- The paper introduces NP with solutions (NP-S) and solution-embedding reductions, preserving the structure of feasible solutions rather than only yes/no answers.
- The paper proves that multilevel adversarial selection, interdiction, and robust optimization variants reach Σ_k^p or Σ_{2k−1} completeness at fixed depth and PSPACE-completeness with unbounded stages.
- The paper unifies and extends hardness results across more than 26 problems, while showing that complexity depends on the chosen solution representation and formulation.
Overview
This paper develops a general framework for proving completeness in the polynomial hierarchy (PH) and PSPACE for multilevel optimization problems derived from classical NP-complete problems, such as interdiction, protection-interdiction games, and multi-stage adjustable robust optimization (2602.12350). The central observation is that many textbook NP-completeness reductions preserve more structure than mere many-one reducibility: solutions of the source problem correspond one-to-one to elements of solutions of the target problem. The authors formalize this via a refinement of NP, called NP with solutions (NP-S), and a restricted reduction type called solution-embedding (SE) reductions. From these they derive meta-theorems showing that if a problem is NP-S-complete, then its natural two-level extensions are PSPACE0-complete, its PSPACE1-level extensions are PSPACE2-complete, and unbounded-level variants are PSPACE3-complete.
The practical significance is substantial: prior work established PSPACE4-completeness one problem at a time — e.g., maximum clique interdiction [Rutenburg], interdiction knapsack [Caprara et al.], and two-stage adjustable TSP/vertex cover/independent set [Goerigk et al.] — each via bespoke reductions from quantified satisfiability. The present framework subsumes essentially all known such results and yields new ones simultaneously for a large catalog of problems, including Clique, Vertex Cover, Knapsack, TSP, Steiner Tree, Dominating Set, Set Cover, Feedback Arc Set, facility location, p-Center, p-Median, 3-Dimensional Matching, Hamiltonian path/cycle variants, disjoint paths, and Scheduling. This demonstrates that high computational complexity is a generic feature of multilevel extensions of PSPACE5-complete problems, rather than an artifact of individual constructions.
The class NP-S and solution-embedding reductions
A problem in PSPACE6-PSPACE7 is a triple PSPACE8: instances PSPACE9, a polynomially bounded universe NP0 per instance, and a solution set NP1 recognized by a polynomial-time verifier; the decision question is whether NP2. The role of NP3-NP4 is not to add computational power — every NP5-NP6 problem embeds into NP7 by encoding the verifier's description number into the instance — but to make solution structure explicit so that reductions can preserve it.
An SE reduction from NP8 to NP9 consists of a standard many-one map NP0 plus, for each instance, an injective universe embedding NP1 satisfying the SE property:
NP2
where NP3. In words, solutions of NP4 extend to solutions of NP5 exactly on the embedded universe, and every solution of NP6 restricts to a solution of NP7. A key lemma shows this extends from full solutions to partial solutions, which is precisely what is needed when players in multilevel games act element-by-element. SE reductions are transitive, and the Cook–Levin construction itself is shown to be an SE reduction, establishing that both variable-based and literal-based Satisfiability are NP8-NP9-complete.
The paper is careful about modeling choices: different universes or verifiers for the same underlying NP0 language yield distinct NP1-NP2 problems with potentially different lifted complexities. A striking example concerns Partition: the symmetric variant (where NP3 is a solution whenever NP4 is) provably admits no SE reduction from Satisfiability and hence cannot be NP5-NP6-complete, whereas the asymmetric variant requiring a fixed element in the solution is NP7-NP8-complete. Similarly, IP-based models of 3-Coloring and Bin Packing are "universe covering" and therefore not NP9-NP0-complete, but equivalence-relation models restore completeness.
The lifting mechanism rests on a proof technique the authors call dual mimicking. Given an SE reduction NP1 from Sat-V to some NP2-NP3-complete NP4, and given a multilevel game operator NP5, one shows that NP6 embeds as a sub-instance of NP7: moves in the Sat game are mapped through NP8, and optimal strategies transfer in both directions. Since adversarial selection on Sat-V is exactly quantified Boolean satisfiability (NP9), which Stockmeyer and Wrathall showed complete for NP0 (and NP1-complete for unbounded alternation), hardness transfers to all NP2-NP3-complete problems at once.
