noiseless case. For $n$3 with a single randomly selected traitor and 1000 shots, the paper maps the observed success rates to simulated “equivalent noise levels,” reporting EM+DD $n$4, DD alone $n$5, and unmitigated $n$6. The simulation is noted to be symmetric about a $n$7 noise level, where $n$8 corresponds to no correlation and values above $n$9 correspond to anti-correlation serving as an analogous resource (Prest et al., 2023).

Scaling data in the DBA experiments show that the benefits of T-REx are real but do not remove the protocol’s intrinsic shot complexity. For fixed $Z$00 and varying numbers of traitors $Z$01, with 1000 iterations and 10,000 shots per setting, EM+DD significantly improves success probability across non-trivial cases relative to no mitigation. For a single traitor and varying $Z$02, again with 1000 iterations and 10,000 shots per setting, the paper reports that with EM+DD, 10,000 shots suffice to reach perfect success $Z$03 up to $Z$04. A separate sweep from 100 to $Z$05 shots shows that, as $Z$06 increases, the shots needed to reach near-perfect success grow exponentially; EM+DD reduces the shot requirement compared to unmitigated runs and exhibits significantly faster convergence to high success probability (Prest et al., 2023).

6. Assumptions, limitations, and relation to adjacent mitigation methods

The scope of T-REx is defined by both its economy and its assumptions. In the VQE study, the principal assumption is that readout errors can be treated qubit-wise for the targeted $Z$07-type observables: the readout channel is modeled as a tensor product of single-qubit confusion matrices, and correlated readout is neglected. The paper states that, when correlations are present, standard MEM with block-factorizations or local correlation models can be layered on top, but T-REx’s main benefit—bias extinction via symmetrization—still applies per qubit. It also notes several limitations: if multi-qubit correlations are significant, per-qubit $Z$08 may not capture all effects; $Z$09 drifts require periodic recalibration; the method as applied assumes binary final measurements; and twirling partitions shots across $Z$10 patterns, so increasing $Z$11 reduces systematic bias but increases statistical variance unless total shots are increased (Belaloui et al., 20 Aug 2025).

The DBA study presents a complementary set of limitations. It emphasizes that T-REx does not require assuming a specific noise form, but it does not analyze correlated readout errors or provide a formal model of measurement noise on the tested device. It does not discuss failure modes of T-REx, nor does it separate T-REx’s impact on correlated readout errors versus independent errors. Theory is also limited: the paper does not present formal propositions, lemmas, or proofs about T-REx’s bias extinction, variance reduction, or convergence guarantees. Its support for T-REx is empirical within the DBA use case (Prest et al., 2023).

Both papers position T-REx relative to broader quantum error-mitigation practice. The VQE paper contrasts it with confusion-matrix MEM, where calibration cost scales exponentially with $Z$12 and inversion can be ill-conditioned under drift or correlations, and with ZNE, which adds multiple “noise-scaled” circuit executions per measurement but does not correct readout bias. It also situates T-REx near Clifford-based and randomized-compiling methods, which target coherent gate noise rather than readout bias. The DBA paper, by contrast, presents T-REx primarily as a runtime-accessible mitigation that is most effective when paired with DD. Taken together, these results suggest a division of labor: T-REx directly targets readout bias and the contraction of $Z$13-type correlators, while other techniques address decoherence, coherent gate error, or extrapolation of gate noise (Belaloui et al., 20 Aug 2025).

A plausible implication is that T-REx is best understood not as a universal replacement for MEM, but as a structurally economical mitigation for final-measurement observables, especially when a task is dominated by readout asymmetry and when the observable algebra is largely $Z$14-type. In one setting, this yields a simple per-qubit scalar correction inside a VQE loop; in the other, it functions as a built-in measurement-side mitigation whose chief value is the empirical preservation of the correlations required for quantum consensus under NISQ conditions (Prest et al., 2023).