Noise Overview

• The phenomenon of noise and its effect on

analog circuits.
• Noise characteristics in the frequency and time
domains.
– Thermal noise
– Shot noise (in BJT)
– Flicker noise
• Methods of representing noise in circuits.
• Noise in single-stage and differential amplifiers.
 Statistical Characteristics of Noise

Noise is a random process, which means the
value of noise cannot be predicted at any
time.

How can we incorporate noise in circuit
analysis? This is accomplished by observing
the noise for a long time and using the
measured results to construct a “statistical
model” for the noise. While the
instantaneous amplitude of noise cannot be
predicted, a statistical model provides
knowledge about some other important
properties of the noise that prove useful and
adequate in circuit analysis.
Average Power of Random Signals
Since the signal are not periodic, the measurement must be carried out over a
long time:
1 +T / 2 x 2 (t )
Pav = lim ∫−T / 2 dt
T →∞ T RL
where x(t) is a voltage quantity.

Low-power random signal High-power random signal

 Average Noise Power
To simplify calculations, we write the definition of Pav as
1 +T / 2
Pav = lim ∫−T / 2 x 2 (t )dt
T →∞ T

where Pav is expressed in V2 rather than W.

In analogy with deterministic signals, we can also define a root-mean-square
(rms) voltage for noise as Pav .
 Noise Spectrum
• Calculation of noise spectrum

Power spectral density (PSD):

The spectrum shows how much power the signal
carries at each frequency. More specifically, the
PSD, SX(f), of a noise waveform x(t) is defined as
the average power carried by x(t) in a one-hertz
bandwidth around f. SX(f) is expresses in V2/Hz.

We can apply x(t) to a bandpass filter with center
frequency f1 and 1-Hz bandwidth, square the output,
and calculate the average over a long time to obtain
SX(f1). Repeating the procedure for different center

In summary, the spectrum shows the power
frequencies, we arrive at the overall shape of SX(f).
carried in a small bandwidth at each frequency,

It is also common to take the square root of SX(f), revealing how fast the waveform is expected
expressing the result in V / Hz . to vary in the time domain.
 Noise Shaping
• White spectrum (white noise)

• Noise shaping by a transfer function

SY ( f ) = S X ( f ) H ( f )
2

• Example: Spectral shaping by telephone BW

 Spectrum Power
• Two-sided and one-sided noise spectra

Since SX(f) is an even function of f for real x(t), the total power carried by
x(t) in the frequency range [f1 f2] is equal to
− f1 + f2 + f2
Pf1 , f 2 = ∫− f S X ( f )df + ∫+ f S X ( f )df = 2 ∫+ f S X ( f )df
2 1 1

• Folded white spectrum

 Amplitude Distribution
• Probability density function (PDF)

Probability density function (PDF): By observing the noise waveform for a long
time, we can construct a “distribution” of the amplitude, indicating how often each
value occurs. The distribution of x(t) is defined as
pdf(x)dx = probability of x < X < x +dx,
where X is the measured value of x(t) at some point in time.
− (x − m )
2
1

Gaussian PDF is defined as pdf ( x) = exp ,
σ 2π 2σ 2

where σ and m are the standard deviation and mean of the distribution,
respectively.
 Correlated and Uncorrelated Sources
We add two noise waveforms and take average of the resulting power:
1
[ ]
+T / 2
Pav = lim
T →∞ T ∫−T / 2 x 1 (t ) + x 2 (t ) 2
dt

1 +T / 2 1 +T / 2 1 +T / 2
= lim ∫ x12 (t )dt + lim ∫ x 22 (t )dt + lim ∫ 2 x1 (t )x 2 (t )dt
T →∞ T −T / 2 T →∞ T −T / 2 T →∞ T −T / 2

1 +T / 2
= Pav 1 + Pav 2 + lim ∫ 2 x1 (t )x 2 (t )dt
T →∞ T −T / 2

correlation

Uncorrelated noise Correlated noise

 Resistor Thermal Noise (1/3)
• Thermal noise of a resistor

The thermal noise of a resistor R can be modeled by a series voltage source, with the
one-sided spectral density
Vn2 = Sv(f) = 4kTR, f ≥ 0,
where k = 1.38×10−23 J/K is the Boltzmann constant and Sv(f) is expressed in V2/Hz.
 Resistor Thermal Noise (2/3)
• Example: low-pass filter

Vout
We compute the transfer function from VR to Vout: (s ) = 1
VR RCs + 1
2
V 1
From the theorem, we have S out ( f ) = S R ( f ) out ( f ) = 4kTR 2 2 2 2 .
VR 4π R C f + 1

The total noise power at the output:

∞ 4kTR 2kT −1 u = ∞ kT
Pn ,out = ∫0 df = tan u = (V2)
4π 2 R 2C 2 f 2 + 1 πC u=0 C
 Resistor Thermal Noise (3/3)
• Representation of resistor thermal noise by a current
source
Vn2 4kT
I n2 = 2= (A2/Hz)
R R

• Example

Since the two noise sources are uncorrelated, we add the powers:
1 1 
I n2,tot = I n21 + I n22 = 4kT  + 
 R1 R2 

The equivalent noise voltage is given by
Vn2,tot = I n2,tot (R1 R2 ) = 4kT (R1 R2 )
2
 MOSFET Thermal Noise
• Thermal noise of a MOSFET
For the MOS devices operating in saturation, the channel noise
can be modeled by a current source connected between the drain
and source terminals with a spectral density: I n2 = 4kTγg m ,
where γ is equal to 2/3 for long-channel transistors and may be a
large value (2) for submicron MOSFETs.