Three families of results are established:
Adversarial selection games. For any NP4-NP5-complete NP6, the NP7-move game in which Alice and Bob alternately select subsets of partitioned universe blocks, with the last mover winning iff the union forms a solution, is NP8-complete; it is NP9-complete when S0 is part of the input.
Protection-interdiction games. For any S1-S2-complete S3, plain interdiction (does there exist a blocker S4 of size at most S5 hitting every solution?) is S6-complete, as is minimum-cost interdiction. More generally, the combinatorial S7-move protection-interdiction game — where protector and interdictor alternately place tokens under nested access sets S8, and the protector must finally find a solution avoiding all interdiction tokens — is S9-complete for constant NP0 and NP1-complete when NP2 grows, for both global and per-move budget variants. The hardness proof for Sat-V is the technical core: a reduction from the adversarial selection game first to a restricted interdiction game with four behavioral rules, then to the unrestricted game via cheat-detection gadgets (variables NP3 and helper variables ensuring that the first player to violate a rule loses). Notably, the authors point out that vertex-cover interdiction formulated with actual vertex deletion lies in coNP, so no general NP4-completeness can hold in that formulation; their result applies to forbidding elements from solutions.
Multi-stage adjustable robust optimization. Under discrete budgeted uncertainty, the combinatorial NP5-stage adjustable robust version of any NP6-NP7-complete problem is NP8-complete for constant NP9 and S0-complete when S1 is part of the input; the cost-function version follows since 0/1 costs encode blocking. Because each stage contributes an existential decision-maker move followed by a universal adversary move, the level index is S2 rather than S3. The same restricted-game-plus-gadgets methodology applies, with a single cheat-detection variable sufficing here. As a corollary, online optimization problems with estimates (e.g., online knapsack with estimates) inherit these lower bounds when modeled as multi-stage adjustable problems.
Robustness of the framework
A dedicated section examines alternative S4-S5 models. Complement-based models (solutions specified by variables set false), literal-based models (universes containing both literals), and duals of literal-based problems (e.g., literal-based Independent Set as the dual of literal-based Vertex Cover) are all shown to be S6-S7-complete whenever the base model is. Universe-covering obstacles are circumvented for coloring and packing problems via equivalence-relation models (yielding strongly S8-S9-complete Bin Packing and 4-Partition) and subset-generation models (Covering by Triangles, Partition into Triangles). Finally, a "2-step verification" technique — where a partially specified solution is completed by a polynomial-time algorithm — yields an PSPACE00-PSPACE01-complete model of Sequencing to Minimize Tardy Tasks, whose natural IP-based model the authors conjecture is not PSPACE02-PSPACE03-complete. These results indicate the framework captures a structural phenomenon persisting across representation choices rather than depending on one canonical encoding.
Limitations and open questions
Several caveats are stated plainly. The framework requires the underlying PSPACE04-complete problem to admit an SE reduction; not all PSPACE05-complete problems qualify, and the paper itself exhibits counterexamples (symmetric Partition, universe-covering formulations of 3-Coloring and Bin Packing, and the conjectured non-completeness of the IP model of Sequencing to Minimize Tardy Tasks). The interdiction results apply only to the formulation where blocked elements are forbidden from solutions rather than deleted from the instance, since deletion can place the problem in coNP. The results are purely complexity-theoretic: no algorithmic consequences, approximation guarantees, or parameterized analyses are derived, and the authors explicitly leave open how these hardness results should inform algorithm design. They also note that refining the PSPACE06-PSPACE07 abstraction or finding alternatives capturing further multilevel settings remains open, as does extending the structural techniques to other lifting scenarios beyond the polynomial hierarchy and PSPACE08.
Conclusion
The paper provides a unified complexity-theoretic account of multilevel optimization over PSPACE09-complete problems. By isolating solution structure in the class PSPACE10-PSPACE11 and formalizing solution-embedding reductions, it converts what were previously lengthy, problem-specific reductions from quantified satisfiability into single verifications of the SE property, after which PSPACE12- and PSPACE13-completeness follow systematically for adversarial selection, interdiction, protection-interdiction, and multi-stage adjustable robust variants. The compendium of 26-plus PSPACE14-PSPACE15-complete problems, together with the demonstrated sensitivity of lifted complexity to the choice of universe and verifier, establishes that PSPACE16-hardness alone is often insufficient for understanding the difficulty of multilevel problems, and that high complexity at higher levels of the hierarchy is the generic case rather than the exception.
