Dr.

Ariel Luzzatto
ANALOG TO DIGITAL CONVERSION
Quantization noise for equal level spacing
Let a zero-average signal to be quantized into M equally-spaced levels

a/2
a

a = P / M = 2V / M

V = Ma / 2 , A = ( M - 1 )a

3a/2

Aj +a/2
Aj

Aj -a/2

Any signal v in the range A j - a / 2 v A j + a / 2 will be quantized (approximated) as v A j ,

thus v = A j + e , where e a / 2 . This process produces errors that introduces an additive
quantization noise to the original signal. However, the quantization noise is limited to a / 2 ,
regardless of the amplitude of the signal. If we assume that all the possible error values of e are
equally likely to occur, namely the quantization error is uniformly distributed, then probability
distribution of e will be p( e ) = 1 / a . Denoting by E[x] the expected value of x one gets

m = E[e ] = 1

a/ 2

e d e = 0 , s 2 = E[( e - m )2 ] = E[e 2 ] = 1

-a / 2

a/ 2

e 2 de = e

-a / 2

3 a/ 2

3a a/ 2

=a
12

The peak signal is V peak = V = Ma / 2 , and if the signal is considered sinusoidal, the rms voltage
is

Vrms = V peak / 2 = Ma / 2 2 . Denoting the full-scale signal-to-noise ratio SNR of the

2 2
2
2
quantized signal by SNR = Vrms
/ s 2 we get SNR = Vrms
/ s 2 = M 2a / 8 = 3 M 2 .
2
a / 12
If now the number of quantization levels is taken to be a power of two, namely M = 2b , where b
is the number of bits used for the quantization, then

SNR dB = 10 log( 3 / 2 ) + 10 log( 22b ) 1.76 + 6b [dB ]

It follows that each quantization bit improves the SNR by 6dB.

Dr. Ariel Luzzatto

When working with radio receivers, one would like to sample and quantize the received signals,
but would also prevent the quantization noise from further deteriorate the SNR already limited by
the NF of the receiver, so we are not worsening the existing limiting factors. However, from the
computation done before, it follows that achieving this goal would require increasing the number
of bits b to a sometimes impractically large value. So one should find other ways of increasing the
effective number of bits while leaving the physical number of bits reasonably limited. Let us
see what figures we would come up if we just increase b.
Example: Assume that a radio has a baseband bandwidth of 9 kHz, and the receiver IMR3 (thirdorder intermodulation rejection) is specified to be 70dB. The required signal to noise at minimum
operating signal (at sensitivity) must be at least SNRO = 10dB. Assume that you allow a
sensitivity degradation of 3 dB, namely, you allow the ADC to produce a quantization noise equal
to the thermal noise at the sampling point. How many bits should the ADC have approximately?
Solution
In order to achieve a total SNRO = 10dB , the ADC must provide SNRADC = 13dB at the smallest
working signal. Since the IMR3 requires us to handle two signals of equal amplitude, each 70dB
higher than the smallest, namely a signal of peak 76dB higher that the smallest, then the ADC
must provide an overall SNRADC = 13 +76 = 89 dB at the largest signal, which yields
89 1.76 + 6b b ( 89 - 1.76 ) / 6 15 bit
The required number of bits may be reduced, while keeping the same SNR O by over-sampling the
signal, namely, sampling it at a rate which is a multiple of the Nyquist rate. This will end up in
keeping the same effective number of bits, while the physical number of bits will be less. To
understand how over-sampling works here, we need to refresh the concept of sample mean.
Sample mean
Consider n statistically-independent and identically-distributed random variables {X i }, i = 1,..,n ,
each with mean m x and variance s x2 . Let us denote by Sample mean the normalized sum Y
n

Y = 1 Xi
n
i =1

Clearly for a finite value of n, Y estimates is some respect the average of each variable X i . In fact,
the mean of Y yields
n

m y = E[Y ] = 1 E[X i ] = E[X i ] = m x .

n

i =1

Using the above result we get

n

s 2y = E[( Y - m y )2 ] = E[( 1 X i - m y )2 ] = E[( 1 X i )2 +m 2y - 2m y 1 X i ]

n

i =1

i =1

i =1

= E[( 1 X i )2 ] + E[m 2y ] - 2 E[m y 1 X i ] = E[( 1 X i )2 ] + m 2y - 2m y m x

n
n
n
i =1
n

i =1

i =1
n

= E[ 12 X i X j ] - m 2y = 12 E[X i2 ] + 12 E[X i X j ] - m x2
n i =1 j =1
n i =1
n i=1 j i

