Figure 4.

The agglomerative information bottleneck method merges clusters $z_k^{*}$ and $z_{k\text{'}}^{*}$ in partitioning Z_m to form a coarser partitioning $(z_k,z_{k\text{'}})$. The resulting partitioning Z_m − 1 is such that there is minimal loss of information – i.e., $I(\overline{X},Z_{m-1})\approx I(\overline{X},Z_{m-1})\approx I(\overline{X},X)$