Martin's post has it right, but let me offer another perspective that also might help.
When you look at a bandwidth graph that indicates something like 10% utilization over 5 minutes on a FastEthernet interface, does this mean the interface only transmitted at 10 Mbps? Of course not, since it always transmits at 100 Mbps. What it means is the interface was only active 10% of the time, although we don't know how the 10% was distributed over the 5 minutes.
Time periods are very important. If the 10% over five minutes, above, was all done in the first 30 seconds of a 5 minute period, but if we measured usage now on 1 minute intervals, the first minute would show 50% utilization, and the remaining 4 minutes would each show only 0% utilization.
Over a 30 second measured periods, again with data all transmitted in the first 30 seconds, the first would show 100% and the next 9 also 0%.
If instead of a 100 Mbps interface, we set the speed down to 10 Mbps, the interface would
actually be busy 100% of the time for all 5 minutes to transmit the same number of bits. It would also measure at 100% regardless of how we subdivide the 5 minute period.
If instead of setting the interface to 10 Mbps, we shape or police at 10 Mbps. Does this now simulate a 10 Mbps interface? Not exactly, but how close it does is related to the burst size which sets the measurement window.
If we set the burst size large enough that 5 minute period is defined, the shaper or policer will not allow more bits to be sent during the whole 5 minutes than would exceed 10% of the total time. Once again, we don't know where in the 5 minutes the bits are transmitted.
If we set the burst size such that the time period is 1 second, again the shaper or policer will insure only 10% of the time interval is used for transmission. If there were 30 seconds of 100 Mbps traffic awaiting in the queue, an average rate policer would transmit for 100 ms and drop all the remaining traffic. An average rate shaper would transmit for 100 ms, pause 900 ms, and repeat until all the data was transmitted, which would take 5 minutes (and now more closely simulates a 10 Mbps interface -- over time).
The above is simplified. Normal burst sizes and their corresponding time windows are often down in the 25 to 40 ms range. Since we work with frames and not bits, effective rates can be slower than expected (or as defined). Not discussed has been peak rate vs. average rates.
In practice, a policer or rate-limiter, simulates a lower speed interface without any queuing. A shaper queues traffic, but its queue depth and queue management might differ from a physical interface you're attempting to simulate. If you're dealing with a downstream provider that "caps" excess bandwidth, it's often advantageous to implement like "caps" where you can control the traffic and see what's happening. In this latter instance, important to find out what all the parameters the provider is using for their "cap", i.e. more than just a CIR or PIR rate.
