Fig. 4: Experimental implementation and noiseless vs. experimental results.

a Parameters for the three circuits implemented in this experiment. Values in parentheses are calculated ignoring contributions due to virtual Z gates. b The depth of the each circuit and number of entangling layers (dark gray) scales to accommodate all 67 features of the input data. c The test accuracy for hardware QKM is competitive with the noiseless simulations even in the case of relatively low circuit fidelity, across multiple choices of qubit counts (the simulated test accuracies for n = 10, 14 were statistically indistinguishable from optimized RBF performance, similarly to Fig. 3 for n = 17). The presence of hardware noise significantly reduces the ability of the model to overfit the data. Error bars on simulated data represent standard deviation of accuracy for an ensemble of SVM classifiers trained on 10 size-m downsampled kernel matrices and tested on size-v downsampled test sets (no replacement). Dataset sampling errors are propagated to the hardware outcomes but lack of larger hardware training/test sets prevents characterization of generalization error (e.g. using bootstrapping techniques⁴⁰).