How does surrogate accuracy requirements differ between exploratory design optimization and high-stakes parameter inference?

# Surrogate Accuracy in Design vs. Inference

These are fundamentally different problems with opposite accuracy demands.

Exploratory Design Optimization

In design optimization, you're exploring a large parameter space to find configurations that *work well enough* [AG-2024.07-2351]. The surrogate's job is to be fast enough to evaluate many candidates, not to be perfect. A surrogate with ~20% error can still reliably guide you toward good designs [AG-2025.03-206]. This is because design is forgiving: if you're searching for the best jet tagger architecture or quantum network configuration, an approximate surrogate steers you toward promising regions and you can validate the top candidates with the expensive simulator later.

The payoff of using surrogates in design is speed. You can parallelize searches across quantum processors or batch-evaluate hundreds of detector configurations on a GPU [AG-2025.01-653], which would be impossible with the original simulator alone.

High-Stakes Parameter Inference

Parameter inference is unforgiving. Here you are extracting fundamental physics parameters (masses, couplings, cosmological constants) from data, and the posterior probability distribution you recover *must* be accurate [AG-2025.03-206]. A biased or inflated surrogate can systematically shift your inferred values or underestimate uncertainty.

The key difference: design asks "which design is better?", while inference asks "what is the true value and its error?" The latter requires the surrogate to preserve the *shape* of the likelihood surface, not just find its peak. Research shows that likelihood-ratio methods outperform direct likelihood estimation in this setting [AG-2025.03-1432], and that you can tolerate surrogate error only up to the point where information-theoretic bounds on posterior distortion remain acceptable [AG-2025.03-206].

One concrete requirement: in particle physics parameter fits, the surrogate must capture the full high-dimensional distribution of jet properties, not just binned histograms, or it will bias your extracted parameters [AG-2025.03-1432].

The provided preprints don't directly compare quantitative accuracy thresholds between these two contexts, though they illustrate both use cases.