Вы находитесь на странице: 1из 10

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO.

XXX, XXX 2004

Characterization of WSSUS Channels: Normalized Mean Square Covariance


Do-Sik Yoo, Member, IEEE, and Wayne E Stark, Fellow, IEEE

Abstract A set of second order statistics collectively called normalized mean square covariance (NMSV) is dened to characterize the frequency and/or the time selectivity of wide-sense stationary uncorrelated scattering (WSSUS) channels. Normalized frequency mean square covariance (NFMSV) quanties the frequency selectivity, while normalized time mean square covariance (NTMSV) characterizes the time selectivity. Normalized frequency-time mean square covariance is dened to characterize the combined effect of frequency and time selectivities. The NMSVs of a WSSUS channel can easily be computed from the scattering function. We show that there is a very close relationship between the NMSV of a WSSUS channel and the performance of various diversity combing scheme. Also we discuss, with practical system design problems, how useful the parameters are for efcient system design. Index Terms normalized mean square covariance (NMSV), frequency selectivity, time selectivity, root mean square (rms) delay spread, coherence bandwidth, correlation time, wide-sense stationary uncorrelated scattering (WSSUS) channel

I. I NTRODUCTION Increasing demand for high speed mobile communications has motivated rapid development of wideband wireless networks. One of the most important performance degrading factors in such a network is signal fading, which is usually divided into shadowing and multipath fading. Shadowing is a phenomenon due to obstacles immediate to the receiver, while multipath fading is a cumulative effect of scattering, diffraction and reection due to various objects surrounding transmitter and receiver. In typical packet-based mobile communications, the effect of shadowing usually lasts longer than the duration of a packet so that its effect can be treated separately from that of multipath fading as an attenuation factor. The effect of shadowing can usually be described by a single randomly varying parameter with a suitable choice of probabilistic distribution such as the log-normal distribution. In contrast, multipath fading requires signicantly more complex descriptions due to its frequency and time dispersive nature. In this paper, we propose a set of parameters to characterize effectively the qualities of multipath fading channels. Since a radio link acts as a linear medium for electromagnetic waves, a multipath fading channel can be described by a time-variant linear system and hence characterized effectively
Manuscript received January xx, 2003. xxx D. Yoo with WESCOMM LLC and W. E. Stark is with the Dept. Electrical Engineering and Computer Science, The University of Michigan, Ann Arbor, MI, 48109-2122 dyoo@umich.edu and stark@eecs.umich.edu. This research was supported in part by Ericsson Inc., in part by DARPA under grant SRA588510732, and in part by National Science Foundation under grant ECS9979347.

by its system functions. Because of the practical unpredictability and the effectiveness in system performance analysis, the system functions are often regarded to be randomly timevariant and described mathematically by random processes. For intuitive physical interpretations, we often divide system functions into specular and diffuse components. The specular component is dened as the mean of the system function and hence is a deterministic function. The diffuse component describes the randomly varying uctuations of the system function from the specular component. In general, the characterization of a multipath fading channel demands the specication of the specular component and the statistical properties of the diffuse component. A non-zero specular component usually implies the existence of reliable signal paths in a fading environment. Consequently, the case of a zero specular component generally implies the worst case scenario for system performance. In fact, such adverse situations occur frequently in practical mobile communication environments. In this paper, we shall conne ourselves to the characterization of multipath fading channels with zero specular components.1 We will assume that the diffuse components satisfy the wide-sense stationary uncorrelated scattering (WSSUS) conditions and have Gaussian probability distribution. Consequently, we will consider only (complex) Gaussian wide-sense stationary uncorrelated scattering (WSSUS) channels [2]. Either the WSSUS assumption or the Gaussian approximation may not be valid in practical channels. For example, as the signal bandwidth or the time duration becomes larger, the WSSUS assumption becomes less valid. Also the Gaussian approximation can be inaccurate for various reasons. To resolve such discrepancy with practical situations, more realistic channel models can be developed. However, the performance analysis under such models are usually more difcult and provide only isolated results rather than drawing complete picture of system performance in terms of channel quality. Consequently, to be able to focus on the relation between channel characteristics and the system performance, we conne ourselves to the class of complex Gaussian WSSUS channels rather than delving into more realistic but isolated channel models. Complex Gaussian WSSUS channels can be specied completely by their scattering functions. However, with innite variety of scattering functions, it is impractical to investigate the direct relation between system performance and scattering functions. Consequently, derived quantities such as rms (root
1 Characterization of multipath fading channels with non-zero specular components is discussed in [1].

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

mean square) delay spread or correlation time are often used to represent the channel quality [3]. The rms delay spread of a channel is dened to be the variance of the time delay with the (normalized) delay power prole of the channel as the probability density function. Because frequency selectivity is a result of time dispersion, it makes some sense to consider the rms delay spread as its measure. However, as shown in Section IV-A, there is not a close relationship between rms delay spread and system performance when the bandwidth of the channel is wide. The coherence bandwidth is a more direct measure of frequency selectivity. There are several ways to dene the coherence bandwidth of a channel. We can dene the coherence bandwidth of level- by the largest frequency separation up to which the correlation of the channel frequency response is greater than . While the number can be chosen to be any number between 0 and 1, it is usually taken to be close to 1 so that the coherence bandwidth actually means the bandwidth over which the channel response is highly correlated. In many cases, we often use the following denition of coherence bandwidth: k Bc = (1)

V, we provide a few examples to illustrate the practical utility of the parameters in efcient system design. Finally, we draw conclusions in Section VI. II. N ORMALIZED M EAN S QUARE C OVARIANCE In this section, we consider a multipath fading channel described as a randomly time-variant linear system [2]. We briey describe various system functions to establish notational conventions and then dene the normalized mean square covariances. A. Randomly Time-Variant Linear System Model We denote by h(, t) the complex representation of the channel response at time t due to a unit impulse at time t . Then, the output y(t) of the channel due to an input signal x(t) is given by y(t) =
R

h(, t)x(t ) d.

