Optimal and Efficient Online Inverse Optimization

Get email updates

Anupam Gupta, Guru Guruganesh, Honghao Lin, Vahab Mirrokni, Renato Paes Leme, David P. Woodruff

A deterministic algorithm attains O(d)O(\sqrt{d}) regret for every horizon T in online inverse linear optimization and runs in time polynomial in the dimension d and the horizon T in the paper’s optimization model.

What was already known

Sakaue established the optimal regret order with a randomized method needing exponentially many linear optimizations in the dimension. He asked whether polynomial time was possible.

AI involvement AI-autonomous

The authors credit the Gemini 4 Colosseum system with obtaining the initial proof end to end. They verified it and edited the exposition.

Importance Notable

The result preserves the optimal regret order while making the running time polynomial.

Technical verification Formal (Lean + human)

The authors report reviewing the proof and releasing a Lean 4 verification of the results. The repository describes the main formal statement and how it captures the regret analysis.

Independent checking

Verified the initial end-to-end proof.

Checker
Paper authors
Scope
A deterministic algorithm attains O(d)O(\sqrt{d}) regret for every horizon T in online inverse linear optimization and runs in time polynomial in the dimension d and the horizon T in the paper’s optimization model.
Basis
Reported by the linked checker

Formal checking

Checked the paper’s formalized results for online inverse optimization.

Checker
Lean 4 kernel
Scope
A deterministic algorithm attains O(d)O(\sqrt{d}) regret for every horizon T in online inverse linear optimization and runs in time polynomial in the dimension d and the horizon T in the paper’s optimization model.
Basis
Reported by the linked checker

Human understanding Few

The authors report verification and editorial changes. A human reconstruction of the model’s metric-revocation argument is undocumented.

Other useful sources

Correction and publication history