Why does interpolating through intermediate geometries work better than sampling random configurations?

Your question asks about a specific algorithmic principle—that interpolating through intermediate geometries outperforms random sampling—but the preprints provided don't directly address this comparison in the context you're asking about.

However, I can point to one relevant result: [AG-2024.07-1770] demonstrates that convex interpolation dramatically improves trainability in variational quantum algorithms by evaluating ground-state properties through a carefully chosen subset of basis states rather than random sampling. The authors show this "close-to-solution initialization" strategy, combined with a "regular cost function landscape," enables a 40-qubit problem to be solved using only 10 qubits—a gain that comes precisely from *structured* rather than *random* exploration of the solution space.

The intuition is that interpolation creates a smoother path through parameter space: random configurations jump unpredictably between distant points in the configuration landscape, often landing in regions with poor gradient information (a problem called "barren plateaus" in quantum machine learning). By contrast, interpolating through intermediate geometries traces a continuous trajectory where each step informs the next, preserving gradient signal and keeping the optimization grounded in regions of relevance to the problem.

The provided preprints don't contain a direct head-to-head comparison of interpolation versus random sampling in the specific geometry-optimization context your question implies.