The elbow method has a simple theoretical justification. If the data actually has k well-separated clusters, the gains from splitting halves of clusters should

The elbow method has a simple theoretical justification. If the data actually has k well-separated clusters, the gains from splitting halves of clusters should be dramatically less than from splitting distinct clusters. If that isn't true, the data doesn't have well-separated clusters anyway.