8.

2-1
8.2 Start With CPFSK

8.2.1 CPFSK Signals

• Three pictures of the same CPFSK signal:

instantaneous frequency instantaneous phase

trajectory
Trajectory is a circle with radius A = P .

CPFSK moves at a steady speed around

the circle, with reversals.

The starting phase of each symbol is determined by past bits. This

memory can improve detection by giving additional hints about past
symbol values.

8.2-1
 8.2-2
• Relation between bandpass signal and complex baseband:
⎡ ⎤
s (t ) = 2 Re ⎢ Ae jφ(t ) e j 2 πfc t ⎥ = 2 A cos ( 2πf c t + φ(t ) )
⎢⎣ s (t ) ⎥⎦
o Instantaneous values:
¾ total phase: 2πf c t + φ(t )

¾ frequency (rad/s): 2πf c + φ(t ) = 2πf c ± 2πf d

¾ frequency (Hz): f c ± f d

o To shift frequency up or shift frequency down:

phase increases linearly s (t ) rotates in positive direction,
constant speed

or phase decreases linearly or s (t ) rotates in negative direction,

constant speed

8.2-2
 8.2-3
• PSK carries information in phase, FSK in the direction of travel:
PSK FSK

• The product f d T has a strong effect on power spectrum:

o Generally, a greater max frequency deviation f d produces a wider

spectrum.

o Multiples of 1/2 create discrete spectral components (tones). Consider

1
the value f d T = cycle :
2
One half-cycle (half-circle) forward or
backward in each bit. Start on the same
phase, end on the same phase. No phase
memory.

8.2-3
 8.2-4
1
o Continuing with the f d T = cycle example, suppose the carrier is
2
f c = 1 T (for illustration only, f c is not relevant to presence of tones).
1 3
Then f c − f d = and f c + f d = , and the bandpass signal is
2T 2T

A tuned circuit will ring at f c − f d and another one will ring at

f c + f d , hence there are tones in the signal.

o Ringing occurs if the total phase 2πf c t + φ(t ) (the cosine argument) is
the same at the end of a symbol (i.e., the start of the next symbol), no
matter which data value was sent:
2π ( f c + f d ) T = 2π ( f c − f d ) T + 2πn , n an integer

n
fdT =
2
fd n 1
So we get tones if f d T = = is an integer multiple of .
Rs 2 2
This extends to multilevel (i.e., non-binary) FSK, too.

8.2-4
 8.2-5
o Why we don’t like tones:

¾ They mix in nonlinearities to produce unexpected and

unwelcome images.

¾ They consume power without carrying information.

• Formalize all these observations as expressions defining CPFSK.

o Use I n = 2m − 1 − M , for m = 1,…, M as the data at symbol time n.

Odd integers from − M + 1 to M − 1 (e.g., -5, -3, -1, +1, +3, +5)

o Alternative possibilities for the symbol (if M = 2 ):

frequency phase

The modulation index h is the accumulated phase difference (in cycles)

between the alternatives over one symbol. It is dimensionless. Integer
h causes tones.

8.2-5
 8.2-6
The modulation index h is related to the spacing Δf of CPFSK signal
frequencies and the symbol rate R :

o Waveforms: Instantaneous frequency is a rectangular pulse of height

f d I , but we’d like to use h ( f d = h 2T ). So normalize, by defining a
frequency pulse g (t ) with area 1/2.

Then instantaneous frequency

is f (t ) = h I g (t ) = 2 f d T I g (t ) .
Larger h ⇔ greater frequency swing.

Extending this to a sequence of pulses,

f (t ) = h∑ I n g (t − nT ) .
n

8.2-6
 8.2-7
t
• Next, the instantaneous phase φ(t ) = 2π ∫ f (α) d α (radians) :
−∞

φ(t ) = 2π h ∑ I k q (t − kT ) =4π f d T ∑ I k q (t − kT )
k k

t
where the phase pulse q (t ) = ∫ g (α) d α :
−∞

The saturation value of 1/2 means that the phase shift caused by I n has a
continuing effect in the future; i.e., the modulation has memory.

• The usual CPFSK defining equation writes phase in interval n as

n
φ(t , I ) = 2πh ∑ I k q(t − kT ) + φo , nT ≤ t ≤ (n + 1)T
k =−∞

where φo is the phase at the beginning of time (but is zero if the receiver’s
phase tracking works) and I is the symbol sequence .
But all q (t − kT ) up to k = n − 1 have already saturated at 1/2, so

⎛ t − nT ⎞
φ(t , I ) = θn + 2π h ⎜ ⎟ In in nT ≤ t ≤ (n + 1)T
⎝ 2T ⎠
n −1
θn = π h ∑ Ik
k =−∞

8.2-7
 8.2-8

• Examples of phase trajectories:

↑ Above: binary ( M = 2 ) CPFSK

Right: quaternary ( M = 4 ) →
CPFSK

Note phase states.

Looks like a tree, but…

8.2-8
 8.2-9

• …the phases are modulo 2π , so the signals wrap around a cylinder in time.

• If h = 2 f d T is rational, then phase state values recur, and we have a finite

set of phase state values.