(2)

where k is some constant and is the rms delay spread [4], [5]. Though mathematically less compelling, this denition gives a convenient measure of the coherence bandwidth as a function of the rms delay spread. However, regardless of the denition, there is no close relationship between the coherence bandwidth and the system performance, when the channel bandwidth is wide relative to the coherence bandwidth. Such poor relations between existing parameters and system performance stem from the fact that they fail to characterize the correlation property of a channel over the whole region of interest. For example, the coherence bandwidth is a local measure of the correlation function and hence cannot make an overall measure of the frequency selectivity of a channel when the coherence bandwidth is small compared to the whole bandwidth. Observing this, we propose, as a measure of frequency selectivity, a parameter called normalized frequency mean square covariance (NFMSV), which is the absolutesquared auto-covariance of the frequency response averaged over all frequency regions in consideration so that it can represent the overall covariance of the channel. We propose similar parameters called normalized time mean square covariance (NTMSV) and normalized frequency-time mean square covariance (NFTMSV) for the characterization of the time selectivity and the combined frequency and time selectivity of a given channel, respectively. The three parameters, namely, NFMSV, NTMSV and NFTMSV, will collectively be called the normalized mean square covariances (NMSVs) of the channel. The rest of this paper is organized as follows. In Section II, we dene NFMSV, NTMSV, and NFTMSV for general multipath fading channels. In Section III, we provide simple formulae that express the NMSVs in terms of the correlation functions of WSSUS channels. In Section IV, we show how closely the NMSV is related to the performance of practical diversity combining schemes through simulations. In Section

It is in many ways useful to analyze the spectral content of the impulse response h(, t) in both t and variables. As usual, the time variant frequency response or the transfer function H(f, t) is dened by H(f, t) =
R

h(, t) ej2f d.

(3)

By manipulating (2) and (3), we can obtain y(t) =


R

H(f, t)X(f )ej2f t df.

(4)

Similarly, we dene the delay Doppler spread function k(, ) and the Doppler spread function K(f, t), respectively, by k(, ) =
R

h(, t)ej2t dt k(, )ej2f d.


R

(5)

and K(f, ) = (6)

To model the randomly time-varying behavior of the channel, we regard the system functions as stochastic processes. We assume that the time variant frequency response H(f, t) has well dened rst and second order statistics2 and denote by VH (f, t; f , t ) the covariance function of H(f, t), i.e., VH (f, t; f , t ) = E[H(f, t)H (f , t )] E[H(f, t)]E[H (f , t )]. (7)

Before proceding, we note the following Cauchy-Schwarz inequality: |VH (f, t; f , t )|2 VH (f, t; f, t) VH (f , t ; f , t ). B. Denition In this subsection, we dene parameters collectively called normalized mean square covariances for the channel described above.
2 Note that other system functions such as k(, ) are often assumed to have impulsive second order statistics. However, in most applications, we assume that H(f, t) has well-behaved rst and second order statistics.

(8)

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

1) Normalized Frequency Mean Square Covariance (NFMSV): To characterize the frequency selectivity of the channel over a particular frequency region of interest at a particular time, we dene the normalized frequency mean square covariance (NFMSV). Let represent the frequency region of interest such that 0<

VH (f, t; f, t) df <

(9)

at time t. Then, we dene by |VH (f, t; f , t)|2 df df Vf (; t) =


2

(10)

VH (f, t; f, t) df

the normalized frequency mean square covariance of the channel over the frequency region at time t. If the relevant communication signal is transmitted over the frequency band between frequencies A and B, then we usually choose = [A, B]. For a system such as a frequency hopping spread spectrum system, the region of interest can be chosen to be a union of intervals to designate the frequency hopping region. 2) Normalized Time Mean Square Covariance (NTMSV): Similarly, we introduce the normalized time mean square covariance (NTMSV) to describe the time selectivity of the channel over a certain time region at a certain frequency. Let T be the time region of interest such that 0<
T

As in the case of the time or the frequency region of interest, the frequency-time region often denotes a rectangle or a union of rectangles. It is very important to note that the NFMSV, NTMSV, and NFTMSV depend not only on the covariance function VH but also on the region of interest, which implies that two different systems using different frequency/time regions can experience, under the same physical channel, different frequency/time selectivity. This property can be exploited judiciously in channel resource allocation. Note that the amount of channel resource used by a given system is described by the area of the region of interest. Before proceeding, consider the case when = T . If VH (f, t; f , t ) does not change appreciably over time in T or frequency in , then Vf t () is approximately the same as Vf (; t) (t T ) or Vt (f ; T ) (f ), respectively. In particular, we have lim Vf t { (t , t + )} = Vf (, t)
0

(15)

and lim Vf t {(f , f + ) T } = Vt (f ; T ).


