Consider a Turing machine M that, after one or more transitions, again has a blank (all-0) tape. Let C_i be the state that M is after the last such shift is made. Create a machine M' that is identical to M but starts in state C_i. Clearly, M' will behave exactly as M does. Since M' has a different first transition than does M, M' is non-isomorphic to M and, based on the completeness of normalization, will be generated (possibly by proxy through a machine representing its behaviour) at some time during the generation of the tree. Hence, empty tape machines need not be considered.

Notice that machines whose first action (i.e the action take from the start state when reading a 0) is anything other than "write a 1 on the tape" are empty tape machines. Therefore, by defining the first action taken by any machine to be "write a 1," we can immediately eliminate an enormous number of empty tape machines. The image below shows the expanded root node after applying this optimization.

Of course in addition to all machines which do not have a first write 1, we also eliminate any other empty tape machines which fit the above specification by implementing a generalized solution to the problem. To do this, we subsequent to the first transition, simply track how many non-blank symbols are on the tape; if this number ever reaches 0, then we discard the machine and its successors in the tree as blank-tape machines. Since the first move has been defined to be "write a 1", we are assured that all blank-tape machines have been eliminated from consideration.