When should I use TREX?#

Advantages#

Model-free: Unlike REM, TREX does not require an explicit confusion matrix or prior knowledge of readout error rates. It automatically estimates and corrects readout errors using calibration circuits.
Handles correlated readout errors: The readout twirling procedure diagonalizes the readout error channel, which means it works even when readout errors are correlated between qubits.
Simple integration: TREX only requires that the executor returns raw bitstrings (MeasurementResult). It can be combined with other error mitigation techniques that address different noise sources (e.g., PT for gate errors, ZNE for noise extrapolation).
Mathematically rigorous: TREX provides provable error bounds on the corrected expectation values when the measurement errors are small.

Execution overhead: TREX requires running additional circuits for both the readout twirling (multiple randomizations of the original circuit) and calibration (identity circuits with twirling). The total number of circuit executions scales as num_randomizations * (num_groups + 1), where num_groups is the number of commuting groups in the observable.
Shot overhead: Each randomization requires its own set of measurement shots. For a fixed total shot budget, increasing the number of randomizations decreases the shots per randomization, which introduces a trade-off between twirling quality and per-circuit statistical precision.
Readout error sensitivity: When readout errors are very large, the calibration factors can become small, leading to noisy corrected values. In extreme cases, the correction may amplify statistical noise.

Feature	TREX	REM
Requires confusion matrix	No	Yes
Handles correlated errors	Yes	Only with full confusion matrix
Additional circuits needed	Yes (calibration + twirling)	No
Scaling	Linear in randomizations	Depends on matrix inversion

Find more information on TREX on the QEM Zoo.