0

(16)

III. NMSV FOR WSSUS C HANNELS The normalized mean square covariances are dened for a general randomly time-variant linear channel. However, the parameters are the most useful for channels with zero specular components.3 Consequently, we shall only consider channels with zero specular components from now on. We further assume that the diffuse components satisfy the WSSUS channel model. Hence, in the following discussion, we consider only WSSUS multipath fading channels [6], [7], [8].4 In this section, we show how the NMSVs are represented by various correlation functions of WSSUS channels. We start with brief descriptions of WSSUS channel models for notations and denitions before studying the relations between the NMSVs and the correlation functions. A. WSSCUS Channel Consider a WSSUS channel with impulse response h(, t). Then, E[h(, t)] = 0 (17) and E[h(, t)h ( , t )] = p(, t t )( ) (18)

VH (f, t; f, t) dt <

(11)

at frequency f . Then, the normalized time mean-square covariance (NTMSV) of the channel over the time region T at frequency f is dened by |VH (f, t; f, t )|2 dt dt Vt (f ; T ) =
T T 2

(12)

VH (f, t; f, t) dt
T

Usually, the time region T of interest is chosen to be the region over which the relevant packet is transmitted. In many cases, it is just an interval of time. However, it can represent a union of time intervals in systems such as time-division multiplex access (TDMA) systems. 3) Normalized Frequency-Time Mean Square Covariance (NFTMSV): NFMSV and NTMSV are dened to characterize frequency and time selectivity separately. To quantify the combined effect of frequency and time selectivity, we introduce the normalized frequency-time mean square covariance (NFMSV). Let be the frequency-time region of interest such that 0<

where p(, t) is the delay cross-power spectral density of the channel [2]. Note that p(, 0) is the delay power prole that describes how the average received power is distributed with respect to the time delay.
3 The specular components of a fading channel is dened by the mean functions of the system function. For example, the specular component of the impulse response h(, t) is dened by E[h(, t)], i.e., by the mean function of h(, t). The diffuse component is then dened by h(, t) E[h(, t)]. 4 Wide-sense stationarity (in time t) and uncorrelatedness (between delay ) assumptions become less valid as the observed time duration become longer and the channel bandwidth become wider, respectively. However, the idealized assumptions of WSSUS models provide mathematical simplicity and enable insightful system performance analysis. WSSUS assumptions are particularly meaningful for complex Gaussian channels for which the rst and the second order statistics provide complex statistical specications.

VH (f, t; f, t) df dt < .

(13)

Then, we dene the normalized frequency-time mean square covariance (NFTMSV) Vf t () of the channel over the frequency-time region by |VH (f, t; f , t )|2 df dt df dt Vf t () =
2

(14)

VH (f, t; f, t) df dt

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

It is easy to see that the time-variant transfer function H(f, t) satises that E[H(f, t)H (f , t )] = P (f f , t t ) (19)

where P (f, t) is the time-frequency correlation function dened by P (f, t) =


R

p(, t)ej2f d.

(20)

Note that the uncorrelatedness in the delay implies the widesense stationarity in the frequency f . The converse is also true, since the Fourier transform is invertible. Note that the transfer function H(f, t) is wide-sense stationary in both time and frequency for a WSSUS channel. Similarly, the Doppler delay spread function k(, ) satises E[k(, )k ( , )] = q(, )( )( ) where q(, ) is the scattering function dened by q(, ) =
R

Since the channel is wide-sense stationary in both time and frequency, NFMSV and NTMSV are independent of the time t and the frequency f , respectively. Consequently, we shall write, from now on, Vf () and Vt (T ) for Vf (; t) and Vt (f ; T ), respectively. It is not difcult to rewrite (27) - (29) in terms of p(, t), q(, ) or Q(f, ) using (20), (25) and (26). Of these variations, the formulae with the scattering function q(, ) are the most useful because of the ease of physical interpretation and evident duality. To manifest the symmetry and the role of the scattering function, we dene a set of kernel functions Kf (; ), Kt (; T ) and Kf t (; , ) by5
2

ej2f df Kf (; ) =
2

,
2

(30)

(21)

df ej2t dt

p(, t)ej2t dt.

(22)

Kt (; T )

T 2

(31)

We see the wide-sense stationarity of h(, t) in time t implies the uncorrelatedness of k(, ) in frequency . Finally, we see that the Doppler-spread function K(f, ) satises E[K(f, )K (f , )] = Q(f f , )( ) (23)

dt
T

and ej2(f t) df dt Kf t (; , ) =
2

(32)

df dt

where Q(f, ) is the Doppler cross-power spectral density dened by Q(f, ) =


R

q(, )ej2f d.

(24)

Note that P (f, t) =


R2

q(, )ej2(f t) d d Q(f, )ej2t d.


R

(25) (26)

B. NMSVs and Correlation Functions In this subsection, we show how we can represent Vf (; t), Vt (f ; T ), and Vf t () in terms of various correlation functions. First of all, since VH (f, t; f , t ) = P (f f , t t ), we have P (f f , 0) df df Vf (; t) =
2 2

respectively. Then, as shown in Appendix I, Vf (), Vt (T ), and Vf t () can be rewritten as in (33), (34), and (35) (see next page). Note that R q(, ) d = p(, 0) is the delay power prole of the channel. Consequently, NFMSV can easily be obtained from the delay power prole. Similarly, NTMSV can be obtained from the Doppler power sepctrum Q(0, ) = q(, ) d of the channel. R From (33) - (35), we see that the NMSVs are dependent not only on the scattering function but also on the regions of interest through the kernel functions. Although the kernel functions can be calculated in principle for any well-shaped regions, we are mostly interested in some rectangular shaped regions. As an example, consider a frequency region r dened by
K1

(27)

r =
k=0

[A + k(v + w), A + k(v + w) + w].