Dr. Ariel Luzzatto

Recalling that E[X i2 ] = s x2 + m x2 , and that E[X i X j ] = E[X i ] E[X j ] = m x2 from independence, we
get
n

s 2y = 1 ( s x2 + m x2 ) + 12 E[X i X j ] - m x2 = 1 ( s x2 + m x2 ) + 12 m x2 1 - m x2
n

n i =1 j i

i =1 j i

s
= 1 ( s x2 + m x2 ) + 1 m x2 n( n - 1 ) - m x2 = 1 s x2 + 1 m x2 + m x2 - 1 m x2 - m x2 = x
n
n
n
n
n
n2
s2

Summarizing, we got s 2y = x , namely, as the number of samples grows, the estimate Y of m x

n
becomes closer and closer to the exact average of X .
Note that if X is the sum of a deterministic signal embedded in a zero-mean noise, then m x is just
the average value of the deterministic signal over the sampling interval, and therefore Y is an
estimate of that average.

Effect of over-sampling on ADC accuracy

The sample mean yields the following conclusion: if we over-sample a signal together with a
zero-mean noise added as before, as the number of samples increases (as the over-sampling
becomes faster) the estimator Y approaches the average of the signal over the sampling interval.
Note that, if the signal is constant over the sampling interval, then s x2 is just the variance of the
quantization noise, which will be larger if one has fewer bits, and thus more samples (a higher
over-sampling) will be required to reduce it. Over-sampling K times can be viewed as squeezing
the quantization noise in a bandwidth K times wider that the Nyquist bandwidth.
From the relation s 2y = s x2 / n we see that the quantization noise power is reduced by 6dB for
each 4x over-sampling , namely when n 4n. In our previous example, if we used a baseband
sampling rate of Nyquist x 16 = (2x9)x16 Ksamples/sec =288 Ksamples/sec, we could use only
13 bits for the ADC, and obtain the same SNRO. In fact, if the over-sampling could be high
enough, one could obtain any SNR equivalent, by using only one bit for the ADC.
The effect of over-sampling, however, does not reduce the quantization noise fast enough, and
leads to over-sampling rates that are not practical for some applications. For instance, if one
wishes to use a one-bit quantizer, and obtain 4-bit effective resolution, the required over-sampling
rate to obtain the 3 additional effective bits would be 43 = 64, which looks reasonable, but to get a
16 bit resolution would require a 415 over-sampling rate, which is not realizable.
This obstacle can be overcome by re-shaping the quantization noise so that more power will fall
out of the Nyquist bandwidth during the over-sampling process. This process is carried out with
an ADC architecture known as Sigma-Delta.

Dr. Ariel Luzzatto

First-order Sigma-delta ADC
Comparator

Integrator

f (t )

Fn +1

+
-

One-bit
output

Latch
Clk

S( t )
DAC

Referring to figure, let f(t) be an analog signal of bandwidth much smaller than the acquisition
frequency 1/T, and with | f(t)| < K .
Denoting the clock period by t, and the acquisition time by T = Nt , N >>1, let Fn be the
value of the output of the integrator at the instant t = nt at which the nth clock occurs.
Assuming that the comparator outputs a logic 1 if Fn 0 or a logic 0 if Fn < 0, the DAC
outputs +K or -K in correspondence to Q = 1 or Q = 0 respectively, we get
T

FN - F0 = [f ( t ) - S( t )] dt = e ( t )dt , e ( t ) = f ( t ) - S( t )
where e ( t ) is the quantization error at the output at any instant t , and S n is the 1-bit quantized
(bipolar) output value at the same instant, with
S( t ) = Sn c n ( t )
S n = K sign( Fn )
1 , nt t < ( n + 1 )t
0 , else

cn( t ) =
Then we may write
FN - F0 =
T

N -1( n +1 )t

n =0

[f ( t ) - Sn ]dt =

nt

= f ( t )dt - t

N -1( n +1 )t

n =0

f ( t )dt -

nt

N -1

N -1

n =0

n =0

N -1

n=0

( n +1 )t

Sn

dt

nt

Sn = f ( t )dt - T N1 Sn

For any since | f(t)| < K , we note that the sequence {Fn} is bounded, for if Fn 0 then the input
to the integrator will be negative, thus Fn+1 < Fn , and similarly, if Fn < 0 the input to the
integrator will be positive, thus Fn+1 > Fn. Therefore, no matter where it started, {Fn} will end-up
oscillating near zero with alternating sign. Since | f(t)| < K and Fn = |Fn|sign( Fn ), we get

