The dual functional of the work function is quasiconvex with the same alignment
ProvedKServer.workFnU_dual_quasiconvexFix a point and consider the dual functional of a -server work function ,
the object at the heart of the Koutsoupias--Papadimitriou duality lemma: its minimisers are where the extended cost of a request at is realised, and its minima drive the potential of Coester and Koutsoupias. The theorem says is quasiconvex in the same pairing sense as itself, with the same alignment: for any two configurations there is a permutation such that for every split of the coordinates into two blocks, the two hybrid configurations (taking one block from and the complementary block from , and vice versa) satisfy
Why it is free
Quasiconvexity is a genuine theorem for --- it encodes the exchange structure of optimal offline solutions. For the dual functional no new argument is needed, because the subtracted term is modular: is a sum of per-coordinate distances, and each coordinate of and of appears in exactly one of the two hybrids. Hence
exactly (after re-indexing the -part along , under which the sum is invariant), and subtracting this identity from the quasiconvexity of gives the claim --- with the same , which is what makes the lemma directly usable: any exchange argument for runs verbatim for .
Role
In the tree analysis of Coester and Koutsoupias, the anchors of the potential are located greedily with respect to dual functionals: their Lemma 25 finds the first anchor among the minimisers of , an instance of minimising under a containment constraint, and the greedy and substitution lemmas that justify such choices are exchange arguments of exactly the pairing form above. This theorem is what licenses running them against rather than .
Formalization note
Configurations are functions and the hybrids are written with an if i ∈ t selector over a Finset (Fin k); is workFnU, the work function of the unlabelled configuration. The statement quantifies the split inside the existential, exactly as in the underlying workFnU_quasiconvex, so a single alignment witnesses all splits at once.
import Mathlib import Definitions.Def_KServer_workfunctionU
namespace KServer
theorem workFnU_dual_quasiconvex (k : ℕ) (hk : 1 ≤ k) (M : Type) [MetricSpace M]
(C₀ : Config k M) (σ : List M) (s : M) (X Y : Config k M) :
∃ π : Equiv.Perm (Fin k), ∀ t : Finset (Fin k),
(workFnU C₀ σ (fun i => if i ∈ t then X i else Y (π i))
- ∑ i, dist s ((fun i => if i ∈ t then X i else Y (π i)) i))
+ (workFnU C₀ σ (fun i => if i ∈ t then Y (π i) else X i)
- ∑ i, dist s ((fun i => if i ∈ t then Y (π i) else X i) i))
≤ (workFnU C₀ σ X - ∑ i, dist s (X i))
+ (workFnU C₀ σ Y - ∑ i, dist s (Y i)) := by sorry
end KServer