(36)

P (0, 0)

df
2

P (0, t t ) dt dt Vt (f ; T ) =
T T 2

(28)

P (0, 0)
T

dt

and P (f f , t t ) df dt df dt Vf t () =
2 2

where A, v and w are non-negative real numbers. So r is a union of K intervals of length w. Consequently, the sum of the lengths of the intervals is Kw regardless of the choice of v. In other words, the total frequency resource occupied by this frequency region is Kw regardless of the value of v. In Section V-A, we consider the problem of determining the suitable value for v. It is not difcult to show (see Appendix II) Kf (r ; ) =
5 Note

. (29)

sin{K(w + v) } K sin{(w + v) }

sinc2 (w ).

(37)

P (0, 0)

df dt

that if = T , then Kf t (; , ) = Kf (; )Kt (; T ).

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

Vf () =

R2

Kf (, )p(, 0)p ( , 0) d d
2

(33)

p(, 0) d
R

Vt (T ) =

R2

Kt ( ; T )Q(0, )Q (0, ) d d
2

(34)

Q(0, ) d
R

Vf t () =

R4

Kf t (; , )q(, )q ( , ) d d d d
2

(35)

q(, ) d d
R2

In particular, Kf ([A, B]; ) = sinc2 {(B A) }. (38)

By denition, NFMSV and NTMSV are properties at a specic time and at a specic frequency, respectively, while NFTMSV is an overall channel property. Normally there is not a simple relation between these three parameters. However, one simple case is worth mentioning in which we can factorize the scattering function q(, ) into delay power prole T ( ) and Doppler power spectrum N () as 6 q(, ) = T ( )N (). In this case, it follows that Vf () =
R2

(39)

Kf (; )T ( )T ( ) d d
2

(40)

T ( ) d
R

Vt (T ) =

R2

Kt ( ; T )N ()N ( ) d d
2

(41)

N () d
R

and Vf t ( T ) = Vf ()Vt (T ). (42) Although this factorization is usually not applicable to practical channels, this shows the product relationship between NFMSV, NTMSV and NFTMSV implying that a channel has signicantly higher frequency-time selectivity when it exhibits both frequency and time selectivities. IV. NMSV AND S YSTEM P ERFORMANCE The NMSVs are dened as average covariances over regions of interest (with proper normalization) to obtain useful insight on the quality of multipath fading channels. Since the quality of a channel means the reliability of communications
6 Delay power prole is a function of the distribution of the signal travel distance, while Doppler power spectrum is related to the angular distribution of the received signal power. Consequently, such a factorization is possible only if the angular distribution of the scatters is independent of the signal travel distance.

it permits to a system, it is essential to investigate the relationship between NMSVs and system performance. In this section, we verify that there are meaningful relations between NMSVs and system performance of popular diversity combining schemes. Since purely analytical performance evaluations of practical systems under practical channel environments are usually very difcult, we rst investigate such relationship through simulations. In the upcoming paper [9], conceptual foundations and idealized mathematical analysis are provided for more eidetic grounds for such relations. For simulations, we consider frequency hopping and direct sequence spread spectrum systems. Throughout this section, we consider only discrete-time circularly symmetric complex Gaussian WSSUS channels with nite number of resolvable paths. As shown in Section III, such channels can be described by the scattering functions. The fading levels at different times for a given path are realizations of a zero mean circularly symmetric complex Gaussian random variable and are correlated according to the Doppler power spectrum corresponding to the path. All the systems considered employ binary phase shift keying (BPSK) modulation. We assume there exist additive white Gaussian noise at the receiver and that the channel fading levels corresponding to different packets are independent. We assume that the receivers know the exact channel impulse responses for coherent demodulation. A. Frequency Hopping Spread Spectrum Systems In this subsection, we consider time non-selective but frequency selective channels and study the relations between NFMSV and performance of frequency hopping spread spectrum (FHSS) systems. For numerical analysis, we consider 71 channels described by discrete-time delay power proles. The delay power proles were obtained through actual measurements7 performed in four aireld-type and nine urban/suburban-type environments. Various combinations of antenna heights between 1.5 and 10m and carrier frequencies among 40.7, 77.5, 237.5, 386.95 and 1814MHz with 20MHz
7 These measurements were performed by the SPAWAR Systems Command in San Diego, CA in USA.

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

0.7

10

0.6

10

Eb/N0 = 10dB

0.5

E /N = 12dB
0.4
NFMSV
BER
b 0
3

10

0.3

Eb/N0 = 14dB
0.2
10
4

0.1

0.5

1.5 2 2.5 rms delay spread [ sec]

3.5

10

0.1

0.2

0.3 NFMSV

0.4

0.5

Fig. 1. Relation between NFMSV and rms delay spread. Each point corresponds to the (rms delay spread, NFMSV) pair for one of 71 frequency selective fading channels. There is not a close relation between rms delay spread and NFMSV.

