Determine whether the MNIST accuracy advantage persists under expanded training conditions

Determine whether the observed accuracy advantage of amplitude damping over the noiseless four-qubit MNIST classifier at damping strength p=0.1 and depth L=8 persists when the training budget or dataset size is increased.

Background

At p=0.1 and L=8, the amplitude-damped classifier exceeds the noiseless classifier in exact-expectation-value simulations under the paper’s fixed training budget and eight initializations. The authors explicitly characterize this as an observation tied to those experimental conditions.

The unresolved issue is whether the apparent advantage reflects a genuine noise-induced generalization effect or instead results from limited optimization, limited data, or the finite number of random initializations.

References

We report it as an observation at a fixed training budget and eight initializations; we have not tested whether it persists with longer training or more data.

— Zero- Versus Infinite-Temperature Damping in Variational Quantum Circuits: Feature Scale, Sampling Cost, and Frame Gauge  (2610.01466 - Nguyen et al., 1 Oct 2026) in Section 5.1, “Sampling cost”