0.

1 Feedback
0.1.1 Block algebra
In order to introduce feedback more easily we start by introducing the algebra of blocks.
It is a very useful and elementary way of representing LTI transfer functions.

LTI transfer functions

8
Start by considering an input xi and an output xo , related by a LTI transfer function k.
We can represent the relation:
xo = kxi
through the following block representation:

xi xo
k
We have that:

• boxes represent LTI transfer functions

• entering arrows represent inputs for the LTI transfer function

• exiting arrows represent outputs for the LTI transfer function

summing junctions

We can represent sums and subtractions like:

7
through summing junctions:
xo = x1 − x2 + x3

x1
x2 xo

x3

In this convention:

• circles represent sums

• entering arrows with a + sign on them represent terms that are added

1
 • entering arrows with a - sign represent terms that are subtracted

• the only leaving arrow represents the result of the sum/subtraction

Note: If no + or - signs are provided they represent sums

sampling nodes

8
When the same variable is used as an input for more than one block, so for more than one
operation, we use sampling nodes.
The sampling node:

xo

xi xo

Such sampling node replicates its input xi on every output xo ( xo = xi ).

In this convention:

• the only entering line without an arrow represents the data that is being sampled

• the solid dot identifies the sampling operation

• exiting arrows represent all the outputs that replicate the input

0.1.2 Feedback
Consider a basic amplifier. Such amplifier can be descibed through a transfer function a. If
we denote with xi the input quantity of the amplifier and with xo the output quantity of the
amplifier we can write the following equation:

8
xo = a · xi

which is also valid for dynamic LTI systems in the Laplace domain.
Such equation admits the block representation:

xi xo
a
In such kind of amplifier the output quantity xo is uniquely determined by the input quantity
xi through the transfer function a.
These amplifiers are very sensitive on variations of the internal parameters, some of which can
variate a lot.

In order to solve such problem, feedback amplifiers are introduced.

Feedback consists in having the output being brought back through another transfer function
and summed to the input (feedback transfer function).
This way, the input of the amplifier is ”corrected” according to the actual value of the output,

2
so that a more stable output may be provided by the whole device (including both the amplifier

32
and the feedback network).
The equations that describe a feedback amplifier can be represented through the block repre-
sentation:

xfb xo
h
xi xε xo xo
k
In this basic architecture, the output is

• picked by a sampling node

• ”weighted” by a transfer function h

• combined to the input through a summing junction

in order to create a backward relation from the output of the amplifier to its input.
Notice that we have two ways of combining the feedback signal to the input signal:

• positive feedback: the feedback signal is summed to the input signal

• negative feedback: the feedback signal is subtracted from the input signal

Notice that a negative feedback doesn’t necessary have to involve a negative sign on the sum-
ming junction, it is enough for the h block to have a negative transfer function for the output
signal to get subtracted from the input one!

The operation of combining the feedback signal to the input signal is also called mixing (done
through a summing junction).

The whole feedback amplifier, consisting of both the amplifier k and the feedback network h
can be described by a transfer function.
We can derive the expression for the transfer function of a feedback loop in both cases of positive
and negative feedback.

3
negative feedback

32
In negative feedback the output signal is subtracted from the input signal.

The structure can be represented in the following way:

xfb xo
h
xi xε xo xo
k
We can derive the transfer function between the input xi and the output xo .
Let’s start from the relation given by the transfer function k:
xo = k xε
Substitute the relations given by the summing junction and the feedback network:
xo = k (xi − xf b )
xo = k (xi − h xo )
xo = k xi − k h xo
xo + k h xo = k xi
(1 + k h) xo = k xi
k
xo = xi
1+kh
| {z }
Af b

So, we have that the transfer function between the input of the feedback system and its output
is:
k
Af b− =
1+kh
positive feedback

32
In positive feedback the output signal is added to the input signal.

The structure can be represented in the following way:

xfb xo
h
xi xε xo xo
k
4
We can derive the transfer function between the input xi and the output xo .
Let’s start from the relation given by the transfer function k:
xo = k xε
Substitute the relations given by the summing junction and the feedback network:
xo = k (xi + xf b )
xo = k (xi + h xo )
xo = k xi + k h xo
xo − k h xo = k xi
(1 − k h) xo = k xi
k
xo = xi
1−kh
| {z }
Af b