Dr. Ariel Luzzatto

( n +1 )t

Fn +1 = Fn +

[f ( t ) - S n ]dt

nt
( n +1 )t

= Fn sign( Fn ) +

f ( t )dt - K sign( Fn ) t

nt

= ( Fn - Kt )sign( Fn ) +

( n +1 )t

f ( t )dt < Fn - Kt + Kt

nt

Which implies that, {|Fn |} decreases monotonically with n as long as Fn > 2 Kt . Indeed
Fn = 2 Kt + x 2 Fn+1 < Fn - Kt + Kt = 2 Kt + x 2 Fn+1 < Fn
Moreover, for any F0 < , there always exists a value of n 0 for which Fn 2 Kt , since
( n +1 )t

Fn +1 - Fn =

( n +1 )t

[f ( t ) - Sn ]dt

nt

( n +1 )t

f ( t ) - Sn dt

nt

[ f ( t ) + K ]dt < 2 Kt

nt

and therefore there is no possibility that Fn will bypass the gap [ - 2 Kt , 2 Kt ] . Once Fn enters
the range FN 2 Kt , then it remains bounded . Indeed
Fn 2 Kt Fn +1 < Fn - Kt + Kt < 2 Kt
Therefore, if we started with F0 = 0+ , or after several clocks, {Fn} remains bounded. In other
words for N large enough FN = O( 2 Kt ) , thus, from the equation for FN we get
T

f ( t )dt - T 1
N

N -1

Sn = O( 2Kt ) = O( 2KT / N )

n =0

which, normalizing over the acquisition time yields

1
T

f ( t )dt = 1
N

N -1

N -1

n =0

n =0

Sn + O( 2K / N ) = K N1 sign( Fn ) + O( 2 / N )

In other words, for any given acquisition interval [t0,t0+T], the average value for f(t) is K times
the average of the comparator signs, with a relative error of the order of magnitude of 2/N.
If N is taken to be a power of two, say N = 2m, all we need to do in order to obtain the A/D
conversion value, is to attach an up/down counter to the clock, with the up/down control
connected to the Q bit of the latch. Then denoting by {Q0 ,Q1 ,..,QD -1} the output word of the
counter, the output word of the D/A is taken to be its mth left-shift, which is the same as the value
obtained by dividing the result by N. The figure shows an exemplary simulation of the

Dr. Ariel Luzzatto

convergence of the sigma-delta A/D converter as described, with K = 1, N = 128, F1 = 0.67, and
f(t0) = - 0.41. The approximation error is 2.8% which is in good agreement with the order of
magnitude 2/128 = 1.6%.

Output value

Clock No.

Output value after

128 clocks: -0.3984

If we increase N about ten times, namely N = 1024, then the output value becomes - 0.40918, and
the approximation error reduces to 0.2% in close agreement with 2/1024 = 0.195% .
From the previous analysis we see that, at any clock shift k and any set {Sn+ k } , as the sampling
rate increases, the average of the bipolar values over a time T at the output tends to the limit

m S ,k lim 1
N

N -1

n =0

Sn+ k = 1
T

kt +T

f ( kt + t )dt

kt

and assuming that all the random processes we deal with are ergodic, then m S ,k = E Sn,k ,
where E[ ] denotes the expected value.
Now, form the previous results, and setting Yk = 1
N
kt +T
N -1

E f ( t )dt - T 1 Sn + k
N

n=0

kt

N -1

Sn+k we get

n =0

= E[( F
2
2
N + k - Fk ) ] = E[( TYk - T m S ,k ) ]

Assuming that {Fn } are statistically independent zero-mean random variables, we get
E[( FN + k - Fk )2 ] = E[FN2 + k ] + E[Fk2 ] = 2 E[Fn2 ]
and since Fn 2 Kt , we estimate the rms value at the output of the integrator as
E[Fn2 ] = 2( Kt )2
from which it follows that

2Kt , thus

Dr. Ariel Luzzatto

E[( FN + k - Fk )2 ] = 2 E[Fn2 ] = ( 2 Kt )2
Substituting the above results, we obtain T 2 E[( Yk - m S ,k )2 ] = ( 2 Kt )2 and therefore

s Y2 E[( Yk - m S ,k )2 ] =

( 2 Kt )2
T

= 4K
N2

where we used the fact that T = Nt . Also we note that E[Yk ] = 1

N

N -1

E[Sn+ k ] = m S ,k .

n =0

If now we average N consecutive values of Yk ,Yk +1 , ,Yk + N -1 , which, as we show later, is

equivalent to low-pass filtering the sequence {Yk } , we obtain the sample mean estimator
Zk = 1
N

