What Feedback Does to Noise and Distortion
Close a loop around one op-amp and measure what improves: distortion falls by 1 + T, input noise does not fall at all.
You Will Be Able To
- Explain why feedback leaves the input-referred noise of an amplifier unchanged.
- Predict the noise that a feedback network adds from its two resistors.
- Predict a closed-loop harmonic from the open-loop harmonic and the loop gain at its frequency.
- Measure harmonic levels from a transient run, and state the numerical floor of the method.
- Lesson
- E10.6
- Time
- 35 min
- Runs on
- ngspice
- Check
- 5 questions
A sensor team closed a loop around its op-amp to clean up the signal. The second harmonic fell by 35 dB, but the noise at the input went from 62.0 to 65.3 nV/√Hz.
- open loop
- closed loop, gain −1
Read the plot as a table
| Frequency | open loop | closed loop, gain −1 |
|---|---|---|
| 100 kHz | 0 dB | 0 dB |
| 150 kHz | -29.53 dB | -47.06 dB |
| 200 kHz | -59.06 dB | -94.13 dB |
| 250 kHz | -70.07 dB | -103.5 dB |
| 300 kHz | -81.08 dB | -112.8 dB |
| 350 kHz | -102.9 dB | -129.2 dB |
| 400 kHz | -124.6 dB | -145.6 dB |
| 450 kHz | -124.1 dB | -148.8 dB |
| 500 kHz | -123.6 dB | -152 dB |
Only the five harmonic frequencies are measured. Levels below −140 dBc are close to the floor of the method.
Feedback treats the two differently, and this lesson measures why.
Feedback Leaves the Input Noise Where It Was
- The loop
RF and RG return half of the output to the − input. The loop gain is T = A/2.
- The noise test
Drive the + input with vp. The gain is 1 + RF/RG = 2. A second loop, with RG = 5.56 kΩ, has a gain of 10.
- The distortion test
Drive RG with vin. The gain is −RF/RG = −1, and both op-amp inputs stay near 0.9 V.
Read the circuit as text
- v_in voltage source: + →
n1, − →0 - R_G resistor (50 kΩ): pin 1 →
n2, pin 2 →n1 - v_p voltage source: + →
n3, − →0 - OPA-B op-amp: −in →
n2, +in →n3, out →n4 - R_F resistor (50 kΩ): pin 1 →
n4, pin 2 →n2 - C_L capacitor (2 pF): pin 1 →
n4, pin 2 →0 - out port: terminal →
n4
The lab also keeps one op-amp with its loop open, loaded by the same 50 kΩ.
Input-referred noise is the output noise divided by the gain. It is the noise a signal at the input must beat.
- open loop
- gain 2
- gain 10
- network alone
- Frequency 1 kHzopen loop = 409.6 nV/rtHz · gain 2 = 410.1 nV/rtHz · gain 10 = 409.7 nV/rtHz · network alone = 20.36 nV/rtHz
At 1 kHz flicker noise dominates. Open loop and gain 2 both read 410 nV/√Hz.
- Frequency 100 kHzopen loop = 62.02 nV/rtHz · gain 2 = 65.27 nV/rtHz · gain 10 = 62.68 nV/rtHz · network alone = 20.36 nV/rtHz
At 100 kHz the open loop reads 62.0 nV/√Hz, gain 10 reads 62.7 and gain 2 reads 65.3.
- Frequency 1 MHzopen loop = 31.47 nV/rtHz · gain 2 = 37.47 nV/rtHz · gain 10 = 32.76 nV/rtHz · network alone = 20.36 nV/rtHz
At 1 MHz the network matters more: 31.5 nV/√Hz open, 37.5 at gain 2.
Read the plot as a table
| Frequency | open loop | gain 2 | gain 10 | network alone |
|---|---|---|---|---|
| 100 Hz | 1.097 µV/rtHz | 1.097 µV/rtHz | 1.097 µV/rtHz | 20.36 nV/rtHz |
| 1 kHz | 409.6 nV/rtHz | 410.1 nV/rtHz | 409.7 nV/rtHz | 20.36 nV/rtHz |
| 10 kHz | 155 nV/rtHz | 156.3 nV/rtHz | 155.3 nV/rtHz | 20.36 nV/rtHz |
| 100 kHz | 62.02 nV/rtHz | 65.27 nV/rtHz | 62.68 nV/rtHz | 20.36 nV/rtHz |
| 1 MHz | 31.47 nV/rtHz | 37.47 nV/rtHz | 32.76 nV/rtHz | 20.36 nV/rtHz |
| 10 MHz | 26.34 nV/rtHz | 33.19 nV/rtHz | 27.84 nV/rtHz | 20.36 nV/rtHz |
Closing the loop never puts a curve below the open-loop one.
Feedback divides the gain by 1 + T. It divides the output noise by the same factor. Their ratio, the input-referred noise, stays where it was.
