---
title: 'QPRC: Quasi-Probabilistic Readout Correction'
url: https://www.emergentmind.com/topics/quasi-probabilistic-readout-correction-qprc
type: topic
---

# QPRC: Quasi-Probabilistic Readout Correction

Quasi-Probabilistic Readout Correction (QPRC) is a protocol for the mitigation of readout errors in quantum measurements, particularly relevant to experiments on near-term quantum devices where measurement noise—often both state-dependent and correlated—can be a dominant source of infidelity. QPRC leverages classical post-processing using information from quantum detector tomography (QDT), or, in scalable variants, from randomized compiling and measurement twirling, to recover unbiased estimators of expectation values without requiring exponentially large calibration or computational overheads. The corrected statistics referred to as "quasi-probabilities" may lie outside the probability simplex, but ensure unbiased estimation of linear functionals such as expectation values of observables or algorithmic success probabilities. QPRC has demonstrated substantial improvements in quantum state and process tomography, non-projective measurement realization, and quantum algorithms such as Grover’s search and Bernstein–Vazirani, and has been extended to efficient correction in multi-qubit and adaptive, mid-circuit measurement scenarios [1907.08518][2312.14139].

## 1. Quantum Measurement Noise: Theoretical Model

A quantum measurement with $n$ outcomes is formally described by a positive-operator valued measure (POVM) $\{E_i\}_{i=1}^n$, with $E_i\ge0$ and $\sum_iE_i=I$. Ideally, measurement of a state $\rho$ yields outcome probabilities $p_i^{\rm true} = \operatorname{Tr}[\rho E_i]$. In practice, readout errors—often originating from classical noise or hardware imperfections—distort the observed empirical distribution $p^{\rm obs}$. In the simplest classical noise model, these effects are modeled by a left-stochastic matrix $M\in\mathbb{R}^{n\times n}$ such that
$$
p^{\rm obs}= M p^{\rm true}\,.
$$
Quantum Detector Tomography (QDT) experimentally reconstructs $M$ by preparing a tomographically complete set of input states and measuring the observed outcome statistics, facilitating the estimation of $E_i^{\rm noisy}$ and extracting $M$ as the transition matrix between the ideal and noisy measurements [1907.08518].

In modern multi-qubit devices, readout noise may include non-unital, state-dependent errors and non-local correlations (crosstalk). These effects can be "twirled" into a Pauli-stochastic channel via randomized compiling and measurement twirling, greatly simplifying the effective noise model to a stochastic bit-flip channel characterized by a single set of bit-flip probabilities per qubit, independent of the initial state [2312.14139]. This approach ensures all elements of the empirical confusion matrix arise from stochastic bit-flips uncorrelated across qubits.

## 2. Quasi-Probabilistic Correction Schemes

The central procedure of QPRC involves inverting the observed noisy statistics to reconstruct unbiased estimators for the true statistics, even though the corrected quasi-probabilities $p^{\rm corr}$ may be negative or exceed unity.

### Matrix Inversion Protocol (Maciejewski et al.)

If $M$ is invertible, one directly computes
$$
p^{\rm corr}=M^{-1} p^{\rm obs}
$$
for each experiment, and any linear functional $f(\rho)=\sum_i f_i p^{\rm true}_i$ is estimated unbiasedly as $\hat f=f^T p^{\rm corr} = f^T M^{-1} p^{\rm obs}$.

However, the sampling variance is amplified by the $\ell_1$-norm of $M^{-1}$, and care is needed when $M$ is ill-conditioned. If some entries of $p^{\rm corr}$ fall outside the simplex, one may project back using least-squares or maximum-entropy procedures, but unbiasedness is preserved for linear functions [1907.08518].

### Scalable QPRC via Measurement Randomized Compiling

QPRC based on measurement twirling and measurement randomized compiling (MRC) eliminates the need for full tomographic inversion. By inserting random single-qubit Paulis before measurement and averaging, the noise channel is symmetrized to a stochastic Pauli channel. One then needs to characterize the bit-flip rates by measuring only a single basis state, and the resulting confusion matrix becomes a simple bit-flip model. Correction is performed by constructing explicit quasi-probability weights that correct for the stochastic flips, avoiding exponential matrix inversion:
- First-order and higher-order analytic expressions for the correction weights $q^{(k)}$ are given, with rapid suppression of higher-order error terms.
- Any noisy distribution $b_x$ can be corrected by quasi-probabilistic reweighting (convolution with the $q$ weights), and expectation values are similarly reconstructed [2312.14139].

