Vous êtes sur la page 1sur 4

University of Illinois

Spring 2011

ECE 361: Lecture 15: Intersymbol Interference in Band-Limited Channels


Thus far in this course, we have been treating the communication channel as having no eect on the signal, or at worst simply as attenuating the transmitted signal by some known factor. Thus, the energy E that has been the subject of much discussion could be referred to as the received energy (which is what it is) or as the transmitted energy since the two quantities were identical, or at worst, had a known linear relationship. If we were to model such a channel as a linear time-invariant system with impulse response g(t), then g(t) would be taken to be (t) (or (t) where < 1 is a known constant) where (t) denotes the unit impulse. Thus, the channel output would be the same as the input, or the input attenuated by a factor . In practice, all channels change the transmitted signal in ways other than simple attenuation, but when the channel has bandwidth much larger than that of the transmitted signal, and G(f ) is essentially a constant over the frequency band of interest, then it is a reasonable approximation to model g(t) as (t) or (t). Otherwise, when the channel transfer function varies signicantly over the frequency band of interest, the eect of the channel on the transmitted signal needs to be taken into account. Such channels are called band-limited or bandwidth-limited channels and they cause a phenomenon called intersymbol interference. As the name implies, intersymbol interference (ISI) means that each sample value in the receiver depends not just on the symbol being demodulated but also on other symbols being transmitted. The presence of these extraneous symbols interferes with the demodulation process. For example, the designs for optimum receivers for signals received over AWGN channels that we have been studying thus far do not take ISI into account at all, and when ISI is present, their performance can be quite poor. In this Lecture and the next few, we shall study how ISI arises, and how to to mitigate its eects on the performance of communication systems operating over band-limited channels.

15.1.

Band-Limited Channels

Wireline channels as opposed to wireless channels are the primary source of examples of band-limited channels. Three common types are the telephone channel which connects an antique object called a telephone handset to the local telephone exchange or central oce, the Ethernet channel used in computer-to-computer (or computer-to-router) communication, and the cable channel that brings television signals, and increasingly more commonly, data signals from the Internet into the home. For nearly a hundred years, one telephone company AT&T a heavily regulated monopoly provided almost all the telephone channels in the U.S., and these channels typically carried voice signals which are low-frequency signals contained below 10 kHz. A telephone channel had a nominal bandwidth of 4 kHz, though typically only the frequency band below 3300 Hz was really useable. Naturally, telephonequality speech was not (and is not) high-delity speech. The bandwidth limitation was enforced by a lter either in the handset or at the central oce. In actuality, the wire connection between the home and the central oce has a lowpass bandwidth of about 1 to 2 MHz depending on the distance of the home from the central oce and the condition and quality of the wiring especially inside the walls of the house. Once upon a time, the telephone company put in high-quality wiring in all houses being constructed. Following de-regulation, any electrician could wire up the telephone jacks in a house under construction, and far too many chose to use cheap quad wire consisting of four straight wires in a plastic sheath. More recently, high-quality wiring has returned to home construction because of the demand for Ethernet connections inside the house. Ethernet wire (as well as old-fashioned telephone wire) consists of a bundle of twisted pairs of wires in which the two conductors comprising a pair cross over each other (and wires in other pairs) so as to minimize electromagnetic interference from external sources and coupling between pairs (which results in crosstalk). This is a signicant improvement over quad wire. Ethernet wire and its use is more

or less standardized. With Cat6 cable, which consists of four pairs of copper wires twisted together, the maximum length of cable between two devices is 100 metres of which up to 90 metres is allowed between wall-jacks, with 10 metres being allocated for connections between wall-jacks and the actual devices. The available bandwidth is about 250 MHz. Cable television signals are delivered over coaxial cable which consists of a central wire conductor running along the axis of a cylindrical conducting sheath and separated from the sheath by insulating spacers. The outer conducting sheath is usually wrapped in additional insulation. The RG-6 cable that is used for cable TV is much better in quality that the RG-59 cable often used to connect UHF antennas to television sets, and has a bandwidth of about 500 MHz. A key feature of a band-limited channel with impulse response g(t) and transfer function G(f ) is that the impulse response is spread out in time. that is, an input signal is dispersed in time. Physical models for g(t) usually lead to exponentially decaying functions for g(t), that is, the support for g(t) extends to , but g(t) is nonetheless essentially time-limited and we shall make the simplifying assumption that g(t) is small enough to be ignored for t > Td , where Td can be regarded as the dispersion of the channel. Thus, if the transmitter sends a pulse of duration T over the channel, then the channel response is small enough to be ignored outside a time interval of duration T + Td . The next section uses this notion in developing a model for communication over band-limited channels.

15.2.

A Model for Communication over Band-Limited Channels

We begin with the same kinds of signals that we used for signaling over the AWGN channel. Let (t) denote a unit-energy pulse of nite duration T . More specically, (t) = 0 for t < 0 or t > T , and also

|(t)|2 dt =

T 0

|(t)|2 dt = 1.