• Output noise voltage

Vn2 = I n2 ro2 = 4kTγg m ro2

 Flicker Noise (1/2)
• Dangling bonds at the oxide-silicon interface

As charge carriers move at the interface, some are
randomly trapped and later released by such energy
states, introducing flicker noise in the drain current.

The flicker noise is modeled as a voltage source in
series with the gate and roughly given by
K 1
Vn2 = ⋅ , where K is a process-dependent
CoxWL f
constant on the order of 10−25 V2F.

The flicker noise is also called 1/f noise, and it does
not depend on the bias current or the temperature.

It is believed that PMOS devices exhibit less 1/f
noise than NMOS transistors because the former
carry the holes in a “buried channel”, i.e., at some
distance from the oxide-silicon interface.
 Flicker Noise (2/2)
• Concept of flicker noise corner frequency
For 1/f noise, the drain noise current per unit bandwidth
K 1
is I n2,1 / f = Vn2,1 / f ⋅ g m2 = ⋅ ⋅ g m2 .
CoxWL f
For the thermal noise, the drain noise current per unit
2 
bandwidth is I n2,th = 4kT  g m  .
3 
Thus, the 1/f noise corner, fC, of the output current is
K 3
determined by f C = gm , which depends on
CoxWL 8kT
device dimensions and bias current.

  

   .   /)
/

  
MOS Noise

  

   .   /)
/

  
 Noise in Circuits (1/2)
• How to quantify the effect of noise?
The natural approach would be to set the
input to zero and calculate the total noise
at the output due to various sources of
noise in the circuit.

( )
Vn2,out = I n2,th + I n2,1 / f + I n2, RD RD2
 2 K 1 4kT  2
=  4kT g m + ⋅ ⋅ g m2 +  RD
 3 CoxWL f RD 
• Example
 Noise in Circuits (2/2)
• Determination of input-referred noise voltage

If the voltage gain is Av, then we have Vn2,out = Av2 Vn2,in , that is, the input-referred
noise voltage is given by the output noise voltage divided by the gain.

The input-referred noise indicates how much the input signal is corrupted by the
circuit’s noise, i.e., how small an input the circuit can detect with acceptable SNR.

The input-referred noise is a fictitious quantity in that it cannot be measured at the
input of the circuit.
 Common-Source Stage (1/3)
 2 K 1 4kT  2
Vn2,out =  4kT g m + ⋅ ⋅ g m2 +  RD
• CS stage  3 C oxWL f R D 

Vn2,out  2 1  K 1
V 2
n ,in = = 4kT  + 2  +
Av2  3 g m g m RD  CoxWL f

How can we reduce the input-referred noise

voltage? It implies that the transconductance
Voltage Amplification Current Generation of M1 must be maximized.

The transconductance must be maximized if
the transistor is to amplify a voltage signal
applied to its gate [Fig.(a)] whereas it must
be minimized if the transistor operates as a
• Discussion current source [Fig.(b)].
 Common-Source Stage (2/3)
• Example: calculate the input-referred thermal noise voltage of the amplifier

2 2 
Thermal noise: Vn2,out = 4kT  g m1 + g m 2 (ro1 ro 2 )
2

3 3 
Voltage gain: |Av| = gm1(ro1||ro2)
The total noise voltage referred to the gate of M1 is
2 2  1  2 2g 
Vn2,in = 4kT  g m1 + g m 2  2 = 4kT  + m2 2 
3 3  g m1  3 g m1 3 g m1 
It reveals the dependence of Vn2,in upon gm1 and gm2, confirming that gm2
must be minimized because M2 serves as a current source.
 Common-Source Stage (3/3)
• How to design a CS stage for low-noise
operation?
Vn2,out  2 1  K 1
V 2
n ,in = = 4kT  + 2  +
Av2  3 g m g m RD  CoxWL f

For thermal noise, we must maximize gm by increasing the drain current or the device
width. A higher ID translates to greater power dissipation and limited output voltage
swings while a wider device leads to larger input and output capacitance. We can also
increase RD, but at the cost of limiting the voltage headroom and lowering the speed.

For 1/f noise, the primary approach is to increase the area of the transistor. If WL is
increased while W/L remains constant, then the device gm and its thermal noise do not
change but the device capacitances increase.

These observations point to the trade-offs between noise, power dissipation,
voltage headroom, and speed.
 Differential Pairs (1/2)

• Differential pair circuit

• ISS contribute common-mode
noise only
• Circuit including input-referred noise source

For low-frequency operation,

the magnitude of I n2,in is
typical negligible.
 Differential Pairs (2/2)
• Calculation of input-referred noise of a differential pair
With the inputs shorted together, we have
I I
Vn ,out M1 = n1 RD1 + n1 RD 2
2 2
Vn2,out M1 = I n21 RD2 (if RD1 = RD2 = RD)

Similarly, Vn2,out M2 = I n22 RD2 .

⇒ Vn2,out M1 , M 2 ( )
= I n21 + I n22 RD2
Taking into account the noise of RD1 and RD2, the total output noise:
Vn2,out M1 , M 2 ( )
= I n21 + I n22 RD2 + 2(4kTRD )
2 
= 8kT  g m RD2 + RD 
3 
And, |Av| = gmRD, we have
Vn2,out  2 1 
Vn2,in = 2
= 8kT  + 2  ≈ 2Vn2,in CS stage
Av  3 g m g m RD 
 Noise Bandwidth
• Output noise spectrum of a circuit

The total output noise:

∞
Vn2,out ,tot = ∫0 Vn2,out df
∞
and V02 ⋅ Bn = ∫0 Vn2,out df
Noise bandwidth (Bn): Bn allows a fair comparison of circuits that exhibit the same
low-frequency noise, V02, but different high-frequency transfer
functions.