## 3. Practical Implementation and Workflow

The QPRC implementation proceeds via the following steps depending on the physical context:

**(a) Detector Tomography-Based Approach [1907.08518]:**
1. Perform QDT on each readout register or small correlated block.
2. Extract the classical transition matrix $M$ from the diagonal part of the reconstructed POVM elements.
3. Invert $M \rightarrow M^{-1}$.
4. For each experiment, collect observed frequencies $p^{\rm obs}$.
5. Apply $p^{\rm corr} = M^{-1} p^{\rm obs}$.
6. If necessary, project $p^{\rm corr}$ onto the probability simplex.
7. Use $p^{\rm corr}$ for all estimates.

**(b) Measurement Twirling/Randomized Compiling-Based Approach [2312.14139]:**
1. Insert random Pauli gates before each measurement and perform classical bit-flip corrections where appropriate.
2. Prepare a single reference state and characterize the bit-flip rates via repeated measurement (avoiding full tomography).
3. Construct analytic quasi-probability correction weights for all possible output strings.
4. Apply these weights in post-processing to reweight observed outcomes.
5. For mid-circuit measurements (MCM) or adaptive protocols, insert randomized Paulis (MRC) and combine measurement results using signed correction protocols according to the number of inserted Pauli-X's.

In both approaches, the protocol is integrated with other error-mitigation strategies such as zero-noise extrapolation or circuit-level quasi-probabilistic error mitigation, enabling stacked error cancellation procedures.

## 4. Performance, Overheads, and Limitations

The QPRC protocol introduces only classical post-processing overhead, which, for $n$-outcome measurements, is $O(n^2)$ (matrix inversion), or, in the scalable MRC-based scheme, becomes $\mathcal{O}(1)$ with respect to the number of qubits because characterization reduces to estimation of single-qubit bit-flip rates. Experimental application to IBM's 5-qubit device reported:
- Single-qubit state tomography fidelities improved from $\sim90\%$–$95\%$ up to $\sim98\%$–$99\%$.
- Single-qubit process tomography infidelity cut roughly in half.
- Success probability for Grover’s search on two qubits increased from $\approx0.47$ to $\approx0.85$.
- Bernstein–Vazirani algorithm success rate improved from $\approx0.33$ to $\approx0.72$.
- For five-qubit output distributions, total-variation distance to target reduced by factors of 2–3 when including correlated qubits [1907.08518].

Sampling variance is increased by the norm of the correction matrix or weights; for typical few-percent readout errors, the overhead is modest. However, if $M$ is ill-conditioned or raw readout fidelities fall below 50%, noise amplification becomes prohibitive, requiring block-wise modeling or higher-order correction schemes [1907.08518][2312.14139].

QPRC assumes dominance of classical (stochastic) noise over coherent (off-diagonal) measurement errors; systematic bias from neglected coherent contributions can be bounded by the operational distance between the full and classicalized POVM. For mid-circuit measurements, QPRC has demonstrated effective cancellation in adaptive circuits without exponential calibration [2312.14139].

| Experimental Task                    | Pre-QPRC Performance   | Post-QPRC Performance    |
|--------------------------------------|-----------------------|--------------------------|
| Single-qubit state tomography        | $\sim$90–95% fidelity | $\sim$98–99% fidelity    |
| Grover search (2 qubits)             | $\approx$0.47         | $\approx$0.85            |
| Bernstein–Vazirani (3 bits)          | $\approx$0.33         | $\approx$0.72            |

## 5. Extensions: Multi-Qubit and Adaptive Measurement Correction

QPRC extends efficiently to large-scale and mid-circuit measurement (MCM) scenarios:
- By using measurement randomized compiling, the calibration and post-processing overhead is rendered independent of the number of qubits; a scalable protocol samples only a single reference state, avoiding exponential confusion matrices [2312.14139].
- For adaptive protocols with real-time feedback based on ancilla measurement and resets, QPRC combines randomized Pauli insertions before each MCM, characterization of bit-flip error probability, and shot-wise quasi-probabilistic correction. This yields effective cancellation of state-dependent or crosstalk mid-circuit measurement errors with negligible additional overhead.

Experimental results on frequency-crowded 8-qubit transmon rings demonstrated QPRC’s superiority over both local and full-matrix inversion correction, attaining lower total-variation distance to ideal output in $>90\%$ of cases, and improved memory protection in bit-flip correction experiments [2312.14139].

## 6. Open Challenges and Future Directions

Outstanding challenges include handling scenarios where bit-flip error rates are dramatically higher, leading to situations where $p_0<50\%$ and traditional correction becomes unstable. Extensions of QPRC beyond purely stochastic bit-flip models to accommodate general non-diagonal measurement noise and the integration with fully fault-tolerant error correction stacks—especially syndrome readout correction on-the-fly—are active areas of research. A plausible implication is that QPRC, tailored for scalable, hardware-agnostic quantum devices, will remain a key tool in the NISQ era for reducing readout infidelities and enabling reliable quantum computation as architectures and error models continue to evolve [2312.14139][1907.08518].

Source: https://www.emergentmind.com/topics/quasi-probabilistic-readout-correction-qprc