N -1

Yn+ k

n=0

, E[Z k ] = E[Yk ] = E[Sn + k ] = m S ,k

By the properties of the sample mean, the variance s Z2 E[( Z k - m S ,k )2 ] satisfies

s Y2

= 4 K3
N
N
2
and s Z is the quantization noise power if f(t) can be considered constant over a time period T.
For a full-scale signal of peak value V peak = K , and f(t) can be considered constant during the

s Z2 =

acquisition time, the peak signal to noise ratio obtained using the estimator Z is
SNR =

2
V peak

3
N
4
s Z2

Now if the Nyquist rate is Bs = 1 / T , and the sampling rate is Bos = 1 / t , then
Bos T
= =N
Bs t
For instance, for 4x over-sampling we get N = 4, and then
SNR dB 10 log( 43 / 4 ) 12dB
Since the SNR is proportional to N 3 , there is a 9dB improvement each time the sampling rate is
doubled, as opposed to the case of simple over-sampling, where the SNR is proportional to N and
there is an improvement of only 3dB (6dB for a 4x sampling rate increase).
Now, we show that averaging N samples f(nt), of an arbitrary function f(t) of bandwidth
A=2p(1/2t ) where T = nt , is roughly equivalent to take the of the same function f(t) after
filtering it using a filter of bandwidth B =2p(1/2T) . To see this let us write f(nt) with the help of
its inverse Fourier transform

Dr. Ariel Luzzatto

1
N

N -1

n =0

= 1
2p
= 1
2p
1
2p

f ( nt + kT ) = 1
N

N -1

n =0

1
2p

-A

f ( w )e jw nt e jw kT dw = 1
2p

-A

f ( w )e jw kt 1
N

N -1

e jw nT dw

n =0

jw Nt
jw Nt
- 1 dw = 1
- 1 dw
f ( w )e jw kT 1 e
f ( w )e jw kT 1 e

2p
N e jwt - 1
N e jwt - 1
-A
-A

sin( wT / 2 ) jw ( kT +T / 2 )
jw kT jw Nt / 2 sin( w Nt / 2 )
dw
f ( w )e ee jwt / 2 N sin( wt / 2 ) dw 21p f ( w ) wT / 2 e
-A
-A
B

-B

sin( wT / 2 )
f ( w )e jw ( kT +T / 2 ) dw g( kT + T / 2 ) , g(
w ) = f ( w )
wT / 2

sin( wT / 2 ) 1
=
wT = p Tf 1.392 f 0.44 1 , and also
wT / 2
2
T
2T
2
we approximated , in the range w B
where we used the fact that

sin( wt / 2 ) wt / 2 , N >> 1
- jwt / 2
1
e

wt / 2 = wT / 2 N p / 2 N << 1

Therefore, the Sigma-Delta operation, can be see as if we spread the quantization noise over a
bandwidth much larger than the signal bandwidth, reshape it so most noise energy lies outside the
signal band (in the higher portion of the spectrum) and then we apply low-pass filtering at signal
bandwidth, thus leaving out more noise than would be possible by just over-sampling.

Dr. Ariel Luzzatto

OSCILLATORS

+
Vi ( w )
-

+
V0 ( w )
-

G(V )
I0 ( w )
Z( w )

Oscillator components
I) A voltage-controlled non-linear current source driven by a controlling input voltage
vi(t) = Vcos(wit), where V can be assumed constant over time as compared to the angular
frequency wi.
Then I0(w) = G(V)Vi(w), where G(V) is a non-linear function of the peak amplitude V,
monotonically decreasing in absolute value as V increases.
II) A trans-impedance resonant linear feedback network with transfer function Z(w)
V (w )
Z( w ) 0
= R /( 1 + j 2Q Dw / w0 ), w > 0
I0 ( w )
In general, the larger the value of Q, the better the oscillator noise performance.
V0(w)can be written in the form
V0 ( w ) = Z( w )I 0 ( w ) = Z( w ) G(V ) Vi ( w )
If G(V) can adjust itself so that V0(w) = Vi( w) for some value of w and V, then the output of the
feedback network can be connected to the input of the current source, and the output current I0( w)
will continue to exist without the need of an external input signal Vi(w), which, in other words,
means that oscillations occur.
In other words an oscillation state must satisfy the complex-valued non-linear equation known as
the Barkhausen criterion
Im [Z( w )G(V )] = 0
Z( w ) G(V ) = 1
Re [Z( w )G(V )] = 1

37

Dr. Ariel Luzzatto

p - topology , Colpitts & common input equivalence
Z2
G(V )
Z3

Z1

p - topology

Z2

G(V)
Z3

G(V)

Z2

Z3

Z1

Z1

Colpitts
Z2

Z1

Z2

G(V)

G(V)
Z3

Z3

Z1

Common Input

- The non-linear transfer function G(V) determines the operating point of the oscillator.
- The critical value to design for, is the peak oscillating voltage at the limiting port of the nonlinear current source.
- The oscillating voltage amplitude at the limiting port, determines the far-out phase noise of the
oscillator (the noise floor).
- The quality factor of the transfer function Z(w) including input and output loading (the loaded
Q), determines the close-in phase noise.

