Why does the separator-based partitioning strategy reduce inter-device communication better than pivot selection?

Your question compares two partitioning strategies, but the preprints provided don't actually discuss a direct comparison between separator-based and pivot-selection approaches. Let me explain what the archive does cover.

The separator-based strategy appears in work on distributed quantum circuits. In the Deferred Constraint Quantum Divide and Conquer algorithm [AG-2024.05-1606], a "separator" is a set of vertices whose removal partitions a graph into disconnected components. The key insight is that by isolating these separator vertices and deferring constraints associated with them, you can simplify the circuit topology dramatically — reducing the number of inter-device operations needed. Think of it like identifying a bottleneck in a supply chain: if you know which nodes are critical connectors, you can reorganize work to minimize goods crossing that bottleneck.

The same archive includes a QUBO-based partitioning method [AG-2024.07-1784] that frames the partitioning problem as a graph minimum-cut problem (finding the smallest set of edges to cut to separate a graph), which reduces quantum state transmissions between QPUs. There's also a reinforcement-learning approach to circuit partitioning [AG-2024.01-2111], but none of these explicitly contrast their method against a "pivot selection" baseline.

The preprints don't contain a head-to-head comparison of separator-based partitioning versus pivot selection strategies.