I've been working with precision op amps. Some have good linearity. Some are excellent. Some have high ZOUT. Some have low. Some are bipolar, others are CMOS. The figure shows the basic test I've been using for linearity to exercise the output to 20 V p-p by applying a ±10-V sine or triangle wave to the signal input.

Meanwhile, the noise gain of 1000 (R2/R1) magnifies the input voltage V( ) by a factor of 1000 so the scope can see (on a 5-mV scale) a 5- µ V signal in cross-plot mode. The noise can be as low as 5 µV p-p, and you can see the distortion of just a microvolt or two riding along under the noise.

It's true that many engineers are interested in the total harmonic distortion plus noise (THD+N). However, the distortion riding along under the noise is sometimes important, even though an Audio Precision can most easily measure THD+N.

I did some evaluation on the LM4562 precision audio op amp. Its noise for an audio bandwidth is down near 0.4 µVRMS, and at least that's easy to measure. But I still had to measure the distortion at 1 kHz. The test circuit in the figure can't show the linearity of the gain at 1 kHz, only at 5 or 10 Hz.

That's because the ac error is so big, it isn't easy to see the distortion. For example, the LM4562's ac gain at 1 kHz is about 60,000. Not bad, but for a full output, the summing point error is 333 µV, and it's hard to see if that is linear within a few microvolts.

CHEATERS EVENTUALLY PROSPER
So I decided to cheat. I used a small variable capacitor—a few inches of twisted pair, often called a "gimmick," using teflon wires. I connected this from the VIN to the input of the op amp. As I wound up the wires, the ac component of the error voltage shrank a lot.

I kept increasing the frequency as well as the capacitance. Finally at 1 kHz, I got it down under 1 µV of signal plus a few µVRMS of noise. This provided good insight into what the 1-kHz gain looked like. But why was the noise so big?

I realized that I'd been using this lazy man's gain test for so long, I wasn't paying attention to the way the noise of the 1k resistor (about 4 nV/√Hz was bigger than the op amp's noise. So, it was time to cut the impedance levels! I didn't rewire the circuit. Instead, I just slapped in 20k across each 1 M½ and 20 Ω across the 1k. Of course, the capacitance had to be scaled up too, so I put in about 140 pF on top of the 3-pF gimmick.

This provided a definitely improved view of the distortion, with an improved noise floor. I could see that the ac distortion, even at 1 kHz, was somewhere well below 1/2 µV p-p. But I still couldn't see exactly how low. So I got mad and fed this signal into our HP3561A spectrum analyzer. This plainly showed the amount of the distortion, such as 71.45 nV at 2.2 kHz, with a 10k load. (It degraded to 200 nV with a 1k load.)

The combination of the subtracting and self-amplifying effects of my circuit, plus the ac cancellation, plus the high resolution of the spectrum analyzer, showed ?159 dB of distortion at 2.2 kHz ( second harmonic) when running the LM4562 at a 20-V p-p sine output at 1.1 kHz. This was the best distortion I have ever seen, and fortunately the best test circuit I have ever seen, or we wouldn't have been able to measure it.

Comments invited!
rap@galaxy.nsc.com —or: Mail Stop D2597A, National Semiconductor P.O. Box 58090, Santa Clara, CA 95052-8090