7.3.2 Running Time
We want a measure of the running time of a procedure that satisfies two properties: (1) it should be robust to ephemeral properties of a particular execution or computer, and (2) it should provide insights into how long it takes evaluate the procedure on a wide range of inputs.
To estimate the running time of an evaluation, we use the number of steps required to perform the evaluation. The actual number of steps depends on the details of how much work can be done on each step. For any particular processor, both the time it takes to perform a step and the amount of work that can be done in one step varies. When we analyze procedures, however, we usually don't want to deal with these details. Instead, what we care about is how the running time changes as the input size increases. This means we can count anything we want as a "step" as long as each step is the approximately same size and the time a step requires does not depend on the size of the input.
The clearest and simplest definition of a step is to use one Turing Machine step. We have a precise definition of exactly what a Turing Machine can do in one step: it can read the symbol in the current square, write a symbol into that square, transition its internal state number, and move one square to the left or right. Counting Turing Machine steps is very precise, but difficult because we do not usually start with a Turing Machine description of a procedure and creating one is tedious.
Time makes more converts than reason. -- Thomas Paine
Instead, we usually reason directly from a Scheme procedure (or any precise description of a procedure) using larger steps. As long as we can claim that whatever we consider a step could be simulated using a constant number of steps on a Turing Machine, our larger steps will produce the same answer within the asymptotic operators. One possibility is to count the number of times an evaluation rule is used in an evaluation of an application of the procedure. The amount of work in each evaluation rule may vary slightly (for example, the evaluation rule for an if expression seems more complex than the rule for a primitive) but does not depend on the input size.
Hence, it is reasonable to assume all the evaluation rules to take constant time. This does not include any additional evaluation rules that are needed to apply one rule. For example, the evaluation rule for application expressions includes evaluating every subexpression. Evaluating an application constitutes one work unit for the application rule itself, plus all the work required to evaluate the subexpressions. In cases where the bigger steps are unclear, we can always return to our precise definition of a step as one step of a Turing Machine.