So, we have that the transfer function between the input of the feedback system and its output
is:
k
Af b+ =
1−kh
positive VS negative feedback

32
Consider the positive feedback system case.
Because of the sum at the left side of the diagram:

xfb xo
h
xi xε xo xo
k
we have that, if the initial conditions for y are nonzero, at each cycle (we can consider this
feedback loop as continuously repeated infinitely many times) the signal xε always increases,
making the output of the system unbounded.
This is because the system is not stable, which means that transients do not die out in time.
Because of this reason such architecture does not have any application in amplifiers.

32
Consider the negative feedback system case.
Because of the difference at the left side of the diagram:

xfb xo
h
xi xε xo xo
k
5
we have that, if the initial conditions for y are nonzero, at each cycle (we can consider this
feedback loop as continuously repeated infinitely many times) the signal xε decreaseas, until it
stabilizes to a fixed (steady state) value.
So, it is much easier to get stable structures when having negative feedback!

Conclusion:
Therefore, the only feedback structure actually employed in feedback amplifiers is negative feed-
back

parameters of the feedback loop

32
Consider now the negative feedback case, which is the only practically useful one in this course:

xfb xo
h
xi xε xo xo
k
We can define a characteristic parameter of the feedback loop, which is the loop gain:

T =hk
xo
We can then define k = the open-loop gain, since it is the gain the feedback loop has when
xε
h = 0, so when the feedback network is not working.
xo
We can then define Af b = the closed-loop gain, since it is the gain the feedback loop has
xi
when the feedback network is present and working.

We can then rewrite the closed loop gain as:

k
Af b =
1+hk
1 kh
=
h1+hk
1 T
=
h1+T
1 1
=
h 1
1+
T
We may also define the additional parameter appearing in these formulae as:
1
A∞ = lim =
k→∞ h
The result is that the previous formula reduces to:
T
Af b = A∞
1+T

6
Note: A∞ is not the same parameter that will appear in Rosenstark’s method!

Note: We will sometimes represent negative feedback with a plus sign in the summing junction.
In such cases the change of sign is already performed by the feedback block.
When this happens we still define the loop gain as positive! So, we will define T = −h k, so
that the resulting formula found from the block diagram is the same ( the minus sign in the
denominator of Af b given by the positive sign on the summing junction is canceled by the one
in the definition of T .
The final result must always be:
k
Af b =
1+hk

0.1.3 Advantages of feedback

Advantage 1
The relative sensitivity to fluctuations of the open-loop gain k is reduced:
  
Af b  ∂Af b Af b 
Sk =   
∂k k 
    
∂ k k 1 
=  ·
∂ 1+hk 1+hk k 
   
 (1 + h k) − h k 1 
= 
2

(1 + h k) 1+hk 
   
 1 1 
=  
(1 + h k)2 1+hk 
 
 1 
=  
(1 + h k) 
1
=
1+hk
1
=
1+T
We can see that, since we consider h ≥ 0, we have:
A 1
Sk f b = ≤1
1+hk
Since 1 is the sensitivity when h = 0 it is the sensitivity when feedback is not present.
Therefore, we can see that the sensitivity on the open loop gain k is reduced when feedback is
more intense (for higher values of h).

Since the sensitivity is given by:

A 1
Sk f b =
1+T
we call its inverse (1 + T ) the desensitivity factor.

We can notice that, for every feedback h > 0 (nonzero) for very high gains k the sensitivity
function goes to zero:
A
lim S f b = 0
a→∞ k
b>0
So, for high gains of the amplifier k the feedback becomes not sensitive to fluctuations of the
same gain k

7
Advantage 2

The input signal swing is increased by a factor (1 + T ).

Consider the amplifier. Its gain k is valid just in linearity conditions (linearity region for a BJT,
saturation for a MOS), but this condition will not be satisfied for every possible value of the
input.
The amplifier is able to provide just a fixed range of output values while still working as a linear
amplifier.
Out of such range the amplifier saturates and the output quantity remains the same for every
input that requires an output out of the range!
We can see this from the output-input characteristic of the amplifier:

xo
xo,max

xε,min
xε,max
xi
xo,min

So, just a certain range of inputs is allowed in order to have a linear amplifier!
We define such range of inputs as the input signal swing.