The transmitter signal is a sequence of pulses of dierent amplitudes, with each pulse delayed by T seconds from the previous pulse. Thus, the transmitted signal is s(t) =
i

Xi (t iT )

(15.1)

and consists of a pulse train of non-overlapping pulses of various amplitudes. This signal passes through the channel which is modeled as a linear time-invariant system with impulse response g(t). As was noted in previous Lectures, what is called the channel noise actually arises in the front end of the receiver and gets added to the signal output of the channel. The receiver lter has impulse response hR (t), and for simplicity we assume that hR (t) is a unit-energy signal with nite support [0, T ). Hence, sample values of the receiver output spaced T seconds (or more) apart in time are independent Gaussian random variables with variance 2 = N0 /2. We model the whole system as shown in the diagram below.

h(t) Xi (t iT )
i

(t)

g(t) N(t)

hR (t)
i

Xi h(t iT )

The input in the gure is shown as an impulse train of impulses of varying magnitudes occurring once every T seconds. The transmitter lter has impulse response (t) and thus converts the impulse train into the pulse train s(t) dened in (15.1). Note that these pulses do not overlap in time. However, s(t) passes 2

through the channel whose impulse response is g(t) of essentially nite duration Td , and thus each (t iT ) is transformed into an (t iT ) where ( ) =

( t)g(t) dt =

Td 0

( t)g(t) dt

is 0 if < 0 or if > T + Td . Note that the pulse (t iT ) overlaps with several pulses (t (i + 1)T ), (t (i + 2)T ), . . . that follow it in time. The signal i Xi (t iT ) of overlapping pulses arrives at the receiver, gets corrupted by the noise N(t) and further dispersed by the receiver lter. Using h(t) to denote the overall impulse response (note that h = g hR ), the received signal is y(t) =
i

Xi h(t iT )

(15.2)

where the pulse h(t) has support (0, 2T + Td ). Note that the pulse h(t iT ) also overlaps with several pulses h(t (i + 1)T ), h(t (i + 2)T ), . . . that follow it in time. Note also that the noise also get processed through the receiver lter and corrupts the received signal.

15.3.

The Discrete-Time Model

For additive white Gaussian noise channels, we used a matched lter and the sampling time was exactly when the lter output peaked in response to the transmitted pulse. This made sense because the channel noise had the same variance no matter when the sampling occurred, and the maximum benet could be extracted by sampling at the peak. Let T0 denote the location of the global extremum of h(t) which we assume without loss of generality to be positive. Thus, h(T0 ) |h(t)| for all t. This is the time when the response of the system to the impulse X0 (t) is maximum. Other sampling times are spaced at intervals of T seconds from T0 . In digital signal processing circles, it is conventional to sample the continuous-time signal x(t) every T seconds and use x[m] to denote the m-th sample value x(mT ). We borrow this notation but bend it to our purpose. Let us dene h[m] = h(T0 + mT ) and note that h[0] is the largest of these numbers. Now, since h(t) has support (0, 2T + Td ), it follows that h[m] = 0 for m < L where L = T0 1, and also that h[m] = 0 T for m > M where M = Td T0 + 1. Similarly, let us dene y[m] = y(T0 + mT ) where y(t) is as shown in T (15.2). Then, we have that
M

y[m] = y(T0 + mT ) =
i

Xi h(T0 + mT iT ) =
i

Xi h[m i] =
j=L

h[j]Xmj .

Changing notation slightly, and including the eects of the noise, we can dene the discrete-time model of the band-limited channel as a system whose input is a sequence {X[m]} and whose output is a sequence {Y[m]} where
M

Y[m] =
j=L

h[j]X[m j] + N[m]
M 1

= h[0]X[m] +
j=1 M

h[j]X[m j] +
j=L

h[j]X[m j] + N[m]

(15.3)

= h[0]X[m] +
j=L

h[j]X[m j] + N[m]

(15.4)

where N[m] is a N (0, 2 ) noise variable independent of all other N[m ] and in (15.4) means that the j = 0 term is excluded from the sum. This sum is called the intersymbol interference (ISI) and it detracts 3

from the demodulation of the rst term in (15.4) which is the desired symbol X[m] with weighting h[0], the global maximum of h(t). Writing out the sums in (15.3) in detail, we see that
M 1

Y[m] = h[0]X[m] +
j=1

h[j]X[m j] +
j=L

h[j]X[m j] + N[m]

= h[0]X[m]+ h[1]X[m 1] + + h[M ]X[m M ] + h[1]X[m + 1] + + h[L]X[m + L] + N[m]. Thus, the m-th sample contains the largest contribution from X[m] together with smaller contributions from M past signals X[m 1], X[m 2], . . . , X[m M ] and also from L future signals X[m + 1], X[m + 2], . . . , X[m + L]!! There is no failure of causality here but just an artifact of the way we have dened the sample values. At time T0 + mT , the past pulses Xm1 h(t (m 1)T ), Xm2 h(t (m 2)T ), . . . , XmM h(t(mM )T ) in (15.2) are all past their peak and are decaying away. In fact, the last in the list will have disappeared from the system at the next sample. Meanwhile, the future pulses Xm+1 h(t (m + 1)T ), Xm+2 h(t(m+2)T ), . . . , Xm+L h(t(m+L)T ) in (15.2) are still building up to their respective peaks which will occur T , 2T , . . . , LT seconds later. The discrete-time model for signaling over band-limited Gaussian noise channels can thus be thought of as follows. The input to the transmitter is a discrete-time sequence {X[m]} with one sample occurring every T seconds. The sample sequence {Y[m]} observed at the receiver output is the result of passing {X[m]} through a nite-impulse-response (FIR) digital lter whose unit pulse response is h[L], h[L + 1], . . . , h[1], h[0], h[1], . . . , h[M ] plus the omnipresent Gaussian noise. The FIR lter is in fact a causal lter, but appears to be noncausal because we are ignoring a z L factor in the z-transforms for the simple reason that it is convenient to dene the m-th output sample Y[m] as the one to which the transmitter input X[m] makes the maximum contribution. See the transversal lter in the gure below.

X[m + L]

X[m + L 1]

X[m]

X[m M ]

h[L]

h[L + 1]

h[1]

h[0]

h[M 1]

h[M ]

h[j]X[m j]
j=L

N[m]

Y[m]

In the next few Lectures, we shall study the eect of ISI on the communication process, and how to mitigate, if not entirely eliminate, the eects of ISI on the reliability of the communication schemes.

Vous aimerez peut-être aussi