If h = m p (and m, p are coprime), then

o for even m, we get p different phase states

⎧ πm πm πm ⎫
θn ∈ ⎨0, ,2 ,…,( p − 1) ⎬
⎩ p p p ⎭

o for odd m, we get 2 p different phase states, alternating even and odd

⎧ πm πm πm⎫
θn ∈ ⎨0, ,2 ,…,(2 p − 1) ⎬
⎩ p p p ⎭
8.2-9
 8.2-10
• Finally, note that CPFSK is not a linear modulation.

The phase is linear in the transmitted symbols

φ(t ) = 2π h ∑ I n q(t − nT )
n

but the actual signal is

s (t , I ) = P e jφ(t , I )

8.2-10
 8.2-11
8.2.2 CPFSK Signal Space, States and Trellis

• A set of basis functions for CPFSK:

o In interval n, the complex lowpass signal is

⎡ ⎛ t − nT ⎞ ⎤
s (t ) = P e jθn exp ⎢ j 2π h I n ⎜ ⎟ ⎥ , nT ≤ t < (n + 1)T
⎣ ⎝ 2 T ⎠⎦

o Therefore, a basis set is the time translates of

gi (t ) = exp ( j πh ( 2i − M − 1) t T ) , 0 ≤ t < T , i = 1,…, M

These basis functions are not orthogonal, except for some choices of h,
but they are linearly independent. For example, for M = 4 and
h =1 2:

o The coefficients P e jθn with respect to this basis carry the memory of
past data values, so the current waveform sheds light on the past.
Ignore this memory during detection (e.g., by use of a differential
detector) and you pay a price of several dB in lost SNR margin.
8.2-11
 8.2-12

o Implementations either use a set of filters matched to the gi (t ) or they

compute those inner products implicitly, as part the metric calculations
in tree or trellis search.

• The influence of the past on the response of a system is summarized in its

state. Here θn acts as the state:

θn +1 = θn + π h I n state transition equation

⎡ ⎛ t − nT ⎞⎤
s (t ) = P e jθn exp ⎢ j 2π h I n ⎜ ⎟ ⎥ output equation in interval n
⎣ ⎝ 2T ⎠⎦

o If h is rational, then there is a finite set of values for the phase state,
and we can summarize the recurrence with a state diagram. For
example, for h = 1 2 and M = 2 or 4 ,

8.2-12
 8.2-13
• The state transition diagram can be “unrolled” to form a trellis. Use the
same example ( h = 1 2 and M = 2 or 4 ).
For M = 2 (binary signals):
compacted or starting from a known state

alternating subsets, even and odd

and for M = 4 (quaternary signals):
Shown starting from state 0 at
time 0.

Note the parallel transitions.

o For M=2, the sequence of states σn uniquely determines the data

sequence I n .

o For M = 4 , the parallel transitions mean the state sequence only

partially determines the data. Also, which of the two I n values in a
(e.g.) +1, -3 transition leaves no trace in the memory (the state), so no
help from context (the future and past) in deciding which it was.

8.2-13
 8.2-14
8.2.3 MSK – A Special Case

• FSK and QPSK are quite different – or are they?

For one selection of parameters – MSK – CPFSK is equivalent to pre-
coding applied to offset QPSK (OQPSK).

• MSK (minimum shift keying): M = 2 and h = 1 2 .

Re [ s (t ) ] = P ∑ J n p (t − nT )
n even

Im [ s (t )] = P ∑ J n p(t − nT )
n odd

where the J n ∈ {−1, +1} , too.

Looks like 2PAM on each axis independently.

8.2-14
 8.2-15
• It can also be considered as offset QPSK, in which the imaginary
component is delayed by half a symbol. If
⎛ ⎞
s (t ) = P ⎜ ∑ J n p(t − nT ) + j ∑ J k p(t − kT ) ⎟
⎝ n even k odd ⎠
then s (t ) is like 4QAM (or QPSK) of symbol duration 2T

P ∑ bn p(t − nT ), bn = J n + j J n +1
n even

but with the imaginary part delayed by half a symbol (i.e., by T), so it’s
offset QPSK (OQPSK).
QPSK Offset QPSK

• If the receiver is coherent, it can tell Re [ s (t ) ] from Im [ s (t ) ] , and it’s easy

to detect the J n values, as the signs of the matched filter output samples.

But what is the relation between these J n values and the I n values of the
MSK sequence, the ones we want to detect?

8.2-15
 8.2-16
• How to get the I n values from the J n values?

Recall that I n = +1 is anticlockwise rotation, I n = −1 is clockwise. So:

o for even n, I n = J n J n +1 ;
o for odd n, I n = − J n J n +1 ;

o hence, I n = (−1) n J n J n +1 .
o It’s differential decoding with an alternating sign flip. Easy.

• From this, the BER follows as Pb = 2 Q ( )

2 Eb N 0 = 2 Q ( )
2 PT N 0 .

If we had detected I n using the signal in interval n only, it would have been
2FSK with non-orthogonal signals, with the two alternative signals starting
from the same unknown phase. See Assignment 2 from 2005 for its BER:

Pb = Q ( Eb N 0 ) (MSK detection without making use of memory)

8.2-16
 8.2-17
• The similarity with QPSK also gives us an easy calculation of the MSK
power spectrum: they are the same. See Section 8.6.

8.2-17