A loop cannot tell noise from signal. Both arrive at the same input.
The Network Adds Its Own Noise
Each resistor of the network makes thermal noise. Seen from the input, RF and RG act as one resistor RF ∥ RG in series with the − input.
| Loop | RF ∥ RG | Network alone | Input noise at 100 kHz | Above open loop |
|---|---|---|---|---|
| Open loop | none | none | 62.02 nV/√Hz | 0 |
| Gain 10 | 5 kΩ | 9.1 nV/√Hz | 62.68 nV/√Hz | 1.1 % |
| Gain 2 | 25 kΩ | 20.34 nV/√Hz | 65.27 nV/√Hz | 5.2 % |
| Gain 2, 500 kΩ resistors | 250 kΩ | 64.3 nV/√Hz | 89.38 nV/√Hz | 44 % |
Densities in V/√Hz. Noise powers add: 62.0² + 20.3² = 65.3².
The network alone gives √(4kT·RF ∥ RG). For 25 kΩ that is 20.3 nV/√Hz, and the lab reads the same from the two resistors.
Distortion Falls by 1 + T at the Harmonic
An amplifier that bends its output adds harmonics. HD2 is the second harmonic relative to the tone, in dBc, and HD3 is the third.
A harmonic is an error made inside the loop, after the gain. The loop senses it at the output and corrects it. It can correct by 1 + T, taken at the harmonic's own frequency.
- Source
- I/O
- Device
- Data
- Noise joins at the input
Op-amp noise and network noise arrive where the signal arrives. The loop cannot tell them apart.
- So input noise is not divided
The loop treats the noise as signal. The two noise powers add: 65.3 nV/√Hz at gain 2.
- Harmonics join after the gain
The output stage makes the harmonic. The network returns it, and the loop divides it by 1 + T.
Read the diagram as text
- Input signal: at the + input
- Op-amp noise: 62.0 nV/√Hz at 100 kHz
- Op-amp inputs: signal minus returned
- Network noise: 20.3 nV/√Hz for 25 kΩ
- Gain A: T = A/2
- Output stage: bends: adds harmonics
- RF and RG: return half of the output
- Input signal → Op-amp inputs: signal
- Op-amp noise → Op-amp inputs: adds
- Network noise → Op-amp inputs: adds
- Op-amp inputs → Gain A: error
- Gain A → Output stage
- Output stage → RF and RG: output
- RF and RG → Op-amp inputs: returned
An error ahead of the gain passes as signal. An error after the gain is divided by 1 + T.
| Loop | Harmonic | |1 + T| there | Rule | Run |
|---|---|---|---|---|
| Open | HD2, 200 kHz | none | none | −59.06 dBc |
| Gain −1 | HD2, 200 kHz | 35.52 dB | −94.58 dBc | −94.13 dBc |
| Gain −9 | HD2, 200 kHz | 21.61 dB | −80.66 dBc | −80.23 dBc |
| Open | HD3, 300 kHz | none | none | −81.08 dBc |
| Gain −1 | HD3, 300 kHz | 32.03 dB | −113.11 dBc | −112.82 dBc |
| Gain −9 | HD3, 300 kHz | 18.14 dB | −99.22 dBc | −99.04 dBc |
The rule is the open-loop level less |1 + T|. Every circuit has 0.2 V of output swing.
The rule and the run agree within 0.5 dB in all four cases. T is complex, so the lab adds 1 before it takes the magnitude.
The gain −9 loop has a fifth of the loop gain, and its harmonics are 14 dB higher.
The rule treats each harmonic alone. A strong HD2 can also mix with the tone inside the loop and add some HD3.
Compare at equal output swing. The output stage then works equally hard in every circuit, so it makes the same error.
Why Distortion Returns Near the Unity-Gain Frequency
- HD2
- HD3
- T at twice the tone
- Tone frequency 100 kHzHD2 = -94.13 dB · HD3 = -112.8 dB · T at twice the tone = 59.65
At 100 kHz the second harmonic sees T = 59.7. HD2 is −94.1 dBc.
- Tone frequency 2 MHzHD2 = -58.16 dB · HD3 = -69.27 dB · T at twice the tone = 2.97
At 2 MHz the second harmonic is at 4 MHz, where T is only 2.97. HD2 is −58.2 dBc.
Read the plot as a table
| Tone frequency | HD2 | HD3 | T at twice the tone |
|---|---|---|---|
| 100 kHz | -94.13 dB | -112.8 dB | 59.65 |
| 1 MHz | -71.62 dB | -89.9 dB | 5.968 |
| 2 MHz | -58.16 dB | -69.27 dB | 2.97 |
Six separate runs. The lab pins the 100 kHz and 2 MHz points.
T falls with frequency, and it reaches 1 near the unity-gain frequency fu. A harmonic above that frequency gets almost no correction.
