Skip to content
PreprintJuly 2026

Proof-Carrying Optimality for Finite Identification under Bounded Adversarial Answer Errors

Independently checkable upper and lower certificates for exact finite-identification query complexity

Vikram Lex

Public preprint posted on Zenodo in July 2026. The journal manuscript is under review at the Journal of Machine Learning Research.

Zenodo · Under Review at JMLR

Formal MethodsCertified SearchRobust Identification
30/30
Primary Cells Proved
Two-sided optimality certificates at error budgets 0 and 1
300/303
Sweep Cells Proved
Three declared cells remain explicit resource-limit outcomes
84
Maximum Adaptive Gain
Probe reduction at error budget 2
29/30
Replay Faster
Timed primary cells where checking beats synthesis
Abstract

Paper Summary

Finite identification asks a learner to determine an unknown behaviour using probes from a fixed alphabet when an adversary may corrupt up to b answers. Exact worst-case query complexity is the value of a large minimax game, so an optimum can be expensive to reproduce. We develop proof-carrying methods for many-valued behaviour tables. Portable certificates combine a noiseless strategy with a small lower witness to prove an affine identity OPT_A(b) = αb + β for every budget, with proof size independent of b. Expanded certificates pair a robust strategy at one budget with an adversary DAG closed by analytic lower bounds. We also provide independently checkable lower bounds and certificates for the non-adaptive optimum. All 30 cells in a 15-instance primary matrix at b ∈ {0,1} receive two-sided optimality proofs. A 101-instance sweep across b ∈ {0,1,2} proves 300 of 303 declared cells, leaving three explicit resource-limit outcomes. At b = 2, adaptivity strictly improves on the exact non-adaptive optimum in 91 completed cells, by as many as 84 probes. Replay is faster than synthesis in 29 of 30 timed primary cells, although the largest accepted lower certificate contains 8.9 million moves. These results make query-complexity claims auditable while exposing proof construction and proof size as bottlenecks.

Certificate framework

Proof-carrying Optimality

Untrusted search constructs candidate strategies and witnesses; a separate verifier recomputes and replays the supplied evidence without repeating synthesis.

Finite model

Declared Identification Game

The input is an explicit many-valued behaviour table with a fixed probe alphabet and a declared adversarial error budget. The target is exact worst-case adaptive or non-adaptive query complexity.

Upper evidence

Portable and Expanded Strategies

Portable certificates prove affine all-budget identities from a noiseless strategy and lower witness. Expanded certificates encode a robust strategy for one budget when a compact all-budget proof is unavailable.

Lower evidence

Adversary and Cover Certificates

Replayable adversary DAGs, analytic closures, and non-adaptive cover evidence independently establish matching lower bounds, allowing the verifier to accept an optimum or reject with a reason.

Certified evidence

Proof Coverage

Completed cells carry matching upper and lower evidence; incomplete cells remain explicitly identified as resource limits rather than inferred values.

Primary Matrix

15 instances at adversarial error budgets 0 and 1

Two-sided optimality30/30 cells

Every declared primary cell receives both an accepted strategy certificate and a matching lower-bound certificate.

Replay versus synthesis29/30 faster

Certificate replay is faster than constructing the result in 29 of 30 timed primary cells; timing observations are descriptive and come from two machines.

Largest lower certificate8.9M moves

The largest accepted lower certificate contains 8.9 million moves, exposing proof size as a practical bottleneck even when verification avoids search.

Declared Sweep

101 instances across adversarial error budgets 0, 1, and 2

Completed proofs300/303 cells

The sweep proves 300 of 303 declared cells. The remaining three are reported as resource-limit outcomes, not estimated optima.

Strict adaptive improvements91 cells

At error budget 2, adaptivity strictly improves on the exact non-adaptive optimum in 91 completed cells.

Largest certified improvement84 probes

The maximum certified adaptive advantage is 84 probes relative to the exact non-adaptive optimum.

Verification Scope

What the certificates do and do not establish

Explicit table modelDeclared alphabet

All claims concern an explicit finite behaviour table and its declared probe alphabet; the largest primary table has 49 classes and 220 probes.

Checker complexityPolynomial

Verification is polynomial in the explicit table and supplied proof, not necessarily in a succinct description of the underlying game.

Unresolved cells3 explicit

Resource exhaustion remains a first-class reported outcome rather than being converted into an unsupported query-complexity value.

Citation

Cite This Paper

@misc{lex2026proofcarrying,
  title={Proof-Carrying Optimality for Finite Identification under Bounded
         Adversarial Answer Errors},
  author={Lex, Vikram},
  year={2026},
  month={jul},
  publisher={Zenodo},
  doi={10.5281/zenodo.21651833},
  url={https://doi.org/10.5281/zenodo.21651833},
  note={Preprint; manuscript under review at the Journal of Machine
        Learning Research}
}