38

Dr. Ariel Luzzatto

Basic feedback network
In the following, we assume high-Q behavior of the resonant circuit, therefore the transfer
function of the network in the figure can be readily computed with the help of zero-pole diagram
approximation technique (left as an exercise)

I0 ( w )

V0+( w )

CV

CI

V (w )
w L
1
1
Z( w ) = 0
, QL = 0 , w02 = 1 ( 1 + 1 )
2
I0 ( w )
w
w
r
L CV CI
0
w0 CI CV r 1 + j 2Q
L

w0

At w = w0 the transfer function is real, thus the first condition of Barkhausen criterion is satisfied,
and since Z(w0) = -1/(w02CICVr), if there exists a value of V that satisfies of the non-linear
equation
G(V ) = -w02 CI CV r
the second condition of Barkhausen criterion is also satisfied, and therefore oscillations will occur
and self-stabilize at that value.
- Note that the negative sign implies that we must use a phase-inverting active device.
- Since |G(V)| is a monotonically decreasing function, whenever V is smaller than the self-limiting
value, the amplitude of the oscillations will continue to grow, which assures oscillator start-up
following the presence of any small disturbance such as thermally-generated noise.
Example: large-signal oscillations
+ Vcc

C
L

Cb

vosc ( t )

Ce

I dc

Vbb +
-

39

Dr. Ariel Luzzatto

Denoting the base-emitter oscillating voltage by vosc ( t ) = Vdc + Vosc cos( w0t ) , and assuming that
neither collector saturation or reverse junction breakdown occur, and that r represent all the
resistive losses in the circuit, and given that Vosc >> 26mV , compute approximately Vosc
Solution: The ac-equivalent circuit is shown below

L
I osc cos( w0t )

+
Vosc cos( w0t )

Ce

Cb

Recall that for a bipolar transistor

I (x)

iC iE = I dc 1 + 2 n
cos ( nw0t )

n=1 I 0 ( x )

According to the voltage and current polarity given in the figure

I (x)
V
I osc = -2 I dc 1
, x = osc
I0 ( x )
VT
At resonance
V
Ce + Cb
1
, w0 =
Z( w0 ) = osc = I osc
LCe Cb
w02 CeCb r
and therefore

I ( x)
Vosc = - 2 1
-2 I dc 1
w C C r
I 0 ( x )
0 e b
Thus, we get a non-linear equation for x
x=

I dc / VT

w02CeCb r

I (x)
I (x)
gm
=
2 1
2 1
2
I0 ( x ) w C C r I0 ( x )
0 e b

I ( x)
For large values of x we may approximate 1
1 and therefore
I0 ( x )
2 I dc
Vosc
, Vosc >> 26mV
2
w0 CeCb r

40

Dr. Ariel Luzzatto

Example: oscillator start-up
In the previous example, what is the smallest value of I dc for which oscillations will start?
Solution: At the limit where oscillations just start, the value of x will be small, thus, we may
write
I /V
I ( x ) I dc / VT

x = dc T 2 1
x , x << 1
2
I
w C C r 0 ( x ) w 2C C r
0 e b

0 e b

Therefore, the requirement for oscillation start is

I dc / VT

w02CeCb r

1 I dc VT w02CeCb r

Exercise:
Neglecting the ac loading of Rb ,Rc and Re , assuming that the transistor base-emitter capacitance
is 5pF, and assuming that the base-emitter dc voltage is about 0.6Vdc compute approximately:
1) The oscillating frequency f0
2) The peak oscillating base-emitter voltage Vosc for r = 30W
3) The smallest value of Vcc for which oscillations will start
Circuit values:
Vcc = 10V

Rb

L = 130nH
Ce = 20 pF

+ Vcc

Rc

C
C

Cb = 15 pF
Rb = Rc = 33k W
Re = 5k W

hFE = 300

Cb

vosc ( t )

Ce

Re

Answers:
1) f0 139.6 MHz
2) Vosc 186mV
3) Vcc = 3.7V

The simulation is shown below. The simulated amplitude is about 80% of the computed due to
the ac loading of the resistors that has been neglected, which reflects in effectively increasing r.

41

Dr. Ariel Luzzatto

7.3 nS

290 mV

42