APPLICATIONS / 30 MIN
Quantum kernels & responsible benchmarking
K(x,y) = |⟨φ(x)|φ(y)⟩|²
DERIVATION 1
Feature maps define a kernel
Encode classical x into |φ(x)⟩=U(x)|0⟩. Overlaps can be estimated by applying U(y)†U(x) and measuring the all-zero probability. For normalized feature states K(x,x)=1. Encoding choices determine whether the kernel is informative and hard to simulate.
DERIVATION 2
Training is still a pipeline
A classical learner can use the measured kernel matrix. Finite-shot estimates may require care to restore positive semidefiniteness. Evaluate data leakage, train/test separation, hyperparameter budgets and measurement cost across all matrix entries.
DERIVATION 3
Do not equate accuracy with advantage
Compare strong classical kernels and equally tuned baselines, including data encoding, queue latency and shot counts. Concentrated kernels and poor encodings can erase useful signal. Report out-of-sample performance and uncertainty; quantum hardware alone proves neither better predictions nor speedup.
Assessment
0 / 2 answered · 0 saved attempts
