Isabelle/HOL: Formalizing Modern Submodular Greedy Algorithms
Problem: Formalize in Isabelle/HOL the approximation theory for monotone nonnegative submodular maximization under a cardinality constraint, covering classical greedy, LazyGreedy, and StochasticGreedy, together with their approximation guarantees, algorithmic constructions, and oracle-cost analyses.
Methods: A locale-based modular organization that separates abstract assumptions, algorithmic constructions, step-spec / recurrence proofs, concrete instance interpretations, and complexity / oracle-cost layers, so that the proof structure remains clear and extensible.
Current stage: The classical greedy baseline is complete; LazyGreedy now has a full stateful, approximation, and oracle-cost development; StochasticGreedy has completed a concrete uniform with-replacement theorem line and reached the key milestone of packaging the final approximation theorem.
Next step: Continue the clean-up and refinement: unifying names and comments in the stochastic part, pruning dependencies, strengthening the cost-comparison narrative, and improving the presentation toward an AFP / submission-style version.
Current status: Ongoing (the core theorem lines are complete; current work is focused on clean-up and presentation quality).
Duality and Lower Bounds under Invariances
Problem: For certain distance-fitting objectives defined through graph-relative differences, natural dual formulations may collapse because of translation invariance or related non-identifiability.
Methods: Comparison of Fenchel/DC and Lagrangian routes, followed by structural analysis of where the degeneration comes from and what extra constraints or modifications may be needed.
Current stage: I have worked through the main derivations and clarified where the trivial lower bound appears.
Next step: Test whether gauge-fixing, augmentation, or tighter constraints are enough to recover a nontrivial bound.
Current status: Ongoing.
Metalcore Piano Lab | Audio / MIDI to Chart (WIP)
This prototype is doing something fairly concrete: taking continuous audio or MIDI input and turning it into playable rhythm-game chart events. In practice that means onset detection, beat tracking, segmentation, alignment, and a set of playability constraints that are not yet fully settled.
I have built a browser version mainly as a way to test where those assumptions break on real inputs. What interests me most at the moment is how chart quality should be evaluated, and which constraints are genuinely necessary rather than just convenient for implementation.
Current status: Work in progress.