LoRA (Low‑Rank Adaptation) has become a key mechanism for efficiently specializing foundation models into many task‑ and user‑specific variants under a shared resource budget. Compressing a fleet of LoRA adapters requires allocating rank both across layers within each adapter and across adapters in the fleet, forming a two‑level allocation problem.\ \ Although trained LoRAs exhibit substantial spectral redundancy, spectral reconstruction alone does not determine how much functional degradation an adapter can tolerate. We therefore define a functional rank‑distortion frontier $D(r)$ for each adapter, describing the functional loss at rank $r$, and perform fleet‑level minimax allocation under a common total rank budget.\ \ Given exact frontiers, this decomposition is equivalent to joint optimization over all layer ranks. Based on this insight we introduce Functional Rank Allocation (FRA), an unlabeled, data‑driven method for constructing practical frontiers. FRA proceeds as follows:\
- Measure end‑to‑end functional loss for isolated module truncations;\
- Treat these measurements as an additive surrogate and solve the rank selection problem via dynamic programming;\
- Re‑measure every complete candidate to correct surrogate error;\
- Take the lower‑risk envelope with a complementary spectral proposal;\
- Perform exact minimax allocation over the constructed frontiers.\ \ Across three public fleets comprising 78 adapters, FRA improves worst‑case and tail utility under constrained budgets, outperforming uniform spectral compression. Fixed‑budget probes show that isolated cost rankings rank candidates well, but their calibration shifts with proposal shapes; thus end‑to‑end verification and the spectral proposal together bound functional‑search error.\ \ Review