Fig. 2. BER vs NFMSV for a FHSS system with a (31, 15) RS code. Each point corresponds to the (NFMSV, BER) pair for one of 71 frequency selective fading channel at Eb /N0 = 10, 12 or 14 dB. We observe a roughly monotonic relation between the NFMSV and the BER.
10
1

bandwidth were used to measure the characteristics in the thirteen different physical locations resulting in total of 71 different delay power proles. Among the 71 delay power proles, 41 have 240 tabs, 3 have 300 tabs and remaining 27 have 400 tabs. For all the 71 proles, the tab-spacing is 31.25nsec. We assume that a contiguous 10.375MHz is allocated to the systems and it is sliced into 332 equal bandwidth frequency slots. We assume that during a packet duration the carrier hops many times over the 332 slots using pseudo-randomly generated hopping patterns. We dene the frequency region of interest by the whole frequency band of 10.375MHz. Figure 1 shows the relation between NFMSV and rms delay spread for these 71 channels. We nd that there is a very weak relation between the two parameters implying that at least one of the two parameters is bound to exhibit a weak relation with the system performance. Since frequency hopping alone cannot exploit the frequency selectivity, we need to employ channel coding for meaningful results. A (31, 15) Reed-Solomon (RS) code and a rate 1/2 convolutional code are used as examples of block and trellis codes, respectively. Since both codes as well as the channel are linear, we assume that all zero information bits are transmitted. The simulations generate at least 20, 000 packets, that is, at least 20, 000 realizations of channel fading levels and then proceed to generate more until we count 1, 000 bit errors before computing the bit error rate (BER). 1) Reed-Solomon Code: First, we consider a (31, 15) RS code [10] as an example of block codes. In this coding scheme, a set of 15 32-ary symbols are encoded to a block of 31 32-ary symbols. Consequently, each codeword consists of 155 coded bits for 75 information bits. We assume that each packet consists of 20 codewords so that there are 3100 coded bits in a packet, which is RS-symbol interleaved by a block interleaver of size 62 by 10. The coded symbols are written column-wise and are read row-wise. Each row of coded symbols is transmitted over the same frequency slot

Eb/N0 = 5dB
10
2

Eb/N0 = 7dB

BER

10

Eb/N0 = 9dB

10

10

0.1

0.2

0.3 NFMSV

0.4

0.5

Fig. 3. BER vs NFMSV for a FHSS system with a (r=1/2, K=5) convolutional code. Each point corresponds to the (BER, NFMSV) pair for one of 71 frequency selective fading channel at Eb /N0 = 5, 7 or 9 dB.We observe a roughly monotonic relation between the NFMSV and the BER.

but different rows are transmitted over randomly and independently chosen frequency slots. This implies the hopping rate is 62 hops/packet and that the hopping pattern is random. We assume that bounded distance decoding algorithm is used with hard decisions made on the encoded bits. Figure 2 depicts the relationship between NFMSV and BER of the system. Each solid dot corresponds to the (NFMSV, BER) pair for one of the 71 channels at a particular signal to noise ratio (SNR). We clearly see that the BER is strongly related to the value of NFMSV. 2) Convolutional Code: As an example of trellis codes, we consider the rate 1/2, constraint length 5 convolutional code with 23 and 35 as the generators in octal from [6]. Each packet is again assumed to have 3100 coded bits which correspond

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

10

10

E /N = 3dB
b 0

E /N = 8dB
b 0

10
BER

10

Eb/N0 = 13dB

10

10

0.1

0.2

0.3

0.4

0.5 NFTMSV

0.6

0.7

0.8

0.9

Fig. 4. BER vs NFTMSV for a DSSS system with 2-stage SIC. Each point corresponds to the (NFTMSV, BER) pair for one of 200 channels at Eb /N0 = 3, 8 or 13 dB. Circles correspond to the channels with time selectivity only, while x-mark to the channels with frequency selectivity only. The solid dots are results for the channels with both frequency and time selectivity. We observe a roughly monotonic relation between the NFTMSV and the BER

to 1550 information bits, and the output of the convolutional encoder is bit interleaved by a block interleaver of size 62 by 50. Coded bits are written column-wise and are read rowwise. Again each row of coded bits is transmitted over the same frequency slot and different rows will be transmitted over randomly and independently chosen frequency slots. The receiver uses soft-decision Viterbi algorithm for decoding. Figure 3 shows the simulation results. Again we nd there is also very a strong relation between NFMSV and BER. B. Direct Sequence Spread Spectrum System In the previous subsection, we studied the relationship between NFMSV and the performance of FHSS systems under frequency selective but time non-selective channels.8 In this subsection, we consider doubly selective (namely frequency and time selective) channels as well as frequency-only selective and time-only selective channels and investigate the relationship between NFTMSV and performance of a direct sequence spread spectrum (DSSS) system with successive interference cancellation (SIC). By spreading, DSSS systems obtain more resolvable paths for rake reception. This is the way how DSSS benets from the frequency selectivity. Time selectivity can also be exploited by interleaving and channel coding. Consequently, a practical DSSS system can exploit both frequency and time selectivity. We show that there exists a nice relationship between NFTMSV and the performance of DSSS systems considered regardless whether channels exhibit frequency (or time) selectivity. In the DSSS system considered, a multi-stage successive interference cancellation (SIC) is employed to mitigate the detriis not unreasonable to expect similar relations between NFTMSV and the performance of FHSS systems if channel state varies during a packet duration (but not during a hop duration.) For DSSS systems, the combined effect of frequency and time selectivity is less obvious.
8 It

