How Reviewers Pushed a LogSumExp Optimality Paper to Arbitrary Dimensions

prof_grimmer · x · 2026-11-18

profgrimmer shares how peer review matured their Journal of Complexity paper on softmax/LogSumExp smoothing. After previously proving LogSumExp strictly suboptimal for d=2,3, reviewers asked for results in any dimension; the authors articulated a parameterized family of smoothings containing LogSumExp, with their strictly improved smoothing optimizing over the family. They also added a vignette showing how to construct a large family of "elementary" lower-bound proofs, with theirs the optimum within it.

Original post →

More from Research

Research channel →