The funny thing about all this talk about smoothing is that it is a function of the software related to the visual display. One guy wants a nice smooth curve like the one he saw from an OE for a stock engine in a new car brochure, the next guy is tuning the living daylights out of his race car and wants precision details to see the best picture of what is actually going on during the run. The software engineers who design the dyno software have to create some balance that makes everyone happy.
Now, from a perspective of a dyno manufacturer, and since that's a field I'm sort of involved in, there are some deeper things going on than the smoothing bar sufrace that was scratched here. Every dyno manufacturer faces the same issue: Accurately converting analog datastreams collected from the machine and the vehicle into measured horsepower and torque and displaying them in a meaningful format that can be easily interpreted.
Depending on the dyno manufacturer, this data can be at a rate of 8hz to 100hz. The obvious thought would be that the more data you can collect, the more accurate the graph will be. Unfortunately, this is not the case. If you displayed the raw data on the screen, the power curve would look like a fuzzy bold line and show the power to be in a closely oscilating range of up to 200-300hp at any given engine speed on a 750hp car. The fact is, with increased accuracy, you can actually measure the spikes from each individual combustion event, and in fact see the angular accelleration in the engine itself. This is really cool if you are looking for a misfire or something truly bizarre like the cause of repeated cracked crankshafts in your combination, but thats another subject.
The key to this entire argument is that software (and in some cases firmware and hardware) engineers have created filters that reduce the amount of measured data to as little as 1/100th of the actual data collected. Also, to show the curve as it is generally dispayed by dyno manufacturers, there are averaging calculations built into the software that take high and low values of the oscillation and find a median average.
The data you see when you look at a dyno sheet is actually very much controled by the company that produces the dyno. Every dyno manufacturer has it's own staff of engineers that decide for us what we see from their equipment, and this is why the argument exists over which dyno produces the most accurate numbers. The simple fact is, no dyno is 100% accurate as a result of filtering and averaging, and depending on the company, how the data is presented. The best you can hope for is a good "sample" of the data that you can use to your benefit in some way or other.