mental effect (namely intersymbol interference) of frequency selectivity. Each packet consists of 2, 000 binary symbols which are spread by a pseudo-random spreading code with spreading gain 25. Each set of 50, 000 chips is interleaved by a pseudo random interleaver before transmission.9 Since chip interleaving is employed, the spreading/despreading process, that can be regarded as a process of repetition coding/decoding, benets from the time selectivity of channels. The receiver is assumed to employ rake reception to exploit the frequency selectivity of channels. The signalling rate is assumed to be 5 mega-chips per second. The multistage SIC is executed as follows. After detecting the rst symbol using the usual threshold test, its (estimated) contribution is cancelled from the received signal (assuming the decision on the rst symbol is correct) and then the second symbol is detected. After detecting the second symbol, its contribution is cancelled similarly and then detection is made for the third symbol and so on. This is the rst stage of the SIC employed in the system. After detecting the last (i.e. 2000th ) symbol and subtracting its contribution, noise estimates are obtained that equals to the actual additive noise if all symbols were correctly detected. The second stage begins with the addition to the noise estimates of the contribution due to the rst symbol assuming that the detection in the rst stage is correct. At this point, we have noise estimates plus the signals corresponding to the rst symbol regardless whether the decision for the rst symbol in the rst stage is correct or not.10 Now, assuming this signal as the received signal, the rake receiver detects the rst symbol and subtracts its contribution again. Now, new noise estimates are obtained to which the contribution is added due to the second symbol. The second stage continues until this process reaches the last symbol. Similarly, higher stages can be processed. In this paper, we stop at the end of the second stage and compute the BER. Instead of the measured delay power proles considered in the previous subsection, we use 200 randomly generated scattering functions to model doubly selective channels. To generate the 200 scattering functions, we rst generate 500 delay power proles randomly and then choose 150 delay power proles with wide range of NFMSV values. Then, 50 scattering functions are directly dened by 50 of the 150 delay power proles assuming time non-selectivity. With the remaining 100 delay power proles, we generate 100 frequency and time selective scattering functions by multiplying the relative gain for each path by a randomly generated Doppler spectrum. Remaining 50 are dened to be timeonly selective scattering functions by randomly generating 50 Doppler spectra. Consequently, 50 channels are time nonselective, 50 channels frequency non-selective, and remaining 100 channels doubly selective. The process of random generation of the delay power proles and the Doppler spectra is described as follows. First,
this system, no error correction coding is employed. this reason, multi-stage SIC performs decently in many cases [11], [12]. This is also consistent with the simulation results shown in Figure 4 since no manifest evidence of detrimental effect of frequency selectivity is displayed.
10 For 9 For

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

the number of resolvable paths are chosen uniformly among 1 30 and the delay spreads are chosen uniformly between 0 and 20 microseconds. The relative amplitude gains for the paths (or collectively delay power proles) are chosen by rn |X|b where n is the index indicating the order of the path (i.e. nth path), r one of 0.8, 0.9 and 0.95, X a standard normal random variable, and b one of 1, 2 and 3. Finally, the resultant delay power proles are properly normalized. To generate the Doppler spectrum for each path, we rst consider the Jakes spectrum [13] of uniformly distributed scatterers with Doppler frequency fm = 200 Hz and randomly choose a scale factor s between 0 and 1. Then, we truncate from the Jakes spectrum the region outside the frequencies between sfm and sfm , which is then properly normalized. The resultant Doppler Spectrum S(f ) is 2 arcsin s for |f fc | < sfm 2 fm (f fc )2 S(f ) = (43) 0 otherwise, where fc is the carrier frequency of the signal. When the carrier frequency is 2GHz, the Doppler frequency 200Hz corresponds to the vehicular speed 30m/s. For these 200 channels, the NFTMSV is computed from the formula (35) assuming the system occupies an interval of 5MHz frequency band. The BER of the system is computed by Monte Carlo simulations under each of the 200 channels for three different SNRs. We rst generate 1000 packets and enumerate the number of total symbol errors. Then, the process of packet generation continues until the number of bit errors exceed 1000 before computing the BER. The results of the simulations are depicted in Figure 4. In the gure, Eb stands for the energy per bit and N0 for the one-sided noise power spectral density (PSD) of the additive white Gaussian noise at the receiver. Each point corresponds to the (NFTMSV, BER) pair for each of the 200 channel at Eb /N0 = 3, 8 or 13 dB. We observe that the BER does not depends heavily on the detailed shape of the scattering functions if the NFTMSVs are the same. In particular, the performance does not depend much on whether the channel has time or frequency selectivity if the NFTMSV is the same. We can also see there exists a roughly monotonically increasing relationship between the NFTMSV and the BER. For this reason, we propose the NFTMSV as a measure of channel quality. 11 V. NMSV AND S YSTEM D ESIGN Differences in service requirements and channel characteristics have led to a variety of communication systems. The optimality of system performance is especially crucial in mobile communications due to various stringent requirements such as high energy and spectral efciency. However, the diversity in the characteristics of mobile communication channels makes it difcult to investigate the optimality of communication systems. In the past, it was very difcult to evaluate the
11 As a minor point, we note that the BER is slightly higher with the channel with time selectivity only. As mentioned earlier the system exploits the time selectivity by the interleaver and the repetition. Hence, the results imply the random interleaver considered is not effective. In contrast, we can conclude that the 2-stage SIC is quite effective in treating the intersymbol interference.