At 2 MHz the lab prints hd2_g1_2meg = -5.816173e+01. That is about the open-loop level at 100 kHz, −59.1 dBc, with the loop closed.
A second effect adds to it. As T falls, the error voltage between the op-amp inputs grows, so the input pair starts to distort too.
Asked in interviews
Does feedback improve the signal-to-noise ratio of an amplifier?
No. Feedback divides the signal gain and the output noise by the same 1 + T, so the input-referred noise is unchanged. The feedback resistors then add their own thermal noise. Here 62.0 nV/√Hz became 65.3.
How much noise does a feedback network add?
The thermal noise of RF in parallel with RG, in series with the input: √(4kT·RF ∥ RG). It adds in power to the amplifier's own noise. For 25 kΩ it is 20.3 nV/√Hz.
By how much does feedback reduce harmonic distortion?
By about 1 + T, with T taken at the frequency of the harmonic, and compared at the same output swing. The lab's loop with T = 59.7 at 200 kHz lowers HD2 from −59.1 to −94.1 dBc.
Why does a feedback amplifier distort more at high frequency?
The loop gain falls with frequency, so harmonics near and above the unity-gain frequency are hardly corrected. The error voltage at the input also grows, and the input stage adds distortion of its own.
Try It
Part 1 reads the open-loop gain A. The loop gain is A/2 in loop A and A/10 in loop B.
drop is |1 + T| in dB. T stays complex until the magnitude is taken.
Part 2 runs noise four times. The last 1 on the second run also saves each device's share, which gives n_network.
Part 3 is one transient at 100 kHz. It skips 120 µs of start-up and keeps four periods. linearize puts them on a uniform grid.
The foreach loop multiplies the output by a cosine and a sine at each harmonic and integrates. That gives each harmonic's amplitude.
Part 4 moves the tone to 2 MHz with alter and repeats the measurement on loop A.
Predict before each change. Set RGB to 2.5k and VB to 10 mV: where do hd2_g9 and n_g10 move? Then remove the linearize line of part 3: which harmonic reading changes most?
Run the netlist to draw its plots here.
a_100k = 2.371481e+02 a_4meg = 5.941985e+00 drop2_g1 = 3.552342e+01 drop3_g1 = 3.202503e+01 drop2_g9 = 2.160703e+01 drop3_g9 = 1.814076e+01 n_open = 6.201608e-08 n_g2 = 6.527070e-08 n_g10 = 6.268064e-08 n_network = 2.034394e-08 n_g2_500k = 8.938033e-08 hd2_open = -5.905527e+01 hd3_open = -8.108169e+01 hd2_g1 = -9.412515e+01 hd3_g1 = -1.128244e+02 hd2_g9 = -8.023079e+01 hd3_g9 = -9.904270e+01 rule2_g1 = -9.457869e+01 rule3_g1 = -1.131067e+02 rule2_g9 = -8.066230e+01 rule3_g9 = -9.922245e+01 hd2_g1_2meg = -5.816173e+01 hd3_g1_2meg = -6.926799e+01Check Yourself
Common Mistakes
Wrong Close a loop, or raise its loop gain, to lower an amplifier's noise.
The loop divides signal and noise by the same factor, so the input-referred noise does not change.
Fix Lower noise where it is made: larger or better-biased input devices, and a smaller feedback network.
Wrong Choose large feedback resistors to save current, without a noise check.
RF ∥ RG sits in series with the input and adds √(4kT·R) of thermal noise.
Fix Compare the network's noise with the amplifier's own. In the lab a 500 kΩ pair raised the input noise by 44 %.
Wrong Predict a closed-loop harmonic with the loop gain at the tone.
The loop corrects the harmonic at the harmonic's frequency, where T is lower.
Fix Read T at 2f for HD2 and at 3f for HD3.
Wrong Read harmonics from a window that is not a whole number of periods.
The cut-off piece of the tone then leaks into every harmonic and can hide the real distortion.
Fix Save a whole number of periods and put them on a uniform grid before you correlate.
tran 40n 160u 120u 40n
let d=v(oo)-mean(v(oo))tran 40n 160u 120u 40n
linearize v(oo)
let d=v(oo)-mean(v(oo))Remember
- Feedback divides gain and output noise by the same 1 + T. Input-referred noise stays: 62.0 nV/√Hz open, never lower when closed.
- The network adds √(4kT·RF ∥ RG) in series with the input. Powers add: 25 kΩ took 62.0 to 65.3 nV/√Hz.
- A harmonic falls by about 1 + T at its own frequency, at equal output swing. The lab agrees within 0.5 dB.
- T falls toward fu, so distortion returns: HD2 went from −94.1 dBc at 100 kHz to −58.2 dBc at 2 MHz.
- Read harmonics over whole periods on a uniform grid, and measure the floor on a linear circuit.