The Signal
Chicago–IBM Team Maps Backaction with Three Measurements
A University of Chicago and IBM team reports a protocol that learns how midcircuit measurements both record results and alter qubit states using strings of three repeated measurements. Tests on IBM processors found that the learned model pr
Quantum.world Team · Written with AI · fact-checked ·
“Consequently, realistic MCMs display both readout and backaction errors, including measurement-induced excitation or decay, which cannot be captured by a confusion matrix alone.”
Chia-Tung Chu and his coauthors, the paper’s authors, write.
A midcircuit measurement does two things: it records a classical result and changes the quantum state that continues through the circuit. A standard confusion matrix tracks whether the reported result is wrong, but not how the measurement itself disturbs the qubit. In a paper published October 26 in Physical Review Letters, researchers from the University of Chicago and IBM Quantum describe a protocol that learns both effects from strings of just three repeated measurements.
The distinction matters in dynamic circuits, where a qubit may be measured and then reused—for example, during syndrome extraction or measurement-based reset. Errors in the state left behind can accumulate across later operations even when the readout record appears accurate. The team’s method captures correlations between the reported bit and the qubit’s postmeasurement state, including whether measurement pushes the qubit toward the zero or one state.
Three outcomes constrain a hidden measurement process
The researchers model a single-qubit measurement as a pair of outcome-conditioned state-transition matrices. Each matrix describes the chances of a reported result and a resulting computational-basis state, given the state before measurement. To learn this reduced model, they apply three repeated, randomized measurements to a maximally mixed input and use the probabilities of the resulting bit strings.
Three repetitions identify all learnable parameters in that model except one gauge degree of freedom: different mathematical descriptions can produce the same observed strings. The authors bound that ambiguity by requiring the reconstructed probabilities to remain physically valid, reporting ranges rather than pretending the data pick one unique description. Their protocol uses Z-twirling, which simplifies the process while retaining population transitions and readout–backaction correlations.
IBM tests span 60 selected qubits
The team tested the method on 20 selected qubits apiece from three 156-qubit IBM Heron processors: ibm_kingston, ibm_pittsburgh and ibm_boston. Learning circuits had depth three. For each qubit, the researchers used four input-randomization settings and collected one million shots per setting, for four million learning shots per qubit in total.
Across the processors, the inferred measurement-induced decay rate—from one to zero—was one to two orders of magnitude larger than the excitation rate, from zero to one. The authors say the imbalance is consistent with T1 relaxation, or spontaneous emission, dominating the backaction in their reduced model. The physically allowed ranges for the inferred error rates were typically less than 10% wide relative to their midpoints.
Prediction improves, with assumptions attached
For an independent test, the researchers ran sequences of 10 repeated X-gate and measurement blocks. The full learned model predicted the Pauli-Z observable with deviations one to two orders of magnitude smaller than a predictor based on a confusion matrix, which does not update the qubit’s state after measurement. The paper’s abstract summarizes the improvement as about 100-fold.
The result applies to a specific effective model, not every possible measurement process: the analysis assumes stationary, Markovian, leakage-free behavior and deliberately removes coherence-sensitive information through twirling. Longer sequences can check whether measured bit-string probabilities remain consistent with those assumptions. In lower-shot independent IBM runs, 99.4% of reconstructions succeeded with 10,000 shots per Pauli input setting; the authors also warn that some instruments can make the estimators more sensitive to finite-shot noise.
By retaining the state change that follows each reported bit, the protocol supplies information a readout-only calibration omits. That distinction is especially relevant when circuits reuse measured qubits, where accumulated decay and excitation—not just a wrong recorded result—shape the next operation.