Unverified50% confidenceFactExact time
In the presence of corruption, IAFTRL's regret upper bound is on the order of ~O(C* log T + √(C T)), where C is the total corruption.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe best upper bound previously given by Liu et al. in the strong corruption setting was O~(√(d^3 T) + dC), whose leading term has an extra √d factor compared with the corruption-free optimal rate O~(d√T).60% similarUnverifiedThe current best regret upper bound scales as √(Tγ_T) up to logarithmic factors.54% similarUnverifiedFloating-point addition is not associative, and the bf16 format has only a 7-bit mantissa, so rounding errors accumulate across multi-layer multiply-add operations and may cause the argmax result to flip.52% similarUnverified2^31 is the boundary of a 32-bit signed integer; when the total number of matrix elements M*N reaches 2^31, linear indexing with int32 overflows (wraps around to a negative number), which may lead to incorrect results or illegal memory access.52% similarUnverifiedUnder mild conditions satisfied by common kernel functions, minimizing the kernel-specific upper bound yields a number of replications of Θ(log T).52% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1128910API
curl https://kongchang.com/api/v1/knowledge/claims/1128910MCP
get_claim(id=1128910)