Adversarial and Cost-Aware Guarantees for Adaptive Learned Data Structures

Characterize the behavior of adaptive learned data structures, including ALEX and related model-based indexes, under adversarial update sequences; determine suitable rebuild policies; and establish end-to-end guarantees that account for retraining and memory-movement costs.

Background

The survey distinguishes static learned-index guarantees from the empirical or formal update costs of adaptive structures such as ALEX and the PGM-index. Static models can provide certified local-search error bounds, but workload drift, insertions, deletions, retraining, and data movement introduce additional costs that are not captured by ordinary query-time analyses.

The unresolved direction is to develop guarantees that remain meaningful under adversarial update sequences while also specifying when and how an index should be rebuilt and how retraining and memory movement contribute to end-to-end performance. A portfolio of learned models would require an additional state-migration analysis before model-selection regret could yield an index-level guarantee.