Minimax bounds for watermarked and masked recursive discrete distribution estimation
Watermarks only help this estimator if detection misses disappear asymptotically.
Kanabar and Gastpar analyze recursive discrete distribution estimation when real and synthetic samples are mixed without metadata. Their lower bound says that, as the real-sample fraction vanishes, watermarking cannot improve performance unless the detector’s false negative rate also goes to zero. They also give simple deterministic estimators whose worst-case losses match the bounds up to constants in most regimes, and propose masking to narrow the remaining gap. ArXiv · AI/CL/LG's note
Kanabar and Gastpar analyze recursive discrete distribution estimation when real and synthetic samples are mixed without metadata. Their lower bound says that, as the real-sample fraction vanishes, watermarking cannot improve performance unless the detector’s false negative rate also goes to zero. They also give simple deterministic estimators whose worst-case losses match the bounds up to constants in most regimes, and propose masking to narrow the remaining gap. ArXiv · AI/CL/LG's note
score 4