I want a .1%-accurate standards lab

September 4, 2012 1:49:31 PM CDT

Over the past year, I've caught a bad case of the analog circuit design bug. One symptom of that is a fascination with making precise measurements.

Most circuits don't bother with precision. Resistors are the 'precise' components, and they usually have error tolerances of 1-5%. Capacitor tolerances are usually 20%, and if you work with semiconductors you can expect critical values to vary by a factor of 3 or more. 50% of the tools in a circuit designer's bag of tricks involve ways to sidestep or ignore those errors.. give or take 20%.

That's generally a good thing. The world is a noisy, messy place, and circuits that only operate for specific values tend to be worthless. Digital circuits have taken the wide-tolerances thing to extremes, reducing all signals to two values that are far enough apart that you won't mistake one for the other in a million years (seriously.. bit-error-rate calculations tell you how wide your tolerances have to be to get that kind of reliability).

Still, we live in an analog world, and there are times when precision matters.

That leads to the chicken-and-egg problem inherent in all standards work though: it's easy to define error as deviation from a standard value, but that assumes you have a standard value to use as your reference. How do you measure the straightness of a line if you don't have a straight line to compare it with? More generally, if you have two almost-straight lines, how do you decide which one is straighter?

As it turns out, there are tricks for making errors cancel themselves out, or for moving them to places where you can ignore them. Knowing how to apply those tricks is valuable whether you're doing precision work or not. Making your error tolerances bigger is always easier than making them smaller.

The good news is that once you've created an accurate measurement, you can use it as a reference to test others. A collection of such reference measurements is a standards lab.

Eventually, all standards work becomes a question of temperature control. Every measurement in the world varies with temperature, and the smaller your tolerances get, the more important those thermal variations become.

Frankly, I don't have the bug that bad (yet). I want a craftsman's set of references, not a set that requires you to wait for three days while the temperatures equalize to within .01° of nominal.

For now, I want measurements that are accurate to four significant digits. If I measure 1 volt, I want to know the value is closer to 1000mV than to 999mV or 1001mV. It's not the most ambitious goal in the world, but it's good place to start. The tools and techniques that get me to that point will allow me to go farther if I want to.

Random brain cookies:

The algorithm to do that is extremely nasty. You might want to mug someone with it. -- M. Devine, Computer Science 340