Dave,
Noise can not be controlled. Noise can not be minimized. Noise is a state of nature.
But the photographer can control the signal level.
There are two primary sources of noise. Shot noise is fundamental to the nature of light. Bright regions like sky have significant shot noise. The camera adds electronic noise, a.k.a. read noise. The sensor itself and the analog to digital converter contribute to the read noise. Signal amplification after the shutter closes is called ISO in digital cameras and in newer camera designs ISO amplification is not a source of noise.
The only thing you can control is the signal level, or the exposure. Your concern is actually how to maximize signal-to-noise ratio, (SNR) which means you have to maximize exposure.
To maximize exposure:
Use base ISO when possible (the native ISO of the sensor)
Otherwise, use the slowest ISO above base ISO.
(These steps will also maximize dynamic range.)
Let highlights that are not important to the purpose of the photograph overexpose. This maximizes the SNR of shadow regions.
This strategy can be confused with the Expose To The Right, ETTR, method. But it is not ETTR nor is the motivation related to the original rationale for ETTR. It is only a coincidence the histogram when you maximize exposure resembles a ETTR histogram for some scenes.
Because it sells software, digital photography misplaces emphasis on post-production noise reduction. Noise reduction is a great marketing term, but like most marketing it does not tell the whole story.
Noise reduction is actually noise filtering. Mathematical filtering averages pixels with low signal together with nearby pixels with more signal. This increases the information in the low signal pixels but it also decreases the information of the high signal pixels. The total images noise content remains constant.
If anyone can figure out how to decrease the total noise content of data after a single measurement is over, they will become famous and wealthy overnight.
Of course, noise filtering is useful because a productive compromise or balance is easy to achieve. We know from experience the filtering algorithms can be very effective. Often the information loss is insignificant compared to the information gain of the low signal pixels.
But nothing beats having data with a the highest possible SNR in the first place.