If we use our amplifier k in a feedback loop the gain between the input signal xi and the output
signal xo becomes:
k
Af b =
1+hk
While the gain of the basic amplifying block is k.

Given the saturation limits xo,max and xo,min we can proceed to determine the input swing for
the signal.
For simplicity consider the simplified setting of symmetric saturation limits, that is:
xo,max = −xo,min
We can then determine the maximum input swing of the block k by applying its characteristic
equation. We can work with absolute values if we consider symmetric saturation limits:
| xε k |≤ xo,max
xo,max
| xε |≤
k

8
So, the input swing of the amplifying block k alone is:
xo,max xo,max
− ≤ xε ≤
k k
We can now repeat the same computation with the same output (subjected to the same con-
straints) but with as an input the input of the feedback amplifier.
We have that:
k
| xi |≤ xo,max
1+hk
xo,max
| xi |≤ (1 + h k)
k
So, the input swing of the feedback amplifier is:
xo,max xo,max
− (1 + h k) ≤ xi ≤ (1 + h k)
k k
We can notice that the input swing of the feedback amplifier is (1 + h k) times larger than the
one of the block k!
Graphically, we will have:

xo
xo,max

xi,min
xi,max
xi
xo,min

The result is that, despite the gain decreasing, the input swing is improved (with the same
constraints on the output).

Advantage 3
We get a bandwidth extension for the amplifier.

Let’s consider a dominant pole approximation for the transfer function k of the amplifier:
k0
k(jf ) =
1 + jf /fk
With:
• k0 : low-frequency gain

• fk : dominant pole of k

9
Consider now a simple constant as a feedback transfer function h.
The gain of the feedback system will be:
k(jf )
Af b (jf ) =
1 + k(jf )h
k0
1 + jf /fk
=
k0
1+ h
1 + jf /fk
k0
1 + jf /fk
=
k0 h + 1 + jf /fk
1 + jf /fk
k0
=
k0 h + 1 + jf /fk
k
=  0 
jf
(k0 h + 1) 1 +
fk (1 + k0 h)
k0 1
= · 
(k0 h + 1) jf
1+
fk (1 + k0 h)
We can see how the first term is the low-frequency gain of the feedback loop, so we define:
k0
Af b,0 =
1 + k0 h
The second term, on the other hand, is nothing but a simple pole.
We can define the new pole as:
fk0 = fk (1 + k0 h)
After defining the two quantities, the gain of the feedback system becomes:
1
Af b (jf ) = Af b,0 ·
1 + jf /fk0
Notice that the pole of Af b (jf ) is at a frequency (1 + k0 h) times higher than the one of k(jf ).

⇒ So, we get a bandwidth extension of (1 + k0 h) = T times.

Despite this, the cutoff frequency of the feedback system, also known as gain-bandwidth product,
is still the same as the one of the amplifier k(jf ).
To show this, we can write down the expression for the cutoff frequency of the feedback system:
k0
fT,f b = Af b,0 fk0 = fk (1 + k0 h) = k0 fk = fT
1 + k0 h
With:
• fT,f b : gain-bandwidth product (cutoff frequency) of the feedback system
• fT : cutoff frequency of the amplifier k(jf )
So, the gain reduction gives an extension of the bandwidth of (1 + T ).

10
The following magnitude Bode diagram shows how feedback reduces the low-frequency gain but,
at the same time, improves the bandwidth and how the cutoff frequency fT does not change
from one to the other:

|x|dB

k0
k(jf)
Afb0

Afb(jf)
0dB
fk fk ' fT
log10(f/f0)

0.1.4 Sampling and mixing

Creating a feedback loop involves sampling the output quantity, transforming it with a suitable
feedback transfer function and adding it to the input quantity.
But until now we have always considered two possible choices for the input type:

• voltage input

• current input

and two possible choices for the output type:

• voltage output

• current output

So, we can have four combinations for feedback:

sampling mixing
voltage voltage
current current
voltage current
current voltage

The simplest transfer function for the feedback network is a constant.

We can easily model such transfer function with controlled sources.

We will now see which are the ideal characteristics for the sampling and mixing of both types.

