What if quantum parameter tuning were constrained by limited coherence time—would Bayesian optimization still converge?

This is a sharp question about a real constraint in quantum computing—and the archive has directly relevant material.

The Short Answer

Yes, Bayesian optimization can still converge under limited coherence time, but it requires careful design. The key is that Bayesian methods don't need *many* function evaluations; they're built to learn efficiently from *few* expensive ones. Coherence time limits the total number of circuit runs, not the ability of the optimizer to learn from them.

Why This Matters

Coherence time—the window during which a quantum system stays in a usable superposition—is a hard physical constraint [AG-2024.04-1942]. Every parameter update requires a fresh circuit run, and running out of coherence means you literally cannot measure your objective function anymore. So the question becomes: can Bayesian optimization find good parameters *before* decoherence kills you?

What The Archive Shows

Bayesian optimization was designed for exactly this scenario: expensive black-box functions where you want to minimize the number of calls. In variational quantum algorithms (VQAs), each circuit evaluation costs quantum time and is noisy, which mirrors coherence-time constraints. Researchers have shown that modified Bayesian approaches can work within this budget:

• Shot efficiency matters first. The classical optimizer should be able to extract maximum information from few quantum measurements [AG-2024.06-2126]. If you're wasting shots, you waste coherence time faster than necessary.

• Physics-informed priors help convergence. By encoding what you know about the quantum circuit's structure into the Bayesian model, you can dramatically reduce the posterior uncertainty after just a few evaluations [AG-2024.06-1689]. This means fewer total circuit runs to converge.

• Hardware adaptation is crucial. The optimal parameter-tuning strategy depends on your specific quantum device's noise profile and gate fidelities [AG-2024.04-1942]. A generic optimizer might thrash around in a way that wastes coherence; a customized one won't.

• Surface topology priors help with molecular problems. For variational quantum eigensolvers (a key application), priors over the structure of the solution landscape let Bayesian optimization converge in fewer shots [AG-2024.07-1887].

The Catch

All of this assumes the optimizer *doesn't have to restart the quantum system continuously*. If each parameter update requires re-preparation of the initial state and you're cycling through your coherence budget with each attempt, then no classical optimizer can save you—the problem is physical, not algorithmic. But if you can keep the quantum system active while the classical loop tunes parameters, Bayesian optimization's sample efficiency becomes your ally: it's *designed* to converge with minimal function calls, which translates to minimal coherence consumption.

The preprints don't specifically address what happens if coherence time runs out *mid-optimization*—that would be an edge case worth studying.