Sharp Bounds for Sums of Random Variables: Resolving the Feige Conjecture
This presentation explores a fundamental breakthrough in probability theory: the resolution of Feige's conjecture for bounded independent random variables. We examine how the authors establish sharp, dimension-dependent lower bounds on small-deviation probabilities using an elegant synthesis of probabilistic inequalities and convex geometry, particularly Grünbaum's centroid theorem. The work proves that the optimal constant is exactly e to the negative 1 for deviations of at least 1, settling a two-decade-old open problem and demonstrating the deep connections between geometric measure theory and probabilistic bounds.Script
When you add up independent nonnegative random variables, how likely is the sum to fall just below its expected value? For two decades, the Feige conjecture predicted the answer should be at least e to the negative 1, but proving it required breakthrough techniques combining probability and geometry.
The journey began in 2004 when Feige conjectured a sharp universal bound. Early results gave conservative constants around 0.13, then 0.18, but always under restrictive assumptions like identical distributions or log-concavity. This paper eliminates all such constraints and hits the target exactly.
The proof hinges on translating the probability question into convex geometry. By calibrating the sum distribution with the Dirichlet distribution on a simplex, the authors apply Grünbaum's centroid theorem, which bounds the volume of halfspaces not containing the centroid. This geometric insight directly yields the sharp probabilistic inequality.
Sharpness is demonstrated by explicit extremal examples: two-point distributions that concentrate mass at zero and n plus delta. These constructions prove the bound cannot be improved for any dimension n or deviation delta greater than or equal to 1, confirming the result is not just sufficient but optimal.
For deviations at least 1, the bound converges to e to the negative 1 as dimension grows, resolving Feige's conjecture in that regime. For smaller deviations, the bound remains explicit and universal but does not match the conjectured optimum, leaving an intriguing gap for future work.
This breakthrough opens doors across algorithm analysis, statistical inference, and high-dimensional probability, showing that sharp constants matter and that geometry can unlock them. To dive deeper into how probabilistic and geometric tools unite, visit EmergentMind.com and create your own exploration of cutting-edge research.