11
Ideal voltage sampling

Voltage sampling, because of its nature, has to be done by a parallel connection.

We must be able to sample the voltage without affecting the current absorbed from the amplifier,
since that would alter the voltage because of the presence of the output resistance.
The only ideal component that samples the voltage without absorbing any current is the ideal
open circuit.
Therefore, voltage sampling will be done by the feedback network through an open circuit:

io

Ro
vo RL
Rmii

vsmp

By a KVL at the output side we obtain:

vsmp = vo

So the sampled voltage is exactly the output voltage.

The relation between the output current and the output voltage remains the same as the one
we had with the load only, as we can see by deriving it:

vo = −RL io

So this sampling does not influence the operating point of the output port of the amplifier!

This sampling has multiple names:

• voltage sampling: because it samples a voltage

• parallel sampling: because it is obtained through a parallel with the load and with the
output port of the amplifier

• shunt sampling: because shunt is a synonim for parallel

12
Ideal current sampling

Current sampling, because of its nature, has to be done by a series connection.

We must be able to sample the current without affecting the voltage required from the amplifier,
since that would alter the current because of the presence of the output resistance.
The only ideal component that samples the voltage without absorbing any current is the ideal
short circuit.
Therefore, current sampling will be done by the feedback network through a short circuit:

io

Go vo RL
Aiii

ismp

By a KCL at the output node we obtain:

ismp = io

So the sampled current is exactly the output current.

The relation between the output current and the output voltage remains the same as the one
we had with the load only, as we can see by deriving it:

vo = −RL ismp = −RL io

So this sampling does not influence the operating point of the output port of the amplifier!

This sampling has multiple names:

• current sampling: because it samples a current

• series sampling: because it is obtained through a series with the load

13
Ideal voltage mixing

Voltage mixing, because of its nature, has to be done through a series connection.
We must be able to mix the voltage without imposing any additional constraint between the
input current and the input voltage, in order to obtain a pure voltage mixing.
The only ideal component that gives a voltage without setting any constraint on the current
flowing through it is a voltage source.
Therefore, voltage mixing will be done through a series between the real generator and a voltage
source (one that is controlled by the sampled quantity).
We still need to choose the orientation of the controlled voltage source appropriately in order
to obtain positive/negative feedback.
The following architecture implements negative voltage mixing:

ii

RGEN
vS vi Gi

vfb

By a KVL at the input side we obtain:

vi = vs − Rgen ii − vf b

Therefore, we can see how the feedback voltage is subtracted from the one of the real voltage
source, which confirms that it is a negative feedback.

This mixing has multiple names:

• voltage mixing: because it mixes a voltage

• series mixing: because it is obtained through a series with the generator

14
Ideal current mixing

Current mixing, because of its nature, has to be done through a parallel connection.
We must be able to mix the current without imposing any additional constraint between the
input voltage and the input current, in order to obtain a pure current mixing.
The only ideal component that gives a current flowing through it without setting any constraint
on the voltage across its terminals is a current source.
Therefore, current mixing will be done through a parallel between the real generator and a
current source (one that is controlled by the sampled quantity).
We still need to choose the orientation of the controlled current source appropriately in order
to obtain positive/negative feedback.
The following architecture implements negative current mixing:

ii

iS RGEN vi Ri

ifb

By a KCL at the input side we obtain:

vi
ii = is − − if b
Rgen

Therefore, we can see how the feedback current is subtracted from the one of the real current
source, which confirms that it is a negative feedback.

This mixing has multiple names:

• current mixing: because it mixes a current

• parallel mixing: because it is obtained through a parallel with the generator and with the
input port of the amplifier

• shunt mixing: because shunt is a synonim of parallel

15
Real voltage sampling

In general, since ideal controlled sources (whose inputs behave as perfect short circuits or open
circuits and whose output resistance is zero or infinity) do not exist, we will always have to use
other components to implement the feedback.

The easiest choice is a resistor, but any other choice is acceptable. We will study it by consider-

9
ing the case of a general electrical network, which will present an equivalent impedance towards
the ground terminal Zf b .

