On the minimum of the mean-squared error in 2-means clustering. (English) Zbl 1406.62067

Bodmann, Bernhard G.; George, Craig J.

Involve 12, No. 2, 301-319 (2019).

Summary: We study the minimum mean-squared error for 2-means clustering when the outcomes of the vector-valued random variable to be clustered are on two spheres, that is, the surface of two touching balls of unit radius in \(n\)-dimensional Euclidean space, and the underlying probability distribution is the normalized surface measure. For simplicity, we only consider the asymptotics of large sample sizes and replace empirical samples by the probability measure. The concrete question addressed here is whether a minimizer for the mean-squared error identifies the two individual spheres as clusters. Indeed, in dimensions \(n \geq 3\), the minimum of the mean-squared error is achieved by a partition obtained from a separating hyperplane tangent to both spheres at the point where they touch. In dimension \(n=2\), however, the minimizer fails to identify the individual spheres; an optimal partition is associated with a hyperplane that does not contain the intersection of the two spheres.

MSC:

62H30	Classification and discrimination; cluster analysis (statistical aspects)
62G05	Nonparametric estimation

Keywords:

\(k\)-means clustering; performance guarantees; mean-squared error

References:

[1]	10.1109/TIT.1982.1056486 · Zbl 0476.94013
[2]	10.1109/SFFCS.1999.814639
[3]	10.1137/S0036144599352836 · Zbl 0983.65021
[4]	10.1007/978-1-4615-3626-0
[5]	10.1007/BFb0103945 · Zbl 0951.60003
[6]	10.1007/s10107-016-1097-0 · Zbl 1377.65012
[7]	10.1109/TIT.1982.1056482 · Zbl 0525.94006
[8]	10.1109/TIT.1982.1056489 · Zbl 0504.94015
[9]	10.1109/ITW.2016.7606826
[10]	10.1016/0022-1236(82)90069-6 · Zbl 0506.46022
[11]	10.1137/050641983 · Zbl 1146.90046
[12]	10.1214/aop/1176993713 · Zbl 0502.62055
[13]	10.1109/TPAMI.1984.4767478 · Zbl 0546.62037
[14]	; Steinhaus, Bull. Acad. Polon. Sci. Cl. III., 4, 801 (1956)
[15]	10.1007/s00454-011-9340-1 · Zbl 1218.68088