10

10

NFMSV

10

10

v=0 v=w v=2w v=5w 10 20 30 40 50 60 70

Fig. 5. The usage of NFMSV in the frequency allocation. We see that we can lower the overall NFMSV of the channel for a FHSS system by separating each hop slots by inserting some amount of guard band. However, the NFMSV does not decrease dramatically as the separation bandwidth become greater than the bandwidth of each hop slot.

system optimality based on performance analysis under a small number of communication channels. However, it is not very difcult to see that the close relationship between NMSV and system performance can greatly simplify the system design process in many cases. Most of all, such close relationship directly enables us to investigate the optimality of various systems based on the performance analysis under a nite number of channels with diverse NMSV values. Moreover such relationship provides us with useful insight for system design and optimization. In this section, we consider two simple examples to illustrate the utility of NMSV. We rst consider the problem of frequency allocation for an FHSS system. Next we consider the effect of frequency hopping rate in an FHSS system. A. Frequency Allocation for a FHSS System In this subsection, we consider the problem of frequency band allocation for an FHSS system. In Section IV-A, we assumed that the 332 frequency slots are allocated to form an interval of frequency band of size 10.375 MHz. In other words, frequency hopping is made within the interval in this system. However, by introducing a space between adjacent slots, we can lower the NFMSV of the resultant frequency region keeping the total bandwidth the same. Assume there are K frequency slots of bandwidth w with a space of bandwidth v between adjacent slots. This frequency region is represented by the r given by (36) and the NFMSV V(r ) can be calculated from (33) with (37). By calculating V(r ) for the 71 channels in introduced Section IV-A, we obtain Figure 5. Here, we assume that K = 332 and w = 31.25kHz. From the gure, we can tell that we have smaller NFMSV with larger v, namely, with larger spacing. Since the BER performance is closely related to NFMSV, we can also tell, from the BER v.s. NFMSV curves (in Section IV-A), how much gain is obtained by introducing a space between slots. The space between slots

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004

BER for Channel 1 10


2

10
BER

10

10

10

20

30

40

50

60

70

80

90

100

NFMSV for Channel 1 6

(NFMSV)

0 10 20 30 40 50 Hops / Packet 60 70 80 90 100

Fig. 6. Bit error rate (BER) and 1/NFMSV versus number of hops for Channel 1. The NFMSV of the total frequency band for Channel 1 is 0.2. We note that the two curves exhibit similar saturation tendency.
BER for Channel 2

10

10

10

10

20

30

40

50

60

70

80

90

100

NFMSV for Channel 2 20

15
(NFMSV)
1

10

0 10 20 30 40 50 Hops / Packet 60 70 80 90 100

Fig. 7. Bit error rate (BER) and 1/NFMSV versus number of hops for Channel 2. The NFMSV of the total frequency band for Channel 1 is 0.06. We note that the two curves exhibit similar saturation tendency.

can be allocated to other systems. For example, if we choose v = w, then we can allocate slots to uplink and downlink systems alternately. We can allow more space by considering neighboring cells together in the allocation of frequency slots without wasting frequency resources. However, increasing v larger than 5w does not decrease V(r ) appreciably. Consequently, we expect saturation of the performance enhancement after v becomes larger than 5w. B. Choice of Frequency Hopping Rate Choosing optimal frequency hopping rate is an important problem in FHSS system design. In this subsection, we show how we can use NMSV to solve this problem. To illustrate the basic idea, we consider a frequency hopping system with (255, 127) Reed-Solomon code [10] under two different frequency selective and time non-selective WSSUS channels

called Channel 1 and Channel 2. We assume that each packet consists of one codeword, namely, 2040 binary digits. The total bandwidth of the channel is assumed to be 10.24MHz which is divided into N = 255 slots of bandwidth 40kHz. We assume that the fading is at over each of the 255 slots. For deniteness, we assume the system is allocated the frequency region between A = 2000 and B = 2010.24[MHz]. The two channels are so chosen that Vf ([A, B]) = 0.20 and Vf ([A, B]) = 0.06 for Channel 1 and Channel 2, respectively. The number K of frequency hopping per packet is chosen between 1 and 255. The hopping pattern is chosen to be roughly uniformly spaced. For example, if K = 7, the 1st , the 37th , the 73th , the 110th , the 146th , the 183th , and the 219th slots are used for transmission. We denote by K the frequency region consisting of the K slots on which the system hops. Similarly, the 255 symbols are divided roughly equally so that the 1st the 36th symbols are transmitted over the 1st slot, the 37th the 72th symbols over the 37th slot and so on. As a performance measure of the system, we consider the bit error rate (BER). For performance evaluation, we use a systematic (255, 127) Reed Solomon code and assume that the all zero information bits are transmitted. The receiver rst makes a hard-decision on each coded symbol. Then, it counts the number of coded-symbol errors and assume the errors are not corrected by the decoder if the number of symbol errors exceeds the error correcting capability, in which case the number of information bit errors in the systematic parts are counted toward the total errors. By simulations, we obtain the BER vs. hopping rate. Similarly, we plot V(K ) as a function of K. Figures 6 and 7 show the results. For the two channels, different SNRs are chosen to make the saturated BER roughly the same. For Channel 1, V(K ) is virtually the same for K > 20 and we observe that the BER of the system is saturated around K = 20. Similar coincidence between the tendencies of NFMSV and BER can also be observed for Channel 2, although it is very difcult to nd out the point of obvious saturation. Generally, it is much more time consuming to evaluate the system performance compared to the computation of the NFMSV. However, as shown in the gures, we can predict the performance saturation point with the NFMSV saturation point. VI. S UMMARY AND C ONCLUSION In this paper, we dened three parameters called normalized frequency mean square covariance (NFMSV), normalized time mean square covariance (NTMSV), and normalized frequencytime mean square covariance (NFTMSV) to characterize WSSUS channels. From the denitions, it is easy to understand how these parameters characterize the overall frequency and/or time selectivities of WSSUS channels. We showed that there exists very close relationship between NMSVs and performance of practical systems such as FHSS or DSSS systems. Due to the close relationship with the system performance, the NMSVs can be regarded as useful measures of channel quality and are very useful for efcient system design. In the upcoming paper [9], conceptual foundations and idealized mathematical analysis are provided for more eidetic grounds.

