No extra CNOT needed: trainable single-qubit rotations before readout lift Iris three-class peak from 60% to 83.33%
Synopsis
Variational quantum classifiers usually buy accuracy with more entangling gates, yet CNOTs err and run far slower than single-qubit gates. OQMD folds the readout map into training, using trainable single-qubit rotations to choose the measurement basis without adding any CNOT. On a 30-point stratified Iris test split, the best 0-CNOT OQMD configuration reaches 83.33% against 56.67% for the best 18-CNOT control; at 18 CNOTs OQMD raises the peak from 56.67% to 66.67%, though the pooled mean gap is small.
Interpretation
The readout stage of a quantum classifier can now be trained jointly with the variational circuit instead of being fixed to computational-basis measurement plus parity mapping. Earlier practice left the map from measurement outcomes to class labels to default parity rules and sought accuracy mainly by deepening entanglement. On the 30-point stratified Iris test split, the best 0-CNOT OQMD configuration reaches 83.33% versus 60.00% for the best control; at 18 CNOTs OQMD peaks at 66.67% versus 56.67%.
The concrete realization adds three trainable single-qubit rotations per qubit before measurement, making the measurement basis itself a learned parameter. This readout layer is called OQMD, Optimal Quantum Measurement Decoding, and its rotation angles are updated by the same classical optimizer as the circuit parameters. Experiments run three-class classification on a noiseless statevector simulator with one fixed Iris split, 50 random seeds per configuration, and 3,600 runs across the minimal and complex designs.
The size of the gain depends on circuit architecture rather than being uniform. The peak rises by 23.33 percentage points on minimal circuits and by 10.00 points at 18 CNOTs, while the 18-CNOT pooled mean moves only from 40.22% to 41.57%. Mann–Whitney U tests on pooled runs give p below 10 to the minus 100 on minimal circuits and about 0.03 at 18 CNOTs.
Perspective
The findings rest on the Iris dataset, a single stratified split, a noiseless statevector simulator, and a fixed parity multiclass readout; the comparison concerns quantum classifiers inside a gate budget, not overall competitiveness with classical models. The authors propose extending the same peak-focused protocol to additional datasets and calibrated noise models.
The authors explicitly avoid calling this a universal lift: effects are magnitude-dependent, and the pooled mean gap at 18 CNOTs is small; the six-CNOT ladder point uses only the ZXZ rotation code and the COBYLA optimizer, leaving intermediate-depth rotation sweeps to future work; the exploratory correlation between CNOT count and peak margin is about negative 0.45 with p about 0.37, below significance.