arXiv:2610.03690
Universal Bounds for Out-of-Distribution Unitary Learning
Joachim Favre, Armando Angrisani, Zoë Holmes
Abstract
How much can the action of an unknown unitary dynamics on one family of states reveal about its action elsewhere? We develop a general framework for out-of-distribution quantum dynamics learning based on the first and second moments of the input distribution. These define a notion of the bias and expressivity of an ensemble, which together control generalization between arbitrary mixed-state training and testing distributions. Our framework shows that only studying the action of a unitary on random real states can learn its action on random complex states, that tensor products of unbiased pure-state ensembles preserve expressivity, and that single-qubit risks are independent of bias. It also recovers earlier locally scrambling results as loose special cases and yields tight product-to-Haar bounds. We conclude with a series of no-go results and limited generalization bounds for learning certain channels, which further serve to underscore the distinctive role of unitarity in enabling out-of-distribution generalization.
AI Summary
The paper asks when small average error in learning a quantum process on one input ensemble guarantees small error on another. It studies squared Hilbert–Schmidt output loss, equal to infidelity for unitary processes on pure inputs; finite-sample learning is treated separately. For unitary targets and models, vectorization expresses risk through $M(\mathcal P)=2^n\mathbb E[\rho\otimes\rho^{\mathsf T}]$. Separating identity and traceless components defines bias $\nu$ and expressivity $\eta=\bar\gamma-\lambda_{\max}(\mathcal S)$, where $\bar\gamma$ is average purity. Spectral bounds give $R_{\rm test}\leq(R_{\rm train}+\nu_{\rm train})/\eta_{\rm train}+\nu_{\rm test}$ when $\eta_{\rm train}>0$. Jointly Pauli-diagonal, unbiased ensembles admit tight eigenvalue-ratio comparisons. Applications include real-Haar training controlling complex-Haar risk with tight overhead $1+1/(2^n+1)$, preservation of expressivity under tensor products of unbiased pure-state ensembles, and tight product-Haar-to-global-Haar overhead approaching $3/2$ rather than the previous $2$. Single-qubit risk is independent of bias. Tight noisy-to-clean comparisons have constant overhead for local depolarizing rate $O(1/n)$, polynomial overhead for $O(\log n/n)$, and exponential overhead at fixed positive rate. Explicit channel counterexamples require exponential global-Haar-to-product-Haar overhead, or show zero real-input risk despite constant complex-input risk. For Pauli channels with combined support $s$, arbitrary-input risk is bounded by $s(1+2^{-n})$ times global-Haar risk or $s^2$ times product-Haar risk. Except for these sparse-channel guarantees, results concern average rather than worst-case performance; appreciable bias or low expressivity can make the general bounds uninformative.
In the past month, ChatGPT helped us find no-go results to rule out the hoped-for general channel extension. It also helped with the final writing sprint.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 13+62 pages, 1+1 figures. Edit: Added some results from the appendices to the main text, added a visualization of expressivity to the appendices, and fixed some typos