BER

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. X, NO. XXX, XXX 2004 2

10

ej2f ( Vf t () =
R4

)+j2( )(tt )

df dt q(, )q ( , ) d d d d
2

(44)

q(, ) d d
R2

df dt

A PPENDIX I NMSV IN TERMS OF S CATTERING F UNCTION In this section, we derive (33) - (35). From the denition of P (f, t) in (20), we have P (f f , t t )P (f f , t t ) df dt df dt

=
R4

ej2(f f

)( )+j2( )(tt )

q(, )q ( , ) d d d d df dt df dt
2

(45)

=
R4

ej2f (

)+j2( )(tt )

df dt (46) (47)

q(, )q ( , ) d d d d P (0, 0) =
R2

and q(, ) d d.

[5] M. J. Gans, A power spectral theory of propagation in the mobile raido environment, IEEE Transactions on Vehicular Technology, vol. 21, pp. 27 38, Feb. 1972. [6] J. G. Proakis, Digital Communications, 3rd ed. McGraw-Hill, 1995. [7] R. S. Kennedy, Fading Dispersive Communication Channels. John Wiley & Sons, 1969. [8] R. L. Peterson, R. E. Ziemer, and D. E. Borth, Introduction to Spread Spectrum Communications. Prentice Hall, 1995. [9] D.-S. Yoo and W. E. Stark, Normalized mean square covariance and diversity combining, IEEE Transactions on Wireless Communications, submitted. [10] F. J. MacWilliams and N. J. A. Sloane, The Thoery of Error-Correcting Codes. North-Holland, 1977. [11] D.-S. Yoo, A. Hafeez, and W. E. Stark, Trellis-based multiuser detection for ds-cdma systems with frequency-selective fading, IEEE Wireless Communications and Networking Conference, vol. 2, pp. 829 833, 1999. [12] D.-S. Yoo and W. E. Stark, Interference cancellation for multirate multiuser systems, IEEE Vehicular Technology Conference, Spring 2001. [13] W. C. Jakes, Microwave Mobile Communications. IEEE Press, 1994.

Consequently, by plugging (46) and (47) into (29), we obtain (44) for Vf t () (see above). For Vf (; t) and Vt (f ; T ), choose = (t , t + ) and = (f , f + ) T and take the limit 0 in (44). A PPENDIX II C ALCULATION OF A K ERNEL Let Ik = [A + k(v + w), A + k(v + w) + w]. Then,
2

ej2f df
K1 K1

(48) ej2f df ej2f


Ik

=
k=1 k =1 K1 K1 Ik

df

(49)

Do-Sik Yoo received B.S. degree in electrical engineering and M.S. degree in physical from Seoul National University, Seoul, Korea in 1990 and 1994, respectively. He received M.S. and Ph.D. degrees in electrical engineering from the University of Michigan, Ann Arbor, in 1998 and 2002, respectively. His research interests consist of diverse aspects of communications and networking including adaptive signal processing, coding and modulation, information theory, multiple access and resource allocation, and wireless networking. In the last few years, he has developed a turbo coded FH-OFDM receiver with interpolated joint channel estimation and decoding and power control algorithms that tracks timing errors and Doppler shifts.

=
k=1 k =1

w2 sinc2 (w )ej2 (w+v)(kk ) (50)


K1 2

= w2 sinc2 (w )
k=0

ej2 (w+v)k sin{K(w + v) } sin{(w + v) }


2

(51) . (52)

= w2 sinc2 (w ) Since

df = Kw, we have (37). R EFERENCES

[1] D.-S. Yoo and W. E. Stark, Characterization of multipath fading channels: Channels with specular components, IEEE Transactions on Wireless Communications, submitted. [2] P. A. Bello, Characterization of randomly time-variant linear channels, IEEE Transactions on Communication Systems, vol. CS-11, pp. 360 393, Dec. 1963. [3] J. G. Proakis, Digital Communications, 4th ed. McGraw-Hill, 2000. [4] W. C. Y. Lee, Mobile Communications Engineering, 2nd ed. McGrawHill Book Company, 1997.

Wayne E. Stark received the B.S. (with highest honors), M. S., and Ph.D. degrees in electrical engineering from the University of Illinois, UrbanaChampaign, in 1978, 1979, and 1982, respectively. Since September 1982, he has been a Faculty Member in the Department of Electrical Engineering and Computer Science at the University of Michigan, Ann Arbor, where he is currently a Professor. His research interests include the areas of coding and communication theory, especially for spreadspectrum and wireless communication networks. Dr. Stark is a member of Eta Kappa Nu, Phi Kappa Phi, and Tau Beta Phi. He was involved in the planning and organization of the 1986 International Symposium on Information Theory of the IEEE TRANSACTIONS ON COMMUNICATIONS in the area of spread-spectrum communications. He was selected by the National Scicence Foundation as a 1985 Presidential Young Investigator. He was the Principal Investigator of an Army Research Ofce Multidisciplinary University Research Initiative (MURI) Project on low-energy mobile communications.

Вам также может понравиться