Voltage sampling is implemented through an architecture like the following:

out
(C)

vsmp
ZFB
in (B)

(E)
FB

RE

We can notice by writing a KVL at the output side that:

vsmp = vout
However, the feedback network will not behave as a perfect open circuit, which means that it
will affect the output current.
More in general, we can identify voltage sampling through the following criterion:

identifying VOLTAGE SAMPLING

When the feedback network is connected to the output terminal we have voltage sampling.

Real current sampling

In general, we have to avoid adding the feedback network in series to the load, since then a
two-terminal network wouldn’t be enough (we would need at least one more terminal for the
mixing, which would make it more complicated).
Therefore, we will use a ”trick” in order to obtain a current flowing through the feedback net-
work that is proportional to the output current.

16
 9
Current sampling is implemented through an architecture like the following:

io
out
(C) iC

in (B)

ismp
(E)
iE

FB ZFB RE

We can say that the collector current of the bipolar transistor splits according to a current
divider between the load and whatever is connected to the collector.
Therefore, we will have a linear relation:

iC = α iout α constant

Moreover, we have the following relation between the collector current and the emitter current:
β+1
iE = iC = γiC γ constant
β
Which can be approximated to iE ≈ iC for high values of β.
We then have a current divider at the emitter between the feedback network and whatever is
connected to the emitter.
so we have the linear relation:

ismp = δ iE δ constant

Therefore, putting all pieces together, we obtain:

ismp = δ γ α iout = µ iout µ constant

So, the sampled current is proportional to the output current.

Even though it is not equal it is not a problem since this can be corrected by taking it into
account when choosing the feedback gain.
Moreover, this kind of sampling is again not ideal since it affects the output voltage.

More in general, we can identify current sampling through the following criterion:

identifying CURRENT SAMPLING

When the feedback network is connected to any terminal of the output stage but the
output terminal we have current sampling.

17
Real voltage mixing

Since in general our feedback networks will be two-terminal elements we will have one terminal
used for the sampling and one used for the mixing.
So, in general, we will not be able to implement a series connection.

9
We can anyway exploit another ”trick” to mix the voltage.

Voltage mixing is implemented through an architecture like the following:

out
(C)

in (B)

(E)
ZFB
vBE

RE vmx
FB

We can notice by writing a KVL at the input side that:

vin = vBE + vmx

We can observe that the controlling quantity of the amplifying stage is vBE and that we can
rewrite the KVL as:
vBE = vin − vmx
Therefore, the voltage vmx that we are mixing is being subtracted from the input voltage.
According to the actual sign of vmx and the sign of the output quantity we will either have
negative or positive feedback.

Remark: This means that if a positive voltage is subtracted from the input quantity to give
the controlling quantity, but the amplifier is inverting, the output signal is actually negative, so
there is one more sign inversion in the feedback block, but this results in a quantity proportional
to the output quantity being summed to the input voltage to give the controlling quantity! So
that would actually be a positive feedback, even though a positive quantity is subtracted from
the input quantity.

The feedback network will not behave as a perfect voltage source, so it will introduce additional
constraints on the currents at the input side.
.
More in general, we can identify voltage mixing through the following criterion:

18
 identifying VOLTAGE MIXING
When the feedback network is connected to any terminal of the input stage but the
input terminal we have voltage mixing.

9
Real current mixing

We can implement current mixing by obtaining a KCL at the input node

Current mixing is implemented through an architecture like the following:

out
(C)

iin iB
in (B)
imx
ZFB (E)

RE
FB

We can notice by writing a KCL at the input side that:

iB = iin − imx

Therefore, the current imx that we are mixing is being subtracted from the input current.
According to the actual sign of imx and the sign of the output quantity we will either have
negative or positive feedback.

Remark: This means that if a positive current is subtracted from the input quantity to give
the controlling quantity, but the amplifier is inverting, the output signal is actually negative, so
there is one more sign inversion in the feedback block, but this results in a quantity proportional
to the output quantity being summed to the input current to give the controlling quantity! So
that would actually be a positive feedback, even though a positive quantity is subtracted from
the input quantity.

The feedback network will not behave as a perfect current source, so it will introduce additional
constraints on the currents at the input side.

More in general, we can identify voltage mixing through the following criterion:

19
identifying CURRENT MIXING
When the feedback network is connected to the input terminal we have current mixing.

20