Counterexamples show e-means clustering can lose unboundedly to k-means

RexDouglass · x · 2026-11-18

An arXiv paper, "Two Counterexamples for e-means," challenges the e-means clustering algorithm (Bessa et al., 2026), which combines exponential averaging with cluster revival. The author constructs two counterexamples where, starting from identical initial centers, following the published recipe yields a higher final loss than standard k-means (Lloyd, 1982). The paper further shows that mass-based revival can destroy a global optimum, and the ratio of expected losses can be unbounded even with k-means++ seeding.

Original post →

More from Research

Research channel →