Decision Feedback (DFE)

yn + un Ân
feedforward detector
−

Lecture 3 feedback

• Equalization: CSE8.6.2
– Nonlinear equalization X
0 X
N2
un = qm yn−m − qm Ân−m
∗ DFE: 8.6.2 m=−N1 m=1
∗ MLSD: 8.6.2, 8.5.2, (9.7.2) • Try to cancel ISI from “old” symbols by feeding back previous
detector outputs. This, in effect, modifies the system filter {xm }
and leaves a “nicer” ISI for the feedforward filter to handle. . .
• The coefficients {qm }N
m=−N1 can be determined based on the ZF
2

or MMSE criterion.

Adv. Digital Communications 3.1 c M. Skoglund

Adv. Digital Communications 3.3 c M. Skoglund

Simulation Example

Nonlinear Equalization
−1 −1
Solid: X(z) = 1−0.9z Dashed: X(z) = 1−0.1z
0
10

• Main problem with linear methods: A filter X(z) with zeros

−1
10
outside or on (or inside, but close to) the unit circle cannot be
“inverted” by another linear filter!
• Nonlinear methods necessary for severe ISI! −2

e
10

P
−1
X(z)=1−0.9z , no eq.
−1
X(z)=1−0.9z , ZF
• Two main approaches: −1
X(z)=1−0.9z , MMSE
−1
X(z)=1−0.9z , DFE
– Decision feedback (DFE) −3
10 −1
X(z)=1−0.1z , no eq.
−1
X(z)=1−0.1z , ZF.
– Maximum likelihood sequence detection (MLSD) X(z)=1−0.1z −1, MMSE
−1
X(z)=1−0.1z , DFE

−4
10
0 1 2 3 4 5 6 7 8 9 10
SNR 2E /N
b 0

Adv. Digital Communications 3.2 c M. Skoglund

Adv. Digital Communications 3.4 c M. Skoglund

Maximum Likelihood Sequence Detection
ISI Channels versus Convolutional
(MLSD) Codes
• MLSD: Based on observing y1N = (y1 , . . . , yN ), choose the
• Convolutional code, recap. (example),
sequence aN
1 that maximizes the conditional pdf
cm1 = bm + bm−2
f (y1N |AN
1 = aN
1 )
bm
• When {xm } is an FIR filter and when zn is white, an efficient cm2 = bm + bm−1 + bm−2
method based on the Viterbi algorithm exists.
– Information bits bm ∈ {0, 1}, coded bits cm1 , cm2
– The noise zn can be made uncorrelated either by choosing
gR (t) properly or by using a (discrete-time) whitening filter . . . – Generator sequences: g1 = (101) and g2 = (111)
– Two (binary) FIR filters!
• Often used in practice (radio channels are FIR filters). . .

Adv. Digital Communications 3.5 c M. Skoglund

Adv. Digital Communications 3.7 c M. Skoglund

• Principle behind using the Viterbi algorithm: Assume xm = 0 for

• FIR “channel taps” {xm }, MLSD equalization
m < 0 and m ≥ M (an “M-tap channel”) and that zn is white ⇒
channel
X
M −1 zn
yn = An−m xm + zn
m=0
An yn Ân
{xm } MLSD
• Let an = (an , an−1 , . . . , an−M +1 ) be the state at time n ⇒

X
N
max log f (y1N |An1 = aN
1 ) ⇐⇒ max log f (yn |An = an ) • Code with generators {gm }, MLSD soft decoding
n=1
channel
X
N X
M −1
⇐⇒ min (yn − an−m xm ) 2 zn
n=1 m=0
bn yn b̂n
• Use a trellis to keep track of how possible state-transitions {gm } MLSD
P
ak−1 → ak influence the possibel values for (yk − m ak−m xm )2

Adv. Digital Communications 3.6 c M. Skoglund

Adv. Digital Communications 3.8 c M. Skoglund

 Generalizations

• Fractionally spaced
– Do equalization based on several samples per symbol interval
(“fractionally” and not “symbol” -spaced sampling). . .
• Adaptive
– When the channel impulse response varies in time, the
parameters of the equalizer have to be updated regularly. . .

Can be applied in both linear and nonlinear equalization.

Adv. Digital Communications 3.9 c M. Skoglund